PETSc Manual
PETSc_manual
User Manual: Pdf
Open the PDF directly: View PDF .
Page Count: 266 [warning: Documents this large are best viewed by clicking the View PDF Link!]
- Abstract
- I Introduction to PETSc
- II Programming with PETSc
- Vectors and Parallel Data
- Matrices
- KSP: Linear System Solvers
- SNES: Nonlinear Solvers
- TS: Scalable ODE and DAE Solvers
- Basic TS Options
- Using Implicit-Explicit (IMEX) Methods
- Using fully implicit methods
- Using the Explicit Runge-Kutta timestepper with variable timesteps
- Special Cases
- Monitoring and visualizing solutions
- Error control via variable time-stepping
- Handling of discontinuities
- Using TChem from PETSc
- Using Sundials from PETSc
- Performing sensitivity analysis
- Solving Steady-State Problems with Pseudo-Timestepping
- High Level Support for Multigrid with KSPSetDM() and SNESSetDM()
- DMPlex: Unstructured Grids in PETSc
- III Additional Information
- PETSc for Fortran Users
- Using MATLAB with PETSc
- Profiling
- Hints for Performance Tuning
- Other PETSc Features
- Makefiles
- Unimportant and Advanced Features of Matrices and Solvers
- Index
- Bibliography
![](asset-1.png)
Argonne National Laboratory is a U.S. Department of Energy
laboratory managed by UChicago Argonne, LLC
Mathematics and Computer Science Division
Argonne National Laboratory
9700 South Cass Avenue, Bldg. 240
Argonne, IL 60439
www.anl.gov
ANL-95/11 Rev 3.8
PETSc Users Manual
Revision 3.8
Mathematics and Computer Science Division
About Argonne National Laboratory
Argonne is a U.S. Department of Energy laboratory managed by UChicago Argonne, LLC
under contract DE-AC02-06CH11357. The Laboratory’s main facility is outside Chicago,
at 9700 South Cass Avenue, Argonne, Illinois 60439. For information about Argonne
and its pioneering science and technology programs, see www.anl.gov.
DOCUMENT AVAILABILITY
Online Access: U.S. Department of Energy (DOE) reports produced after 1991 and a
growing number of pre-1991 documents are available free via DOE's SciTech Connect
(http://www.osti.gov/scitech/)
Reports not in digital format may be purchased by the public from the National
Technical Information Service (NTIS):
U.S. Department of Commerce
National Technical Information Service
5301 Shawnee Rd
Alexandra, VA 22312
www.ntis.gov
Phone: (800) 553-NTIS (6847) or (703) 605-6000
Fax: (703) 605-6900
Email: orders@ntis.gov
Reports not in digital format are available to DOE and DOE contractors from the Office of
Scientific and Technical Information (OSTI):
U.S. Department of Energy
Office of Scientific and Technical Information
P.O. Box 62
Oak Ridge, TN 37831-0062
www.osti.gov
Phone: (865) 576-8401
Fax: (865) 576-5728
Email: reports@osti.gov
Disclaimer
This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States
Government nor any agency thereof, nor UChicago Argonne, LLC, nor any of their employees or officers, makes any warranty, express
or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus,
product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific
commercial product, process, or service by trade name, trademark, manufacturer, or otherwise, does not necessarily constitute or imply
its endorsement, recommendation, or favoring by the United States Government or any agency thereof. The views and opinions of
document authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof,
Argonne National Laboratory, or UChicago Argonne, LLC.
![](asset-3.png)
ANL-95/11 Rev 3.8
PETSc Users Manual
Revision 3.8
Prepared by
S. Balay1, S. Abhyankar2, M. Adams3, J. Brown1, P. Brune1, K. Buschelman1, L.
Dalcin4, V. Eijkhout6, W. Gropp1, D. Karpeyev1, D. Kaushik1, M. Knepley1, D. May7,
L. Curfman McInnes1, T. Munson1, K. Rupp1, P. Sanan8, B. Smith1, S. Zampini4, H.
Zhang5, and H. Zhang1
1Mathematics and Computer Science Division, Argonne National Laboratory
2Energy Systems Division, Argonne National Laboratory
3Computational Research, Lawrence Berkeley National Laboratory
4Extreme Computing Research Center, King Abdullah University of Science and Technology
5Computer Science Department, Illinois Institute of Technology
6Texas Advanced Computing Center, University of Texas at Austin
7Department of Earth Sciences, University of Oxford
8Institute of Geophysics, ETH Zurich
September 2017
This work was supported by the Office of Advanced Scientific Computing Research,
Office of Science, U.S. Department of Energy, under Contract DE-AC02-06CH11357.
![](asset-5.png)
PETSc 3.8 March 24, 2018
Abstract:
This manual describes the use of PETSc for the numerical solution of partial differential equa-
tions and related problems on high-performance computers. The Portable, Extensible Toolkit for
Scientific Computation (PETSc) is a suite of data structures and routines that provide the building
blocks for the implementation of large-scale application codes on parallel (and serial) computers.
PETSc uses the MPI standard for all message-passing communication.
PETSc includes an expanding suite of parallel linear solvers, nonlinear solvers, and time in-
tegrators that may be used in application codes written in Fortran, C, C++, and Python (via
petsc4py). PETSc provides many of the mechanisms needed within parallel application codes, such
as parallel matrix and vector assembly routines. The library is organized hierarchically, enabling
users to employ the level of abstraction that is most appropriate for a particular problem. By using
techniques of object-oriented programming, PETSc provides enormous flexibility for users.
PETSc is a sophisticated set of software tools; as such, for some users it initially has a much
steeper learning curve than a simple subroutine library. In particular, for individuals without some
computer science background, experience programming in C, C++ or Fortran and experience using
a debugger such as gdb or dbx, it may require a significant amount of time to take full advantage
of the features that enable efficient software use. However, the power of the PETSc design and
the algorithms it incorporates may make the efficient implementation of many application codes
simpler than “rolling them” yourself.
•For many tasks a package such as MATLAB is often the best tool; PETSc is not intended for
the classes of problems for which effective MATLAB code can be written.
•PETSc should not be used to attempt to provide a “parallel linear solver” in an otherwise
sequential code. Certainly all parts of a previously sequential code need not be parallelized
but the matrix generation portion must be parallelized to expect any kind of reasonable
performance. Do not expect to generate your matrix sequentially and then “use PETSc” to
solve the linear system in parallel.
Since PETSc is under continued development, small changes in usage and calling sequences
of routines will occur. PETSc is supported; see the web site http://www.mcs.anl.gov/petsc for
information on contacting support.
A list of publications and web sites that feature work involving PETSc may be found at
http://www.mcs.anl.gov/petsc/publications.
We welcome any reports of corrections for this document.
5
![](asset-6.png)
PETSc 3.8 March 24, 2018
6
![](asset-7.png)
PETSc 3.8 March 24, 2018
Getting Information on PETSc:
Online:
•Manual pages and example usage : http://www.mcs.anl.gov/petsc/documentation
•Installing PETSc : http://www.mcs.anl.gov/petsc/documentation/installation.html
•Tutorials : https://www.mcs.anl.gov/petsc/documentation/tutorials/index.html
In this manual:
•Complete contents, page 13
•Basic introduction, page 21
•Assembling vectors, page 49; and matrices, page 65
•Linear solvers, page 79
•Nonlinear solvers, page 105
•Timestepping (ODE) solvers, page 143
•Structured grids, page 54
•Unstructured meshes, pages 59 and 163
•Index, page 247
7
![](asset-8.png)
PETSc 3.8 March 24, 2018
8
![](asset-9.png)
PETSc 3.8 March 24, 2018
Acknowledgments:
We thank all PETSc users for their many suggestions, bug reports, and encouragement.
Recent contributors to PETSc can be seen by visualizing the history of the PETSc git repository,
for example at github.com/petsc/petsc/graphs/contributors.
Earlier contributors to PETSc include:
•Asbjorn Hoiland Aarrestad - the explicit Runge-Kutta implementations (TSRK);
•G. Anciaux and J. Roman - the interfaces to the partitioning packages PTScotch, Chaco, and
Party;
•Allison Baker - the flexible GMRES (KSPFGMRES) and LGMRES (KSPLGMRES) code;
•Chad Carroll - Win32 graphics;
•Ethan Coon - the PetscBag and many bug fixes;
•Cameron Cooper - portions of the VecScatter routines;
•Patrick Farrell - the nleqerr line search for SNES;
•Paulo Goldfeld - the balancing Neumann-Neumann preconditioner (PCNN);
•Matt Hille;
•Joel Malard - the BICGStab(l) implementation (KSPBCGSL);
•Paul Mullowney, enhancements to portions of the Nvidia GPU interface;
•Dave May - the GCR implementation (KSPGCR);
•Peter Mell - portions of the DMDA routines;
•Richard Mills - the AIJPERM matrix format (MATAIJPERM) for the Cray X1 and universal F90
array interface;
•Victor Minden - the NVIDIA GPU interface;
•Todd Munson - the LUSOL (sparse solver in MINOS) interface (MATSOLVERLUSOL) and several
Krylov methods;
•Adam Powell - the PETSc Debian package;
•Robert Scheichl - the MINRES implementation (KSPMINRES);
•Kerry Stevens - the pthread-based Vec and Mat classes plus the various thread pools (depre-
cated);
•Karen Toonen - design and implementation of much of the PETSc web pages;
•Desire Nuentsa Wakam - the deflated GMRES implementation (KSPDGMRES);
•Liyang Xu - the interface to PVODE, now SUNDIALS/CVODE (TSSUNDIALS);
9
![](asset-a.png)
PETSc 3.8 March 24, 2018
PETSc source code contains modified routines from the following public domain software packages:
•LINPACK - dense matrix factorization and solve; converted to C using f2c and then hand-
optimized for small matrix sizes, for block matrix data structures;
•MINPACK - see page 138; sequential matrix coloring routines for finite difference Jacobian
evaluations; converted to C using f2c;
•SPARSPAK - see page 88; matrix reordering routines, converted to C using f2c;
•libtfs - the efficient, parallel direct solver developed by Henry Tufo and Paul Fischer for the
direct solution of a coarse grid problem (a linear system with very few degrees of freedom per
processor).
PETSc interfaces to the following external software:
•BLAS and LAPACK - numerical linear algebra;
•Chaco - A graph partitioning package;
http://www.cs.sandia.gov/CRF/chac.html
•ESSL - IBM’s math library for fast sparse direct LU factorization;
•Elemental - Jack Poulson’s parallel dense matrix solver package;
http://libelemental.org/
•HDF5 - the data model, library, and file format for storing and managing data,
https://support.hdfgroup.org/HDF5/
•hypre - the LLNL preconditioner library;
http://www.llnl.gov/CASC/hypre
•LUSOL - sparse LU factorization code (part of MINOS) developed by Michael Saunders,
Systems Optimization Laboratory, Stanford University;
http://www.sbsi-sol-optimize.com/
•MATLAB - see page 191;
•MUMPS - see page 102, MUltifrontal Massively Parallel sparse direct Solver developed by
Patrick Amestoy, Iain Duff, Jacko Koster, and Jean-Yves L’Excellent;
http://www.enseeiht.fr/lima/apo/MUMPS/credits.html
•Metis/ParMeTiS - see page 77, parallel graph partitioner,
http://www-users.cs.umn.edu/˜karypis/metis/
•Party - A graph partitioning package;
http://www2.cs.uni-paderborn.de/cs/ag-monien/PERSONAL/ROBSY/party.html
•PaStiX - Parallel LU and Cholesky solvers;
http://pastix.gforge.inria.fr/
•PTScotch - A graph partitioning package;
http://www.labri.fr/Perso/˜pelegrin/scotch/
10
![](asset-b.png)
PETSc 3.8 March 24, 2018
•SPAI - for parallel sparse approximate inverse preconditioning;
https://cccs.unibas.ch/lehre/software-packages/
•SuiteSparse - see page 102, developed by Timothy A. Davis;
http://faculty.cse.tamu.edu/davis/suitesparse.html
•SUNDIALS/CVODE - see page 153, parallel ODE integrator;
http://www.llnl.gov/CASC/sundials/
•SuperLU and SuperLU Dist - see page 102, the efficient sparse LU codes developed by Jim
Demmel, Xiaoye S. Li, and John Gilbert;
http://crd-legacy.lbl.gov/˜xiaoye/SuperLU
•STRUMPACK - the STRUctured Matrix Package;
http://portal.nersc.gov/project/sparse/strumpack/
•Triangle and Tetgen - mesh generation packages;
https://www.cs.cmu.edu/˜quake/triangle.html
http://wias-berlin.de/software/tetgen/
•Trilinos/ML - Sandia’s main multigrid preconditioning package;
http://software.sandia.gov/trilinos/,
•Zoltan - graph partitioners from Sandia National Laboratory;
http://www.cs.sandia.gov/zoltan/
These are all optional packages and do not need to be installed to use PETSc.
PETSc software is developed and maintained using
•Emacs editor
•Git revision control system
•Python
PETSc documentation has been generated using
•Sowing text processing tools developed by Bill Gropp
http://wgropp.cs.illinois.edu/projects/software/sowing/
•c2html
•pdflatex
•python
11
![](asset-c.png)
PETSc 3.8 March 24, 2018
12
Contents
Abstract 5
I Introduction to PETSc 19
1 Getting Started 21
1.1 Suggested Reading ..................................... 22
1.2 Running PETSc Programs ................................. 23
1.3 Running PETSc Tests ................................... 24
1.4 Writing PETSc Programs ................................. 25
1.5 Simple PETSc Examples .................................. 26
1.6 Citing PETSc ........................................ 41
1.7 Directory Structure ..................................... 42
II Programming with PETSc 45
2 Vectors and Parallel Data 47
2.1 Creating and Assembling Vectors ............................. 47
2.2 Basic Vector Operations .................................. 49
2.3 Indexing and Ordering ................................... 51
2.3.1 Application Orderings ............................... 51
2.3.2 Local to Global Mappings ............................. 53
2.4 Structured Grids Using Distributed Arrays ....................... 54
2.4.1 Creating Distributed Arrays ............................ 54
2.4.2 Local/Global Vectors and Scatters ........................ 55
2.4.3 Local (Ghosted) Work Vectors .......................... 56
2.4.4 Accessing the Vector Entries for DMDA Vectors ................ 57
2.4.5 Grid Information .................................. 58
2.5 Vectors Related to Unstructured Grids .......................... 59
2.5.1 Index Sets ...................................... 59
2.5.2 Scatters and Gathers ................................ 60
2.5.3 Scattering Ghost Values .............................. 62
2.5.4 Vectors with Locations for Ghost Values ..................... 62
3 Matrices 65
3.1 Creating and Assembling Matrices ............................ 65
3.1.1 Sparse Matrices ................................... 67
3.1.2 Dense Matrices ................................... 70
13
![](asset-e.png)
CONTENTS PETSc 3.8 March 24, 2018
3.1.3 Block Matrices ................................... 71
3.2 Basic Matrix Operations .................................. 73
3.3 Matrix-Free Matrices .................................... 73
3.4 Other Matrix Operations ................................. 75
3.5 Partitioning ......................................... 77
4 KSP: Linear System Solvers 79
4.1 Using KSP ......................................... 79
4.2 Solving Successive Linear Systems ............................ 80
4.3 Krylov Methods ...................................... 81
4.3.1 Preconditioning within KSP ............................ 82
4.3.2 Convergence Tests ................................. 84
4.3.3 Convergence Monitoring .............................. 84
4.3.4 Understanding the Operator’s Spectrum ..................... 85
4.3.5 Other KSP Options ................................ 86
4.4 Preconditioners ....................................... 86
4.4.1 ILU and ICC Preconditioners ........................... 87
4.4.2 SOR and SSOR Preconditioners ......................... 88
4.4.3 LU Factorization .................................. 88
4.4.4 Block Jacobi and Overlapping Additive Schwarz Preconditioners ....... 89
4.4.5 Algebraic Multigrid (AMG) Preconditioners ................... 91
4.4.6 Balancing Domain Decomposition by Constraints ................ 93
4.4.7 Shell Preconditioners ................................ 95
4.4.8 Combining Preconditioners ............................ 95
4.4.9 Multigrid Preconditioners ............................. 97
4.5 Solving Block Matrices ................................... 99
4.6 Solving Singular Systems .................................102
4.7 Using External Linear Solvers ...............................102
5 SNES: Nonlinear Solvers 105
5.1 Basic SNES Usage .....................................105
5.1.1 Nonlinear Function Evaluation ..........................113
5.1.2 Jacobian Evaluation ................................113
5.2 The Nonlinear Solvers ...................................114
5.2.1 Line Search Newton ................................114
5.2.2 Trust Region Methods ...............................116
5.2.3 Nonlinear Krylov Methods ............................116
5.2.4 Quasi-Newton Methods ..............................117
5.2.5 The Full Approximation Scheme .........................117
5.2.6 Nonlinear Additive Schwarz ............................118
5.3 General Options ......................................118
5.3.1 Convergence Tests .................................118
5.3.2 Convergence Monitoring ..............................119
5.3.3 Checking Accuracy of Derivatives .........................120
5.4 Inexact Newton-like Methods ...............................120
5.5 Matrix-Free Methods ....................................121
5.6 Finite Difference Jacobian Approximations .......................138
5.7 Variational Inequalities ...................................140
5.8 Nonlinear Preconditioning .................................140
14
![](asset-f.png)
CONTENTS PETSc 3.8 March 24, 2018
6 TS: Scalable ODE and DAE Solvers 143
6.1 Basic TS Options ......................................145
6.2 Using Implicit-Explicit (IMEX) Methods .........................146
6.3 Using fully implicit methods ................................150
6.4 Using the Explicit Runge-Kutta timestepper with variable timesteps .........150
6.5 Special Cases ........................................150
6.6 Monitoring and visualizing solutions ...........................151
6.7 Error control via variable time-stepping .........................151
6.8 Handling of discontinuities .................................152
6.9 Using TChem from PETSc ................................153
6.10 Using Sundials from PETSc ................................153
7 Performing sensitivity analysis 155
7.1 Using the discrete adjoint methods ............................155
7.2 Checkpointing ........................................158
8 Solving Steady-State Problems with Pseudo-Timestepping 159
9 High Level Support for Multigrid with KSPSetDM() and SNESSetDM() 161
10 DMPlex: Unstructured Grids in PETSc 163
10.1 Representing Unstructured Grids .............................163
10.2 Data on Unstructured Grids ................................165
10.2.1 Data Layout ....................................165
10.2.2 Partitioning and Ordering .............................166
10.3 Evaluating Residuals ....................................166
10.4 Networks ..........................................167
10.4.1 Application flow ..................................168
10.4.2 Utility functions ..................................169
10.4.3 Retrieving components ...............................169
III Additional Information 171
11 PETSc for Fortran Users 173
11.1 C vs. Fortran Interfaces ..................................173
11.1.1 Include Files ....................................173
11.1.2 Error Checking ...................................173
11.1.3 Calling Fortran Routines from C (and C Routines from Fortran) .......174
11.1.4 Passing Null Pointers ...............................174
11.1.5 Duplicating Multiple Vectors ...........................175
11.1.6 Matrix, Vector and IS Indices ...........................175
11.1.7 Setting Routines ..................................175
11.1.8 Compiling and Linking Fortran Programs ....................175
11.1.9 Routines with Different Fortran Interfaces ....................176
11.2 Sample Fortran Programs .................................176
11.2.1 Array Arguments ..................................189
15
![](asset-10.png)
CONTENTS PETSc 3.8 March 24, 2018
12 Using MATLAB with PETSc 191
12.1 Dumping Data for MATLAB ...............................191
12.1.1 Dumping ASCII MATLAB data .........................191
12.1.2 Dumping Binary Data for MATLAB .......................191
12.2 Sending Data to an Interactive MATLAB Session ....................192
12.3 Using the MATLAB Compute Engine ..........................192
13 Profiling 195
13.1 Basic Profiling Information ................................195
13.1.1 Interpreting -log view Output: The Basics ...................195
13.1.2 Interpreting -log view Output: Parallel Performance .............197
13.1.3 Using -log mpe with Jumpshot ..........................199
13.2 Profiling Application Codes ................................199
13.3 Profiling Multiple Sections of Code ............................200
13.4 Restricting Event Logging .................................200
13.5 Interpreting -log info Output: Informative Messages .................201
13.6 Time .............................................202
13.7 Saving Output to a File ..................................202
13.8 Accurate Profiling and Paging Overheads ........................202
14 Hints for Performance Tuning 205
14.1 Maximizing Memory Bandwidth .............................205
14.1.1 Memory Bandwidth vs. Processes ........................205
14.1.2 Non-Uniform Memory Access (NUMA) and Process Placement ........207
14.2 Performance Pitfalls and Advice .............................211
14.2.1 Debug vs. Optimized Builds ............................211
14.2.2 Profiling .......................................212
14.2.3 Aggregation .....................................212
14.2.4 Memory Allocation for Sparse Matrix Assembly .................213
14.2.5 Memory Allocation for Sparse Matrix Factorization ..............213
14.2.6 Detecting Memory Allocation Problems .....................213
14.2.7 Data Structure Reuse ...............................214
14.2.8 Numerical Experiments ..............................214
14.2.9 Tips for Efficient Use of Linear Solvers ......................215
14.2.10 System-Related Problems .............................215
15 Other PETSc Features 217
15.1 PETSc on a process subset ................................217
15.2 Runtime Options ......................................217
15.2.1 The Options Database ...............................217
15.2.2 Options Prefixes ..................................218
15.2.3 User-Defined PetscOptions ............................218
15.2.4 Keeping Track of Options .............................219
15.3 Viewers: Looking at PETSc Objects ...........................219
15.3.1 Viewing From Options ...............................221
15.3.2 Using Viewers to Check Load Imbalance .....................221
15.4 Using SAWs with PETSc .................................221
15.5 Debugging ..........................................222
15.6 Error Handling .......................................222
16
![](asset-11.png)
CONTENTS PETSc 3.8 March 24, 2018
15.7 Numbers ...........................................224
15.8 Parallel Communication ..................................224
15.9 Graphics ...........................................224
15.9.1 Windows as PetscViewers .............................225
15.9.2 Simple PetscDrawing ................................225
15.9.3 Line Graphs .....................................226
15.9.4 Graphical Convergence Monitor ..........................229
15.9.5 Disabling Graphics at Compile Time .......................229
15.10Emacs Users ........................................229
15.11Vi and Vim Users .....................................230
15.12Eclipse Users ........................................230
15.13Qt Creator Users ......................................231
15.14Visual Studio Users .....................................233
15.15XCode Users (The Apple GUI Development System) ..................233
15.15.1 Mac OS X ......................................233
15.15.2 iPhone/iPad iOS ..................................233
16 Makefiles 235
16.1 Makefile System ......................................235
16.1.1 Makefile Commands ................................235
16.1.2 Customized Makefiles ...............................236
16.2 PETSc Flags ........................................236
16.2.1 Sample Makefiles ..................................236
16.3 Limitations .........................................239
17 Unimportant and Advanced Features of Matrices and Solvers 241
17.1 Extracting Submatrices ..................................241
17.2 Matrix Factorization ....................................241
17.3 Unimportant Details of KSP ...............................243
17.4 Unimportant Details of PC ................................244
Index 247
Bibliography 263
17
![](asset-12.png)
CONTENTS PETSc 3.8 March 24, 2018
18
Part I
Introduction to PETSc
19
Chapter 1
Getting Started
The Portable, Extensible Toolkit for Scientific Computation (PETSc) has successfully demonstrated
that the use of modern programming paradigms can ease the development of large-scale scientific
application codes in Fortran, C, C++, and Python. Begun over 20 years ago, the software has
evolved into a powerful set of tools for the numerical solution of partial differential equations and
related problems on high-performance computers.
PETSc consists of a variety of libraries (similar to classes in C++), which are discussed in detail
in Parts II and III of the users manual. Each library manipulates a particular family of objects
(for instance, vectors) and the operations one would like to perform on the objects. The objects
and operations in PETSc are derived from our long experiences with scientific computation. Some
of the PETSc modules deal with
•index sets (IS), including permutations, for indexing into vectors, renumbering, etc;
•vectors (Vec);
•matrices (Mat) (generally sparse);
•managing interactions between mesh data structures and vectors and matrices (DM);
•over thirty Krylov subspace methods (KSP);
•dozens of preconditioners, including multigrid, block solvers, and sparse direct solvers (PC);
•nonlinear solvers (SNES); and
•timesteppers for solving time-dependent (nonlinear) PDEs, including support for differential
algebraic equations, and the computation of adjoints (sensitivities/gradients of the solutions)
(TS)
Each consists of an abstract interface (simply a set of calling sequences) and one or more imple-
mentations using particular data structures. Thus, PETSc provides clean and effective codes for
the various phases of solving PDEs, with a uniform approach for each class of problem. This design
enables easy comparison and use of different algorithms (for example, to experiment with different
Krylov subspace methods, preconditioners, or truncated Newton methods). Hence, PETSc pro-
vides a rich environment for modeling scientific applications as well as for rapid algorithm design
and prototyping.
The libraries enable easy customization and extension of both algorithms and implementations.
This approach promotes code reuse and flexibility, and separates the issues of parallelism from
the choice of algorithms. The PETSc infrastructure creates a foundation for building large-scale
applications.
It is useful to consider the interrelationships among different pieces of PETSc. Figure 1is a
diagram of some of these pieces. The figure illustrates the library’s hierarchical organization, which
enables users to employ the level of abstraction that is most appropriate for a particular problem.
21
![](asset-16.png)
1.1. SUGGESTED READING PETSc 3.8 March 24, 2018
Increasing Level of Abstraction
BLAS/LAPACK MPI . . .
Vec (Vectors)
Standard CUDA CUSP ViennaCL . . .
IS (Index Sets)
General Block Stride
Mat (Operators)
Compressed
Sparse Row
Block
CSR
Symmetric
Block CSR Dense CUSPARSE CUSP ViennaCL FFT Shell . . .
PC (Preconditioners)
Additive
Schwarz
Block
Jacobi Jacobi ICC ILU LU SOR MG AMG BDDC Shell . . .
KSP (Krylov Subspace Methods)
GMRES Richardson CG CGS Bi-CGStab TFQMR MINRES GCR Chebyshev Pipelined
CG . . .
SNES (Nonlinear Solvers)
Newton
Line Search
Newton
Trust Region FAS NGMRES NASM ASPIN . . .
TAO (Optimization)
Newton Levenberg-
Marquardt . . .
TS (Time Steppers)
Euler Backward
Euler RK BDF SSP ARKIMEX Rosenbrock-
W. . .
DM (Domain Management)
Distributed
Array
Plex (Un-
structured) . . .
PETSc
Application Codes Higher-Level Libraries . . .
Figure 1: Numerical libraries of PETSc
1.1 Suggested Reading
The manual is divided into three parts:
•Part I - Introduction to PETSc
•Part II - Programming with PETSc
•Part III - Additional Information
Part I describes the basic procedure for using the PETSc library and presents two simple exam-
ples of solving linear systems with PETSc. This section conveys the typical style used throughout
the library and enables the application programmer to begin using the software immediately. Part
I is also distributed separately for individuals interested in an overview of the PETSc software, ex-
cluding the details of library usage. Readers of this separate distribution of Part I should note that
all references within the text to particular chapters and sections indicate locations in the complete
users manual.
Part II explains in detail the use of the various PETSc libraries, such as vectors, matrices, index
sets, linear and nonlinear solvers, and graphics. Part III describes a variety of useful information,
including profiling, the options database, viewers, error handling, makefiles, and some details of
PETSc design.
PETSc has evolved to become quite a comprehensive package, and therefore the PETSc Users
Manual can be rather intimidating for new users. We recommend that one initially read the entire
document before proceeding with serious use of PETSc, but bear in mind that PETSc can be used
22
![](asset-17.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
efficiently before one understands all of the material presented here. Furthermore, the definitive
reference for any PETSc function is always the online manual page (“manpage”).
Within the PETSc distribution, the directory ${PETSC_DIR}/docs contains all documentation.
Manual pages for all PETSc functions can be accessed at www.mcs.anl.gov/petsc/documentation.
The manual pages provide hyperlinked indices (organized by both concept and routine name) to
the tutorial examples and enable easy movement among related topics.
Emacs and Vi/Vim users may find the etags/ctags option to be extremely useful for exploring
the PETSc source code. Details of this feature are provided in Section 15.10.
The file manual.pdf contains the complete PETSc Users Manual in the portable document
format (PDF), while intro.pdf includes only the introductory segment, Part I. The complete
PETSc distribution, users manual, manual pages, and additional information are also available via
the PETSc home page at www.mcs.anl.gov/petsc. The PETSc home page also contains details
regarding installation, new features and changes in recent versions of PETSc, machines that we
currently support, and a FAQ list for frequently asked questions.
Note to Fortran Programmers: In most of the manual, the examples and calling sequences
are given for the C/C++ family of programming languages. We follow this convention because we
recommend that PETSc applications be coded in C or C++. However, pure Fortran programmers
can use most of the functionality of PETSc from Fortran, with only minor differences in the user
interface. Chapter 11 provides a discussion of the differences between using PETSc from Fortran
and C, as well as several complete Fortran examples. This chapter also introduces some routines
that support direct use of Fortran90 pointers.
Note to Python Programmers: To program with PETSc in Python you need to install the
PETSc4py package developed by Lisandro Dalcin. This can be done by configuring PETSc with
the option --download-petsc4py. See the PETSc installation guide for more details:
http://www.mcs.anl.gov/petsc/documentation/installation.html.
1.2 Running PETSc Programs
Before using PETSc, the user must first set the environmental variable PETSC_DIR, indicating the
full path of the PETSc home directory. For example, under the UNIX bash shell a command of
the form
export PETSC_DIR=$HOME/petsc
can be placed in the user’s .bashrc or other startup file. In addition, the user may need to set
the environment variable PETSC_ARCH to specify a particular configuration of the PETSc libraries.
Note that PETSC_ARCH is just a name selected by the installer to refer to the libraries compiled
for a particular set of compiler options and machine type. Using different values of PETSC_ARCH
allows one to switch between several different sets (say debug and optimized) of libraries easily.
To determine if you need to set PETSC_ARCH, look in the directory indicated by PETSC_DIR, if
there are subdirectories beginning with arch then those subdirectories give the possible values for
PETSC_ARCH.
All PETSc programs use the MPI (Message Passing Interface) standard for message-passing
communication [25]. Thus, to execute PETSc programs, users must know the procedure for be-
ginning MPI jobs on their selected computer system(s). For instance, when using the MPICH
implementation of MPI [16] and many others, the following command initiates a program that uses
eight processors:
23
![](asset-18.png)
1.3. RUNNING PETSC TESTS PETSc 3.8 March 24, 2018
mpiexec -n 8 ./petsc_program_name petsc_options
PETSc also comes with a script that uses the information set in ${PETSC_DIR}/${PETSC_ARCH}/
lib/petsc/conf/petscvariables to automatically use the correct mpiexec for your configuration.
${PETSC_DIR}/bin/petscmpiexec -n 8 ./petsc_program_name petsc_options
All PETSc-compliant programs support the use of the -h or -help option as well as the -v or
-version option.
Certain options are supported by all PETSc programs. We list a few particularly useful ones
below; a complete list can be obtained by running any PETSc program with the option -help.
•-log_view - summarize the program’s performance
•-fp_trap - stop on floating-point exceptions; for example divide by zero
•-malloc_dump - enable memory tracing; dump list of unfreed memory at conclusion of the
run
•-malloc_debug - enable memory tracing (by default this is activated for debugging versions)
•-start_in_debugger [noxterm,gdb,dbx,xxgdb] [-display name] - start all processes in
debugger
•-on_error_attach_debugger [noxterm,gdb,dbx,xxgdb] [-display name] - start debug-
ger only on encountering an error
•-info - print a great deal of information about what the program is doing as it runs
•-options_file filename - read options from a file
See Section 15.5 for more information on debugging PETSc programs.
1.3 Running PETSc Tests
As discussed in Section 1.2, users should set PETSC_DIR and PETSC_ARCH before running the tests,
or can provide them on the command line as below.
To check if the libraries are working do:
make PETSC_DIR=<PETSC_DIR> PETSC_ARCH=<PETSC_ARCH> test
A larger set of tests can be run with
make PETSC_DIR=<PETSC_DIR> PETSC_ARCH=<PETSC_ARCH> alltests
A new testing system is available by running
make -f gmakefile test PETSC_ARCH=<PETSC_ARCH>
The test reporting system classifies them according to the Test Anywhere Protocal (TAP)1. In
brief, the categories are
•ok
The test passed.
•not ok
The test failed.
1See https://testanything.org/tap-specification.html
24
![](asset-19.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
•not ok #SKIP
The test was skipped, usually because build requirements were not met (for example, an
external solver library was required, but not compiled against it).
•ok #TODO
The test is under development by the developers.
The tests are a series of shell scripts, generated by information contained within the test source
file, that are invoked by the makefile system. The tests are run in ${PETSC_DIR}/${PETSC_ARCH}/
tests with the same directory as the source tree underneath (See Section 1.7). A label is used to
denote where it can be found within the source tree. For example, test vec_vec_tutorials_runex6
denotes the shell script:
${PETSC_DIR}/${PETSC_ARCH}/tests/src/vec/vec/examples/tutorials/runex6.sh
These shell scripts can be run independently in those directories, and take arguments to show the
commands run, change arguments, etc. Use the -h option to the shell script to see these options.
Often, you want to run only a subset of tests. Our makefiles use gmake’s wildcard syntax.
In this syntax, %is a wild card character and is passed in using the search argument. Two
wildcard characters cannot be used in a search, so the searchin argument is used to provide the
equivalent of %pattern% search. The default examples have default arguments, and we often wish
to test examples with various arguments; we use the argsearch argument for these searches. Like
searchin, it does not use wildcards, but rather whether the string is within the arguments.
Some examples are:
make -f gmakefile test search=’ts%’ # Run all TS examples
make -f gmakefile test searchin=’tutorials’ # Run all tutorials
make -f gmakefile test search ’ts%’ searchin=’tutorials’ # Run all TS tutorials
make -f gmakefile test argsearch=’cuda’ # Run examples with
cuda in arguments
It is useful before invoking the tests to see what targets will be run. The print-test target
helps with this:
make -f gmakefile print-test argsearch=’cuda’
To see all of the test targets, this command can be used:
make -f gmakefile print-test
For more help, run
make -f gmakefile help-test
To learn more about the test system details, one can look at the Developer’s Guide.
1.4 Writing PETSc Programs
Most PETSc programs begin with a call to
PetscInitialize(int *argc,char ***argv,char *file,char *help);
which initializes PETSc and MPI. The arguments argc and argv are the command line arguments
delivered in all C and C++ programs. The argument file optionally indicates an alternative
name for the PETSc options file, .petscrc, which resides by default in the user’s home directory.
25
![](asset-1a.png)
1.5. SIMPLE PETSC EXAMPLES PETSc 3.8 March 24, 2018
Section 15.2 provides details regarding this file and the PETSc options database, which can be used
for runtime customization. The final argument, help, is an optional character string that will be
printed if the program is run with the -help option. In Fortran the initialization command has
the form
call PetscInitialize(character(*) file,integer ierr)
PetscInitialize() automatically calls MPI_Init() if MPI has not been not previously initialized.
In certain circumstances in which MPI needs to be initialized directly (or is initialized by some
other library), the user can first call MPI_Init() (or have the other library do it), and then call
PetscInitialize(). By default, PetscInitialize() sets the PETSc “world” communicator,
given by PETSC_COMM_WORLD, to MPI_COMM_WORLD.
For those not familiar with MPI, a communicator is a way of indicating a collection of processes
that will be involved together in a calculation or communication. Communicators have the variable
type MPI_Comm. In most cases users can employ the communicator PETSC_COMM_WORLD to indicate
all processes in a given run and PETSC_COMM_SELF to indicate a single process.
MPI provides routines for generating new communicators consisting of subsets of processors,
though most users rarely need to use these. The book Using MPI, by Lusk, Gropp, and Skjel-
lum [17] provides an excellent introduction to the concepts in MPI. See also the MPI homepage
http://www.mcs.anl.gov/mpi/. Note that PETSc users need not program much message passing
directly with MPI, but they must be familiar with the basic concepts of message passing and
distributed memory computing.
All PETSc routines return a PetscErrorCode, which is an integer indicating whether an error
has occurred during the call. The error code is set to be nonzero if an error has been detected;
otherwise, it is zero. For the C/C++ interface, the error variable is the routine’s return value, while
for the Fortran version, each PETSc routine has as its final argument an integer error variable. Error
tracebacks are discussed in the following section.
All PETSc programs should call PetscFinalize() as their final (or nearly final) statement, as
given below in the C/C++ and Fortran formats, respectively:
PetscFinalize();
call PetscFinalize(ierr)
This routine handles options to be called at the conclusion of the program, and calls MPI_Finalize()
if PetscInitialize() began MPI. If MPI was initiated externally from PETSc (by either the user
or another software package), the user is responsible for calling MPI_Finalize().
1.5 Simple PETSc Examples
To help the user start using PETSc immediately, we begin with a simple uniprocessor example in
Figure 2that solves the one-dimensional Laplacian problem with finite differences. This sequential
code, which can be found in $PETSC_DIR/src/ksp/ksp/examples/tutorials/ex1.c, illustrates
the solution of a linear system with KSP, the interface to the preconditioners, Krylov subspace
methods, and direct linear solvers of PETSc. Following the code we highlight a few of the most
important parts of this example.
static char help[] = "Solves a tridiagonal linear system with KSP.\n\n";
/*T
26
![](asset-1b.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
Concepts: KSP^solving a system of linear equations
Processors: 1
T*/
/*
Include "petscksp.h" so that we can use KSP solvers. Note that this file
automatically includes:
petscsys.h - base PETSc routines petscvec.h - vectors
petscmat.h - matrices
petscis.h - index sets petscksp.h - Krylov subspace methods
petscviewer.h - viewers petscpc.h - preconditioners
Note: The corresponding parallel example is ex23.c
*/
#include <petscksp.h>
int main(int argc,char **args)
{
Vec x, b, u; /* approx solution, RHS, exact solution */
Mat A; /* linear system matrix */
KSP ksp; /* linear solver context */
PC pc; /* preconditioner context */
PetscReal norm; /* norm of solution error */
PetscErrorCode ierr;
PetscInt i,n = 10,col[3],its;
PetscMPIInt size;
PetscScalar one = 1.0,value[3];
PetscBool nonzeroguess = PETSC FALSE,changepcside = PETSC FALSE;
ierr = PetscInitialize(&argc,&args,(char*)0,help);if (ierr) return ierr;
ierr = MPI Comm size(PETSC COMM WORLD,&size);CHKERRQ(ierr);
if (size != 1) SETERRQ(PETSC COMM WORLD,1,"This is a uniprocessor example only!"
);
ierr = PetscOptionsGetInt(NULL,NULL,"-n",&n,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetBool(NULL,NULL,"-nonzero_guess",&nonzeroguess,NULL);
CHKERRQ(ierr);
/*----------------------------------
Compute the matrix and right-hand-side vector that define
the linear system, Ax = b.
----------------------------------*/
/*
Create vectors. Note that we form 1 vector from scratch and
then duplicate as needed.
*/
ierr = VecCreate(PETSC COMM WORLD,&x);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) x, "Solution");CHKERRQ(ierr);
27
![](asset-1c.png)
1.5. SIMPLE PETSC EXAMPLES PETSc 3.8 March 24, 2018
ierr = VecSetSizes(x,PETSC DECIDE,n);CHKERRQ(ierr);
ierr = VecSetFromOptions(x);CHKERRQ(ierr);
ierr = VecDuplicate(x,&b);CHKERRQ(ierr);
ierr = VecDuplicate(x,&u);CHKERRQ(ierr);
/*
Create matrix. When using MatCreate(), the matrix format can
be specified at runtime.
Performance tuning note: For problems of substantial size,
preallocation of matrix memory is crucial for attaining good
performance. See the matrix chapter of the users manual for details.
*/
ierr = MatCreate(PETSC COMM WORLD,&A);CHKERRQ(ierr);
ierr = MatSetSizes(A,PETSC DECIDE,PETSC DECIDE,n,n);CHKERRQ(ierr);
ierr = MatSetFromOptions(A);CHKERRQ(ierr);
ierr = MatSetUp(A);CHKERRQ(ierr);
/*
Assemble matrix
*/
value[0] = -1.0; value[1] = 2.0; value[2] = -1.0;
for (i=1; i<n-1; i++) {
col[0] = i-1; col[1] = i; col[2] = i+1;
ierr = MatSetValues(A,1,&i,3,col,value,INSERT VALUES);CHKERRQ(ierr);
}
i = n - 1; col[0] = n - 2; col[1] = n - 1;
ierr = MatSetValues(A,1,&i,2,col,value,INSERT VALUES);CHKERRQ(ierr);
i = 0; col[0] = 0; col[1] = 1; value[0] = 2.0; value[1] = -1.0;
ierr = MatSetValues(A,1,&i,2,col,value,INSERT VALUES);CHKERRQ(ierr);
ierr = MatAssemblyBegin(A,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
/*
Set exact solution; then compute right-hand-side vector.
*/
ierr = VecSet(u,one);CHKERRQ(ierr);
ierr = MatMult(A,u,b);CHKERRQ(ierr);
/*----------------------------------
Create the linear solver and set various options
----------------------------------*/
/*
Create linear solver context
*/
ierr = KSPCreate(PETSC COMM WORLD,&ksp);CHKERRQ(ierr);
/*
Set operators. Here the matrix that defines the linear system
28
![](asset-1d.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
also serves as the preconditioning matrix.
*/
ierr = KSPSetOperators(ksp,A,A);CHKERRQ(ierr);
/*
Test if you can change the KSP side and type after they have been previously
set
*/
ierr = PetscOptionsGetBool(NULL,NULL,"-change_pc_side",&changepcside,NULL);
CHKERRQ(ierr);
if (changepcside) {
ierr = KSPSetUp(ksp);CHKERRQ(ierr);
ierr = PetscOptionsInsertString(NULL,"-ksp_norm_type unpreconditioned -
ksp_pc_side right");CHKERRQ(ierr);
ierr = KSPSetFromOptions(ksp);CHKERRQ(ierr);
ierr = KSPSetUp(ksp);CHKERRQ(ierr);
}
/*
Set linear solver defaults for this problem (optional).
- By extracting the KSP and PC contexts from the KSP context,
we can then directly call any KSP and PC routines to set
various options.
- The following four statements are optional; all of these
parameters could alternatively be specified at runtime via
KSPSetFromOptions();
*/
ierr = KSPGetPC(ksp,&pc);CHKERRQ(ierr);
ierr = PCSetType(pc,PCJACOBI);CHKERRQ(ierr);
ierr = KSPSetTolerances(ksp,1.e-5,PETSC DEFAULT,PETSC DEFAULT,PETSC DEFAULT);
CHKERRQ(ierr);
/*
Set runtime options, e.g.,
-ksp_type <type> -pc_type <type> -ksp_monitor -ksp_rtol <rtol>
These options will override those specified above as long as
KSPSetFromOptions() is called _after_ any other customization
routines.
*/
ierr = KSPSetFromOptions(ksp);CHKERRQ(ierr);
if (nonzeroguess) {
PetscScalar p = .5;
ierr = VecSet(x,p);CHKERRQ(ierr);
ierr = KSPSetInitialGuessNonzero(ksp,PETSC TRUE);CHKERRQ(ierr);
}
/*----------------------------------
Solve the linear system
29
![](asset-1e.png)
1.5. SIMPLE PETSC EXAMPLES PETSc 3.8 March 24, 2018
----------------------------------*/
/*
Solve linear system
*/
ierr = KSPSolve(ksp,b,x);CHKERRQ(ierr);
/*
View solver info; we could instead use the option -ksp_view to
print this info to the screen at the conclusion of KSPSolve().
*/
ierr = KSPView(ksp,PETSC VIEWER STDOUT WORLD);CHKERRQ(ierr);
/*----------------------------------
Check solution and clean up
----------------------------------*/
/*
Check the error
*/
ierr = VecAXPY(x,-1.0,u);CHKERRQ(ierr);
ierr = VecNorm(x,NORM 2,&norm);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(ksp,&its);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC COMM WORLD,"Norm of error %g, Iterations %D\n",(double)
norm,its);CHKERRQ(ierr);
/*
Free work space. All PETSc objects should be destroyed when they
are no longer needed.
*/
ierr = VecDestroy(&x);CHKERRQ(ierr); ierr = VecDestroy(&u);CHKERRQ(ierr);
ierr = VecDestroy(&b);CHKERRQ(ierr); ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = KSPDestroy(&ksp);CHKERRQ(ierr);
/*
Always call PetscFinalize() before exiting a program. This routine
- finalizes the PETSc libraries as well as MPI
- provides summary and diagnostic information if certain runtime
options are chosen (e.g., -log_view).
*/
ierr = PetscFinalize();
return ierr;
}
Figure 2: Example of Uniprocessor PETSc Code
Include Files
The C/C++ include files for PETSc should be used via statements such as
#include <petscksp.h>
30
![](asset-1f.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
where petscksp.h is the include file for the linear solver library. Each PETSc program must specify
an include file that corresponds to the highest level PETSc objects needed within the program; all
of the required lower level include files are automatically included within the higher level files.
For example, petscksp.h includes petscmat.h (matrices), petscvec.h (vectors), and petscsys.h
(base PETSc file). The PETSc include files are located in the directory ${PETSC_DIR}/include.
See Section 11.1.1 for a discussion of PETSc include files in Fortran programs.
The Options Database
As shown in Figure 2, the user can input control data at run time using the options database.
In this example the command PetscOptionsGetInt(NULL,NULL,"-n",&n,&flg); checks whether
the user has provided a command line option to set the value of n, the problem dimension. If so,
the variable nis set accordingly; otherwise, nremains unchanged. A complete description of the
options database may be found in Section 15.2.
Vectors
One creates a new parallel or sequential vector, x, of global dimension Mwith the commands
VecCreate(MPI Comm comm,Vec *x);
VecSetSizes(Vec x, PetscInt m, PetscInt M);
where comm denotes the MPI communicator and mis the optional local size which may be PETSC_DECIDE.
The type of storage for the vector may be set with either calls to VecSetType() or VecSetFromOptions().
Additional vectors of the same type can be formed with
VecDuplicate(Vec old,Vec *new);
The commands
VecSet(Vec x,PetscScalar value);
VecSetValues(Vec x,PetscInt n,PetscInt *indices,PetscScalar
*values,INSERT VALUES);
respectively set all the components of a vector to a particular scalar value and assign a different
value to each component. More detailed information about PETSc vectors, including their basic
operations, scattering/gathering, index sets, and distributed arrays, is discussed in Chapter 2.
Note the use of the PETSc variable type PetscScalar in this example. The PetscScalar is
simply defined to be double in C/C++ (or correspondingly double precision in Fortran) for
versions of PETSc that have not been compiled for use with complex numbers. The PetscScalar
data type enables identical code to be used when the PETSc libraries have been compiled for use
with complex numbers. Section 15.7 discusses the use of complex numbers in PETSc programs.
Matrices
Usage of PETSc matrices and vectors is similar. The user can create a new parallel or sequential
matrix, A, which has Mglobal rows and Nglobal columns, with the routines
MatCreate(MPI Comm comm,Mat *A);
MatSetSizes(Mat A,PETSC DECIDE,PETSC DECIDE,PetscInt M,PetscInt N);
where the matrix format can be specified at runtime via the options database. The user could
alternatively specify each processes’ number of local rows and columns using mand n.
31
![](asset-20.png)
1.5. SIMPLE PETSC EXAMPLES PETSc 3.8 March 24, 2018
MatSetSizes(Mat A,PetscInt m,PetscInt n,PETSC DETERMINE,PETSC DETERMINE);
Generally one then sets the “type” of the matrix, with, for example,
MatSetType(A,MATAIJ);
This causes the matrix Ato used the compressed sparse row storage format to store the matrix
entries. See MatType for a list of all matrix types. Values can then be set with the command
MatSetValues(Mat A,PetscInt m,PetscInt *im,PetscInt n,PetscInt *in,PetscScalar
*values,INSERT VALUES);
After all elements have been inserted into the matrix, it must be processed with the pair of com-
mands
MatAssemblyBegin(A,MAT FINAL ASSEMBLY);
MatAssemblyEnd(A,MAT FINAL ASSEMBLY);
Chapter 3discusses various matrix formats as well as the details of some basic matrix manipulation
routines.
Linear Solvers
After creating the matrix and vectors that define a linear system, Ax =b, the user can then use
KSP to solve the system with the following sequence of commands:
KSPCreate(MPI Comm comm,KSP *ksp);
KSPSetOperators(KSP ksp,Mat Amat,Mat Pmat);
KSPSetFromOptions(KSP ksp);
KSPSolve(KSP ksp,Vec b,Vec x);
KSPDestroy(KSP ksp);
The user first creates the KSP context and sets the operators associated with the system (matrix that
defines the linear system, Amat and matrix from which the preconditioner is constructed, Pmat). The
user then sets various options for customized solution, solves the linear system, and finally destroys
the KSP context. We emphasize the command KSPSetFromOptions(), which enables the user to
customize the linear solution method at runtime by using the options database, which is discussed
in Section 15.2. Through this database, the user not only can select an iterative method and
preconditioner, but also can prescribe the convergence tolerance, set various monitoring routines,
etc. (see, e.g., Figure 6).
Chapter 4describes in detail the KSP package, including the PC and KSP packages for precondi-
tioners and Krylov subspace methods.
Nonlinear Solvers
Most PDE problems of interest are inherently nonlinear. PETSc provides an interface to tackle the
nonlinear problems directly called SNES. Chapter 5describes the nonlinear solvers in detail. We
recommend most PETSc users work directly with SNES, rather than using PETSc for the linear
problem within a nonlinear solver.
32
![](asset-21.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
Error Checking
All PETSc routines return an integer indicating whether an error has occurred during the call.
The PETSc macro CHKERRQ(ierr) checks the value of ierr and calls the PETSc error handler
upon error detection. CHKERRQ(ierr) should be used in all subroutines to enable a complete error
traceback. In Figure 3we indicate a traceback generated by error detection within a sample PETSc
program. The error occurred on line 3618 of the file ${PETSC_DIR}/src/mat/impls/aij/seq/
aij.c and was caused by trying to allocate too large an array in memory. The routine was called
in the program ex3.c on line 66. See Section 11.1.2 for details regarding error checking when using
the PETSc Fortran interface.
$ mpiexec -n 1 ./ex3 -m 100000
[0]PETSC ERROR: --------------------- Error Message --------------------------------
[0]PETSC ERROR: Out of memory. This could be due to allocating
[0]PETSC ERROR: too large an object or bleeding by not properly
[0]PETSC ERROR: destroying unneeded objects.
[0]PETSC ERROR: Memory allocated 11282182704 Memory used by process 7075897344
[0]PETSC ERROR: Try running with -malloc_dump or -malloc_log for info.
[0]PETSC ERROR: Memory requested 18446744072169447424
[0]PETSC ERROR: See http://www.mcs.anl.gov/petsc/documentation/faq.html for trouble shooting.
[0]PETSC ERROR: Petsc Development GIT revision: v3.7.1-224-g9c9a9c5 GIT Date: 2016-05-18 22:43:00
-0500
[0]PETSC ERROR: ./ex3 on a arch-darwin-double-debug named Patricks-MacBook-Pro-2.local by patrick
Mon Jun 27 18:04:03 2016
[0]PETSC ERROR: Configure options PETSC_DIR=/Users/patrick/petsc PETSC_ARCH=arch-darwin-double-
debug --download-mpich --download-f2cblaslapack --with-cc=clang --with-cxx=clang++ --with-fc=
gfortran --with-debugging=1 --with-precision=double --with-scalar-type=real --with-viennacl=0
--download-c2html -download-sowing
[0]PETSC ERROR: #1 MatSeqAIJSetPreallocation_SeqAIJ() line 3618 in /Users/patrick/petsc/src/mat/
impls/aij/seq/aij.c
[0]PETSC ERROR: #2 PetscTrMallocDefault() line 188 in /Users/patrick/petsc/src/sys/memory/mtr.c
[0]PETSC ERROR: #3 MatSeqAIJSetPreallocation_SeqAIJ() line 3618 in /Users/patrick/petsc/src/mat/
impls/aij/seq/aij.c
[0]PETSC ERROR: #4 MatSeqAIJSetPreallocation() line 3562 in /Users/patrick/petsc/src/mat/impls/aij
/seq/aij.c
[0]PETSC ERROR: #5 main() line 66 in /Users/patrick/petsc/src/ksp/ksp/examples/tutorials/ex3.c
[0]PETSC ERROR: PETSc Option Table entries:
[0]PETSC ERROR: -m 100000
[0]PETSC ERROR: ----------------End of Error Message ------- send entire error message to petsc-
maint@mcs.anl.gov----------
Figure 3: Example of Error Traceback
When running the debug version of the PETSc libraries, it does a great deal of checking for
memory corruption (writing outside of array bounds etc). The macro CHKMEMQ can be called
anywhere in the code to check the current status of the memory for corruption. By putting several
(or many) of these macros into your code you can usually easily track down in what small segment
of your code the corruption has occured. One can also use Valgrind to track down memory errors;
see the FAQ at www.mcs.anl.gov/petsc/documentation/faq.html
Parallel Programming
Since PETSc uses the message-passing model for parallel programming and employs MPI for all
interprocessor communication, the user is free to employ MPI routines as needed throughout an
application code. However, by default the user is shielded from many of the details of message
33
![](asset-22.png)
1.5. SIMPLE PETSC EXAMPLES PETSc 3.8 March 24, 2018
passing within PETSc, since these are hidden within parallel objects, such as vectors, matrices,
and solvers. In addition, PETSc provides tools such as generalized vector scatters/gathers and
distributed arrays to assist in the management of parallel data.
Recall that the user must specify a communicator upon creation of any PETSc object (such as
a vector, matrix, or solver) to indicate the processors over which the object is to be distributed.
For example, as mentioned above, some commands for matrix, vector, and linear solver creation
are:
MatCreate(MPI Comm comm,Mat *A);
VecCreate(MPI Comm comm,Vec *x);
KSPCreate(MPI Comm comm,KSP *ksp);
The creation routines are collective over all processors in the communicator; thus, all processors in
the communicator must call the creation routine. In addition, if a sequence of collective routines is
being used, they must be called in the same order on each processor.
The next example, given in Figure 4, illustrates the solution of a linear system in parallel. This
code, corresponding to $PETSC_DIR/src/ksp/ksp/examples/tutorials/ex2.c, handles the two-
dimensional Laplacian discretized with finite differences, where the linear system is again solved
with KSP. The code performs the same tasks as the sequential version within Figure 2. Note
that the user interface for initiating the program, creating vectors and matrices, and solving the
linear system is exactly the same for the uniprocessor and multiprocessor examples. The primary
difference between the examples in Figures 2and 4is that each processor forms only its local part
of the matrix and vectors in the parallel case.
static char help[] = "Solves a linear system in parallel with KSP.\n\
Input parameters include:\n\
-random_exact_sol : use a random exact solution vector\n\
-view_exact_sol : write exact solution vector to stdout\n\
-m <mesh_x> : number of mesh points in x-direction\n\
-n <mesh_n> : number of mesh points in y-direction\n\n";
/*T
Concepts: KSP^basic parallel example;
Concepts: KSP^Laplacian, 2d
Concepts: Laplacian, 2d
Processors: n
T*/
/*
Include "petscksp.h" so that we can use KSP solvers. Note that this file
automatically includes:
petscsys.h - base PETSc routines petscvec.h - vectors
petscmat.h - matrices
petscis.h - index sets petscksp.h - Krylov subspace methods
petscviewer.h - viewers petscpc.h - preconditioners
*/
#include <petscksp.h>
34
![](asset-23.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
int main(int argc,char **args)
{
Vec x,b,u; /* approx solution, RHS, exact solution */
Mat A; /* linear system matrix */
KSP ksp; /* linear solver context */
PetscRandom rctx; /* random number generator context */
PetscReal norm; /* norm of solution error */
PetscInt i,j,Ii,J,Istart,Iend,m = 8,n = 7,its;
PetscErrorCode ierr;
PetscBool flg = PETSC FALSE;
PetscScalar v;
#if defined(PETSC_USE_LOG)
PetscLogStage stage;
#endif
ierr = PetscInitialize(&argc,&args,(char*)0,help);if (ierr) return ierr;
ierr = PetscOptionsGetInt(NULL,NULL,"-m",&m,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-n",&n,NULL);CHKERRQ(ierr);
/*----------------------------------
Compute the matrix and right-hand-side vector that define
the linear system, Ax = b.
----------------------------------*/
/*
Create parallel matrix, specifying only its global dimensions.
When using MatCreate(), the matrix format can be specified at
runtime. Also, the parallel partitioning of the matrix is
determined by PETSc at runtime.
Performance tuning note: For problems of substantial size,
preallocation of matrix memory is crucial for attaining good
performance. See the matrix chapter of the users manual for details.
*/
ierr = MatCreate(PETSC COMM WORLD,&A);CHKERRQ(ierr);
ierr = MatSetSizes(A,PETSC DECIDE,PETSC DECIDE,m*n,m*n);CHKERRQ(ierr);
ierr = MatSetFromOptions(A);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(A,5,NULL,5,NULL);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(A,5,NULL);CHKERRQ(ierr);
ierr = MatSeqSBAIJSetPreallocation(A,1,5,NULL);CHKERRQ(ierr);
ierr = MatMPISBAIJSetPreallocation(A,1,5,NULL,5,NULL);CHKERRQ(ierr);
/*
Currently, all PETSc parallel matrix formats are partitioned by
contiguous chunks of rows across the processors. Determine which
rows of the matrix are locally owned.
*/
ierr = MatGetOwnershipRange(A,&Istart,&Iend);CHKERRQ(ierr);
/*
Set matrix elements for the 2-D, five-point stencil in parallel.
35
![](asset-24.png)
1.5. SIMPLE PETSC EXAMPLES PETSc 3.8 March 24, 2018
- Each processor needs to insert only elements that it owns
locally (but any non-local elements will be sent to the
appropriate processor during matrix assembly).
- Always specify global rows and columns of matrix entries.
Note: this uses the less common natural ordering that orders first
all the unknowns for x = h then for x = 2h etc; Hence you see J = Ii +- n
instead of J = I +- m as you might expect. The more standard ordering
would first do all variables for y = h, then y = 2h etc.
*/
ierr = PetscLogStageRegister("Assembly", &stage);CHKERRQ(ierr);
ierr = PetscLogStagePush(stage);CHKERRQ(ierr);
for (Ii=Istart; Ii<Iend; Ii++) {
v = -1.0; i = Ii/n; j = Ii - i*n;
if (i>0) {J = Ii - n; ierr =
MatSetValues(A,1,&Ii,1,&J,&v,ADD VALUES);CHKERRQ(ierr);}
if (i<m-1) {J = Ii + n; ierr =
MatSetValues(A,1,&Ii,1,&J,&v,ADD VALUES);CHKERRQ(ierr);}
if (j>0) {J = Ii - 1; ierr =
MatSetValues(A,1,&Ii,1,&J,&v,ADD VALUES);CHKERRQ(ierr);}
if (j<n-1) {J = Ii + 1; ierr =
MatSetValues(A,1,&Ii,1,&J,&v,ADD VALUES);CHKERRQ(ierr);}
v = 4.0; ierr = MatSetValues(A,1,&Ii,1,&Ii,&v,ADD VALUES);CHKERRQ(ierr);
}
/*
Assemble matrix, using the 2-step process:
MatAssemblyBegin(), MatAssemblyEnd()
Computations can be done while messages are in transition
by placing code between these two statements.
*/
ierr = MatAssemblyBegin(A,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
ierr = PetscLogStagePop();CHKERRQ(ierr);
/* A is symmetric. Set symmetric flag to enable ICC/Cholesky preconditioner */
ierr = MatSetOption(A,MAT SYMMETRIC,PETSC TRUE);CHKERRQ(ierr);
/*
Create parallel vectors.
- We form 1 vector from scratch and then duplicate as needed.
- When using VecCreate(), VecSetSizes and VecSetFromOptions()
in this example, we specify only the
vector’s global dimension; the parallel partitioning is determined
at runtime.
- When solving a linear system, the vectors and matrices MUST
be partitioned accordingly. PETSc automatically generates
appropriately partitioned matrices and vectors when MatCreate()
36
![](asset-25.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
and VecCreate() are used with the same communicator.
- The user can alternatively specify the local vector and matrix
dimensions when more sophisticated partitioning is needed
(replacing the PETSC DECIDE argument in the VecSetSizes() statement
below).
*/
ierr = VecCreate(PETSC COMM WORLD,&u);CHKERRQ(ierr);
ierr = VecSetSizes(u,PETSC DECIDE,m*n);CHKERRQ(ierr);
ierr = VecSetFromOptions(u);CHKERRQ(ierr);
ierr = VecDuplicate(u,&b);CHKERRQ(ierr);
ierr = VecDuplicate(b,&x);CHKERRQ(ierr);
/*
Set exact solution; then compute right-hand-side vector.
By default we use an exact solution of a vector with all
elements of 1.0; Alternatively, using the runtime option
-random_sol forms a solution vector with random components.
*/
ierr =
PetscOptionsGetBool(NULL,NULL,"-random_exact_sol",&flg,NULL);CHKERRQ(ierr);
if (flg) {
ierr = PetscRandomCreate(PETSC COMM WORLD,&rctx);CHKERRQ(ierr);
ierr = PetscRandomSetFromOptions(rctx);CHKERRQ(ierr);
ierr = VecSetRandom(u,rctx);CHKERRQ(ierr);
ierr = PetscRandomDestroy(&rctx);CHKERRQ(ierr);
} else {
ierr = VecSet(u,1.0);CHKERRQ(ierr);
}
ierr = MatMult(A,u,b);CHKERRQ(ierr);
/*
View the exact solution vector if desired
*/
flg = PETSC FALSE;
ierr =
PetscOptionsGetBool(NULL,NULL,"-view_exact_sol",&flg,NULL);CHKERRQ(ierr);
if (flg) {ierr = VecView(u,PETSC VIEWER STDOUT WORLD);CHKERRQ(ierr);}
/*----------------------------------
Create the linear solver and set various options
----------------------------------*/
/*
Create linear solver context
*/
ierr = KSPCreate(PETSC COMM WORLD,&ksp);CHKERRQ(ierr);
/*
Set operators. Here the matrix that defines the linear system
37
![](asset-26.png)
1.5. SIMPLE PETSC EXAMPLES PETSc 3.8 March 24, 2018
also serves as the preconditioning matrix.
*/
ierr = KSPSetOperators(ksp,A,A);CHKERRQ(ierr);
/*
Set linear solver defaults for this problem (optional).
- By extracting the KSP and PC contexts from the KSP context,
we can then directly call any KSP and PC routines to set
various options.
- The following two statements are optional; all of these
parameters could alternatively be specified at runtime via
KSPSetFromOptions(). All of these defaults can be
overridden at runtime, as indicated below.
*/
ierr = KSPSetTolerances(ksp,1.e-2/((m+1)*(n+1)),1.e-50,PETSC DEFAULT,
PETSC DEFAULT);CHKERRQ(ierr);
/*
Set runtime options, e.g.,
-ksp_type <type> -pc_type <type> -ksp_monitor -ksp_rtol <rtol>
These options will override those specified above as long as
KSPSetFromOptions() is called _after_ any other customization
routines.
*/
ierr = KSPSetFromOptions(ksp);CHKERRQ(ierr);
/*----------------------------------
Solve the linear system
----------------------------------*/
ierr = KSPSolve(ksp,b,x);CHKERRQ(ierr);
/*----------------------------------
Check solution and clean up
----------------------------------*/
/*
Check the error
*/
ierr = VecAXPY(x,-1.0,u);CHKERRQ(ierr);
ierr = VecNorm(x,NORM 2,&norm);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(ksp,&its);CHKERRQ(ierr);
/*
Print convergence information. PetscPrintf() produces a single
print statement from all processes that share a communicator.
An alternative is PetscFPrintf(), which prints to a file.
*/
ierr = PetscPrintf(PETSC COMM WORLD,"Norm of error %g iterations
38
![](asset-27.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
%D\n",(double)norm,its);CHKERRQ(ierr);
/*
Free work space. All PETSc objects should be destroyed when they
are no longer needed.
*/
ierr = KSPDestroy(&ksp);CHKERRQ(ierr);
ierr = VecDestroy(&u);CHKERRQ(ierr); ierr = VecDestroy(&x);CHKERRQ(ierr);
ierr = VecDestroy(&b);CHKERRQ(ierr); ierr = MatDestroy(&A);CHKERRQ(ierr);
/*
Always call PetscFinalize() before exiting a program. This routine
- finalizes the PETSc libraries as well as MPI
- provides summary and diagnostic information if certain runtime
options are chosen (e.g., -log_view).
*/
ierr = PetscFinalize();
return ierr;
}
Figure 4: Example of Multiprocessor PETSc Code
Compiling and Running Programs
Figure 5illustrates compiling and running a PETSc program using MPICH on an OS X laptop.
Note that different machines will have compilation commands as determined by the configuration
process. See Chapter 16 for a discussion about compiling PETSc programs. Users who are ex-
periencing difficulties linking PETSc programs should refer to the FAQ on the PETSc website
http://www.mcs.anl.gov/petsc or given in the file $PETSC_DIR/docs/faq.html.
$ make ex2
/Users/patrick/petsc/arch-darwin-double-debug/bin/mpicc -o ex2.o -c -Wall -Wwrite-strings -Wno-
strict-aliasing -Wno-unknown-pragmas -Qunused-arguments -fvisibility=hidden -g3 -I/Users/
patrick/petsc/include -I/Users/patrick/petsc/arch-darwin-double-debug/include -I/opt/X11/
include -I/opt/local/include ‘pwd‘/ex2.c
/Users/patrick/petsc/arch-darwin-double-debug/bin/mpicc -Wl,-multiply_defined,suppress -Wl,-
multiply_defined -Wl,suppress -Wl,-commons,use_dylibs -Wl,-search_paths_first -Wl,-
multiply_defined,suppress -Wl,-multiply_defined -Wl,suppress -Wl,-commons,use_dylibs -Wl,-
search_paths_first -Wall -Wwrite-strings -Wno-strict-aliasing -Wno-unknown-pragmas -Qunused-
arguments -fvisibility=hidden -g3 -o ex2 ex2.o -Wl,-rpath,/Users/patrick/petsc/arch-darwin-
double-debug/lib -L/Users/patrick/petsc/arch-darwin-double-debug/lib -lpetsc -Wl,-rpath,/Users
/patrick/petsc/arch-darwin-double-debug/lib -lf2clapack -lf2cblas -Wl,-rpath,/opt/X11/lib -L/
opt/X11/lib -lX11 -lssl -lcrypto -Wl,-rpath,/Applications/Xcode.app/Contents/Developer/
Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/7.0.2/lib/darwin -L/Applications/Xcode.app/
Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/7.0.2/lib/darwin -lmpifort
-lgfortran -Wl,-rpath,/opt/local/lib/gcc5/gcc/x86_64-apple-darwin14/5.3.0 -L/opt/local/lib/
gcc5/gcc/x86_64-apple-darwin14/5.3.0 -Wl,-rpath,/opt/local/lib/gcc5 -L/opt/local/lib/gcc5 -
lgfortran -lgcc_ext.10.5 -lquadmath -lm -lclang_rt.osx -lmpicxx -lc++ -Wl,-rpath,/Applications/
Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/../lib/clang/7.0.2/lib
/darwin -L/Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/
bin/../lib/clang/7.0.2/lib/darwin -lclang_rt.osx -Wl,-rpath,/Users/patrick/petsc/arch-darwin-
double-debug/lib -L/Users/patrick/petsc/arch-darwin-double-debug/lib -ldl -lmpi -lpmpi -lSystem
-Wl,-rpath,/Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/
39
![](asset-28.png)
1.5. SIMPLE PETSC EXAMPLES PETSc 3.8 March 24, 2018
bin/../lib/clang/7.0.2/lib/darwin -L/Applications/Xcode.app/Contents/Developer/Toolchains/
XcodeDefault.xctoolchain/usr/bin/../lib/clang/7.0.2/lib/darwin -lclang_rt.osx -ldl
/bin/rm -f ex2.o
$ $PETSC_DIR/bin/petscmpiexec -n 1 ./ex2
Norm of error 0.000156044 iterations 6
$ $PETSC_DIR/bin/petscmpiexec -n 2 ./ex2
Norm of error 0.000411674 iterations 7
Figure 5: Running a PETSc Program
As shown in Figure 6, the option -log_view activates printing of a performance summary,
including times, floating point operation (flop) rates, and message-passing activity. Chapter 13
provides details about profiling, including interpretation of the output data within Figure 6. This
particular example involves the solution of a linear system on one processor using GMRES and
ILU. The low floating point operation (flop) rates in this example are due to the fact that the
code solved a tiny system. We include this example merely to demonstrate the ease of extracting
performance information.
$ $PETSC_DIR/bin/petscmpiexec -n 1 ./ex1 -n 1000 -pc_type ilu -ksp_type gmres -ksp_rtol 1.e-7 -log_view
...
------------------------------------------------------------------------------------------------------------------------
Event Count Time (sec) Flops --- Global --- --- Stage --- Total
Max Ratio Max Ratio Max Ratio Mess Avg len Reduct %T %F %M %L %R %T %F %M %L %R Mflop/s
------------------------------------------------------------------------------------------------------------------------
--- Event Stage 0: Main Stage
VecMDot 1 1.0 3.2830e-06 1.0 2.00e+03 1.0 0.0e+00 0.0e+00 0.0e+00 0 5 0 0 0 0 5 0 0 0 609
VecNorm 3 1.0 4.4550e-06 1.0 6.00e+03 1.0 0.0e+00 0.0e+00 0.0e+00 0 14 0 0 0 0 14 0 0 0 1346
VecScale 2 1.0 4.0110e-06 1.0 2.00e+03 1.0 0.0e+00 0.0e+00 0.0e+00 0 5 0 0 0 0 5 0 0 0 499
VecCopy 1 1.0 3.2280e-06 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0
VecSet 11 1.0 2.5537e-05 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 2 0 0 0 0 2 0 0 0 0 0
VecAXPY 2 1.0 2.0930e-06 1.0 4.00e+03 1.0 0.0e+00 0.0e+00 0.0e+00 0 10 0 0 0 0 10 0 0 0 1911
VecMAXPY 2 1.0 1.1280e-06 1.0 4.00e+03 1.0 0.0e+00 0.0e+00 0.0e+00 0 10 0 0 0 0 10 0 0 0 3546
VecNormalize 2 1.0 9.3970e-06 1.0 6.00e+03 1.0 0.0e+00 0.0e+00 0.0e+00 1 14 0 0 0 1 14 0 0 0 638
MatMult 2 1.0 1.1177e-05 1.0 9.99e+03 1.0 0.0e+00 0.0e+00 0.0e+00 1 24 0 0 0 1 24 0 0 0 894
MatSolve 2 1.0 1.9933e-05 1.0 9.99e+03 1.0 0.0e+00 0.0e+00 0.0e+00 1 24 0 0 0 1 24 0 0 0 501
MatLUFactorNum 1 1.0 3.5081e-05 1.0 4.00e+03 1.0 0.0e+00 0.0e+00 0.0e+00 2 10 0 0 0 2 10 0 0 0 114
MatILUFactorSym 1 1.0 4.4259e-05 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 3 0 0 0 0 3 0 0 0 0 0
MatAssemblyBegin 1 1.0 8.2015e-08 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0
MatAssemblyEnd 1 1.0 3.3536e-05 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 2 0 0 0 0 2 0 0 0 0 0
MatGetRowIJ 1 1.0 1.5960e-06 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0
MatGetOrdering 1 1.0 3.9791e-05 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 3 0 0 0 0 3 0 0 0 0 0
MatView 2 1.0 6.7909e-05 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 5 0 0 0 0 5 0 0 0 0 0
KSPGMRESOrthog 1 1.0 7.5970e-06 1.0 4.00e+03 1.0 0.0e+00 0.0e+00 0.0e+00 1 10 0 0 0 1 10 0 0 0 526
KSPSetUp 1 1.0 3.4424e-05 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 2 0 0 0 0 2 0 0 0 0 0
KSPSolve 1 1.0 2.7264e-04 1.0 3.30e+04 1.0 0.0e+00 0.0e+00 0.0e+00 19 79 0 0 0 19 79 0 0 0 121
PCSetUp 1 1.0 1.5234e-04 1.0 4.00e+03 1.0 0.0e+00 0.0e+00 0.0e+00 11 10 0 0 0 11 10 0 0 0 26
PCApply 2 1.0 2.1022e-05 1.0 9.99e+03 1.0 0.0e+00 0.0e+00 0.0e+00 1 24 0 0 0 1 24 0 0 0 475
------------------------------------------------------------------------------------------------------------------------
Memory usage is given in bytes:
Object Type Creations Destructions Memory Descendants’ Mem.
Reports information only for process 0.
--- Event Stage 0: Main Stage
Vector 8 8 76224 0.
Matrix 2 2 134212 0.
Krylov Solver 1 1 18400 0.
Preconditioner 1 1 1032 0.
Index Set 3 3 10328 0.
Viewer 1 0 0 0.
========================================================================================================================
...
Figure 6: Running a PETSc Program with Profiling (partial output)
40
![](asset-29.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
Writing Application Codes with PETSc
The examples throughout the library demonstrate the software usage and can serve as templates
for developing custom applications. We suggest that new PETSc users examine programs in the
directories ${PETSC_DIR}/src/<library>/examples/tutorials where <library> denotes any of
the PETSc libraries (listed in the following section), such as SNES or KSP. The manual pages located
at ${PETSC_DIR}/docs/index.htm or http://www.mcs.anl.gov/petsc/documentation provide links
(organized by both routine names and concepts) to the tutorial examples.
To write a new application program using PETSc, we suggest the following procedure:
1. Install and test PETSc according to the instructions at the PETSc web site.
2. Copy one of the many PETSc examples in the directory that corresponds to the class
of problem of interest (e.g., for linear solvers, see ${PETSC_DIR}/src/ksp/ksp/examples/
tutorials).
3. Copy the corresponding makefile within the example directory; compile and run the example
program.
4. Use the example program as a starting point for developing a custom code.
1.6 Citing PETSc
When citing PETSc in a publication please cite the following:
@Mi=sc{petsc-web-page,
Author = "Satish Balay and Shrirang Abhyankar and Mark~F. Adams and Jed Brown
and Peter Brune and Kris Buschelman and Lisandro Dalcin and Victor Eijkhout
and William~D. Gropp and Dinesh Kaushik and Matthew~G. Knepley and Lois Curfman McInnes
and Karl Rupp and Barry~F. Smith and Stefano Zampini
and Hong Zhang and Hong Zhang",
Title = "{PETS}c {W}eb page",
Note = "http://www.mcs.anl.gov/petsc",
Year = "2017"}
@TechReport{petsc-user-ref,
Author = "Satish Balay and Shrirang Abhyankar and Mark~F. Adams and Jed Brown
and Peter Brune and Kris Buschelman and Lisandro Dalcin and Victor Eijkhout
and Dinesh Kaushik and Matthew~G. Knepley and Dave~A. May and Lois Curfman McInnes
and William~D. Gropp and Karl Rupp and Patrick Sanan and Barry~F. Smith and Stefano Zampini
and Hong Zhang and Hong Zhang",
Title = "PETSc Users Manual",
Number = "ANL-95/11 - Revision 3.8",
Institution = "Argonne National Laboratory",
Year = "2017"}
@InProceedings{petsc-efficient,
Author = "Satish Balay and William D. Gropp and Lois C. McInnes and Barry F. Smith",
Title = "Efficienct Management of Parallelism in Object Oriented
Numerical Software Libraries",
Booktitle = "Modern Software Tools in Scientific Computing",
Editor = "E. Arge and A. M. Bruaset and H. P. Langtangen",
Pages = "163--202",
41
![](asset-2a.png)
1.7. DIRECTORY STRUCTURE PETSc 3.8 March 24, 2018
Publisher = "Birkhauser Press",
Year = "1997"}
1.7 Directory Structure
We conclude this introduction with an overview of the organization of the PETSc software. The
root directory of PETSc contains the following directories:
•docs - All documentation for PETSc. The files manual.pdf contains the hyperlinked users
manual, suitable for printing or on-screen viewering. Includes the subdirectory
-manualpages (on-line manual pages).
•bin - Utilities and short scripts for use with PETSc, including
•petscmpiexec (utility for setting running MPI jobs),
•conf - Base PETSc configuration files that define the standard make variables and rules used
by PETSc
•include - All include files for PETSc that are visible to the user.
•include/petsc/finclude - PETSc include files for Fortran programmers using the .F suffix
(recommended).
•include/petsc/private - Private PETSc include files that should not need to be used by
application programmers.
•share - Some small test matrices in data files
•src - The source code for all PETSc libraries, which currently includes
•vec - vectors,
•is - index sets,
•mat - matrices,
•dm - data management between meshes and vectors and matrices,
•ksp - complete linear equations solvers,
•ksp - Krylov subspace accelerators,
•pc - preconditioners,
•snes - nonlinear solvers
•ts - ODE solvers and timestepping,
•sys - general system-related routines,
•logging - PETSc logging and profiling routines,
•classes - low-level classes
•draw - simple graphics,
•viewer
•bag
•random - random number generators.
•contrib - contributed modules that use PETSc but are not part of the official PETSc
package. We encourage users who have developed such code that they wish to share
with others to let us know by writing to petsc-maint@mcs.anl.gov.
Each PETSc source code library directory has the following subdirectories:
42
![](asset-2b.png)
CHAPTER 1. GETTING STARTED PETSc 3.8 March 24, 2018
•examples - Example programs for the component, including
•tutorials - Programs designed to teach users about PETSc. These codes can serve as
templates for the design of custom applications.
•tests - Programs designed for thorough testing of PETSc. As such, these codes are not
intended for examination by users.
•interface - The calling sequences for the abstract interface to the component. Code here
does not know about particular implementations.
•impls - Source code for one or more implementations.
•utils - Utility routines. Source here may know about the implementations, but ideally will
not know about implementations for other components.
43
![](asset-2c.png)
1.7. DIRECTORY STRUCTURE PETSc 3.8 March 24, 2018
44
Part II
Programming with PETSc
45
![](asset-2f.png)
Chapter 2
Vectors and Parallel Data
The vector (denoted by Vec) is one of the simplest PETSc objects. Vectors are used to store discrete
PDE solutions, right-hand sides for linear systems, etc. This chapter is organized as follows:
•(Vec) Sections 2.1 and 2.2 - basic usage of vectors
•Section 2.3 - management of the various numberings of degrees of freedom, vertices, cells, etc.
•(AO) Mapping between different global numberings
•(ISLocalToGlobalMapping) Mapping between local and global numberings
•(DM) Section 2.4 - management of grids
•(IS,VecScatter) Section 2.5 - management of vectors related to unstructured grids
2.1 Creating and Assembling Vectors
PETSc currently provides two basic vector types: sequential and parallel (MPI-based). To create
a sequential vector with mcomponents, one can use the command
VecCreateSeq(PETSC COMM SELF,PetscInt m,Vec *x);
To create a parallel vector one can either specify the number of components that will be stored on
each process or let PETSc decide. The command
VecCreateMPI(MPI Comm comm,PetscInt m,PetscInt M,Vec *x);
creates a vector distributed over all processes in the communicator, comm, where mindicates the
number of components to store on the local process, and Mis the total number of vector com-
ponents. Either the local or global dimension, but not both, can be set to PETSC_DECIDE or
PETSC_DETERMINE, respectively, to indicate that PETSc should decide or determine it. More
generally, one can use the routines
VecCreate(MPI Comm comm,Vec *v);
VecSetSizes(Vec v, PetscInt m, PetscInt M);
VecSetFromOptions(Vec v);
which automatically generates the appropriate vector type (sequential or parallel) over all pro-
cesses in comm. The option -vec_type mpi can be used in conjunction with VecCreate() and
VecSetFromOptions() to specify the use of MPI vectors even for the uniprocessor case.
47
![](asset-30.png)
2.1. CREATING AND ASSEMBLING VECTORS PETSc 3.8 March 24, 2018
We emphasize that all processes in comm must call the vector creation routines, since these rou-
tines are collective over all processes in the communicator. If you are not familiar with MPI commu-
nicators, see the discussion in Section 1.4 on page 25. In addition, if a sequence of VecCreateXXX()
routines is used, they must be called in the same order on each process in the communicator.
One can assign a single value to all components of a vector with the command
VecSet(Vec x,PetscScalar value);
Assigning values to individual components of the vector is more complicated, in order to make it
possible to write efficient parallel code. Assigning a set of components is a two-step process: one
first calls
VecSetValues(Vec x,PetscInt n,PetscInt *indices,PetscScalar
*values,INSERT VALUES);
any number of times on any or all of the processes. The argument ngives the number of components
being set in this insertion. The integer array indices contains the global component indices, and
values is the array of values to be inserted. Any process can set any components of the vector;
PETSc ensures that they are automatically stored in the correct location. Once all of the values
have been inserted with VecSetValues(), one must call
VecAssemblyBegin(Vec x);
followed by
VecAssemblyEnd(Vec x);
to perform any needed message passing of nonlocal components. In order to allow the overlap of
communication and calculation, the user’s code can perform any series of other actions between
these two calls while the messages are in transition.
Example usage of VecSetValues() may be found in $PETSC_DIR/src/vec/vec/examples/
tutorials/ex2.c or ex2f.F.
Often, rather than inserting elements in a vector, one may wish to add values. This process is
also done with the command
VecSetValues(Vec x,PetscInt n,PetscInt *indices, PetscScalar *values,ADD VALUES);
Again one must call the assembly routines VecAssemblyBegin() and VecAssemblyEnd() after all
of the values have been added. Note that addition and insertion calls to VecSetValues() cannot
be mixed. Instead, one must add and insert vector elements in phases, with intervening calls to
the assembly routines. This phased assembly procedure overcomes the nondeterministic behavior
that would occur if two different processes generated values for the same location, with one process
adding while the other is inserting its value. (In this case the addition and insertion actions could be
performed in either order, thus resulting in different values at the particular location. Since PETSc
does not allow the simultaneous use of INSERT_VALUES and ADD_VALUES this nondeterministic
behavior will not occur in PETSc.)
You can called VecGetValues() to pull local values from a vector (but not off-process values),
an alternative method for extracting some components of a vector are the vector scatter routines.
See Section 2.5.2 for details; see also below for VecGetArray().
One can examine a vector with the command
VecView(Vec x,PetscViewer v);
48
![](asset-31.png)
CHAPTER 2. VECTORS AND PARALLEL DATA PETSc 3.8 March 24, 2018
To print the vector to the screen, one can use the viewer PETSC_VIEWER_STDOUT_WORLD, which
ensures that parallel vectors are printed correctly to stdout. To display the vector in an X-
window, one can use the default X-windows viewer PETSC_VIEWER_DRAW_WORLD, or one can create
a viewer with the routine PetscViewerDrawOpenX(). A variety of viewers are discussed further in
Section 15.3.
To create a new vector of the same format as an existing vector, one uses the command
VecDuplicate(Vec old,Vec *new);
To create several new vectors of the same format as an existing vector, one uses the command
VecDuplicateVecs(Vec old,PetscInt n,Vec **new);
This routine creates an array of pointers to vectors. The two routines are very useful because they
allow one to write library code that does not depend on the particular format of the vectors being
used. Instead, the subroutines can automatically correctly create work vectors based on the specified
existing vector. As discussed in Section 11.1.5, the Fortran interface for VecDuplicateVecs() differs
slightly.
When a vector is no longer needed, it should be destroyed with the command
VecDestroy(Vec *x);
To destroy an array of vectors, use the command
VecDestroyVecs(PetscInt n,Vec **vecs);
Note that the Fortran interface for VecDestroyVecs() differs slightly, as described in Section 11.1.5.
It is also possible to create vectors that use an array provided by the user, rather than having
PETSc internally allocate the array space. Such vectors can be created with the routines
VecCreateSeqWithArray(PETSC COMM SELF,PetscInt bs,PetscInt n,PetscScalar
*array,Vec *V);
and
VecCreateMPIWithArray(MPI Comm comm,PetscInt bs,PetscInt n,PetscInt
N,PetscScalar *array,Vec *vv);
Note that here one must provide the value n; it cannot be PETSC_DECIDE and the user is responsible
for providing enough space in the array; n*sizeof(PetscScalar).
2.2 Basic Vector Operations
As listed in Table 1, we have chosen certain basic vector operations to support within the PETSc
vector library. These operations were selected because they often arise in application codes. The
NormType argument to VecNorm() is one of NORM_1,NORM_2, or NORM_INFINITY. The 1-norm is
Pi|xi|, the 2-norm is (Pix2
i)1/2and the infinity norm is maxi|xi|.
For parallel vectors that are distributed across the processes by ranges, it is possible to determine
a process’s local range with the routine
VecGetOwnershipRange(Vec vec,PetscInt *low,PetscInt *high);
The argument low indicates the first component owned by the local process, while high specifies one
more than the last owned by the local process. This command is useful, for instance, in assembling
parallel vectors.
49
![](asset-32.png)
2.2. BASIC VECTOR OPERATIONS PETSc 3.8 March 24, 2018
Function Name Operation
VecAXPY(Vec y,PetscScalar a,Vec x); y=y+a∗x
VecAYPX(Vec y,PetscScalar a,Vec x); y=x+a∗y
VecWAXPY(Vec w,PetscScalar a,Vec x,Vec y); w=a∗x+y
VecAXPBY(Vec y,PetscScalar a,PetscScalar b,Vec x); y=a∗x+b∗y
VecScale(Vec x, PetscScalar a); x=a∗x
VecDot(Vec x, Vec y, PetscScalar *r); r= ¯x0∗y
VecTDot(Vec x, Vec y, PetscScalar *r); r=x0∗y
VecNorm(Vec x,NormType type, PetscReal *r); r=||x||type
VecSum(Vec x, PetscScalar *r); r=Pxi
VecCopy(Vec x, Vec y); y=x
VecSwap(Vec x, Vec y); y=xwhile x=y
VecPointwiseMult(Vec w,Vec x,Vec y); wi=xi∗yi
VecPointwiseDivide(Vec w,Vec x,Vec y); wi=xi/yi
VecMDot(Vec x,PetscInt n,Vec y[],PetscScalar *r); r[i] = ¯x0∗y[i]
VecMTDot(Vec x,PetscInt n,Vec y[],PetscScalar *r); r[i] = x0∗y[i]
VecMAXPY(Vec y,PetscInt n, PetscScalar *a, Vec x[]); y=y+Piai∗x[i]
VecMax(Vec x, PetscInt *idx, PetscReal *r); r= max xi
VecMin(Vec x, PetscInt *idx, PetscReal *r); r= min xi
VecAbs(Vec x); xi=|xi|
VecReciprocal(Vec x); xi= 1/xi
VecShift(Vec x,PetscScalar s); xi=s+xi
VecSet(Vec x,PetscScalar alpha); xi=α
Table 1: PETSc Vector Operations
On occasion, the user needs to access the actual elements of the vector. The routine VecGetArray()
returns a pointer to the elements local to the process:
VecGetArray(Vec v,PetscScalar **array);
When access to the array is no longer needed, the user should call
VecRestoreArray(Vec v, PetscScalar **array);
If the values do not need to be modified, the routines VecGetArrayRead() and
VecRestoreArrayRead() provide read-only access and should be used instead.
VecGetArrayRead(Vec v, const PetscScalar **array);
VecRestoreArrayRead(Vec v, const PetscScalar **array);
Minor differences exist in the Fortran interface for VecGetArray() and VecRestoreArray(), as
discussed in Section 11.2.1. It is important to note that VecGetArray() and VecRestoreArray()
do not copy the vector elements; they merely give users direct access to the vector elements. Thus,
these routines require essentially no time to call and can be used efficiently.
The number of elements stored locally can be accessed with
VecGetLocalSize(Vec v,PetscInt *size);
The global vector length can be determined by
VecGetSize(Vec v,PetscInt *size);
50
![](asset-33.png)
CHAPTER 2. VECTORS AND PARALLEL DATA PETSc 3.8 March 24, 2018
In addition to VecDot() and VecMDot() and VecNorm(), PETSc provides split phase versions of
these that allow several independent inner products and/or norms to share the same communication
(thus improving parallel efficiency). For example, one may have code such as
VecDot(Vec x,Vec y,PetscScalar *dot);
VecMDot(Vec x,PetscInt nv, Vec y[],PetscScalar *dot);
VecNorm(Vec x,NormType NORM 2,PetscReal *norm2);
VecNorm(Vec x,NormType NORM 1,PetscReal *norm1);
This code works fine, but it performs three separate parallel communication operations. Instead,
one can write
VecDotBegin(Vec x,Vec y,PetscScalar *dot);
VecMDotBegin(Vec x, PetscInt nv,Vec y[],PetscScalar *dot);
VecNormBegin(Vec x,NormType NORM 2,PetscReal *norm2);
VecNormBegin(Vec x,NormType NORM 1,PetscReal *norm1);
VecDotEnd(Vec x,Vec y,PetscScalar *dot);
VecMDotEnd(Vec x, PetscInt nv,Vec y[],PetscScalar *dot);
VecNormEnd(Vec x,NormType NORM 2,PetscReal *norm2);
VecNormEnd(Vec x,NormType NORM 1,PetscReal *norm1);
With this code, the communication is delayed until the first call to VecxxxEnd() at which a single
MPI reduction is used to communicate all the required values. It is required that the calls to
the VecxxxEnd() are performed in the same order as the calls to the VecxxxBegin(); however,
if you mistakenly make the calls in the wrong order, PETSc will generate an error informing you
of this. There are additional routines VecTDotBegin() and VecTDotEnd(),VecMTDotBegin(),
VecMTDotEnd().
Note: these routines use only MPI-1 functionality; they do not allow you to overlap computa-
tion and communication (assuming no threads are spawned within a MPI process). Once MPI-2
implementations are more common we’ll improve these routines to allow overlap of inner product
and norm calculations with other calculations. Also currently these routines only work for the
PETSc built in vector types.
2.3 Indexing and Ordering
When writing parallel PDE codes, there is extra complexity caused by having multiple ways of
indexing (numbering) and ordering objects such as vertices and degrees of freedom. For example,
a grid generator or partitioner may renumber the nodes, requiring adjustment of the other data
structures that refer to these objects; see Figure 8. In addition, local numbering (on a single process)
of objects may be different than the global (cross-process) numbering. PETSc provides a variety
of tools to help to manage the mapping amongst the various numbering systems. The two most
basic are the AO (application ordering), which enables mapping between different global (cross-
process) numbering schemes and the ISLocalToGlobalMapping, which allows mapping between
local (on-process) and global (cross-process) numbering.
2.3.1 Application Orderings
In many applications it is desirable to work with one or more “orderings” (or numberings) of degrees
of freedom, cells, nodes, etc. Doing so in a parallel environment is complicated by the fact that each
process cannot keep complete lists of the mappings between different orderings. In addition, the
51
![](asset-34.png)
2.3. INDEXING AND ORDERING PETSc 3.8 March 24, 2018
orderings used in the PETSc linear algebra routines (often contiguous ranges) may not correspond
to the “natural” orderings for the application.
PETSc provides certain utility routines that allow one to deal cleanly and efficiently with the
various orderings. To define a new application ordering (called an AO in PETSc), one can call the
routine
AOCreateBasic(MPI Comm comm,PetscInt n,const PetscInt apordering[],const
PetscInt petscordering[],AO *ao);
The arrays apordering and petscordering, respectively, contain a list of integers in the application
ordering and their corresponding mapped values in the PETSc ordering. Each process can provide
whatever subset of the ordering it chooses, but multiple processes should never contribute duplicate
values. The argument nindicates the number of local contributed values.
For example, consider a vector of length 5, where node 0 in the application ordering corre-
sponds to node 3 in the PETSc ordering. In addition, nodes 1, 2, 3, and 4 of the application
ordering correspond, respectively, to nodes 2, 1, 4, and 0 of the PETSc ordering. We can write this
correspondence as
{0,1,2,3,4}→{3,2,1,4,0}.
The user can create the PETSc AO mappings in a number of ways. For example, if using two
processes, one could call
AOCreateBasic(PETSC COMM WORLD,2,{0,3},{3,4},&ao);
on the first process and
AOCreateBasic(PETSC COMM WORLD,3,{1,2,4},{2,1,0},&ao);
on the other process.
Once the application ordering has been created, it can be used with either of the commands
AOPetscToApplication(AO ao,PetscInt n,PetscInt *indices);
AOApplicationToPetsc(AO ao,PetscInt n,PetscInt *indices);
Upon input, the n-dimensional array indices specifies the indices to be mapped, while upon output,
indices contains the mapped values. Since we, in general, employ a parallel database for the AO
mappings, it is crucial that all processes that called AOCreateBasic() also call these routines; these
routines cannot be called by just a subset of processes in the MPI communicator that was used in
the call to AOCreateBasic().
An alternative routine to create the application ordering, AO, is
AOCreateBasicIS(IS apordering,IS petscordering,AO *ao);
where index sets (see 2.5.1) are used instead of integer arrays.
The mapping routines
AOPetscToApplicationIS(AO ao,IS indices);
AOApplicationToPetscIS(AO ao,IS indices);
will map index sets (IS objects) between orderings. Both the AOXxxToYyy() and AOXxxToYyyIS()
routines can be used regardless of whether the AO was created with a AOCreateBasic() or
AOCreateBasicIS().
The AO context should be destroyed with AODestroy(AO *ao) and viewed with
AOView(AO ao,PetscViewer viewer).
52
![](asset-35.png)
CHAPTER 2. VECTORS AND PARALLEL DATA PETSc 3.8 March 24, 2018
Although we refer to the two orderings as “PETSc” and “application” orderings, the user is
free to use them both for application orderings and to maintain relationships among a variety of
orderings by employing several AO contexts.
The AOxxToxx() routines allow negative entries in the input integer array. These entries are not
mapped; they simply remain unchanged. This functionality enables, for example, mapping neighbor
lists that use negative numbers to indicate nonexistent neighbors due to boundary conditions, etc.
2.3.2 Local to Global Mappings
In many applications one works with a global representation of a vector (usually on a vector
obtained with VecCreateMPI()) and a local representation of the same vector that includes ghost
points required for local computation. PETSc provides routines to help map indices from a local
numbering scheme to the PETSc global numbering scheme. This is done via the following routines
ISLocalToGlobalMappingCreate(MPI Comm comm,PetscInt bs,PetscInt N,PetscInt*
globalnum,PetscCopyMode mode,ISLocalToGlobalMapping* ctx);
ISLocalToGlobalMappingApply(ISLocalToGlobalMapping ctx,PetscInt n,PetscInt
*in,PetscInt *out);
ISLocalToGlobalMappingApplyIS(ISLocalToGlobalMapping ctx,IS isin,IS* isout);
ISLocalToGlobalMappingDestroy(ISLocalToGlobalMapping *ctx);
Here Ndenotes the number of local indices, globalnum contains the global number of each local
number, and ISLocalToGlobalMapping is the resulting PETSc object that contains the information
needed to apply the mapping with either ISLocalToGlobalMappingApply() or
ISLocalToGlobalMappingApplyIS().
Note that the ISLocalToGlobalMapping routines serve a different purpose than the AO routines.
In the former case they provide a mapping from a local numbering scheme (including ghost points)
to a global numbering scheme, while in the latter they provide a mapping between two global
numbering schemes. In fact, many applications may use both AO and ISLocalToGlobalMapping
routines. The AO routines are first used to map from an application global ordering (that has no
relationship to parallel processing etc.) to the PETSc ordering scheme (where each process has
a contiguous set of indices in the numbering). Then in order to perform function or Jacobian
evaluations locally on each process, one works with a local numbering scheme that includes ghost
points. The mapping from this local numbering scheme back to the global PETSc numbering can
be handled with the ISLocalToGlobalMapping routines.
If one is given a list of block indices in a global numbering, the routine
ISGlobalToLocalMappingApplyBlock(ISLocalToGlobalMapping
ctx,ISGlobalToLocalMappingMode type,PetscInt nin,PetscInt idxin[],PetscInt
*nout,PetscInt idxout[]);
will provide a new list of indices in the local numbering. Again, negative values in idxin are left
unmapped. But, in addition, if type is set to IS_GTOLM_MASK , then nout is set to nin and all
global values in idxin that are not represented in the local to global mapping are replaced by -1.
When type is set to IS_GTOLM_DROP, the values in idxin that are not represented locally in the
mapping are not included in idxout, so that potentially nout is smaller than nin. One must pass in
an array long enough to hold all the indices. One can call ISGlobalToLocalMappingApplyBlock()
with idxout equal to NULL to determine the required length (returned in nout) and then allocate the
required space and call ISGlobalToLocalMappingApplyBlock() a second time to set the values.
Often it is convenient to set elements into a vector using the local node numbering rather than
the global node numbering (e.g., each process may maintain its own sublist of vertices and elements
53
![](asset-36.png)
2.4. STRUCTURED GRIDS USING DISTRIBUTED ARRAYS PETSc 3.8 March 24, 2018
Box-type stencil Star-type stencil
Proc 6
Proc 0 Proc 0Proc 1 Proc 1
Proc 6
Figure 7: Ghost Points for Two Stencil Types on the Seventh Process
and number them locally). To set values into a vector with the local numbering, one must first call
VecSetLocalToGlobalMapping(Vec v,ISLocalToGlobalMapping ctx);
and then call
VecSetValuesLocal(Vec x,PetscInt n,const PetscInt indices[],const PetscScalar
values[],INSERT VALUES);
Now the indices use the local numbering, rather than the global, meaning the entries lie in [0, n)
where nis the local size of the vector.
2.4 Structured Grids Using Distributed Arrays
Distributed arrays (DMDAs), which are used in conjunction with PETSc vectors, are intended for
use with logically regular rectangular grids when communication of nonlocal data is needed before
certain local computations can occur. PETSc distributed arrays are designed only for the case in
which data can be thought of as being stored in a standard multidimensional array; thus, DMDAs are
not intended for parallelizing unstructured grid problems, etc. DAs are intended for communicating
vector (field) information; they are not intended for storing matrices.
For example, a typical situation one encounters in solving PDEs in parallel is that, to evaluate
a local function, f(x), each process requires its local portion of the vector xas well as its ghost
points (the bordering portions of the vector that are owned by neighboring processes). Figure 7
illustrates the ghost points for the seventh process of a two-dimensional, regular parallel grid. Each
box represents a process; the ghost points for the seventh process’s local part of a parallel array
are shown in gray.
2.4.1 Creating Distributed Arrays
The PETSc DMDA object manages the parallel communication required while working with data
stored in regular arrays. The actual data is stored in appropriately sized vector objects; the DMDA
object only contains the parallel data layout information and communication information, however
it may be used to create vectors and matrices with the proper layout.
One creates a distributed array communication data structure in two dimensions with the
command
54
![](asset-37.png)
CHAPTER 2. VECTORS AND PARALLEL DATA PETSc 3.8 March 24, 2018
DMDACreate2d(MPI Comm comm,DMBoundaryType xperiod,DMBoundaryType
yperiod,DMDAStencilType st,PetscInt M, PetscInt N,PetscInt m,PetscInt
n,PetscInt dof,PetscInt s,PetscInt *lx,PetscInt *ly,DM *da);
The arguments Mand Nindicate the global numbers of grid points in each direction, while mand
ndenote the process partition in each direction; m*n must equal the number of processes in the
MPI communicator, comm. Instead of specifying the process layout, one may use PETSC_DECIDE
for mand nso that PETSc will determine the partition using MPI. The type of periodicity of
the array is specified by xperiod and yperiod, which can be DM_BOUNDARY_NONE (no periodic-
ity), DM_BOUNDARY_PERIODIC (periodic in that direction), DM_BOUNDARY_TWIST (periodic in that
direction, but identified in reverse order), DM_BOUNDARY_GHOSTED , or DM_BOUNDARY_MIRROR. The
argument dof indicates the number of degrees of freedom at each array point, and sis the stencil
width (i.e., the width of the ghost point region). The optional arrays lx and ly may contain the
number of nodes along the x and y axis for each cell, i.e. the dimension of lx is mand the dimension
of ly is n; alternately, NULL may be passed in.
Two types of distributed array communication data structures can be created, as specified by
st. Star-type stencils that radiate outward only in the coordinate directions are indicated by
DMDA_STENCIL_STAR, while box-type stencils are specified by DMDA_STENCIL_BOX. For example,
for the two-dimensional case, DMDA_STENCIL_STAR with width 1 corresponds to the standard 5-
point stencil, while DMDA_STENCIL_BOX with width 1 denotes the standard 9-point stencil. In
both instances the ghost points are identical, the only difference being that with star-type stencils
certain ghost points are ignored, decreasing substantially the number of messages sent. Note that
the DMDA_STENCIL_STAR stencils can save interprocess communication in two and three dimensions.
These DMDA stencils have nothing directly to do with any finite difference stencils one might
chose to use for a discretization; they only ensure that the correct values are in place for application
of a user-defined finite difference stencil (or any other discretization technique).
The commands for creating distributed array communication data structures in one and three
dimensions are analogous:
DMDACreate1d(MPI Comm comm,DMBoundaryType xperiod,PetscInt M,PetscInt w,PetscInt
s,PetscInt *lc,DM *inra);
DMDACreate3d=(MPI Comm comm,DMBoundaryType xperiod,DMBoundaryType
yperiod,DMBoundaryType zperiod, DMDAStencilType stencil_type,PetscInt
M,PetscInt N,PetscInt P,PetscInt m,PetscInt n,PetscInt p,PetscInt w,PetscInt
s,PetscInt *lx,PetscInt *ly,PetscInt *lz,DM *inra);
The routines to create distributed arrays are collective, so that all processes in the communicator
comm must call DACreateXXX().
2.4.2 Local/Global Vectors and Scatters
Each DMDA object defines the layout of two vectors: a distributed global vector and a local vector
that includes room for the appropriate ghost points. The DMDA object provides information about
the size and layout of these vectors, but does not internally allocate any associated storage space
for field values. Instead, the user can create vector objects that use the DMDA layout information
with the routines
DMCreateGlobalVector(DM da,Vec *g);
DMCreateLocalVector(DM da,Vec *l);
55
![](asset-38.png)
2.4. STRUCTURED GRIDS USING DISTRIBUTED ARRAYS PETSc 3.8 March 24, 2018
These vectors will generally serve as the building blocks for local and global PDE solutions, etc.
If additional vectors with such layout information are needed in a code, they can be obtained by
duplicating lor gvia VecDuplicate() or VecDuplicateVecs().
We emphasize that a distributed array provides the information needed to communicate the
ghost value information between processes. In most cases, several different vectors can share the
same communication information (or, in other words, can share a given DMDA). The design of the
DMDA object makes this easy, as each DMDA operation may operate on vectors of the appropriate
size, as obtained via DMCreateLocalVector() and DMCreateGlobalVector() or as produced by
VecDuplicate(). As such, the DMDA scatter/gather operations (e.g., DMGlobalToLocalBegin())
require vector input/output arguments, as discussed below.
PETSc currently provides no container for multiple arrays sharing the same distributed array
communication; note, however, that the dof parameter handles many cases of interest.
At certain stages of many applications, there is a need to work on a local portion of the vector,
including the ghost points. This may be done by scattering a global vector into its local parts by
using the two-stage commands
DMGlobalToLocalBegin(DM da,Vec g,InsertMode iora,Vec l);
DMGlobalToLocalEnd(DM da,Vec g,InsertMode iora,Vec l);
which allow the overlap of communication and computation. Since the global and local vec-
tors, given by gand l, respectively, must be compatible with the distributed array, da, they
should be generated by DMCreateGlobalVector() and DMCreateLocalVector() (or be duplicates
of such a vector obtained via VecDuplicate()). The InsertMode can be either ADD_VALUES or
INSERT_VALUES.
One can scatter the local patches into the distributed vector with the commands
DMLocalToGlobalBegin(DM da,Vec l,InsertMode mode,Vec g);
DMLocalToGlobalEnd(DM da,Vec l,InsertMode mode,Vec g);
In general this is used with an InsertMode of ADD_VALUES, because if one wishes to insert values
into the global vector they should just access the global vector directly and put in the values.
A third type of distributed array scatter is from a local vector (including ghost points that
contain irrelevant values) to a local vector with correct ghost point values. This scatter may be
done with the commands
DMDALocalToLocalBegin(DM da,Vec l1,InsertMode iora,Vec l2);
DMDALocalToLocalEnd(DM da,Vec l1,InsertMode iora,Vec l2);
Since both local vectors, l1 and l2, must be compatible with the distributed array, da, they
should be generated by DMCreateLocalVector() (or be duplicates of such vectors obtained via
VecDuplicate()). The InsertMode can be either ADD_VALUES or INSERT_VALUES.
It is possible to directly access the vector scatter contexts (see below) used in the local-to-global
(ltog), global-to-local (gtol), and local-to-local (ltol) scatters with the command
DMDAGetScatter(DM da,VecScatter *ltog,VecScatter *gtol,VecScatter *ltol);
Most users should not need to use these contexts.
2.4.3 Local (Ghosted) Work Vectors
In most applications the local ghosted vectors are only needed during user “function evaluations”.
PETSc provides an easy, light-weight (requiring essentially no CPU time) way to obtain these work
vectors and return them when they are no longer needed. This is done with the routines
56
![](asset-39.png)
CHAPTER 2. VECTORS AND PARALLEL DATA PETSc 3.8 March 24, 2018
DMGetLocalVector(DM da,Vec *l);
... use the local vector l ...
DMRestoreLocalVector(DM da,Vec *l);
2.4.4 Accessing the Vector Entries for DMDA Vectors
PETSc provides an easy way to set values into the DMDA Vectors and access them using the
natural grid indexing. This is done with the routines
DMDAVecGetArray(DM da,Vec l,void *array);
... use the array indexing it with 1 or 2 or 3 dimensions ...
... depending on the dimension of the DMDA ...
DMDAVecRestoreArray(DM da,Vec l,void *array);
DMDAVecGetArrayRead(DM da,Vec l,void *array);
... use the array indexing it with 1 or 2 or 3 dimensions ...
... depending on the dimension of the DMDA ...
DMDAVecRestoreArrayRead(DM da,Vec l,void *array);
and
DMDAVecGetArrayDOF(DM da,Vec l,void *array);
... use the array indexing it with 1 or 2 or 3 dimensions ...
... depending on the dimension of the DMDA ...
DMDAVecRestoreArrayDOF(DM da,Vec l,void *array);
DMDAVecGetArrayDOFRead(DM da,Vec l,void *array);
... use the array indexing it with 1 or 2 or 3 dimensions ...
... depending on the dimension of the DMDA ...
DMDAVecRestoreArrayDOFRead(DM da,Vec l,void *array);
where array is a multidimensional C array with the same dimension as da. The vector lcan be
either a global vector or a local vector. The array is accessed using the usual global indexing on
the entire grid, but the user may only refer to the local and ghost entries of this array as all other
entries are undefined. For example, for a scalar problem in two dimensions one could use
PetscScalar **f,**u;
...
DMDAVecGetArray(DM da,Vec local,&u);
DMDAVecGetArray(DM da,Vec global,&f);
...
f[i][j] = u[i][j] - ...
...
DMDAVecRestoreArray(DM da,Vec local,&u);
DMDAVecRestoreArray(DM da,Vec global,&f);
The recommended approach for multi-component PDEs is to declare a struct representing the
fields defined at each node of the grid, e.g.
typedef struct {
PetscScalar u,v,omega,temperature;
} Node;
57
![](asset-3a.png)
2.4. STRUCTURED GRIDS USING DISTRIBUTED ARRAYS PETSc 3.8 March 24, 2018
and write residual evaluation using
Node **f,**u;
DMDAVecGetArray(DM da,Vec local,&u);
DMDAVecGetArray(DM da,Vec global,&f);
...
f[i][j].omega = ...
...
DMDAVecRestoreArray(DM da,Vec local,&u);
DMDAVecRestoreArray(DM da,Vec global,&f);
See $PETSC_DIR/src/snes/examples/tutorials/ex5.c for a complete example and see $PETSC
_DIR/src/snes/examples/tutorials/ex19.c for an example for a multi-component PDE.
2.4.5 Grid Information
The global indices of the lower left corner of the local portion of the array as well as the local array
size can be obtained with the commands
DMDAGetCorners(DM da,PetscInt *x,PetscInt *y,PetscInt *z,PetscInt *m,PetscInt
*n,PetscInt *p);
DMDAGetGhostCorners(DM da,PetscInt *x,PetscInt *y,PetscInt *z,PetscInt
*m,PetscInt *n,PetscInt *p);
The first version excludes any ghost points, while the second version includes them. The rou-
tine DMDAGetGhostCorners() deals with the fact that subarrays along boundaries of the problem
domain have ghost points only on their interior edges, but not on their boundary edges.
When either type of stencil is used, DMDA_STENCIL_STAR or DMDA_STENCIL_BOX, the local vectors
(with the ghost points) represent rectangular arrays, including the extra corner elements in the
DMDA_STENCIL_STAR case. This configuration provides simple access to the elements by employing
two- (or three-) dimensional indexing. The only difference between the two cases is that when
DMDA_STENCIL_STAR is used, the extra corner components are not scattered between the processes
and thus contain undefined values that should not be used.
To assemble global stiffness matrices, one can use these global indices with MatSetValues() or
MatSetValuesStencil(). Alternately, the global node number of each local node, including the
ghost nodes, can be obtained by calling
DMGetLocalToGlobalMapping(DM da,ISLocalToGlobalMapping *map);
followed by
VecSetLocalToGlobalMapping(Vec v,ISLocalToGlobalMapping map);
MatSetLocalToGlobalMapping(Mat A,ISLocalToGlobalMapping map);
Now entries may be added to the vector and matrix using the local numbering and
VecSetValuesLocal() and MatSetValuesLocal().
Since the global ordering that PETSc uses to manage its parallel vectors (and matrices) does
not usually correspond to the “natural” ordering of a two- or three-dimensional array, the DMDA
structure provides an application ordering AO (see Section 2.3.1) that maps between the natural
ordering on a rectangular grid and the ordering PETSc uses to parallelize. This ordering context
can be obtained with the command
DMDAGetAO(DM da,AO *ao);
58
![](asset-3b.png)
CHAPTER 2. VECTORS AND PARALLEL DATA PETSc 3.8 March 24, 2018
Processor 2 Processor 3
22 23 24 29 3026 27 28 29 30
Processor 2 Processor 3
21 22 23 24 25
16 17 18 19 20
11 12 13 14 15
6 7 8 9 10
1 2 3 4 5
19 20 21 27 28
16 17 18 25 26
Natural Ordering PETSc Ordering
Processor 1Processor 0Processor 1Processor 0
7 8 9 14 15
4 5 6 12 13
1 2 3 10 11
Figure 8: Natural Ordering and PETSc Ordering for a 2D Distributed Array (Four Processes)
In Figure 8we indicate the orderings for a two-dimensional distributed array, divided among four
processes.
The example $PETSC_DIR/src/snes/examples/tutorials/ex5.c illustrates the use of a dis-
tributed array in the solution of a nonlinear problem. The analogous Fortran program is $PETSC
_DIR/src/snes/examples/tutorials/ex5f.F90; see Chapter 5for a discussion of the nonlinear
solvers.
2.5 Vectors Related to Unstructured Grids
2.5.1 Index Sets
To facilitate general vector scatters and gathers used, for example, in updating ghost points for
problems defined on unstructured grids 1, PETSc employs the concept of an index set, via the IS
class. An index set, which is a generalization of a set of integer indices, is used to define scatters,
gathers, and similar operations on vectors and matrices.
The following command creates an index set based on a list of integers:
ISCreateGeneral(MPI Comm comm,PetscInt n,PetscInt *indices,PetscCopyMode mode,
IS *is);
When mode is PETSC_COPY_VALUES, this routine copies the nindices passed to it by the integer
array indices. Thus, the user should be sure to free the integer array indices when it is no longer
needed, perhaps directly after the call to ISCreateGeneral(). The communicator, comm, should
consist of all processes that will be using the IS.
Another standard index set is defined by a starting point (first) and a stride (step), and can
be created with the command
ISCreateStride(MPI Comm comm,PetscInt n,PetscInt first,PetscInt step,IS *is);
1Also see Chapter 10 which describes DMPlex, an abstraction for working with unstructured grids.
59
![](asset-3c.png)
2.5. VECTORS RELATED TO UNSTRUCTURED GRIDS PETSc 3.8 March 24, 2018
Index sets can be destroyed with the command
ISDestroy(IS &is);
On rare occasions the user may need to access information directly from an index set. Several
commands assist in this process:
ISGetSize(IS is,PetscInt *size);
ISStrideGetInfo(IS is,PetscInt *first,PetscInt *stride);
ISGetIndices(IS is,PetscInt **indices);
The function ISGetIndices() returns a pointer to a list of the indices in the index set. For certain
index sets, this may be a temporary array of indices created specifically for a given routine. Thus,
once the user finishes using the array of indices, the routine
ISRestoreIndices(IS is, PetscInt **indices);
should be called to ensure that the system can free the space it may have used to generate the list
of indices.
A blocked version of the index sets can be created with the command
ISCreateBlock(MPI Comm comm,PetscInt bs,PetscInt n,PetscInt
*indices,PetscCopyMode mode, IS *is);
This version is used for defining operations in which each element of the index set refers to a block
of bs vector entries. Related routines analogous to those described above exist as well, including
ISBlockGetIndices(),ISBlockGetSize(),ISBlockGetLocalSize(),ISGetBlockSize(). See
the man pages for details.
2.5.2 Scatters and Gathers
PETSc vectors have full support for general scatters and gathers. One can select any subset of the
components of a vector to insert or add to any subset of the components of another vector. We
refer to these operations as generalized scatters, though they are actually a combination of scatters
and gathers.
To copy selected components from one vector to another, one uses the following set of commands:
VecScatterCreate(Vec x,IS ix,Vec y,IS iy,VecScatter *ctx);
VecScatterBegin(VecScatter ctx,Vec x,Vec y,INSERT VALUES,SCATTER FORWARD);
VecScatterEnd(VecScatter ctx,Vec x,Vec y,INSERT VALUES,SCATTER FORWARD);
VecScatterDestroy(VecScatter *ctx);
Here ix denotes the index set of the first vector, while iy indicates the index set of the destination
vector. The vectors can be parallel or sequential. The only requirements are that the number
of entries in the index set of the first vector, ix, equals the number in the destination index set,
iy, and that the vectors be long enough to contain all the indices referred to in the index sets.
The argument INSERT_VALUES specifies that the vector elements will be inserted into the specified
locations of the destination vector, overwriting any existing values. To add the components, rather
than insert them, the user should select the option ADD_VALUES instead of INSERT_VALUES.
To perform a conventional gather operation, the user simply makes the destination index set,
iy, be a stride index set with a stride of one. Similarly, a conventional scatter can be done with an
initial (sending) index set consisting of a stride. The scatter routines are collective operations (i.e.
all processes that own a parallel vector must call the scatter routines). When scattering from a
60
![](asset-3d.png)
CHAPTER 2. VECTORS AND PARALLEL DATA PETSc 3.8 March 24, 2018
Vec p, x; /* initial vector, destination vector */
VecScatter scatter; /* scatter context */
IS from, to; /* index sets that define the scatter */
PetscScalar *values;
PetscInt idx_from[] = {100,200}, idx_to[] = {0,1};
VecCreateSeq(PETSC COMM SELF,2,&x);
ISCreateGeneral(PETSC COMM SELF,2,idx_from,PETSC COPY VALUES,&from);
ISCreateGeneral(PETSC COMM SELF,2,idx_to,PETSC COPY VALUES,&to);
VecScatterCreate(p,from,x,to,&scatter);
VecScatterBegin(scatter,p,x,INSERT VALUES,SCATTER FORWARD);
VecScatterEnd(scatter,p,x,INSERT VALUES,SCATTER FORWARD);
VecGetArray(x,&values);
ISDestroy(&from);
ISDestroy(&to);
VecScatterDestroy(&scatter);
Figure 9: Example Code for Vector Scatters
parallel vector to sequential vectors, each process has its own sequential vector that receives values
from locations as indicated in its own index set. Similarly, in scattering from sequential vectors to
a parallel vector, each process has its own sequential vector that makes contributions to the parallel
vector.
Caution: When INSERT_VALUES is used, if two different processes contribute different values to
the same component in a parallel vector, either value may end up being inserted. When ADD_VALUES
is used, the correct sum is added to the correct location.
In some cases one may wish to “undo” a scatter, that is perform the scatter backwards, switching
the roles of the sender and receiver. This is done by using
VecScatterBegin(VecScatter ctx,Vec y,Vec x,INSERT VALUES,SCATTER REVERSE);
VecScatterEnd(VecScatter ctx,Vec y,Vec x,INSERT VALUES,SCATTER REVERSE);
Note that the roles of the first two arguments to these routines must be swapped whenever the
SCATTER_REVERSE option is used.
Once a VecScatter object has been created it may be used with any vectors that have the
appropriate parallel data layout. That is, one can call VecScatterBegin() and VecScatterEnd()
with different vectors than used in the call to VecScatterCreate() so long as they have the same
parallel layout (number of elements on each process are the same). Usually, these “different” vectors
would have been obtained via calls to VecDuplicate() from the original vectors used in the call
to VecScatterCreate().
There is a PETSc routine that is nearly the opposite of VecSetValues(), that is, VecGetValues(),
but it can only get local values from the vector. To get off-process values, the user should create
a new vector where the components are to be stored, and then perform the appropriate vector
scatter. For example, if one desires to obtain the values of the 100th and 200th entries of a parallel
vector, p, one could use a code such as that within Figure 9. In this example, the values of the
100th and 200th components are placed in the array values. In this example each process now has
the 100th and 200th component, but obviously each process could gather any elements it needed,
or none by creating an index set with no entries.
The scatter comprises two stages, in order to allow overlap of communication and computation.
61
![](asset-3e.png)
2.5. VECTORS RELATED TO UNSTRUCTURED GRIDS PETSc 3.8 March 24, 2018
The introduction of the VecScatter context allows the communication patterns for the scatter to
be computed once and then reused repeatedly. Generally, even setting up the communication for a
scatter requires communication; hence, it is best to reuse such information when possible.
2.5.3 Scattering Ghost Values
Generalized scatters provide a very general method for managing the communication of required
ghost values for unstructured grid computations. One scatters the global vector into a local
“ghosted” work vector, performs the computation on the local work vectors, and then scatters
back into the global solution vector. In the simplest case this may be written as
VecScatterBegin(VecScatter scatter,Vec globalin,Vec localin,InsertMode
INSERT VALUES,ScatterMode SCATTER FORWARD);
VecScatterEnd(VecScatter scatter,Vec globalin,Vec localin,InsertMode
INSERT VALUES,ScatterMode SCATTER FORWARD);
/* For example, do local calculations from localin to localout */
...
VecScatterBegin(VecScatter scatter,Vec localout,Vec globalout,InsertMode
ADD VALUES,ScatterMode SCATTER REVERSE);
VecScatterEnd(VecScatter scatter,Vec localout,Vec globalout,InsertMode
ADD VALUES,ScatterMode SCATTER REVERSE);
2.5.4 Vectors with Locations for Ghost Values
There are two minor drawbacks to the basic approach described above:
•the extra memory requirement for the local work vector, localin, which duplicates the
memory in globalin, and
•the extra time required to copy the local values from localin to globalin.
An alternative approach is to allocate global vectors with space preallocated for the ghost values;
this may be done with either
VecCreateGhost(MPI Comm comm,PetscInt n,PetscInt N,PetscInt nghost,PetscInt
*ghosts,Vec *vv)
or
VecCreateGhostWithArray(MPI Comm comm,PetscInt n,PetscInt N,PetscInt
nghost,PetscInt *ghosts,PetscScalar *array,Vec *vv)
Here nis the number of local vector entries, Nis the number of global entries (or NULL) and nghost
is the number of ghost entries. The array ghosts is of size nghost and contains the global vector
location for each local ghost location. Using VecDuplicate() or VecDuplicateVecs() on a ghosted
vector will generate additional ghosted vectors.
In many ways, a ghosted vector behaves just like any other MPI vector created by VecCreateMPI().
The difference is that the ghosted vector has an additional “local” representation that allows one
to access the ghost locations. This is done through the call to
VecGhostGetLocalForm(Vec g,Vec *l);
62
![](asset-3f.png)
CHAPTER 2. VECTORS AND PARALLEL DATA PETSc 3.8 March 24, 2018
The vector lis a sequential representation of the parallel vector gthat shares the same array space
(and hence numerical values); but allows one to access the “ghost” values past “the end of the”
array. Note that one access the entries in lusing the local numbering of elements and ghosts, while
they are accessed in gusing the global numbering.
A common usage of a ghosted vector is given by
VecGhostUpdateBegin(Vec globalin,InsertMode INSERT VALUES,ScatterMode
SCATTER FORWARD);
VecGhostUpdateEnd(Vec globalin,InsertMode INSERT VALUES,ScatterMode
SCATTER FORWARD);
VecGhostGetLocalForm(Vec globalin,Vec *localin);
VecGhostGetLocalForm(Vec globalout,Vec *localout);
... Do local calculations from localin to localout ...
VecGhostRestoreLocalForm(Vec globalin,Vec *localin);
VecGhostRestoreLocalForm(Vec globalout,Vec *localout);
VecGhostUpdateBegin(Vec globalout,InsertMode ADD VALUES,ScatterMode
SCATTER REVERSE);
VecGhostUpdateEnd(Vec globalout,InsertMode ADD VALUES,ScatterMode
SCATTER REVERSE);
The routines VecGhostUpdateBegin() and VecGhostUpdateEnd() are equivalent to the rou-
tines VecScatterBegin() and VecScatterEnd() above except that since they are scattering into
the ghost locations, they do not need to copy the local vector values, which are already in place. In
addition, the user does not have to allocate the local work vector, since the ghosted vector already
has allocated slots to contain the ghost values.
The input arguments INSERT_VALUES and SCATTER_FORWARD cause the ghost values to be cor-
rectly updated from the appropriate process. The arguments ADD_VALUES and SCATTER_REVERSE
update the “local” portions of the vector from all the other processes’ ghost values. This would be
appropriate, for example, when performing a finite element assembly of a load vector.
Section 3.5 discusses the important topic of partitioning an unstructured grid.
63
![](asset-40.png)
2.5. VECTORS RELATED TO UNSTRUCTURED GRIDS PETSc 3.8 March 24, 2018
64
![](asset-41.png)
Chapter 3
Matrices
PETSc provides a variety of matrix implementations because no single matrix format is appropriate
for all problems. Currently, we support dense storage and compressed sparse row storage (both
sequential and parallel versions), as well as several specialized formats. Additional formats can be
added.
This chapter describes the basics of using PETSc matrices in general (regardless of the particular
format chosen) and discusses tips for efficient use of the several simple uniprocess and parallel
matrix types. The use of PETSc matrices involves the following actions: create a particular type
of matrix, insert values into it, process the matrix, use the matrix for various computations, and
finally destroy the matrix. The application code does not need to know or care about the particular
storage formats of the matrices.
3.1 Creating and Assembling Matrices
The simplest routine for forming a PETSc matrix, A, is followed by
MatCreate(MPI Comm comm,Mat *A)
MatSetSizes(Mat A,PetscInt m,PetscInt n,PetscInt M,PetscInt N)
This routine generates a sequential matrix when running one process and a parallel matrix for two or
more processes; the particular matrix format is set by the user via options database commands. The
user specifies either the global matrix dimensions, given by Mand Nor the local dimensions, given
by mand nwhile PETSc completely controls memory allocation. This routine facilitates switching
among various matrix types, for example, to determine the format that is most efficient for a certain
application. By default, MatCreate() employs the sparse AIJ format, which is discussed in detail
Section 3.1.1. See the manual pages for further information about available matrix formats.
To insert or add entries to a matrix, one can call a variant of MatSetValues(), either
MatSetValues(Mat A,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt
idxn[],const PetscScalar values[],INSERT VALUES);
or
MatSetValues(Mat A,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt
idxn[],const PetscScalar values[],ADD VALUES);
This routine inserts or adds a logically dense subblock of dimension m*n into the matrix. The
integer indices idxm and idxn, respectively, indicate the global row and column numbers to be
inserted. MatSetValues() uses the standard C convention, where the row and column matrix
65
![](asset-42.png)
3.1. CREATING AND ASSEMBLING MATRICES PETSc 3.8 March 24, 2018
indices begin with zero regardless of the storage format employed. The array values is logically
two-dimensional, containing the values that are to be inserted. By default the values are given in
row major order, which is the opposite of the Fortran convention, meaning that the value to be
put in row idxm[i] and column idxn[j] is located in values[i*n+j]. To allow the insertion of
values in column major order, one can call the command
MatSetOption(Mat A,MAT ROW ORIENTED,PETSC FALSE);
Warning: Several of the sparse implementations do not currently support the column-oriented
option.
This notation should not be a mystery to anyone. For example, to insert one matrix into another
when using MATLAB, one uses the command A(im,in) = B; where im and in contain the indices
for the rows and columns. This action is identical to the calls above to MatSetValues().
When using the block compressed sparse row matrix format (MATSEQBAIJ or MATMPIBAIJ),
one can insert elements more efficiently using the block variant, MatSetValuesBlocked() or
MatSetValuesBlockedLocal().
The function MatSetOption() accepts several other inputs; see the manual page for details.
After the matrix elements have been inserted or added into the matrix, they must be processed
(also called “assembled”) before they can be used. The routines for matrix processing are
MatAssemblyBegin(Mat A,MAT FINAL ASSEMBLY);
MatAssemblyEnd(Mat A,MAT FINAL ASSEMBLY);
By placing other code between these two calls, the user can perform computations while messages
are in transit. Calls to MatSetValues() with the INSERT_VALUES and ADD_VALUES] options
{\em cannot be mixed without intervening calls to the assembly routines. For such intermediate
assembly calls the second routine argument typically should be MAT_FLUSH_ASSEMBLY, which omits
some of the work of the full assembly process. MAT_FINAL_ASSEMBLY is required only in the last
matrix assembly before a matrix is used.
Even though one may insert values into PETSc matrices without regard to which process
eventually stores them, for efficiency reasons we usually recommend generating most entries on the
process where they are destined to be stored. To help the application programmer with this task
for matrices that are distributed across the processes by ranges, the routine
MatGetOwnershipRange(Mat A,PetscInt *first_row,PetscInt *last_row);
informs the user that all rows from first_row to last_row-1 (since the value returned in last_row
is one more than the global index of the last local row) will be stored on the local process.
In the sparse matrix implementations, once the assembly routines have been called, the matrices
are compressed and can be used for matrix-vector multiplication, etc. Any space for preallocated
nonzeros that was not filled by a call to MatSetValues() or a related routine is compressed out by
assembling with MAT_FINAL_ASSEMBLY. If you intend to use that extra space later, be sure to insert
explicit zeros before assembling with MAT_FINAL_ASSEMBLY so the space will not be compressed out.
Once the matrix has been assembled, inserting new values will be expensive since it will require
copies and possible memory allocation.
If one wishes to repeatedly assemble matrices that retain the same nonzero pattern (such as
within a nonlinear or time-dependent problem), the option
MatSetOption(Mat A,MAT NEW NONZERO LOCATIONS,PETSC FALSE);
should be specified after the first matrix has been fully assembled. This option ensures that
certain data structures and communication information will be reused (instead of regenerated)
66
![](asset-43.png)
CHAPTER 3. MATRICES PETSc 3.8 March 24, 2018
during successive steps, thereby increasing efficiency. See $PETSC_DIR/src/ksp/ksp/examples/
tutorials/ex5.c for a simple example of solving two linear systems that use the same matrix
data structure.
3.1.1 Sparse Matrices
The default matrix representation within PETSc is the general sparse AIJ format (also called the
Yale sparse matrix format or compressed sparse row format, CSR). This section discusses tips for
efficiently using this matrix format for large-scale applications. Additional formats (such as block
compressed row and block diagonal storage, which are generally much more efficient for problems
with multiple degrees of freedom per node) are discussed below. Beginning users need not concern
themselves initially with such details and may wish to proceed directly to Section 3.2. However,
when an application code progresses to the point of tuning for efficiency and/or generating timing
results, it is crucial to read this information.
Sequential AIJ Sparse Matrices
In the PETSc AIJ matrix formats, we store the nonzero elements by rows, along with an array of
corresponding column numbers and an array of pointers to the beginning of each row. Note that
the diagonal matrix entries are stored with the rest of the nonzeros (not separately).
To create a sequential AIJ sparse matrix, A, with mrows and ncolumns, one uses the command
MatCreateSeqAIJ(PETSC COMM SELF,PetscInt m,PetscInt n,PetscInt nz,PetscInt
*nnz,Mat *A);
where nz or nnz can be used to preallocate matrix memory, as discussed below. The user can set
nz=0 and nnz=NULL for PETSc to control all matrix memory allocation.
The sequential and parallel AIJ matrix storage formats by default employ i-nodes (identical
nodes) when possible. We search for consecutive rows with the same nonzero structure, thereby
reusing matrix information for increased efficiency. Related options database keys are -mat_no_
inode (do not use inodes) and -mat_inode_limit <limit> (set inode limit (max limit=5)). Note
that problems with a single degree of freedom per grid node will automatically not use I-nodes.
The internal data representation for the AIJ formats employs zero-based indexing.
Preallocation of Memory for Sequential AIJ Sparse Matrices
The dynamic process of allocating new memory and copying from the old storage to the new is
intrinsically very expensive. Thus, to obtain good performance when assembling an AIJ matrix, it
is crucial to preallocate the memory needed for the sparse matrix. The user has two choices for
preallocating matrix memory via MatCreateSeqAIJ().
One can use the scalar nz to specify the expected number of nonzeros for each row. This is
generally fine if the number of nonzeros per row is roughly the same throughout the matrix (or as a
quick and easy first step for preallocation). If one underestimates the actual number of nonzeros in
a given row, then during the assembly process PETSc will automatically allocate additional needed
space. However, this extra memory allocation can slow the computation,
If different rows have very different numbers of nonzeros, one should attempt to indicate (nearly)
the exact number of elements intended for the various rows with the optional array, nnz of length
m, where mis the number of rows, for example
PetscInt nnz[m];
nnz[0] = <nonzeros in row 0>
67
![](asset-44.png)
3.1. CREATING AND ASSEMBLING MATRICES PETSc 3.8 March 24, 2018
nnz[1] = <nonzeros in row 1>
....
nnz[m-1] = <nonzeros in row m-1>
In this case, the assembly process will require no additional memory allocations if the nnz estimates
are correct. If, however, the nnz estimates are incorrect, PETSc will automatically obtain the
additional needed space, at a slight loss of efficiency.
Using the array nnz to preallocate memory is especially important for efficient matrix assembly
if the number of nonzeros varies considerably among the rows. One can generally set nnz either
by knowing in advance the problem structure (e.g., the stencil for finite difference problems on a
structured grid) or by precomputing the information by using a segment of code similar to that
for the regular matrix assembly. The overhead of determining the nnz array will be quite small
compared with the overhead of the inherently expensive mallocs and moves of data that are needed
for dynamic allocation during matrix assembly. Always guess high if an exact value is not known
(extra space is cheaper than too little).
Thus, when assembling a sparse matrix with very different numbers of nonzeros in various rows,
one could proceed as follows for finite difference methods:
1. item Allocate integer array nnz.
2. Loop over grid, counting the expected number of nonzeros for the row(s) associated with the
various grid points.
3. Create the sparse matrix via MatCreateSeqAIJ() or alternative.
4. Loop over the grid, generating matrix entries and inserting in matrix via MatSetValues().
For (vertex-based) finite element type calculations, an analogous procedure is as follows:
1. Allocate integer array nnz.
2. Loop over vertices, computing the number of neighbor vertices, which determines the number
of nonzeros for the corresponding matrix row(s).
3. Create the sparse matrix via MatCreateSeqAIJ() or alternative.
4. Loop over elements, generating matrix entries and inserting in matrix via MatSetValues().
The -info option causes the routines MatAssemblyBegin() and MatAssemblyEnd() to print in-
formation about the success of the preallocation. Consider the following example for the MATSEQAIJ
matrix format:
MatAssemblyEnd_SeqAIJ:Matrix size 10 X 10; storage space:20 unneeded, 100 used
MatAssemblyEnd_SeqAIJ:Number of mallocs during MatSetValues is 0
The first line indicates that the user preallocated 120 spaces but only 100 were used. The second
line indicates that the user preallocated enough space so that PETSc did not have to internally
allocate additional space (an expensive operation). In the next example the user did not preallocate
sufficient space, as indicated by the fact that the number of mallocs is very large (bad for efficiency):
MatAssemblyEnd_SeqAIJ:Matrix size 10 X 10; storage space:47 unneeded, 1000 used
MatAssemblyEnd_SeqAIJ:Number of mallocs during MatSetValues is 40000
Although at first glance such procedures for determining the matrix structure in advance may
seem unusual, they are actually very efficient because they alleviate the need for dynamic construc-
tion of the matrix data structure, which can be very expensive.
68
![](asset-45.png)
CHAPTER 3. MATRICES PETSc 3.8 March 24, 2018
Parallel AIJ Sparse Matrices
Parallel sparse matrices with the AIJ format can be created with the command
MatCreateAIJ=(MPI Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt
d_nz,PetscInt *d_nnz, PetscInt o_nz,PetscInt *o_nnz,Mat *A);
Ais the newly created matrix, while the arguments m,M, and N, indicate the number of local rows
and the number of global rows and columns, respectively. In the PETSc partitioning scheme, all the
matrix columns are local and nis the number of columns corresponding to local part of a parallel
vector. Either the local or global parameters can be replaced with PETSC_DECIDE, so that PETSc
will determine them. The matrix is stored with a fixed number of rows on each process, given by
m, or determined by PETSc if mis PETSC_DECIDE.
If PETSC_DECIDE is not used for the arguments mand n, then the user must ensure that they are
chosen to be compatible with the vectors. To do this, one first considers the matrix-vector product
y=Ax. The mthat is used in the matrix creation routine MatCreateAIJ() must match the local
size used in the vector creation routine VecCreateMPI() for y. Likewise, the nused must match
that used as the local size in VecCreateMPI() for x.
The user must set d_nz=0,o_nz=0,d_nnz=NULL, and o_nnz=NULL for PETSc to control dy-
namic allocation of matrix memory space. Analogous to nz and nnz for the routine
MatCreateSeqAIJ(), these arguments optionally specify nonzero information for the diagonal (d_nz
and d_nnz) and off-diagonal (o_nz and o_nnz) parts of the matrix. For a square global matrix, we
define each process’s diagonal portion to be its local rows and the corresponding columns (a square
submatrix); each process’s off-diagonal portion encompasses the remainder of the local matrix (a
rectangular submatrix). The rank in the MPI communicator determines the absolute ordering of
the blocks. That is, the process with rank 0 in the communicator given to MatCreateAIJ() con-
tains the top rows of the matrix; the ith process in that communicator contains the ith block of the
matrix.
Preallocation of Memory for Parallel AIJ Sparse Matrices
As discussed above, preallocation of memory is critical for achieving good performance during
matrix assembly, as this reduces the number of allocations and copies required. We present an
example for three processes to indicate how this may be done for the MATMPIAIJ matrix format.
Consider the 8 by 8 matrix, which is partitioned by default with three rows on the first process,
three on the second and two on the third.
120|030|0 4
056|700|8 0
9 0 10 |11 0 0 |12 0
13 0 14 |15 16 17 |0 0
0 18 0 |19 20 21 |0 0
000|22 23 0 |24 0
25 26 27 |0 0 28 |29 0
30 0 0 |31 32 33 |0 34
The “diagonal” submatrix, d, on the first process is given by
1 2 0
0 5 6
9 0 10
,
69
![](asset-46.png)
3.1. CREATING AND ASSEMBLING MATRICES PETSc 3.8 March 24, 2018
while the “off-diagonal” submatrix, o, matrix is given by
03004
70080
1100120
.
For the first process one could set d_nz to 2 (since each row has 2 nonzeros) or, alternatively, set
d_nnz to {2,2,2}.The o_nz could be set to 2 since each row of the omatrix has 2 nonzeros, or
o_nnz could be set to {2,2,2}.
For the second process the dsubmatrix is given by
15 16 17
19 20 21
22 23 0
.
Thus, one could set d_nz to 3, since the maximum number of nonzeros in each row is 3, or
alternatively one could set d_nnz to {3,3,2}, thereby indicating that the first two rows will have 3
nonzeros while the third has 2. The corresponding osubmatrix for the second process is
1301400
0 18 0 0 0
0 0 0 24 0
so that one could set o_nz to 2 or o_nnz to {2,1,1}.
Note that the user never directly works with the dand osubmatrices, except when preallocating
storage space as indicated above. Also, the user need not preallocate exactly the correct amount of
space; as long as a sufficiently close estimate is given, the high efficiency for matrix assembly will
remain.
As described above, the option -info will print information about the success of preallocation
during matrix assembly. For the MATMPIAIJ and MATMPIBAIJ formats, PETSc will also list the
number of elements owned by on each process that were generated on a different process. For
example, the statements
MatAssemblyBegin_MPIAIJ:Stash has 10 entries, uses 0 mallocs
MatAssemblyBegin_MPIAIJ:Stash has 3 entries, uses 0 mallocs
MatAssemblyBegin_MPIAIJ:Stash has 5 entries, uses 0 mallocs
indicate that very few values have been generated on different processes. On the other hand, the
statements
MatAssemblyBegin_MPIAIJ:Stash has 100000 entries, uses 100 mallocs
MatAssemblyBegin_MPIAIJ:Stash has 77777 entries, uses 70 mallocs
indicate that many values have been generated on the “wrong” processes. This situation can be
very inefficient, since the transfer of values to the “correct” process is generally expensive. By using
the command MatGetOwnershipRange() in application codes, the user should be able to generate
most entries on the owning process.
Note: It is fine to generate some entries on the “wrong” process. Often this can lead to cleaner,
simpler, less buggy codes. One should never make code overly complicated in order to generate all
values locally. Rather, one should organize the code in such a way that most values are generated
locally.
3.1.2 Dense Matrices
PETSc provides both sequential and parallel dense matrix formats, where each process stores its
entries in a column-major array in the usual Fortran style. To create a sequential, dense PETSc
matrix, Aof dimensions mby n, the user should call
70
![](asset-47.png)
CHAPTER 3. MATRICES PETSc 3.8 March 24, 2018
MatCreateSeqDense(PETSC COMM SELF,PetscInt m,PetscInt n,PetscScalar *data,Mat *A);
The variable data enables the user to optionally provide the location of the data for matrix storage
(intended for Fortran users who wish to allocate their own storage space). Most users should
merely set data to NULL for PETSc to control matrix memory allocation. To create a parallel,
dense matrix, A, the user should call
MatCreateDense(MPI Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt
N,PetscScalar *data,Mat *A)
The arguments m,n,M, and N, indicate the number of local rows and columns and the number of
global rows and columns, respectively. Either the local or global parameters can be replaced with
PETSC_DECIDE, so that PETSc will determine them. The matrix is stored with a fixed number of
rows on each process, given by m, or determined by PETSc if mis PETSC_DECIDE.
PETSc does not provide parallel dense direct solvers, instead interfacing to external packages
that provide these solvers. Our focus is on sparse iterative solvers.
3.1.3 Block Matrices
Block matrices arise when coupling variables with different meaning, especially when solving prob-
lems with constraints (e.g. incompressible flow) and “multi-physics” problems. Usually the number
of blocks is small and each block is partitioned in parallel. We illustrate for a 3 ×3 system with
components labeled a, b, c. With some numbering of unknowns, the matrix could be written as
Aaa Aab Aac
Aba Abb Abc
Aca Acb Acc
.
There are two fundamentally different ways that this matrix could be stored, as a single assembled
sparse matrix where entries from all blocks are merged together (“monolithic”), or as separate
assembled matrices for each block (“nested”). These formats have different performance character-
istics depending on the operation being performed. In particular, many preconditioners require a
monolithic format, but some that are very effective for solving block systems (see Section 4.5) are
more efficient when a nested format is used. In order to stay flexible, we would like to be able to use
the same code to assemble block matrices in both monolithic and nested formats. Additionally, for
software maintainability and testing, especially in a multi-physics context where different groups
might be responsible for assembling each of the blocks, it is desirable to be able to use exactly the
same code to assemble a single block independently as to assemble it as part of a larger system.
To do this, we introduce the four spaces shown in Figure 10.
•The monolithic global space is the space in which the Krylov and Newton solvers operate,
with collective semantics across the entire block system.
•The split global space splits the blocks apart, but each split still has collective semantics.
•The split local space adds ghost points and separates the blocks. Operations in this space can
be performed with no parallel communication. This is often the most natural, and certainly
the most powerful, space for matrix assembly code.
•The monolithic local space can be thought of as adding ghost points to the monolithic global
space, but it is often more natural to use it simply as a concatenation of split local spaces
on each process. It is not common to explicitly manipulate vectors or matrices in this space
71
![](asset-48.png)
3.1. CREATING AND ASSEMBLING MATRICES PETSc 3.8 March 24, 2018
rank 0
rank 2
rank 1
rank 0
rank 1
rank 2
LocalToGlobalMapping
Monolithic Global Monolithic Local
Split Local
GetLocalSubMatrix()
Split Global
GetSubMatrix() / GetSubVector()
LocalToGlobal()
rank 0
rank 1
rank 2
Figure 10: The relationship between spaces used for coupled assembly.
(at least not during assembly), but it is a useful for declaring which part of a matrix is being
assembled.
The key to format-independent assembly is the function
MatGetLocalSubMatrix(Mat A,IS isrow,IS iscol,Mat *submat);
which provides a “view” submat into a matrix Athat operates in the monolithic global space. The
submat transforms from the split local space defined by iscol to the split local space defined by
isrow. The index sets specify the parts of the monolithic local space that submat should operate in.
If a nested matrix format is used, then MatGetLocalSubMatrix() finds the nested block and returns
it without making any copies. In this case, submat is fully functional and has a parallel communica-
tor. If a monolithic matrix format is used, then MatGetLocalSubMatrix() returns a proxy matrix
on PETSC_COMM_SELF that does not provide values or implement MatMult(), but does implement
MatSetValuesLocal() and, if isrow,iscol have a constant block size, MatSetValuesBlockedLocal().
Note that although submat may not be a fully functional matrix and the caller does not even know
a priori which communicator it will reside on, it always implements the local assembly functions
(which are not collective). The index sets isrow,iscol can be obtained using
DMCompositeGetLocalISs() if DMComposite is being used. DMComposite can also be used to
create matrices, in which case the MATNEST format can be specified using -prefix_dm_mat_type
nest and MATAIJ can be specified using -prefix_dm_mat_type aij. See $PETSC_DIR/src/snes/
examples/tutorials/ex28.c for a simple example using this interface.
72
![](asset-49.png)
CHAPTER 3. MATRICES PETSc 3.8 March 24, 2018
3.2 Basic Matrix Operations
Table 2summarizes basic PETSc matrix operations. We briefly discuss a few of these routines in
more detail below.
The parallel matrix can multiply a vector with nlocal entries, returning a vector with mlocal
entries. That is, to form the product
MatMult(Mat A,Vec x,Vec y);
the vectors xand yshould be generated with
VecCreateMPI(MPI Comm comm,n,N,&x);
VecCreateMPI(MPI Comm comm,m,M,&y);
By default, if the user lets PETSc decide the number of components to be stored locally (by passing
in PETSC_DECIDE as the second argument to VecCreateMPI() or using VecCreate()), vectors and
matrices of the same dimension are automatically compatible for parallel matrix-vector operations.
Along with the matrix-vector multiplication routine, there is a version for the transpose of the
matrix,
MatMultTranspose(Mat A,Vec x,Vec y);
There are also versions that add the result to another vector:
MatMultAdd(Mat A,Vec x,Vec y,Vec w);
MatMultTransposeAdd(Mat A,Vec x,Vec y,Vec w);
These routines, respectively, produce w=A∗x+yand w=AT∗x+y. In C it is legal for the
vectors yand wto be identical. In Fortran, this situation is forbidden by the language standard,
but we allow it anyway.
One can print a matrix (sequential or parallel) to the screen with the command
MatView(Mat mat,PETSC VIEWER STDOUT WORLD);
Other viewers can be used as well. For instance, one can draw the nonzero structure of the matrix
into the default X-window with the command
MatView(Mat mat,PETSC VIEWER DRAW WORLD);
Also one can use
MatView(Mat mat,PetscViewer viewer);
where viewer was obtained with PetscViewerDrawOpen(). Additional viewers and options are
given in the MatView() man page and Section 15.3.
The NormType argument to MatNorm() is one of NORM_1,NORM_INFINITY, and NORM_FROBENIUS.
3.3 Matrix-Free Matrices
Some people like to use matrix-free methods, which do not require explicit storage of the matrix, for
the numerical solution of partial differential equations. To support matrix-free methods in PETSc,
one can use the following command to create a Mat structure without ever actually generating the
matrix:
73
![](asset-4a.png)
3.3. MATRIX-FREE MATRICES PETSc 3.8 March 24, 2018
Function Name Operation
MatAXPY(Mat Y, PetscScalar a, Mat X, MatStructure s); Y=Y+a∗X
MatMult(Mat A,Vec x, Vec y); y=A∗x
MatMultAdd(Mat A,Vec x, Vec y,Vec z); z=y+A∗x
MatMultTranspose(Mat A,Vec x, Vec y); y=AT∗x
MatMultTransposeAdd(Mat A, Vec x, Vec y, Vec z); z=y+AT∗x
MatNorm(Mat A,NormType type, PetscReal *r); r=||A||type
MatDiagonalScale(Mat A,Vec l,Vec r); A= diag(l)∗A∗diag(r)
MatScale(Mat A,PetscScalar a); A=a∗A
MatConvert(Mat A, MatType type, Mat *B); B=A
MatCopy(Mat A, Mat B, MatStructure s); B=A
MatGetDiagonal(Mat A, Vec x); x= diag(A)
MatTranspose(Mat A, MatReuse,Mat* B); B=AT
MatZeroEntries(Mat A); A= 0
MatShift(Mat Y, PetscScalar a); Y=Y+a∗I
Table 2: PETSc Matrix Operations
MatCreateShell(MPI Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,void
*ctx,Mat *mat);
Here Mand Nare the global matrix dimensions (rows and columns), mand nare the local matrix
dimensions, and ctx is a pointer to data needed by any user-defined shell matrix operations; the
manual page has additional details about these parameters. Most matrix-free algorithms require
only the application of the linear operator to a vector. To provide this action, the user must write
a routine with the calling sequence
UserMult(Mat mat,Vec x,Vec y);
and then associate it with the matrix, mat, by using the command
MatShellSetOperation(Mat mat,MatOperation MATOP_MULT, (void(*)(void))
PetscErrorCode (*UserMult)(Mat,Vec,Vec));
Here MATOP_MULT is the name of the operation for matrix-vector multiplication. Within each user-
defined routine (such as UserMult()), the user should call MatShellGetContext() to obtain the
user-defined context, ctx, that was set by MatCreateShell(). This shell matrix can be used with
the iterative linear equation solvers discussed in the following chapters.
The routine MatShellSetOperation() can be used to set any other matrix operations as well.
The file $PETSC_DIR/include/petscmat.h provides a complete list of matrix operations, which
have the form MATOP_<OPERATION>, where <OPERATION> is the name (in all capital letters) of
the user interface routine (for example, MatMult() →MATOP_MULT). All user-provided functions
have the same calling sequence as the usual matrix interface routines, since the user-defined func-
tions are intended to be accessed through the same interface, e.g., MatMult(Mat,Vec,Vec)→
UserMult(Mat,Vec,Vec). The final argument for MatShellSetOperation() needs to be cast to a
void *, since the final argument could (depending on the MatOperation) be a variety of different
functions.
Note that MatShellSetOperation() can also be used as a “backdoor” means of introducing
user-defined changes in matrix operations for other storage formats (for example, to override the
74
![](asset-4b.png)
CHAPTER 3. MATRICES PETSc 3.8 March 24, 2018
default LU factorization routine supplied within PETSc for the MATSEQAIJ format). However, we
urge anyone who introduces such changes to use caution, since it would be very easy to accidentally
create a bug in the new routine that could affect other routines as well.
See also Section 5.5 for details on one set of helpful utilities for using the matrix-free approach
for nonlinear solvers.
3.4 Other Matrix Operations
In many iterative calculations (for instance, in a nonlinear equations solver), it is important for
efficiency purposes to reuse the nonzero structure of a matrix, rather than determining it anew
every time the matrix is generated. To retain a given matrix but reinitialize its contents, one can
employ
MatZeroEntries(Mat A);
This routine will zero the matrix entries in the data structure but keep all the data that indicates
where the nonzeros are located. In this way a new matrix assembly will be much less expensive,
since no memory allocations or copies will be needed. Of course, one can also explicitly set selected
matrix elements to zero by calling MatSetValues().
By default, if new entries are made in locations where no nonzeros previously existed, space
will be allocated for the new entries. To prevent the allocation of additional memory and simply
discard those new entries, one can use the option
MatSetOption(Mat A,MAT NEW NONZERO LOCATIONS,PETSC FALSE);
Once the matrix has been assembled, one can factor it numerically without repeating the ordering
or the symbolic factorization. This option can save some computational time, although it does
require that the factorization is not done in-place.
In the numerical solution of elliptic partial differential equations, it can be cumbersome to
deal with Dirichlet boundary conditions. In particular, one would like to assemble the matrix
without regard to boundary conditions and then at the end apply the Dirichlet boundary conditions.
In numerical analysis classes this process is usually presented as moving the known boundary
conditions to the right-hand side and then solving a smaller linear system for the interior unknowns.
Unfortunately, implementing this requires extracting a large submatrix from the original matrix
and creating its corresponding data structures. This process can be expensive in terms of both
time and memory.
One simple way to deal with this difficulty is to replace those rows in the matrix associated with
known boundary conditions, by rows of the identity matrix (or some scaling of it). This action can
be done with the command
MatZeroRows(Mat A,PetscInt numRows,PetscInt rows[],PetscScalar diag_value,Vec
x,Vec b),
or equivalently,
MatZeroRowsIS(Mat A,IS rows,PetscScalar diag_value,Vec x,Vec b);
For sparse matrices this removes the data structures for certain rows of the matrix. If the pointer
diag_value is NULL, it even removes the diagonal entry. If the pointer is not null, it uses that given
value at the pointer location in the diagonal entry of the eliminated rows.
One nice feature of this approach is that when solving a nonlinear problem such that at each
iteration the Dirichlet boundary conditions are in the same positions and the matrix retains the
75
![](asset-4c.png)
3.4. OTHER MATRIX OPERATIONS PETSc 3.8 March 24, 2018
same nonzero structure, the user can call MatZeroRows() in the first iteration. Then, before
generating the matrix in the second iteration the user should call
MatSetOption(Mat A,MAT NEW NONZERO LOCATIONS,PETSC FALSE);
From that point, no new values will be inserted into those (boundary) rows of the matrix.
The functions MatZeroRowsLocal() and MatZeroRowsLocalIS() can also be used if for each
process one provides the Dirichlet locations in the local numbering of the matrix. A drawback of
MatZeroRows() is that it destroys the symmetry of a matrix. Thus one can use
MatZeroRowsColumns(Mat A,PetscInt numRows,PetscInt rows[],PetscScalar
diag_value,Vec x,Vec b),
or equivalently,
MatZeroRowsColumnsIS(Mat A,IS rows,PetscScalar diag_value,Vec x,Vec b);
Note that with all of these for a given assembled matrix it can be only called once to update the
x and b vector. It cannot be used if one wishes to solve multiple right hand side problems for the
same matrix since the matrix entries needed for updating the b vector are removed in its first use.
Once the zeroed rows are removed the new matrix has possibly many rows with only a diagonal
entry affecting the parallel load balancing. The PCREDISTRIBUTE preconditioner removes all the
zeroed rows (and associated columns and adjusts the right hand side based on the removed columns)
and then rebalances the resulting rows of smaller matrix across the processes. Thus one can use
MatZeroRows() to set the Dirichlet points and then solve with the preconditioner PCREDISTRIBUTE.
Note if the original matrix was symmetric the smaller solved matrix will also be symmetric.
Another matrix routine of interest is
MatConvert(Mat mat,MatType newtype,Mat *M)
which converts the matrix mat to new matrix, M, that has either the same or different format.
Set newtype to MATSAME to copy the matrix, keeping the same matrix format. See $PETSC_DIR
/include/petscmat.h for other available matrix types; standard ones are MATSEQDENSE,MATSEQAIJ,
MATMPIAIJ,MATSEQBAIJ and MATMPIBAIJ.
In certain applications it may be necessary for application codes to directly access elements of
a matrix. This may be done by using the the command (for local rows only)
MatGetRow(Mat A,PetscInt row, PetscInt *ncols,const PetscInt (*cols)[],const
PetscScalar (*vals)[]);
The argument ncols returns the number of nonzeros in that row, while cols and vals returns the
column indices (with indices starting at zero) and values in the row. If only the column indices
are needed (and not the corresponding matrix elements), one can use NULL for the vals argument.
Similarly, one can use NULL for the cols argument. The user can only examine the values extracted
with MatGetRow(); the values cannot be altered. To change the matrix entries, one must use
MatSetValues().
Once the user has finished using a row, he or she must call
MatRestoreRow(Mat A,PetscInt row,PetscInt *ncols,PetscInt **cols,PetscScalar
**vals);
to free any space that was allocated during the call to MatGetRow().
76
![](asset-4d.png)
CHAPTER 3. MATRICES PETSc 3.8 March 24, 2018
3.5 Partitioning
For almost all unstructured grid computation, the distribution of portions of the grid across the
process’s work load and memory can have a very large impact on performance. In most PDE
calculations the grid partitioning and distribution across the processes can (and should) be done in
a “pre-processing” step before the numerical computations. However, this does not mean it need be
done in a separate, sequential program; rather, it should be done before one sets up the parallel grid
data structures in the actual program. PETSc provides an interface to the ParMETIS (developed
by George Karypis; see $PETSC_DIR/docs/installation.html. for directions on installing PETSc
to use ParMETIS) to allow the partitioning to be done in parallel. PETSc does not currently
provide directly support for dynamic repartitioning, load balancing by migrating matrix entries
between processes, etc. For problems that require mesh refinement, PETSc uses the “rebuild the
data structure” approach, as opposed to the “maintain dynamic data structures that support the
insertion/deletion of additional vector and matrix rows and columns entries” approach.
Partitioning in PETSc is organized around the MatPartitioning object. One first creates
a parallel matrix that contains the connectivity information about the grid (or other graph-type
object) that is to be partitioned. This is done with the command
MatCreateMPIAdj(MPI Comm comm,int mlocal,PetscInt n,const PetscInt ia[],const
PetscInt ja[],PetscInt *weights,Mat *Adj);
The argument mlocal indicates the number of rows of the graph being provided by the given
process, nis the total number of columns; equal to the sum of all the mlocal. The arguments ia
and ja are the row pointers and column pointers for the given rows; these are the usual format for
parallel compressed sparse row storage, using indices starting at 0, not 1.
1
23
4
5
0
0
12
3
Figure 11: Numbering on Simple Unstructured Grid
This, of course, assumes that one has already distributed the grid (graph) information among
the processes. The details of this initial distribution is not important; it could be simply determined
by assigning to the first process the first n0nodes from a file, the second process the next n1nodes,
etc.
For example, we demonstrate the form of the ia and ja for a triangular grid where we
77
![](asset-4e.png)
3.5. PARTITIONING PETSc 3.8 March 24, 2018
(1) partition by element (triangle)
•Process 0: mlocal = 2,n=4,ja = {2,3, 3},ia = {0,2,3}
•Process 1: mlocal = 2,n=4,ja = {0, 0,1},ia = {0,1,3}
Note that elements are not connected to themselves and we only indicate edge connections (in some
contexts single vertex connections between elements may also be included). We use a space above
to denote the transition between rows in the matrix.
and (2) partition by vertex.
•Process 0: mlocal = 3,n=6,ja = {3,4, 4,5, 3,4,5},ia = {0, 2, 4, 7}
•Process 1: mlocal = 3,n=6,ja = {0,2, 4, 0,1,2,3,5, 1,2,4},ia = {0, 3, 8, 11}.
Once the connectivity matrix has been created the following code will generate the renumbering
required for the new partition
MatPartitioningCreate(MPI Comm comm,MatPartitioning *part);
MatPartitioningSetAdjacency(MatPartitioning part,Mat Adj);
MatPartitioningSetFromOptions(MatPartitioning part);
MatPartitioningApply(MatPartitioning part,IS *is);
MatPartitioningDestroy(MatPartitioning *part);
MatDestroy(Mat *Adj);
ISPartitioningToNumbering(IS is,IS *isg);
The resulting isg contains for each local node the new global number of that node. The resulting
is contains the new process number that each local node has been assigned to.
Now that a new numbering of the nodes has been determined, one must renumber all the nodes
and migrate the grid information to the correct process. The command
AOCreateBasicIS(isg,NULL,&ao);
generates, see Section 2.3.1, an AO object that can be used in conjunction with the is and isg
to move the relevant grid information to the correct process and renumber the nodes etc. In this
context, the new ordering is the “application” ordering so AOPetscToApplication() converts old
global indices to new global indices and AOApplicationToPetsc() converts new global indices back
to old global indices.
PETSc does not currently provide tools that completely manage the migration and node renum-
bering, since it will be dependent on the particular data structure you use to store the grid infor-
mation and the type of grid information that you need for your application. We do plan to include
more support for this in the future, but designing the appropriate general user interface and pro-
viding a scalable implementation that can be used for a wide variety of different grids requires a
great deal of time.
78
![](asset-4f.png)
Chapter 4
KSP: Linear System Solvers
The object KSP is the heart of PETSc, because it provides uniform and efficient access to all of
the package’s linear system solvers, including parallel and sequential, direct and iterative. KSP is
intended for solving nonsingular systems of the form
Ax =b, (4.1)
where Adenotes the matrix representation of a linear operator, bis the right-hand-side vector, and
xis the solution vector. KSP uses the same calling sequence for both direct and iterative solution
of a linear system. In addition, particular solution techniques and their associated options can be
selected at runtime.
The combination of a Krylov subspace method and a preconditioner is at the center of most
modern numerical codes for the iterative solution of linear systems. See, for example, [13] for an
overview of the theory of such methods. KSP creates a simplified interface to the lower-level KSP
and PC modules within the PETSc package. The KSP package, discussed in Section 4.3, provides
many popular Krylov subspace iterative methods; the PC module, described in Section 4.4, includes
a variety of preconditioners. Although both KSP and PC can be used directly, users should employ
the interface of KSP.
4.1 Using KSP
To solve a linear system with KSP, one must first create a solver context with the command
KSPCreate(MPI Comm comm,KSP *ksp);
Here comm is the MPI communicator, and ksp is the newly formed solver context. Before actually
solving a linear system with KSP, the user must call the following routine to set the matrices
associated with the linear system:
KSPSetOperators(KSP ksp,Mat Amat,Mat Pmat);
The argument Amat, representing the matrix that defines the linear system, is a symbolic place
holder for any kind of matrix. In particular, KSP does support matrix-free methods. The rou-
tine MatCreateShell() in Section 3.3 provides further information regarding matrix-free methods.
Typically, the matrix from which the preconditioner is to be constructed, Pmat, is the same as
the matrix that defines the linear system, Amat; however, occasionally these matrices differ (for in-
stance, when a preconditioning matrix is obtained from a lower order method than that employed
to form the linear system matrix).
Much of the power of KSP can be accessed through the single routine
79
![](asset-50.png)
4.2. SOLVING SUCCESSIVE LINEAR SYSTEMS PETSc 3.8 March 24, 2018
KSPSetFromOptions(KSP ksp);
This routine accepts the options -h and -help as well as any of the KSP and PC options discussed
below. To solve a linear system, one sets the rhs and solution vectors using and executes the
command
KSPSolve(KSP ksp,Vec b,Vec x);
where band xrespectively denote the right-hand-side and solution vectors. On return, the iteration
number at which the iterative process stopped can be obtained using
KSPGetIterationNumber(KSP ksp, PetscInt *its);
Note that this does not state that the method converged at this iteration: it can also have reached
the maximum number of iterations, or have diverged.
Section 4.3.2 gives more details regarding convergence testing. Note that multiple linear solves
can be performed by the same KSP context. Once the KSP context is no longer needed, it should be
destroyed with the command
KSPDestroy(KSP *ksp);
The above procedure is sufficient for general use of the KSP package. One additional step is
required for users who wish to customize certain preconditioners (e.g., see Section 4.4.4) or to log
certain performance data using the PETSc profiling facilities (as discussed in Chapter 13). In this
case, the user can optionally explicitly call
KSPSetUp(KSP ksp)
before calling KSPSolve() to perform any setup required for the linear solvers. The explicit call
of this routine enables the separate monitoring of any computations performed during the set up
phase, such as incomplete factorization for the ILU preconditioner.
The default solver within KSP is restarted GMRES, preconditioned for the uniprocess case with
ILU(0), and for the multiprocess case with the block Jacobi method (with one block per process,
each of which is solved with ILU(0)). A variety of other solvers and options are also available. To
allow application programmers to set any of the preconditioner or Krylov subspace options directly
within the code, we provide routines that extract the PC and KSP contexts,
KSPGetPC(KSP ksp,PC *pc);
The application programmer can then directly call any of the PC or KSP routines to modify the
corresponding default options.
To solve a linear system with a direct solver (currently supported by PETSc for sequential
matrices, and by several external solvers through PETSc interfaces (see Section 4.7)) one may use
the options -ksp_type preonly -pc_type lu (see below).
By default, if a direct solver is used, the factorization is not done in-place. This approach
prevents the user from the unexpected surprise of having a corrupted matrix after a linear solve.
The routine PCFactorSetUseInPlace(), discussed below, causes factorization to be done in-place.
4.2 Solving Successive Linear Systems
When solving multiple linear systems of the same size with the same method, several options are
available. To solve successive linear systems having the same preconditioner matrix (i.e., the same
80
![](asset-51.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
data structure with exactly the same matrix elements) but different right-hand-side vectors, the
user should simply call KSPSolve(), multiple times. The preconditioner setup operations (e.g.,
factorization for ILU) will be done during the first call to KSPSolve() only; such operations will
not be repeated for successive solves.
To solve successive linear systems that have different preconditioner matrices (i.e., the matrix
elements and/or the matrix data structure change), the user must call KSPSetOperators() and
KSPSolve() for each solve. See Section 4.1 for a description of various flags for KSPSetOperators()
that can save work for such cases.
4.3 Krylov Methods
The Krylov subspace methods accept a number of options, many of which are discussed below.
First, to set the Krylov subspace method that is to be used, one calls the command
KSPSetType(KSP ksp,KSPType method);
The type can be one of KSPRICHARDSON,KSPCHEBYSHEV,KSPCG,KSPGMRES,KSPTCQMR,KSPBCGS,
KSPCGS,KSPTFQMR,KSPCR,KSPLSQR,KSPBICG,KSPPREONLY. or others; see Table 3or the KSPType
man page for more. The KSP method can also be set with the options database command -
ksp_type, followed by one of the options richardson,chebyshev,cg,gmres,tcqmr,bcgs,cgs,
tfqmr,cr,lsqr,bicg,preonly., or others (see Table 3or the KSPType man page) There are
method-specific options:for instance, for the Richardson, Chebyshev, and GMRES methods:
KSPRichardsonSetScale(KSP ksp,PetscReal scale);
KSPChebyshevSetEigenvalues(KSP ksp,PetscReal emax,PetscReal emin);
KSPGMRESSetRestart(KSP ksp,PetscInt max_steps);
The default parameter values are damping_factor=1.0, emax=0.01, emin=100.0, and max_steps=
30. The GMRES restart and Richardson damping factor can also be set with the options -
ksp_gmres_restart <n> and -ksp_richardson_scale <factor>.
The default technique for orthogonalization of the Hessenberg matrix in GMRES is the unmod-
ified (classical) Gram-Schmidt method, which can be set with
KSPGMRESSetOrthogonalization(KSP
ksp,KSPGMRESClassicalGramSchmidtOrthogonalization);
or the options database command -ksp_gmres_classicalgramschmidt. By default this will not
use iterative refinement to improve the stability of the orthogonalization. This can be changed with
the option
KSPGMRESSetCGSRefinementType(KSP ksp,KSPGMRESCGSRefinementType type)
or via the options database with
-ksp_gmres_cgs_refinement_type none,ifneeded,always
The values for KSPGMRESCGSRefinementType() are KSP_GMRES_CGS_REFINEMENT_NONE,
KSP_GMRES_CGS_REFINEMENT_IFNEEDED and KSP_GMRES_CGS_REFINEMENT_ALWAYS.
One can also use modified Gram-Schmidt, by using the orthogonalization routine
KSPGMRESModifiedGramSchmidtOrthogonalization() or by using the command line option -
ksp_gmres_modifiedgramschmidt.
For the conjugate gradient method with complex numbers, there are two slightly different
algorithms depending on whether the matrix is Hermitian symmetric or truly symmetric (the
81
![](asset-52.png)
4.3. KRYLOV METHODS PETSc 3.8 March 24, 2018
default is to assume that it is Hermitian symmetric). To indicate that it is symmetric, one uses the
command
KSPCGSetType(KSP ksp,KSPCGType KSP CG SYMMETRIC);
Note that this option is not valid for all matrices.
The LSQR algorithm does not involve a preconditioner; any preconditioner set to work with
the KSP object is ignored if KSPLSQR was selected.
By default, KSP assumes an initial guess of zero by zeroing the initial value for the solution
vector that is given; this zeroing is done at the call to KSPSolve() (or KSPSolve()). To use a
nonzero initial guess, the user must call
KSPSetInitialGuessNonzero(KSP ksp,PetscBool flg);
4.3.1 Preconditioning within KSP
Since the rate of convergence of Krylov projection methods for a particular linear system is strongly
dependent on its spectrum, preconditioning is typically used to alter the spectrum and hence
accelerate the convergence rate of iterative techniques. Preconditioning can be applied to the
system (4.1) by
(M−1
LAM−1
R) (MRx) = M−1
Lb, (4.2)
where MLand MRindicate preconditioning matrices (or, matrices from which the preconditioner
is to be constructed). If ML=Iin (4.2), right preconditioning results, and the residual of (4.1),
r≡b−Ax =b−AM−1
RMRx,
is preserved. In contrast, the residual is altered for left (MR=I) and symmetric preconditioning,
as given by
rL≡M−1
Lb−M−1
LAx =M−1
Lr.
By default, most KSP implementations use left preconditioning. Some more naturally use other
options, though. For instance, KSPQCG defaults to use symmetric preconditioning and KSPFGMRES
uses right preconditioning by default. Right preconditioning can be activated for some methods by
using the options database command -ksp_pc_side right or calling the routine
KSPSetPCSide(KSP ksp,PCSide PC RIGHT);
Attempting to use right preconditioning for a method that does not currently support it results in
an error message of the form
KSPSetUp_Richardson:No right preconditioning for KSPRICHARDSON
We summarize the defaults for the residuals used in KSP convergence monitoring within Ta-
ble 3. Details regarding specific convergence tests and monitoring routines are presented in the
following sections. The preconditioned residual is used by default for convergence testing of all left-
preconditioned KSP methods. For the conjugate gradient, Richardson, and Chebyshev methods the
true residual can be used by the options database command ksp_norm_type unpreconditioned
or by calling the routine
KSPSetNormType(KSP ksp,KSP NORM UNPRECONDITIONED);
82
![](asset-53.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
Options
Database
Method KSPType Name
Richardson KSPRICHARDSON richardson
Chebyshev KSPCHEBYSHEV chebyshev
Conjugate Gradient [19]KSPCG cg
Pipelined Conjugate Gradients KSPPIPECG pipecg
Pipelined Conjugate Gradients (Gropp) KSPGROPPCG cgne
Pipelined Conjugate Gradients with Residual Replacement KSPPIPECGRR pipecgrr
Conjugate Gradients for the Normal Equations KSPCGNE cgne
Flexible Conjugate Gradients [26]KSPFCG fcg
Pipelined, Flexible Conjugate Gradient KSPPIPEFCG pipefcg
Conjugate Gradients for Least Squares KSPCGLS cgls
Conjugate Gradients with Constraint (1) KSPNASH nash
Conjugate Gradients with Constraint (2) KSPSTCG stcg
Conjugate Gradients with Constraint (3) KSPGLTR gltr
Conjugate Gradients with Constraint (4) KSPQCG qcg
BiConjugate Gradient KSPBICG bicg
BiCGSTAB [36]KSPBCGS bcgs
Improved BiCGSTAB KSPIBCGS ibcgs
Flexible BiCGSTAB KSPFBCGS fbcgs
Flexible BiCGSTAB (variant) KSPFBCGSR fbcgsr
Enhanced BiCGSTAB(L) KSPBCGSL bcgsl
Minimal Residual Method [27]KSPMINRES minres
Generalized Minimal Residual [32]KSPGMRES gmres
Flexible Generalized Minimal Residual [31]KSPFGMRES fgmres
Deflated Generalized Minimal Residual KSPDGMRES dgmres
Pipelined Generalized Minimal Residual KSPPGMRES pgmres
Pipelined, Flexible Generalized Minimal Residual KSPPIPEFGMRES pipefgmres
Generalized Minimal Residual with Accelerated Restart KSPLGMRES lgmres
Conjugate Residual [12]KSPCR cr
Generalized Conjugate Residual KSPGCR gcr
Pipelined Conjugate Residual KSPPIPECR pipecr
Pipelined, Flexible Conjugate Residual KSPPIPEGCR pipegcr
FETI-DP KSPFETIDP fetidp
Conjugate Gradient Squared [35]KSPCGS cgs
Transpose-Free Quasi-Minimal Residual (1) [14]KSPTFQMR tfqmr
Transpose-Free Quasi-Minimal Residual (2) KSPTCQMR tcqmr
Least Squares Method KSPLSQR lsqr
Symmetric LQ Method [27]KSPSYMMLQ symmlq
TSIRM KSPTSIRM tsirm
Python Shell KSPPYTHON python
Shell for no KSP method KSPPREONLY preonly
Table 3: KSP Objects.
83
![](asset-54.png)
4.3. KRYLOV METHODS PETSc 3.8 March 24, 2018
Note: the bi-conjugate gradient method requires application of both the matrix and its transpose
plus the preconditioner and its transpose. Currently not all matrices and preconditioners provide
this support and thus the KSPBICG cannot always be used.
Note: PETSc implements the FETI-DP (Finite Element Tearing and Interconnecting Dual-
Primal) method as a KSP since it recasts the original problem into a contstrained minimization
one with Lagrange multipliers. The only matrix type supported is MATIS. Support for saddle point
problems is provided. Consult the online documentation for further details.
4.3.2 Convergence Tests
The default convergence test, KSPConvergedDefault(), is based on the l2-norm of the residual.
Convergence (or divergence) is decided by three quantities: the decrease of the residual norm
relative to the norm of the right hand side, rtol, the absolute size of the residual norm, atol, and
the relative increase in the residual, dtol. Convergence is detected at iteration kif
krkk2<max(rtol ∗ kbk2,atol),
where rk=b−Axk. Divergence is detected if
krkk2>dtol ∗ kbk2.
These parameters, as well as the maximum number of allowable iterations, can be set with the
routine
KSPSetTolerances(KSP ksp,PetscReal rtol,PetscReal atol,PetscReal dtol,PetscInt
maxits);
The user can retain the default value of any of these parameters by specifying PETSC_DEFAULT as
the corresponding tolerance; the defaults are rtol=1e-5,atol=1e-50,dtol=1e5, and maxits=1e4.
These parameters can also be set from the options database with the commands -ksp_rtol <rtol>,
-ksp_atol <atol>,-ksp_divtol <dtol>, and -ksp_max_it <its>.
In addition to providing an interface to a simple convergence test, KSP allows the application
programmer the flexibility to provide customized convergence-testing routines. The user can specify
a customized routine with the command
KSPSetConvergenceTest(KSP ksp,PetscErrorCode (*test)(KSP ksp,PetscInt
it,PetscReal rnorm, KSPConvergedReason *reason,void *ctx),void
*ctx,PetscErrorCode (*destroy)(void *ctx));
The final routine argument, ctx, is an optional context for private data for the user-defined conver-
gence routine, test. Other test routine arguments are the iteration number, it, and the residual’s
l2norm, rnorm. The routine for detecting convergence, test, should set reason to positive for con-
vergence, 0 for no convergence, and negative for failure to converge. A full list of possible values for
KSPConvergedReason is given in include/petscksp.h. You can use KSPGetConvergedReason()
after KSPSolve() to see why convergence/divergence was detected.
4.3.3 Convergence Monitoring
By default, the Krylov solvers run silently without displaying information about the iterations.
The user can indicate that the norms of the residuals should be displayed by using -ksp_monitor
within the options database. To display the residual norms in a graphical window (running under X
Windows), one should use -ksp_monitor_lg_residualnorm [x,y,w,h], where either all or none
of the options must be specified. Application programmers can also provide their own routines to
perform the monitoring by using the command
84
![](asset-55.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
KSPMonitorSet(KSP ksp,PetscErrorCode (*mon)(KSP ksp,PetscInt it,PetscReal
rnorm,void *ctx),void *ctx,PetscErrorCode (*mondestroy)(void**));
The final routine argument, ctx, is an optional context for private data for the user-defined moni-
toring routine, mon. Other mon routine arguments are the iteration number (it) and the residual’s l2
norm (rnorm). A helpful routine within user-defined monitors is
PetscObjectGetComm((PetscObject)ksp,MPI_Comm *comm), which returns in comm the MPI com-
municator for the KSP context. See section 1.4 for more discussion of the use of MPI communicators
within PETSc.
Several monitoring routines are supplied with PETSc, including
KSPMonitorDefault(KSP,PetscInt,PetscReal, void *);
KSPMonitorSingularValue(KSP,PetscInt,PetscReal,void *);
KSPMonitorTrueResidualNorm(KSP,PetscInt,PetscReal, void *);
The default monitor simply prints an estimate of the l2-norm of the residual at each iteration.
The routine KSPMonitorSingularValue() is appropriate only for use with the conjugate gradient
method or GMRES, since it prints estimates of the extreme singular values of the preconditioned
operator at each iteration. Since KSPMonitorTrueResidualNorm() prints the true residual at each
iteration by actually computing the residual using the formula r=b−Ax, the routine is slow
and should be used only for testing or convergence studies, not for timing. These monitors may
be accessed with the command line options -ksp_monitor,-ksp_monitor_singular_value, and
-ksp_monitor_true_residual.
To employ the default graphical monitor, one should use the commands
PetscDrawLG lg;
KSPMonitorLGResidualNormCreate(MPI Comm comm,char *display,char *title,PetscInt
x,PetscInt y,PetscInt w,PetscInt h,PetscDrawLG *lg);
KSPMonitorSet(KSP ksp,KSPMonitorLGResidualNorm,lg,0);
When no longer needed, the line graph should be destroyed with the command
PetscDrawLGDestroy(PetscDrawLG *lg);
The user can change aspects of the graphs with the PetscDrawLG*() and PetscDrawAxis*()
routines. One can also access this functionality from the options database with the command
-ksp_monitor_lg_residualnorm [x,y,w,h]. , where x, y, w, h are the optional location and
size of the window.
One can cancel hardwired monitoring routines for KSP at runtime with -ksp_monitor_cancel.
Unless the Krylov method converges so that the residual norm is small, say 10−10, many of
the final digits printed with the -ksp_monitor option are meaningless. Worse, they are different
on different machines; due to different round-off rules used by, say, the IBM RS6000 and the Sun
SPARC. This makes testing between different machines difficult. The option -ksp_monitor_short
causes PETSc to print fewer of the digits of the residual norm as it gets smaller; thus on most of
the machines it will always print the same numbers making cross system testing easier.
4.3.4 Understanding the Operator’s Spectrum
Since the convergence of Krylov subspace methods depends strongly on the spectrum (eigenvalues)
of the preconditioned operator, PETSc has specific routines for eigenvalue approximation via the
Arnoldi or Lanczos iteration. First, before the linear solve one must call
85
![](asset-56.png)
4.4. PRECONDITIONERS PETSc 3.8 March 24, 2018
KSPSetComputeEigenvalues(KSP ksp,PETSC TRUE);
Then after the KSP solve one calls
KSPComputeEigenvalues(KSP ksp,PetscInt n,PetscReal *realpart,PetscReal
*complexpart,PetscInt *neig);
Here, nis the size of the two arrays and the eigenvalues are inserted into those two arrays. neig is
the number of eigenvalues computed; this number depends on the size of the Krylov space generated
during the linear system solution, for GMRES it is never larger than the restart parameter. There
is an additional routine
KSPComputeEigenvaluesExplicitly(KSP ksp, PetscInt n,PetscReal
*realpart,PetscReal *complexpart);
that is useful only for very small problems. It explicitly computes the full representation of the
preconditioned operator and calls LAPACK to compute its eigenvalues. It should be only used for
matrices of size up to a couple hundred. The PetscDrawSP*() routines are very useful for drawing
scatter plots of the eigenvalues.
The eigenvalues may also be computed and displayed graphically with the options data base
commands -ksp_plot_eigenvalues and -ksp_plot_eigenvalues_explicitly. Or they can
be dumped to the screen in ASCII text via -ksp_compute_eigenvalues and -ksp_compute_
eigenvalues_explicitly.
4.3.5 Other KSP Options
To obtain the solution vector and right hand side from a KSP context, one uses
KSPGetSolution(KSP ksp,Vec *x);
KSPGetRhs(KSP ksp,Vec *rhs);
During the iterative process the solution may not yet have been calculated or it may be stored in
a different location. To access the approximate solution during the iterative process, one uses the
command
KSPBuildSolution(KSP ksp,Vec w,Vec *v);
where the solution is returned in v. The user can optionally provide a vector in was the location to
store the vector; however, if wis NULL, space allocated by PETSc in the KSP context is used. One
should not destroy this vector. For certain KSP methods, (e.g., GMRES), the construction of the
solution is expensive, while for many others it doesn’t evenrequire a vector copy.
Access to the residual is done in a similar way with the command
KSPBuildResidual(KSP ksp,Vec t,Vec w,Vec *v);
Again, for GMRES and certain other methods this is an expensive operation.
4.4 Preconditioners
As discussed in Section 4.3.1, Krylov subspace methods are typically used in conjunction with a
preconditioner. To employ a particular preconditioning method, the user can either select it from
the options database using input of the form -pc_type <methodname> or set the method with the
command
86
![](asset-57.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
Method PCType Options Database Name
Jacobi PCJACOBI jacobi
Block Jacobi PCBJACOBI bjacobi
SOR (and SSOR) PCSOR sor
SOR with Eisenstat trick PCEISENSTAT eisenstat
Incomplete Cholesky PCICC icc
Incomplete LU PCILU ilu
Additive Schwarz PCASM asm
Generalized Additive Schwarz PCGASM gasm
Algebraic Multigrid PCGAMG gamg
Balancing Domain Decomposition by Constraints PCBDDC bddc
Linear solver PCKSP ksp
Combination of preconditioners PCCOMPOSITE composite
LU PCLU lu
Cholesky PCCHOLESKY cholesky
No preconditioning PCNONE none
Shell for user-defined PC PCSHELL shell
Table 4: PETSc Preconditioners (partial list)
PCSetType(PC pc,PCType method);
In Table 4we summarize the basic preconditioning methods supported in PETSc. See the PCType
manual page for a complete list. The PCSHELL preconditioner uses a specific, application-provided
preconditioner. The direct preconditioner, PCLU , is, in fact, a direct solver for the linear system
that uses LU factorization. PCLU is included as a preconditioner so that PETSc has a consistent
interface among direct and iterative linear solvers.
Each preconditioner may have associated with it a set of options, which can be set with routines
and options database commands provided for this purpose. Such routine names and commands are
all of the form PC<TYPE><Option> and -pc_<type>_<option> [value]. A complete list can be
found by consulting the PCType manual page; we discuss just a few in the sections below.
4.4.1 ILU and ICC Preconditioners
Some of the options for ILU preconditioner are
PCFactorSetLevels(PC pc,PetscInt levels);
PCFactorSetReuseOrdering(PC pc,PetscBool flag);
PCFactorSetDropTolerance(PC pc,PetscReal dt,PetscReal dtcol,PetscInt dtcount);
PCFactorSetReuseFill(PC pc,PetscBool flag);
PCFactorSetUseInPlace(PC pc,PetscBool flg);
PCFactorSetAllowDiagonalFill(PC pc,PetscBool flg);
When repeatedly solving linear systems with the same KSP context, one can reuse some infor-
mation computed during the first linear solve. In particular, PCFactorSetReuseOrdering() causes
the ordering (for example, set with -pc_factor_mat_ordering_type order) computed in the first
factorization to be reused for later factorizations. PCFactorSetUseInPlace() is often used with
PCASM or PCBJACOBI when zero fill is used, since it reuses the matrix space to store the incomplete
87
![](asset-58.png)
4.4. PRECONDITIONERS PETSc 3.8 March 24, 2018
factorization it saves memory and copying time. Note that in-place factorization is not appropriate
with any ordering besides natural and cannot be used with the drop tolerance factorization. These
options may be set in the database with
•-pc_factor_levels <levels>
•-pc_factor_reuse_ordering
•-pc_factor_reuse_fill
•-pc_factor_in_place
•-pc_factor_nonzeros_along_diagonal
•-pc_factor_diagonal_fill
See Section 14.2.5 for information on preallocation of memory for anticipated fill during factor-
ization. By alleviating the considerable overhead for dynamic memory allocation, such tuning can
significantly enhance performance.
PETSc supports incomplete factorization preconditioners for several matrix types for sequential
matrices (for example MATSEQAIJ,MATSEQBAIJ, and MATSEQSBAIJ).
4.4.2 SOR and SSOR Preconditioners
PETSc only provides only a sequential SOR preconditioner; it can only be used with sequential
matrices or as the subblock preconditioner when using block Jacobi or ASM preconditioning (see
below).
The options for SOR preconditioning with PCSOR are
PCSORSetOmega(PC pc,PetscReal omega);
PCSORSetIterations(PC pc,PetscInt its,PetscInt lits);
PCSORSetSymmetric(PC pc,MatSORType type);
The first of these commands sets the relaxation factor for successive over (under) relaxation. The
second command sets the number of inner iterations its and local iterations lits (the number
of smoothing sweeps on a process before doing a ghost point update from the other processes)
to use between steps of the Krylov space method. The total number of SOR sweeps is given
by its*lits. The third command sets the kind of SOR sweep, where the argument type can
be one of SOR_FORWARD_SWEEP,SOR_BACKWARD_SWEEP or SOR_SYMMETRIC_SWEEP, the default being
SOR_FORWARD_SWEEP. Setting the type to be SOR_SYMMETRIC_SWEEP produces the SSOR method. In
addition, each process can locally and independently perform the specified variant of SOR with the
types SOR_LOCAL_FORWARD_SWEEP,SOR_LOCAL_BACKWARD_SWEEP, and SOR_LOCAL_SYMMETRIC_SWEEP.
These variants can also be set with the options -pc_sor_omega <omega>,-pc_sor_its <its>,
-pc_sor_lits <lits>,-pc_sor_backward,-pc_sor_symmetric,-pc_sor_local_forward,-pc_
sor_local_backward, and -pc_sor_local_symmetric.
The Eisenstat trick [10] for SSOR preconditioning can be employed with the method PCEISENSTAT
(-pc_type eisenstat). By using both left and right preconditioning of the linear system, this vari-
ant of SSOR requires about half of the floating-point operations for conventional SSOR. The option
-pc_eisenstat_no_diagonal_scaling) (or the routine PCEisenstatSetNoDiagonalScaling())
turns off diagonal scaling in conjunction with Eisenstat SSOR method, while the option -pc_
eisenstat_omega <omega> (or the routine PCEisenstatSetOmega(PC pc,PetscReal omega)) sets
the SSOR relaxation coefficient, omega, as discussed above.
4.4.3 LU Factorization
The LU preconditioner provides several options. The first, given by the command
88
![](asset-59.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
PCFactorSetUseInPlace(PC pc,PetscBool flg);
causes the factorization to be performed in-place and hence destroys the original matrix. The
options database variant of this command is -pc_factor_in_place. Another direct preconditioner
option is selecting the ordering of equations with the command -pc_factor_mat_ordering_type
<ordering>. The possible orderings are
•MATORDERINGNATURAL - Natural
•MATORDERINGND - Nested Dissection
•MATORDERING1WD - One-way Dissection
•MATORDERINGRCM - Reverse Cuthill-McKee
•MATORDERINGQMD - Quotient Minimum Degree
These orderings can also be set through the options database by specifying one of the follow-
ing: -pc_factor_mat_ordering_type natural, or nd, or 1wd, or rcm, or qmd. In addition, see
MatGetOrdering(), discussed in Section 17.2.
The sparse LU factorization provided in PETSc does not perform pivoting for numerical stability
(since they are designed to preserve nonzero structure), and thus occasionally a LU factorization
will fail with a zero pivot when, in fact, the matrix is non-singular. The option -pc_factor_
nonzeros_along_diagonal <tol> will often help eliminate the zero pivot, by preprocessing the
column ordering to remove small values from the diagonal. Here, tol is an optional tolerance to
decide if a value is nonzero; by default it is 1.e-10.
In addition, Section 14.2.5 provides information on preallocation of memory for anticipated fill
during factorization. Such tuning can significantly enhance performance, since it eliminates the
considerable overhead for dynamic memory allocation.
4.4.4 Block Jacobi and Overlapping Additive Schwarz Preconditioners
The block Jacobi and overlapping additive Schwarz methods in PETSc are supported in parallel;
however, only the uniprocess version of the block Gauss-Seidel method is currently in place. By
default, the PETSc implementations of these methods employ ILU(0) factorization on each indi-
vidual block (that is, the default solver on each subblock is PCType=PCILU,KSPType=KSPPREONLY);
the user can set alternative linear solvers via the options -sub_ksp_type and -sub_pc_type. In
fact, all of the KSP and PC options can be applied to the subproblems by inserting the prefix -sub_
at the beginning of the option name. These options database commands set the particular options
for all of the blocks within the global problem. In addition, the routines
PCBJacobiGetSubKSP(PC pc,PetscInt *n_local,PetscInt *first_local,KSP **subksp);
PCASMGetSubKSP(PC pc,PetscInt *n_local,PetscInt *first_local,KSP **subksp);
extract the KSP context for each local block. The argument n_local is the number of blocks on
the calling process, and first_local indicates the global number of the first block on the process.
The blocks are numbered successively by processes from zero through bg−1, where bgis the
number of global blocks. The array of KSP contexts for the local blocks is given by subksp. This
mechanism enables the user to set different solvers for the various blocks. To set the appropriate
data structures, the user must explicitly call KSPSetUp() before calling PCBJacobiGetSubKSP() or
PCASMGetSubKSP(). For further details, see the examples $PETSC_DIR/src/ksp/ksp/examples/
tutorials/ex7.c or $PETSC_DIR/src/ksp/ksp/examples/tutorials/ex8.c.
89
![](asset-5a.png)
4.4. PRECONDITIONERS PETSc 3.8 March 24, 2018
The block Jacobi, block Gauss-Seidel, and additive Schwarz preconditioners allow the user to
set the number of blocks into which the problem is divided. The options database commands
to set this value are -pc_bjacobi_blocks n and -pc_bgs_blocks n, and, within a program, the
corresponding routines are
PCBJacobiSetTotalBlocks(PC pc,PetscInt blocks,PetscInt *size);
PCASMSetTotalSubdomains(PC pc,PetscInt n,IS *is,IS *islocal);
PCASMSetType(PC pc,PCASMType type);
The optional argument size is an array indicating the size of each block. Currently, for certain
parallel matrix formats, only a single block per process is supported. However, the MATMPIAIJ
and MATMPIBAIJ formats support the use of general blocks as long as no blocks are shared among
processes. The is argument contains the index sets that define the subdomains.
The object PCASMType is one of PC_ASM_BASIC,PC_ASM_INTERPOLATE,PC_ASM_RESTRICT, or
PC_ASM_NONE and may also be set with the options database -pc_asm_type [basic,interpolate,
restrict,none]. The type PC_ASM_BASIC (or -pc_asm_type basic) corresponds to the stan-
dard additive Schwarz method that uses the full restriction and interpolation operators. The
type PC_ASM_RESTRICT (or -pc_asm_type restrict) uses a full restriction operator, but dur-
ing the interpolation process ignores the off-process values. Similarly, PC_ASM_INTERPOLATE (or
-pc_asm_type interpolate) uses a limited restriction process in conjunction with a full interpo-
lation, while PC_ASM_NONE (or -pc_asm_type none) ignores off-process values for both restriction
and interpolation. The ASM types with limited restriction or interpolation were suggested by
Xiao-Chuan Cai and Marcus Sarkis [6]. PC_ASM_RESTRICT is the PETSc default, as it saves sub-
stantial communication and for many problems has the added benefit of requiring fewer iterations
for convergence than the standard additive Schwarz method.
The user can also set the number of blocks and sizes on a per-process basis with the commands
PCBJacobiSetLocalBlocks(PC pc,PetscInt blocks,PetscInt *size);
PCASMSetLocalSubdomains(PC pc,PetscInt N,IS *is,IS *islocal);
For the ASM preconditioner one can use the following command to set the overlap to compute
in constructing the subdomains.
PCASMSetOverlap(PC pc,PetscInt overlap);
The overlap defaults to 1, so if one desires that no additional overlap be computed beyond what
may have been set with a call to PCASMSetTotalSubdomains() or PCASMSetLocalSubdomains(),
then overlap must be set to be 0. In particular, if one does not explicitly set the subdomains
in an application code, then all overlap would be computed internally by PETSc, and using an
overlap of 0 would result in an ASM variant that is equivalent to the block Jacobi preconditioner.
Note that one can define initial index sets is with any overlap via PCASMSetTotalSubdomains()
or PCASMSetLocalSubdomains(); the routine PCASMSetOverlap() merely allows PETSc to extend
that overlap further if desired.
PCGASM is an experimental generalization of PCASM that allows the user to specify subdomains
that span multiple MPI ranks. This can be useful for problems where small subdomains result in
poor convergence. To be effective, the multirank subproblems must be solved using a sufficient
strong subsolver, such as LU, for which SuperLU_DIST or a similar parallel direct solver could be
used; other choices may include a multigrid solver on the subdomains.
The interface for PCGASM is similar to that of PCASM. In particular, PCGASMType is one of
PC_GASM_BASIC,PC_GASM_INTERPOLATE,PC_GASM_RESTRICT,PC_GASM_NONE. These options have
the same meaning as with PCASM and may also be set with the options database -pc_gasm_type
[basic,interpolate,restrict,none].
90
![](asset-5b.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
Unlike PCASM, however, PCGASM allows the user to define subdomains that span multiple MPI
ranks. The simplest way to do this is using a call to PCGASMSetTotalSubdomains(PC pc,PetscPetscInt
N) with the total number of subdomains Nthat is smaller than the MPI communicator size. In
this case PCGASM will coalesce size/N concecutive single-rank subdomains into a single multi-rank
subdomain. The single-rank subdomains contain the degrees of freedom corresponding to the
locally-owned rows of the PCGASM preconditioning matrix – these are the subdomains PCASM and
PCGASM use by default.
Each of the multirank subdomain subproblems is defined on the subcommunicator that contains
the coalesced PCGASM ranks. In general this might not result in a very good subproblem if the
single-rank problems corresponding to the coalesced ranks are not very strongly connected. In the
future this will be addressed with a hierarchical partitioner that generates well-connected coarse
subdomains first before subpartitioning them into the single-rank subdomains.
In the meantime the user can provide his or her own multi-rank subdomains by calling
PCGASMSetSubdomains(PC,IS[],IS[]) where each of the IS objects on the list defines the inner
(without the overlap) or the outer (including the overlap) subdomain on the subcommunicator of
the IS object. A helper subroutine PCGASMCreateSubdomains2D() is similar to PCASM’s but is ca-
pable of constructing multi-rank subdomains that can be then used with PCGASMSetSubdomains().
An alternative way of creating multi-rank subdomains is by using the underlying DM object, if
it is capable of generating such decompositions via DMCreateDomainDecomposition(). Ordinar-
ily the decomposition specified by the user via PCGASMSetSubdomains() takes precedence, unless
PCGASMSetUseDMSubdomains() instructs PCGASM to prefer DM-created decompositions.
Currently there is no support for increasing the overlap of multi-rank subdomains via
PCGASMSetOverlap() – this functionality works only for subdomains that fit within a single MPI
rank, exactly as in PCASM.
Examples of the described PCGASM usage can be found in ${PETSC_DIR}/src/ksp/ksp/
examples/tutorials/ex62.c. In particular, runex62_superlu_dist illustrates the use of SuperL
U_DIST as the subdomain solver on coalesced multi-rank subdomains. The runex62_2D_* examples
illustrate the use of PCGASMCreateSubdomains2D().
4.4.5 Algebraic Multigrid (AMG) Preconditioners
PETSc has a native algebraic multigrid preconditioner PCGAMG –gamg – and interfaces to two exter-
nal AMG packages: hypre and ML.Hypre is relatively monolithic in that a PETSc matrix is into a
hypre matrix and then hypre is called to do the entire solve. ML is more modular in that PETSc only
has ML generate the coarse grid spaces (columns of the prolongation operator), which is core of an
AMG method, and then constructs a PCMG with Galerkin coarse grid operator construction. GAMG
is designed from the beginning to be modular, to allow for new components to be added easily and
also populates a multigrid preconditioner PCMG so generic multigrid parameters are used. PETSc
provides a fully supported (smoothed) aggregation AMG, (-pc_type gamg -pc_gamg_type agg
or PCSetType(pc,PCGAMG)and PCGAMGSetType(pc,PCGAMGAGG), as well as reference implementa-
tions of a classical AMG method (-pc_gamg_type classical), a hybrid geometric AMG method
(-pc_gamg_type geo), and a 2.5D AMG method DofColumns [20]. GAMG does require the use
of (MPI)AIJ matrices. For instance, BAIJ matrices are not supported. One can use AIJ instead of
BAIJ without changing any code other than the constructor (or the -mat_type from the command
line). For instance, MatSetValuesBlocked works with AIJ matrices.
GAMG provides unsmoothed aggregation (-pc_gamg_agg_nsmooths 0) and smoothed aggre-
gation (-pc_gamg_agg_nsmooths 1 or PCGAMGSetNSmooths(pc,1)). Smoothed aggregation is rec-
ommended for symmetric positive definite systems. Unsmoothed aggregation can be useful for
asymmetric problems and problems where highest eigen estimates are problematic. If poor conver-
91
![](asset-5c.png)
4.4. PRECONDITIONERS PETSc 3.8 March 24, 2018
gence rates are observed using the smoothed version one can test unsmoothed aggregation. AMG
methods requires knowledge of the number of degrees of freedom per vertex, the default is one
(a scalar problem). Vector problems like elasticity should set the block size of the matrix appro-
priately with -mat_block_size bs or MatSetBlockSize(mat,bs). Equations must be ordered in
“vertex-major” ordering (e.g., x1, y1, z1, x2, y2, ...).
Smoothed aggregation requires an explicit representation of the (near) null space of the op-
erator for optimal performance. One can provide an orthonormal set of null space vectors with
MatSetNearNullSpace(). The vector of all ones is the default, for each variable given by the block
size (e.g., the translational rigid body modes). For elasticity, where rotational rigid body modes are
required to complete the near null space you can use MatNullSpaceCreateRigidBody() to create
the null space vectors and then MatSetNearNullSpace().
There are several options to provide parameters to the coarsening algorithm and parallel data
layout. Run a code that uses GAMG with -help to get full listing of GAMG parameters with
short parameter descriptions.
Trouble shooting algebraic multigrid methods: If GAMG,ML, or hypre does not perform
well the first thing to try is one of the other methods. Often the default parameters or just the
strengths of different algorithms can fix performance problems or provide useful information to
guide further debugging.
There are several sources of poor performance of AMG solvers and often special purpose meth-
ods must be developed to achieve the full potential of multigrid. To name just a few sources of
performance degradation that may not be fixed with parameters in PETSc currently: non-elliptic
operators, curl/curl operators, highly stretched grids or highly anisotropic problems, large jumps
in material coefficients with complex geometry (AMG is particularly well suited to jumps in coef-
ficients but it is not a perfect solution), highly incompressible elasticity, not to mention ill-posed
problems, and many others. For Grad-Div and Curl-Curl operators, you may want to try the
Auxilary Maxwell Space (AMS, -pc_type hypre -pc_hypre_set_type ams) or the Auxilary Di-
vergence Space (ADS, -pc_type hypre -pc_hypre_set_type ads) solvers. These solvers need
some additional information on the underlying mesh; specifically, AMS needs the discrete gradient
operator, which can be specified via PCHYPRESetDiscreteGradient(). In addition to the discrete
gradient, ADS also needs the specification of the discrete curl operator, which can be set using
PCHYPRESetDiscreteCurl().
I am converging slowly, what do I do? AMG methods are sensitive to coarsening rates and
methods; for GAMG use -pc_gamg_threshold <x> to regulate coarsening rates and PCGAMGSet-
Threshold, higher values decrease coarsening rate. Squaring the graph is the second mechanism for
increasing coarsening rate. Use -pc_gamg_square_graph <N>,, or PCGAMGSetSquareGraph(pc,N),
to square the graph on the finest N levels. A high threshold (e.g., x= 0.08) will result in an ex-
pensive but potentially powerful preconditioner, and a low threshold (e.g., x= 0.0) will result in
faster coarsening, fewer levels, cheaper solves, and generally worse convergence rates.
One can run with -info and grep for “GAMG” to get some statistics on each level, which
can be used to see if you are coarsening at an appropriate rate. With smoothed aggregation you
generally want to coarse at about a rate of 3:1 in each dimension. Coarsening too slow will result
in large numbers of non-zeros per row on coarse grids (this is reported). The number of non-
zeros can go up very high, say about 300 (times the degrees-of-freedom per vertex) on a 3D hex
mesh. One can also look at the grid complexity, which is also reported (the ration of the total
number of matrix entries for all levels to the number of matrix entries on the fine level). Grid
complexity should be well under 2.0 and preferably around 1.3 or lower. If convergence is poor
and the Galerkin coarse grid construction is much smaller than the time for each solve then one
can safely decrease the coarsening rate. -pc_gamg_threshold 0.0 is the simplest and most robust
option, and is recommended if poor convergence rates are observed, at least until the source of the
92
![](asset-5d.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
problem is discovered. In conclusion, if convergence is slow then decreasing the coarsening rate
(increasing the threshold) should be tried.
A note on Chebyshev smoothers. Chebyshev solvers are attractive as multigrid smoothers
because they can target a specific interval of the spectrum which is the purpose of a smoother.
The spectral bounds for Chebyshev solvers are simple to compute because they rely on the highest
eigenvalue of your (diagonally preconditioned) operator, which is conceptually simple to compute.
However, if this highest eigenvalue estimate is not accurate (too low) then the solvers can fail with
and indefinite preconditioner message. One can run with -info and grep for “GAMG” to get
these estimates or use -ksp_view. These highest eigenvalues are generally between 1.5-3.0. For
symmetric positive definite systems CG is a better eigenvalue estimator -mg_levels_esteig_ksp_
type cg. Indefinite matrix messages are often caused by bad Eigen estimates. Explicitly damped
Jacobi or Krylov smoothers can provide an alternative to Chebyshev and hypre has alternative
smoothers.
Now am I solving alright, can I expect better? If you find that you are getting nearly on
digit in reduction of the residual per iteration and are using a modest number of point smoothing
steps (e.g., 1-4 iterations of SOR), then you may be fairly close to textbook multigrid efficiency.
Although you also need to check the setup costs. This can be determined by running with -
log_view and check that the time for the Galerkin coarse grid construction (MatPtAP) is not
(much) more than the time spent in each solve (KSPSolve). If the MatPtAP time is too large then
one can increase the coarsening rate by decreasing the threshold and squaring the coarsening graph
(-pc_gamg_square_graph <N>, squares the graph on the finest N levels). Likewise if your MatPtAP
time is small and your convergence rate is not ideal then you could decrease the coarsening rate.
PETSc’s AMG solver is constructed as a framework for developers to easily add AMG capa-
bilities, like a new AMG methods or anAMG component like a matrix triple product. Contact us
directly if you are interested in contributing.
4.4.6 Balancing Domain Decomposition by Constraints
PETSc provides the Balancing Domain Decomposition by Constraints (BDDC) method for pre-
conditioning parallel finite element problems stored in unassembled format (see MATIS). BDDC is
a 2-level non-overlapping domain decomposition method which can be easily adapted to different
problems and discretizations by means of few user customizations. The application of the precondi-
tioner to a vector consists in the static condensation of the residual at the interior of the subdomains
by means of local Dirichet solves, followed by an additive combination of Neumann local corrections
and the solution of a global coupled coarse problem. Command line options for the underlying KSP
objects are prefixed by -pc_bddc_dirichlet,-pc_bddc_neumann, and -pc_bddc_coarse respec-
tively.
The current implementation supports any kind of linear system, and assumes a one-to-one
mapping between subdomains and MPI processes. Complex numbers are supported as well. For
non-symmetric problems, use the runtime option -pc_bddc_symmetric 0.
Unlike conventional non-overlapping methods that iterates just on the degrees of freedom at
the interface between subdomain, PCBDDC iterates on the whole set of degrees of freedom, allowing
the use of approximate subdomain solvers. When using approximate solvers, the command line
switches -pc_bddc_dirichlet_approximate and/or -pc_bddc_neumann_approximate should be
used to inform PCBDDC. If any of the local problems is singular, the nullspace of the local operator
should be attached to the local matrix via MatSetNullSpace().
At the basis of the method there’s the analysis of the connected components of the interface
for the detection of vertices, edges and faces equivalence classes. Additional information on the
degrees of freedom can be supplied to PCBDDC by using the following functions:
93
![](asset-5e.png)
4.4. PRECONDITIONERS PETSc 3.8 March 24, 2018
•PCBDDCSetDofsSplitting()
•PCBDDCSetLocalAdjacencyGraph()
•PCBDDCSetPrimalVerticesLocalIS()
•PCBDDCSetNeumannBoundaries()
•PCBDDCSetDirichletBoundaries()
•PCBDDCSetNeumannBoundariesLocal()
•PCBDDCSetDirichletBoundariesLocal()
Crucial for the convergence of the iterative process is the specification of the primal constraints
to be imposed at the interface between subdomains. PCBDDC uses by default vertex continuities
and edge arithmetic averages, which are enough for the three-dimensional Poisson problem with
constant coefficients. The user can switch on and off the usage of vertices, edges or face constraints
by using the command line switches -pc_bddc_use_vertices,-pc_bddc_use_edges,-pc_bddc_
use_faces. A customization of the constraints is available by attaching a MatNullSpace object to
the preconditioning matrix via MatSetNearNullSpace(). The vectors of the MatNullSpace object
should represent the constraints in the form of quadrature rules; quadrature rules for different classes
of the interface can be listed in the same vector. The number of vectors of the MatNullSpace object
corresponds to the maximum number of constraints that can be imposed for each class. Once all the
quadrature rules for a given interface class have been extracted, an SVD operation is performed to
retain the non-singular modes. As an example, the rigid body modes represent an effective choice
for elasticity, even in the almost incompressible case. For particular problems, e.g. edge-based
discretization with Nedelec elements, a user defined change of basis of the degrees of freedom can
be beneficial for PCBDDC; use PCBDDCSetChangeOfBasisMat() to customize the change of basis.
The BDDC method is usually robust with respect to jumps in the material parameters aligned
with the interface; for PDEs with more than one material parameter you may also consider to use
the so-called deluxe scaling, available via the command line switch -pc_bddc_use_deluxe_scaling.
Other scalings are available, see PCISSetSubdomainScalingFactor(),
PCISSetSubdomainDiagonalScaling() or PCISSetUseStiffnessScaling(). However, the con-
vergence properties of the BDDC method degrades in presence of large jumps in the material
coefficients not aligned with the interface; for such cases, PETSc has the capability of adaptively
computing the primal constraints. Adaptive selection of constraints could be requested by specify-
ing a threshold value at command line by using -pc_bddc_adaptive_threshold x. Valid values
for the threshold xranges from 1 to infinity, with smaller values corresponding to more robust
preconditioners. For SPD problems in 2D, or in 3D with only face degrees of freedom (like in the
case of Raviart-Thomas or Brezzi-Douglas-Marini elements), such a threshold is a very accurate
estimator of the condition number of the resulting preconditioned operator. Since the adaptive
selection of constraints for BDDC methods is still an active topic of research, its implementation is
currently limited to SPD problems; moreover, because the technique requires the explicit knowledge
of the local Schur complements, it needs the external package MUMPS.
When solving problems decomposed in thousands of subdomains or more, the solution of the
BDDC coarse problem could become a bottleneck; in order to overcome this issue, the user could
either consider to solve the parallel coarse problem on a subset of the communicator associated
with PCBDDC by using the command line switch -pc_bddc_coarse_redistribute, or instead use
a multilevel approach. The latter can be requested by specifying the number of requested level at
command line (-pc_bddc_levels) or by using PCBDDCSetLevels(). An additional parameter (see
PCBDDCSetCoarseningRatio()) controls the number of subdomains that will be generated at the
next level; the larger the coarsening ratio, the lower the number of coarser subdomains.
For further details, see the example ${PETSC_DIR}/src/ksp/ksp/examples/tutorials/ex59.
94
![](asset-5f.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
cand the online documentation for PCBDDC.
4.4.7 Shell Preconditioners
The shell preconditioner simply uses an application-provided routine to implement the precondi-
tioner. To set this routine, one uses the command
PCShellSetApply(PC pc,PetscErrorCode (*apply)(PC,Vec,Vec));
Often a preconditioner needs access to an application-provided data structured. For this, one should
use
PCShellSetContext(PC pc,void *ctx);
to set this data structure and
PCShellGetContext(PC pc,void **ctx);
to retrieve it in apply. The three routine arguments of apply() are the PC, the input vector, and
the output vector, respectively.
For a preconditioner that requires some sort of “setup” before being used, that requires a new
setup every time the operator is changed, one can provide a routine that is called every time the
operator is changed (usually via KSPSetOperators()).
PCShellSetSetUp(PC pc,PetscErrorCode (*setup)(PC));
The argument to the setup routine is the same PC object which can be used to obtain the op-
erators with PCGetOperators() and the application-provided data structure that was set with
PCShellSetContext().
4.4.8 Combining Preconditioners
The PC type PCCOMPOSITE allows one to form new preconditioners by combining already-defined
preconditioners and solvers. Combining preconditioners usually requires some experimentation to
find a combination of preconditioners that works better than any single method. It is a tricky
business and is not recommended until your application code is complete and running and you
are trying to improve performance. In many cases using a single preconditioner is better than
a combination; an exception is the multigrid/multilevel preconditioners (solvers) that are always
combinations of some sort, see Section 4.4.9.
Let B1and B2represent the application of two preconditioners of type type1 and type2. The
preconditioner B=B1+B2can be obtained with
PCSetType(pc,PCCOMPOSITE);
PCCompositeAddPC(pc,type1);
PCCompositeAddPC(pc,type2);
Any number of preconditioners may added in this way.
This way of combining preconditioners is called additive, since the actions of the preconditioners
are added together. This is the default behavior. An alternative can be set with the option
PCCompositeSetType(PC pc,PCCompositeType PC COMPOSITE MULTIPLICATIVE);
95
![](asset-60.png)
4.4. PRECONDITIONERS PETSc 3.8 March 24, 2018
In this form the new residual is updated after the application of each preconditioner and the next
preconditioner applied to the next residual. For example, with two composed preconditioners: B1
and B2;y=Bx is obtained from
y=B1x
w1=x−Ay
y=y+B2w1
Loosely, this corresponds to a Gauss-Seidel iteration, while additive corresponds to a Jacobi itera-
tion.
Under most circumstances, the multiplicative form requires one-half the number of iterations
as the additive form; however, the multiplicative form does require the application of Ainside the
preconditioner.
In the multiplicative version, the calculation of the residual inside the preconditioner can be
done in two ways: using the original linear system matrix or using the matrix used to build the
preconditioners B1,B2, etc. By default it uses the “preconditioner matrix”, to use the Amat matrix
use the option
PCSetUseAmat(PC pc);
The individual preconditioners can be accessed (in order to set options) via
PCCompositeGetPC(PC pc,PetscInt count,PC *subpc);
For example, to set the first sub preconditioners to use ILU(1)
PC subpc;
PCCompositeGetPC(pc,0,&subpc);
PCFactorSetFill(subpc,1);
These various options can also be set via the options database. For example, -pc_type
composite -pc_composite_pcs jacobi,ilu causes the composite preconditioner to be used with
two preconditioners: Jacobi and ILU. The option -pc_composite_type multiplicative initiates
the multiplicative version of the algorithm, while -pc_composite_type additive the additive ver-
sion. Using the Amat matrix is obtained with the option -pc_use_amat. One sets options for the
sub-preconditioners with the extra prefix -sub_N_ where Nis the number of the sub-preconditioner.
For example, -sub_0_pc_ifactor_fill 0.
PETSc also allows a preconditioner to be a complete linear solver. This is achieved with the
PCKSP type.
PCSetType(PC pc,PCKSP PCKSP);
PCKSPGetKSP(pc,&ksp);
/* set any KSP/PC options */
From the command line one can use 5 iterations of biCG-stab with ILU(0) preconditioning as
the preconditioner with -pc_type ksp -ksp_pc_type ilu -ksp_ksp_max_it 5 -ksp_ksp_type
bcgs.
By default the inner KSP solver uses the outer preconditioner matrix, Pmat, as the matrix to be
solved in the linear system; to use the matrix that defines the linear system, Amat use the option
PCSetUseAmat(PC pc);
96
![](asset-61.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
or at the command line with -pc_use_amat.
Naturally, one can use a PCKSP preconditioner inside a composite preconditioner. For example, -
pc_type composite -pc_composite_pcs ilu,ksp -sub_1_pc_type jacobi -sub_1_ksp_max_it
10 uses two preconditioners: ILU(0) and 10 iterations of GMRES with Jacobi preconditioning.
However, it is not clear whether one would ever wish to do such a thing.
4.4.9 Multigrid Preconditioners
A large suite of routines is available for using geometric multigrid as a preconditioner1. In the
PC framework, the user is required to provide the coarse grid solver, smoothers, restriction and
interpolation operators, and code to calculate residuals. The PC package allows these components
to be encapuslated within a PETSc-compliant preconditioner. We fully support both matrix-free
and matrix-based multigrid solvers.
A multigrid preconditioner is created with the four commands
KSPCreate(MPI Comm comm,KSP *ksp);
KSPGetPC(KSP ksp,PC *pc);
PCSetType(PC pc,PCMG);
PCMGSetLevels(pc,PetscInt levels,MPI Comm *comms);
A large number of parameters affect the multigrid behavior. The command
PCMGSetType(PC pc,PCMGType mode);
indicates which form of multigrid to apply [34].
For standard V or W-cycle multigrids, one sets the mode to be PC_MG_MULTIPLICATIVE; for the
additive form (which in certain cases reduces to the BPX method, or additive multilevel Schwarz, or
multilevel diagonal scaling), one uses PC_MG_ADDITIVE as the mode. For a variant of full multigrid,
one can use PC_MG_FULL, and for the Kaskade algorithm PC_MG_KASKADE. For the multiplicative
and full multigrid options, one can use a W-cycle by calling
PCMGSetCycleType(PC pc,PCMGCycleType ctype);
with a value of PC_MG_CYCLE_W for ctype. The commands above can also be set from the options
database. The option names are -pc_mg_type [multiplicative, additive, full, kaskade],
and -pc_mg_cycle_type <ctype>.
The user can control the amount of pre- and postsmoothing by using either the options
-pc_mg_smoothup m and -pc_mg_smoothdown n or the routines
PCMGSetNumberSmoothUp(PC pc,PetscInt m);
PCMGSetNumberSmoothDown(PC pc,PetscInt n);
The multigrid routines, which determine the solvers and interpolation/restriction operators that
are used, are mandatory. To set the coarse grid solver, one must call
PCMGGetCoarseSolve(PC pc,KSP *ksp);
and set the appropriate options in ksp. Similarly, the smoothers are setcontrolled by first calling
PCMGGetSmoother(PC pc,PetscInt level,KSP *ksp);
and then setting the various options in the ksp. For example,
1See Section 4.4.5 for information on using algebraic multigrid.
97
![](asset-62.png)
4.4. PRECONDITIONERS PETSc 3.8 March 24, 2018
PCMGGetSmoother(pc,1,&ksp);
KSPSetOperators(ksp,A1,A1);
sets the matrix that defines the smoother on level 1 of the multigrid. While
PCMGGetSmoother(pc,1,&ksp);
KSPGetPC(ksp,&pc);
PCSetType(pc,PCSOR);
sets SOR as the smoother to use on level 1.
To use a different pre- or postsmoother, one should call the following routines instead.
PCMGGetSmootherUp(PC pc,PetscInt level,KSP *upksp);
PCMGGetSmootherDown(PC pc,PetscInt level,KSP *downksp);
Use
PCMGSetInterpolation(PC pc,PetscInt level,Mat P);
and
PCMGSetRestriction(PC pc,PetscInt level,Mat R);
to define the intergrid transfer operations. If only one of these is set, its transpose will be used for
the other.
It is possible for these interpolation operations to be matrix free (see Section 3.3); One should
then make sure that these operations are defined for the (matrix-free) matrices passed in. Note
that this system is arranged so that if the interpolation is the transpose of the restriction, you can
pass the same mat argument to both PCMGSetRestriction() and PCMGSetInterpolation().
On each level except the coarsest, one must also set the routine to compute the residual. The
following command suffices:
PCMGSetResidual(PC pc,PetscInt level,PetscErrorCode
(*residual)(Mat,Vec,Vec,Vec),Mat mat);
The residual() function normally does not need to be set if one’s operator is stored in Mat format.
In certain circumstances, where it is much cheaper to calculate the residual directly, rather than
through the usual formula b−Ax, the user may wish to provide an alternative.
Finally, the user may provide three work vectors for each level (except on the finest, where only
the residual work vector is required). The work vectors are set with the commands
PCMGSetRhs(PC pc,PetscInt level,Vec b);
PCMGSetX(PC pc,PetscInt level,Vec x);
PCMGSetR(PC pc,PetscInt level,Vec r);
The PC references these vectors, so you should call VecDestroy() when you are finished with them.
If any of these vectors are not provided, the preconditioner will allocate them.
One can control the KSP and PC options used on the various levels (as well as the coarse
grid) using the prefix mg_levels_ (mg_coarse_ for the coarse grid). For example, -mg_levels_
ksp_type cg will cause the CG method to be used as the Krylov method for each level. Or
-mg_levels_pc_type ilu -mg_levels_pc_factor_levels 2 will cause the ILU preconditioner
to be used on each level with two levels of fill in the incomplete factorization.
98
![](asset-63.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
4.5 Solving Block Matrices
Block matrices represent an important class of problems in numerical linear algebra and offer the
possibility of far more efficient iterative solvers than just treating the entire matrix as black box.
In this section we use the common linear algebra definition of block matrices where matrices are
divided in a small, problem-size independent (two, three or so) number of very large blocks. These
blocks arise naturally from the underlying physics or discretization of the problem, for example,
the velocity and pressure. Under a certain numbering of unknowns the matrix can be written as
A00 A01 A02 A03
A10 A11 A12 A13
A20 A21 A22 A23
A30 A31 A32 A33
,
where each Aij is an entire block. On a parallel computer the matrices are not explicitly stored
this way. Instead, each process will own some of the rows of A0∗,A1∗etc. On a process, the blocks
may be stored one block followed by another
A0000 A0001 A0002 ... A0100 A0102 ...
A0010 A0011 A0012 ... A0110 A0112 ...
A0020 A0021 A0022 ... A0120 A0122 ...
...
A1000 A1001 A1002 ... A1100 A1102 ...
A1010 A1011 A1012 ... A1110 A1112 ...
...
or interlaced, for example with two blocks
A0000 A0100 A0001 A0101 ...
A1000 A1100 A1001 A1101 ...
...
A0010 A0110 A0011 A0111 ...
A1010 A1110 A1011 A1111 ...
...
.
Note that for interlaced storage the number of rows/columns of each block must be the same size.
Matrices obtained with DMCreateMatrix() where the DM is a DMDA are always stored interlaced.
Block matrices can also be stored using the MATNEST format which holds separate assembled blocks.
Each of these nested matrices is itself distributed in parallel. It is more efficient to use MATNEST
with the methods described in this section because there are fewer copies and better formats (e.g.
BAIJ or SBAIJ) can be used for the components, but it is not possible to use many other methods
with MATNEST. See Section 3.1.3 for more on assembling block matrices without depending on a
specific matrix format.
The PETSc PCFIELDSPLIT preconditioner is used to implement the “block” solvers in PETSc.
There are three ways to provide the information that defines the blocks. If the matrices are
stored as interlaced then PCFieldSplitSetFields() can be called repeatedly to indicate which
fields belong to each block. More generally PCFieldSplitSetIS() can be used to indicate exactly
which rows/columns of the matrix belong to a particular block. You can provide names for each
block with these routines, if you do not provide names they are numbered from 0. With these
two approaches the blocks may overlap (though generally they will not). If only one block is
defined then the complement of the matrices is used to define the other block. Finally the option
99
![](asset-64.png)
4.5. SOLVING BLOCK MATRICES PETSc 3.8 March 24, 2018
-pc_fieldsplit_detect_saddle_point causes two diagonal blocks to be found, one associated
with all rows/columns that have zeros on the diagonals and the rest.
For simplicity in the rest of the section we restrict our matrices to two by two blocks. So the
matrix is A00 A01
A10 A11 .
On occasion the user may provide another matrix that is used to construct parts of the precondi-
tioner Ap00 Ap01
Ap10 Ap11 .
For notational simplicity define ksp(A, Ap) to mean approximately solving a linear system using
KSP with operator Aand preconditioner built from matrix Ap.
For matrices defined with any number of blocks there are three “block” algorithms available:
block Jacobi, ksp(A00, Ap00) 0
0 ksp(A11, Ap11)
block Gauss-Seidel, I0
0A−1
11 I0
−A10 I A−1
00 0
0I
which is implemented2as
I0
0 ksp(A11, Ap11)0 0
0I+I0
−A10 −A11 I0
0 0 ksp(A00, Ap00) 0
0I
and symmetric block Gauss-Seidel
A−1
00 0
0I I−A01
0I A00 0
0A−1
11 I0
−A10 I A−1
00 0
0I.
These can be accessed with -pc_fieldsplit_type<additive,multiplicative,
symmetric_multiplicative> or the function PCFieldSplitSetType(). The option prefixes for
the internal KSPs are given by -fieldsplit_name_.
By default blocks A00, A01 and so on are extracted out of Pmat, the matrix that the KSP uses to
build the preconditioner, and not out of Amat (i.e., Aitself). As discussed above in Sec. 4.4.8, how-
ever, it is possible to use Amat instead of Pmat by calling PCSetUseAmat(pc) or using -pc_use_amat
on the command line. Alternatively, you can have PCFieldSplit extract the diagonal blocks
A00, A11 etc. out of Amat by calling PCFieldSplitSetDiagUseAmat(pc,PETSC_TRUE) or supplying
command-line argument -pc_fieldsplit_diag_use_amat. Similarly,
PCFieldSplitSetOffDiagUseAmat(pc,{PETSC_TRUE) or -pc_fieldsplit_off_diag_use_amat will
cause the off-diagonal blocks A01, A10 etc. to be extracted out of Amat.
For two by two blocks only there are another family of solvers, based on Schur complements.
The inverse of the Schur complement factorization is
I0
A10A−1
00 I A00 0
0S I A−1
00 A01
0I−1
I A−1
00 A01
0I−1A−1
00 0
0S−1 I0
A10A−1
00 I−1
2This may seem an odd way to implement since it involves the “extra” multiply by −A11. The reason is this is
implemented this way is that this approach works for any number of blocks that may overlap.
100
![](asset-65.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
I−A−1
00 A01
0I A−1
00 0
0S−1 I0
−A10A−1
00 I
A−1
00 0
0I I−A01
0I A00 0
0S−1 I0
−A10 I A−1
00 0
0I.
The preconditioner is accessed with -pc_fieldsplit_type schur and is implemented as
ksp(A00, Ap00) 0
0I I−A01
0I I0
0 ksp( ˆ
S, ˆ
Sp) I0
−A10ksp(A00, Ap00)I.
Where ˆ
S=A11 −A10ksp(A00, Ap00)A01 is the approximate Schur complement.
There are several variants of the Schur complement preconditioner obtained by dropping some of
the terms, these can be obtained with -pc_fieldsplit_schur_fact_type <diag,lower,upper,
full> or the function PCFieldSplitSetSchurFactType(). Note that the diag form uses the
preconditioner ksp(A00, Ap00) 0
0−ksp( ˆ
S, ˆ
Sp).
This is done to ensure the preconditioner is positive definite for a common class of problems, saddle
points with a positive definite A00: for these the Schur complement is negative definite.
The effectiveness of the Schur complement preconditioner depends on the availability of a good
preconditioner ˆ
Sp for the Schur complement matrix. In general, you are responsible for supplying ˆ
Sp
via PCFieldSplitSchurPrecondition(pc,PC_FIELDSPLIT_SCHUR_PRE_USER,Sp). In the absence
of a good problem-specific ˆ
Sp, you can use some of the built-in options.
Using -pc_fieldsplit_schur_precondition user on the command line activates the matrix
supplied programmatically as explained above.
With -pc_fieldsplit_schur_precondition a11 (default) ˆ
Sp =A11 is used to build a pre-
conditioner for ˆ
S.
Otherwise, -pc_fieldsplit_schur_precondition self will set ˆ
Sp =ˆ
Sand use the Schur
complement matrix itself to build the preconditioner.
The problem with the last approach is that ˆ
Sis used in unassembled, matrix-free form, and
many preconditioners (e.g., ILU) cannot be built out of such matrices. Instead, you can assemble an
approximation to ˆ
Sby inverting A00, but only approximately, so as to ensure the sparsity of ˆ
Sp as
much as possible. Specifically, using -pc_fieldsplit_schur_precondition selfp will assemble
ˆ
Sp =A11 −A10inv(A00)A01.
By default inv(A00) is the inverse of the diagonal of A00, but using -fieldsplit_1_mat_schur_
complement_ainv_type lump will lump A00 first. Option -mat_schur_complement_ainv_type
applies to any matrix of MatSchurComplement type and here it is used with the prefix -fieldsplit_
1of the linear system in the second split.
Finally, you can use the PCLSC preconditioner for the Schur complement with -pc_fieldsplit_
type schur -fieldsplit_1_pc_type lsc. This uses for the preconditioner to ˆ
Sthe operator
ksp(A10A01, A10A01)A10A00A01ksp(A10A01, A10A01)
which, of course, introduces two additional inner solves for each application of the Schur comple-
ment. The options prefix for this inner KSP is -fieldsplit_1_lsc_. Instead of constructing the
matrix A10A01 the user can provide their own matrix. This is done by attaching the matrix/ma-
trices to the Sp matrix they provide with
PetscObjectCompose((PetscObject)Sp,"LSC_L",(PetscObject)L);
PetscObjectCompose((PetscObject)Sp,"LSC_Lp",(PetscObject)Lp);
101
![](asset-66.png)
4.6. SOLVING SINGULAR SYSTEMS PETSc 3.8 March 24, 2018
4.6 Solving Singular Systems
Sometimes one is required to solver singular linear systems. In this case, the system matrix has
a nontrivial null space. For example, the discretization of the Laplacian operator with Neumann
boundary conditions has a null space of the constant functions. PETSc has tools to help solve these
systems.
First, one must know what the null space is and store it using an orthonormal basis in an array
of PETSc Vecs. The constant functions can be handled separately, since they are such a common
cas). Create a MatNullSpace object with the command
MatNullSpaceCreate(MPI Comm,PetscBool hasconstants,PetscInt dim,Vec
*basis,MatNullSpace *nsp);
Her, dim is the number of vectors in basis and hasconstants indicates if the null space contains
the constant functions. If the null space contains the constant functions you do not need to include
it in the basis vectors you provide, nor in the count dim.
One then tells the KSP object you are using what the null space is with the call
MatSetNullSpace(Mat Amat,MatNullSpace nsp);
MatSetTransposeNullSpace(Mat Amat,MatNullSpace nsp);
The Amat should be the first matrix argument used with KSPSetOperators(),SNESSetJacobian(),
or TSSetIJacobian(). You can also use KSPSetNullspace(). The PETSc solvers will now handle
the null space during the solution process.
If one chooses a direct solver (or an incomplete factorization) it may still detect a zero pivot.
You can run with the additional options or -pc_factor_shift_type NONZERO -pc_factor_shift_
amount <dampingfactor> to prevent the zero pivot. A good choice for the dampingfactor is 1.e-
10.
4.7 Using External Linear Solvers
PETSc interfaces to several external linear solvers (see Acknowledgment at the beginning of this
manual). To use these solvers, one may:
1. Run ./configure with the additional options --download-packagename e.g. --download-
superlu_dist --download-parmetis (SuperLU DIST needs ParMetis) or --download-mumps
--download-scalapack (MUMPS requires ScaLAPACK).
2. Build the PETSc libraries.
3. Use the runtime option: -ksp_type preonly -pc_type <pctype> -pc_factor_mat_solver_
package <packagename>. For eg: -ksp_type preonly -pc_type lu -pc_factor_mat_solver_
package superlu_dist.
102
![](asset-67.png)
CHAPTER 4. KSP: LINEAR SYSTEM SOLVERS PETSc 3.8 March 24, 2018
MatType PCType MatSolverPackage Package
(-pc_factor_mat_solver_package)
seqaij lu MATSOLVERESSL essl
seqaij lu MATSOLVERLUSOL lusol
seqaij lu MATSOLVERMATLAB matlab
aij lu MATSOLVERMUMPS mumps
aij cholesky
sbaij cholesky
seqaij lu MATSOLVERSUPERLU superlu
aij lu MATSOLVERSUPERLU_DIST superlu_dist
seqaij lu MATSOLVERUMFPACK umfpack
seqaij cholesky MATSOLVERCHOLMOD cholmod
aij lu MATSOLVERCLIQUE clique
seqaij lu MATSOLVERKLU klu
dense lu MATSOLVERELEMENTAL elemental
dense cholesky
seqaij lu MATSOLVERMKL_PARDISO mkl_pardiso
aij lu MATSOLVERMKL_CPARDISO mkl_cpardiso
aij lu MATSOLVERPASTIX pastix
aij cholesky MATSOLVERBAS bas
aijcusp lu MATSOLVERCUSPARSE cusparse
aijcusparse lu
aijcusp cholesky
aijcusparse cholesky
aij lu MATSOLVERPETSC petsc
baij
aijcrl
aijperm
seqdense
aij cholesky
baij
aijcrl
aijperm
seqdense
Table 5: Options for External Solvers
The default and available input options for each external software can be found by specifying
-help (or -h) at runtime.
As an alternative to using runtime flags to employ these external packages, procedural calls
are provided for some packages. For example, the following procedural calls are equivalent to
runtime options -ksp_type preonly -pc_type lu -pc_factor_mat_solver_package mumps -
mat_mumps_icntl_7 2:
KSPSetType(ksp,KSPPREONLY);
KSPGetPC(ksp,&pc);
PCSetType(pc,PCLU);
PCFactorSetMatSolverPackage(pc,MATSOLVERMUMPS);
PCFactorGetMatrix(pc,&F);
103
![](asset-68.png)
4.7. USING EXTERNAL LINEAR SOLVERS PETSc 3.8 March 24, 2018
icntl=7; ival = 2;
MatMumpsSetIcntl(F,icntl,ival);
One can also create matrices with the appropriate capabilities by calling MatCreate() followed
by MatSetType() specifying the desired matrix type from Table 5. These matrix types inherit
capabilities from their PETSc matrix parents: seqaij,mpiaij, etc. As a result, the preallocation
routines MatSeqAIJSetPreallocation(),MatMPIAIJSetPreallocation(), etc. and any other
type specific routines of the base class are supported. One can also call MatConvert() inplace
to convert the matrix to and from its base class without performing an expensive data copy.
MatConvert() cannot be called on matrices that have already been factored.
In Table 5, the base class aij refers to the fact that inheritance is based on MATSEQAIJ when
constructed with a single process communicator, and from MATMPIAIJ otherwise. The same holds
for baij and sbaij. For codes that are intended to be run as both a single process or with multiple
processes, depending on the mpiexec command, it is recommended that both sets of preallocation
routines are called for these communicator morphing types. The call for the incorrect type will
simply be ignored without any harm or message.
104
Chapter 5
SNES: Nonlinear Solvers
The solution of large-scale nonlinear problems pervades many facets of computational science and
demands robust and flexible solution strategies. The SNES library of PETSc provides a powerful
suite of data-structure-neutral numerical routines for such problems. Built on top of the linear
solvers and data structures discussed in preceding chapters, SNES enables the user to easily cus-
tomize the nonlinear solvers according to the application at hand. Also, the SNES interface is
identical for the uniprocess and parallel cases; the only difference in the parallel version is that each
process typically forms only its local contribution to various matrices and vectors.
The SNES class includes methods for solving systems of nonlinear equations of the form
F(x)=0,(5.1)
where F:<n→ <n. Newton-like methods provide the core of the package, including both line
search and trust region techniques. A suite of nonlinear Krylov methods and methods based upon
problem decomposition are also included. The solvers are discussed further in Section 5.2. Following
the PETSc design philosophy, the interfaces to the various solvers are all virtually identical. In
addition, the SNES software is completely flexible, so that the user can at runtime change any facet
of the solution process.
PETSc’s default method for solving the nonlinear equation is Newton’s method. The general
form of the n-dimensional Newton’s method for solving (5.1) is
xk+1 =xk−J(xk)−1F(xk), k = 0,1,..., (5.2)
where x0is an initial approximation to the solution and J(xk) = F0(xk), the Jacobian, is nonsin-
gular at each iteration. In practice, the Newton iteration (5.2) is implemented by the following two
steps:
1.(Approximately) solve J(xk)∆xk=−F(xk).(5.3)
2.Update xk+1 =xk+ ∆xk.(5.4)
Other defect-correction algorithms can be implemented by using different choices for J(xk).
5.1 Basic SNES Usage
In the simplest usage of the nonlinear solvers, the user must merely provide a C, C++, or Fortran
routine to evaluate the nonlinear function of Equation (5.1). The corresponding Jacobian matrix
can be approximated with finite differences. For codes that are typically more efficient and accurate,
the user can provide a routine to compute the Jacobian; details regarding these application-provided
105
![](asset-6a.png)
5.1. BASIC SNES USAGE PETSc 3.8 March 24, 2018
routines are discussed below. To provide an overview of the use of the nonlinear solvers, we first
introduce a complete and simple example in Figure 12, corresponding to $PETSC_DIR/src/snes/
examples/tutorials/ex1.c.
static char help[] = "Newton’s method for a two-variable system, sequential.\n\n"
;
/*T
Concepts: SNES^basic example
T*/
/*
Include "petscsnes.h" so that we can use SNES solvers. Note that this
file automatically includes:
petscsys.h - base PETSc routines petscvec.h - vectors
petscmat.h - matrices
petscis.h - index sets petscksp.h - Krylov subspace methods
petscviewer.h - viewers petscpc.h - preconditioners
petscksp.h - linear solvers
*/
/*F
This examples solves either
\begin{equation}
F\genfrac{(}{)}{0pt}{}{x_0}{x_1} = \genfrac{(}{)}{0pt}{}{x^2_0 + x_0 x_1 - 3}{
x_0 x_1 + x^2_1 - 6}
\end{equation}
or if the {\tt -hard} options is given
\begin{equation}
F\genfrac{(}{)}{0pt}{}{x_0}{x_1} = \genfrac{(}{)}{0pt}{}{\sin(3 x_0) + x_0}{x_1
}
\end{equation}
F*/
#include <petscsnes.h>
/*
User-defined routines
*/
extern PetscErrorCode FormJacobian1(SNES,Vec,Mat,Mat,void*);
extern PetscErrorCode FormFunction1(SNES,Vec,Vec,void*);
extern PetscErrorCode FormJacobian2(SNES,Vec,Mat,Mat,void*);
extern PetscErrorCode FormFunction2(SNES,Vec,Vec,void*);
int main(int argc,char **argv)
{
SNES snes; /* nonlinear solver context */
KSP ksp; /* linear solver context */
PC pc; /* preconditioner context */
Vec x,r; /* solution, residual vectors */
106
![](asset-6b.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
Mat J; /* Jacobian matrix */
PetscErrorCode ierr;
PetscInt its;
PetscMPIInt size;
PetscScalar pfive = .5,*xx;
PetscBool flg;
ierr = PetscInitialize(&argc,&argv,(char*)0,help);if (ierr) return ierr;
ierr = MPI Comm size(PETSC COMM WORLD,&size);CHKERRQ(ierr);
if (size > 1) SETERRQ(PETSC COMM WORLD,PETSC_ERR_SUP,"Example is only for
sequential runs");
/*----------------------------------
Create nonlinear solver context
----------------------------------*/
ierr = SNESCreate(PETSC COMM WORLD,&snes);CHKERRQ(ierr);
/*-----------------------------------
Create matrix and vector data structures; set corresponding routines
-----------------------------------*/
/*
Create vectors for solution and nonlinear function
*/
ierr = VecCreate(PETSC COMM WORLD,&x);CHKERRQ(ierr);
ierr = VecSetSizes(x,PETSC DECIDE,2);CHKERRQ(ierr);
ierr = VecSetFromOptions(x);CHKERRQ(ierr);
ierr = VecDuplicate(x,&r);CHKERRQ(ierr);
/*
Create Jacobian matrix data structure
*/
ierr = MatCreate(PETSC COMM WORLD,&J);CHKERRQ(ierr);
ierr = MatSetSizes(J,PETSC DECIDE,PETSC DECIDE,2,2);CHKERRQ(ierr);
ierr = MatSetFromOptions(J);CHKERRQ(ierr);
ierr = MatSetUp(J);CHKERRQ(ierr);
ierr = PetscOptionsHasName(NULL,NULL,"-hard",&flg);CHKERRQ(ierr);
if (!flg) {
/*
Set function evaluation routine and vector.
*/
ierr = SNESSetFunction(snes,r,FormFunction1,NULL);CHKERRQ(ierr);
/*
Set Jacobian matrix data structure and Jacobian evaluation routine
*/
ierr = SNESSetJacobian(snes,J,J,FormJacobian1,NULL);CHKERRQ(ierr);
} else {
ierr = SNESSetFunction(snes,r,FormFunction2,NULL);CHKERRQ(ierr);
107
![](asset-6c.png)
5.1. BASIC SNES USAGE PETSc 3.8 March 24, 2018
ierr = SNESSetJacobian(snes,J,J,FormJacobian2,NULL);CHKERRQ(ierr);
}
/*----------------------------------
Customize nonlinear solver; set runtime options
-----------------------------------*/
/*
Set linear solver defaults for this problem. By extracting the
KSP and PC contexts from the SNES context, we can then
directly call any KSP and PC routines to set various options.
*/
ierr = SNESGetKSP(snes,&ksp);CHKERRQ(ierr);
ierr = KSPGetPC(ksp,&pc);CHKERRQ(ierr);
ierr = PCSetType(pc,PCNONE);CHKERRQ(ierr);
ierr = KSPSetTolerances(ksp,1.e-4,PETSC DEFAULT,PETSC DEFAULT,20);CHKERRQ(ierr);
/*
Set SNES/KSP/KSP/PC runtime options, e.g.,
-snes_view -snes_monitor -ksp_type <ksp> -pc_type <pc>
These options will override those specified above as long as
SNESSetFromOptions() is called _after_ any other customization
routines.
*/
ierr = SNESSetFromOptions(snes);CHKERRQ(ierr);
/*----------------------------------
Evaluate initial guess; then solve nonlinear system
-----------------------------------*/
if (!flg) {
ierr = VecSet(x,pfive);CHKERRQ(ierr);
} else {
ierr = VecGetArray(x,&xx);CHKERRQ(ierr);
xx[0] = 2.0; xx[1] = 3.0;
ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr);
}
/*
Note: The user should initialize the vector, x, with the initial guess
for the nonlinear solver prior to calling SNESSolve(). In particular,
to employ an initial guess of zero, the user should explicitly set
this vector to zero by calling VecSet().
*/
ierr = SNESSolve(snes,NULL,x);CHKERRQ(ierr);
ierr = SNESGetIterationNumber(snes,&its);CHKERRQ(ierr);
if (flg) {
Vec f;
ierr = VecView(x,PETSC VIEWER STDOUT WORLD);CHKERRQ(ierr);
ierr = SNESGetFunction(snes,&f,0,0);CHKERRQ(ierr);
ierr = VecView(r,PETSC VIEWER STDOUT WORLD);CHKERRQ(ierr);
108
![](asset-6d.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
}
ierr = PetscPrintf(PETSC COMM WORLD,"Number of SNES iterations = %D\n",its);
CHKERRQ(ierr);
/*----------------------------------
Free work space. All PETSc objects should be destroyed when they
are no longer needed.
-----------------------------------*/
ierr = VecDestroy(&x);CHKERRQ(ierr); ierr = VecDestroy(&r);CHKERRQ(ierr);
ierr = MatDestroy(&J);CHKERRQ(ierr); ierr = SNESDestroy(&snes);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
/* ------------------------------------------------------------------- */
/*
FormFunction1 - Evaluates nonlinear function, F(x).
Input Parameters:
. snes - the SNES context
. x - input vector
. ctx - optional user-defined context
Output Parameter:
. f - function vector
*/
PetscErrorCode FormFunction1(SNES snes,Vec x,Vec f,void *ctx)
{
PetscErrorCode ierr;
const PetscScalar *xx;
PetscScalar *ff;
/*
Get pointers to vector data.
- For default PETSc vectors, VecGetArray() returns a pointer to
the data array. Otherwise, the routine is implementation dependent.
- You MUST call VecRestoreArray() when you no longer need access to
the array.
*/
ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr);
ierr = VecGetArray(f,&ff);CHKERRQ(ierr);
/* Compute function */
ff[0] = xx[0]*xx[0] + xx[0]*xx[1] - 3.0;
ff[1] = xx[0]*xx[1] + xx[1]*xx[1] - 6.0;
/* Restore vectors */
ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr);
109
![](asset-6e.png)
5.1. BASIC SNES USAGE PETSc 3.8 March 24, 2018
ierr = VecRestoreArray(f,&ff);CHKERRQ(ierr);
return 0;
}
/* ------------------------------------------------------------------- */
/*
FormJacobian1 - Evaluates Jacobian matrix.
Input Parameters:
. snes - the SNES context
. x - input vector
. dummy - optional user-defined context (not used here)
Output Parameters:
. jac - Jacobian matrix
. B - optionally different preconditioning matrix
. flag - flag indicating matrix structure
*/
PetscErrorCode FormJacobian1(SNES snes,Vec x,Mat jac,Mat B,void *dummy)
{
const PetscScalar *xx;
PetscScalar A[4];
PetscErrorCode ierr;
PetscInt idx[2] = {0,1};
/*
Get pointer to vector data
*/
ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr);
/*
Compute Jacobian entries and insert into matrix.
- Since this is such a small problem, we set all entries for
the matrix at once.
*/
A[0] = 2.0*xx[0] + xx[1]; A[1] = xx[0];
A[2] = xx[1]; A[3] = xx[0] + 2.0*xx[1];
ierr = MatSetValues(B,2,idx,2,idx,A,INSERT VALUES);CHKERRQ(ierr);
/*
Restore vector
*/
ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr);
/*
Assemble matrix
*/
ierr = MatAssemblyBegin(B,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(B,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
if (jac != B) {
110
![](asset-6f.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
ierr = MatAssemblyBegin(jac,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(jac,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
}
return 0;
}
/* ------------------------------------------------------------------- */
PetscErrorCode FormFunction2(SNES snes,Vec x,Vec f,void *dummy)
{
PetscErrorCode ierr;
const PetscScalar *xx;
PetscScalar *ff;
/*
Get pointers to vector data.
- For default PETSc vectors, VecGetArray() returns a pointer to
the data array. Otherwise, the routine is implementation dependent.
- You MUST call VecRestoreArray() when you no longer need access to
the array.
*/
ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr);
ierr = VecGetArray(f,&ff);CHKERRQ(ierr);
/*
Compute function
*/
ff[0] = PetscSinScalar(3.0*xx[0]) + xx[0];
ff[1] = xx[1];
/*
Restore vectors
*/
ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr);
ierr = VecRestoreArray(f,&ff);CHKERRQ(ierr);
return 0;
}
/* ------------------------------------------------------------------- */
PetscErrorCode FormJacobian2(SNES snes,Vec x,Mat jac,Mat B,void *dummy)
{
const PetscScalar *xx;
PetscScalar A[4];
PetscErrorCode ierr;
PetscInt idx[2] = {0,1};
/*
Get pointer to vector data
*/
ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr);
111
![](asset-70.png)
5.1. BASIC SNES USAGE PETSc 3.8 March 24, 2018
/*
Compute Jacobian entries and insert into matrix.
- Since this is such a small problem, we set all entries for
the matrix at once.
*/
A[0] = 3.0*PetscCosScalar(3.0*xx[0]) + 1.0; A[1] = 0.0;
A[2] = 0.0; A[3] = 1.0;
ierr = MatSetValues(B,2,idx,2,idx,A,INSERT VALUES);CHKERRQ(ierr);
/*
Restore vector
*/
ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr);
/*
Assemble matrix
*/
ierr = MatAssemblyBegin(B,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(B,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
if (jac != B) {
ierr = MatAssemblyBegin(jac,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(jac,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
}
return 0;
}
Figure 12: Example of Uniprocess SNES Code
To create a SNES solver, one must first call SNESCreate() as follows:
SNESCreate(MPI Comm comm,SNES *snes);
The user must then set routines for evaluating the function of equation (5.1) and its associated
Jacobian matrix, as discussed in the following sections.
To choose a nonlinear solution method, the user can either call
SNESSetType(SNES snes,SNESType method);
or use the option -snes_type <method>, where details regarding the available methods are pre-
sented in Section 5.2. The application code can take complete control of the linear and nonlinear
techniques used in the Newton-like method by calling
SNESSetFromOptions(snes);
This routine provides an interface to the PETSc options database, so that at runtime the user can
select a particular nonlinear solver, set various parameters and customized routines (e.g., specialized
line search variants), prescribe the convergence tolerance, and set monitoring routines. With this
routine the user can also control all linear solver options in the KSP, and PC modules, as discussed
in Chapter 4.
After having set these routines and options, the user solves the problem by calling
SNESSolve(SNES snes,Vec b,Vec x);
112
![](asset-71.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
where xindicates the solution vector. The user should initialize this vector to the initial guess for
the nonlinear solver prior to calling SNESSolve(). In particular, to employ an initial guess of zero,
the user should explicitly set this vector to zero by calling VecSet(). Finally, after solving the
nonlinear system (or several systems), the user should destroy the SNES context with
SNESDestroy(SNES *snes);
5.1.1 Nonlinear Function Evaluation
When solving a system of nonlinear equations, the user must provide a vector, f, for storing the
function of Equation (5.1), as well as a routine that evaluates this function at the vector x. This
information should be set with the command
SNESSetFunction(SNES snes,Vec f,PetscErrorCode (*FormFunction)(SNES snes,Vec
x,Vec f,void *ctx),void *ctx);
The argument ctx is an optional user-defined context, which can store any private, application-
specific data required by the function evaluation routine; NULL should be used if such information
is not needed. In C and C++, a user-defined context is merely a structure in which various objects
can be stashed; in Fortran a user context can be an integer array that contains both parame-
ters and pointers to PETSc objects. ${PETSC_DIR}/src/snes/examples/tutorials/ex5.c and
${PETSC_DIR}/src/snes/examples/tutorials/ex5f.F90 give examples of user-defined applica-
tion contexts in C and Fortran, respectively.
5.1.2 Jacobian Evaluation
The user must also specify a routine to form some approximation of the Jacobian matrix, A, at the
current iterate, x, as is typically done with
SNESSetJacobian(SNES snes,Mat Amat,Mat Pmat,PetscErrorCode (*FormJacobian)(SNES
snes, Vec x,Mat A,Mat B,void *ctx),void *ctx);
The arguments of the routine FormJacobian() are the current iterate, x; the (approximate)
Jacobian matrix, Amat; the matrix from which the preconditioner is constructed, Pmat (which is
usually the same as Amat); a flag indicating information about the preconditioner matrix structure;
and an optional user-defined Jacobian context, ctx, for application-specific data. The options for
flag are identical to those for the flag of KSPSetOperators(), discussed in Section 4.1. Note that
the SNES solvers are all data-structure neutral, so the full range of PETSc matrix formats (including
“matrix-free” methods) can be used. Chapter 3discusses information regarding available matrix
formats and options, while Section 5.5 focuses on matrix-free methods in SNES. We briefly touch
on a few details of matrix usage that are particularly important for efficient use of the nonlinear
solvers.
A common usage paradigm is to assemble the problem Jacobian in the preconditioner storage
B, rather than A. In the case where they are identical, as in many simulations, this makes no
difference. However, it allows us to check the analytic Jacobian we construct in FormJacobian()
by passing the -snes_mf_operator flag. This causes PETSc to approximate the Jacobian using
finite differencing of the function evaluation (discussed in section 5.6), and the analytic Jacobian
becomes merely the preconditioner. Even if the analytic Jacobian is incorrect, it is likely that the
finite difference approximation will converge, and thus this is an excellent method to verify the
analytic Jacobian. Moreover, if the analytic Jacobian is incomplete (some terms are missing or
113
![](asset-72.png)
5.2. THE NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
approximate), -snes_mf_operator may be used to obtain the exact solution, where the Jacobian
approximation has been transferred to the preconditioner.
One such approximate Jacobian comes from “Picard linearization” which writes the nonlinear
system as
F(x) = A(x)x−b= 0
where A(x) usually contains the lower-derivative parts of the equation. For example, the nonlinear
diffusion problem
−∇ · (κ(u)∇u)=0
would be linearized as
A(u)v' −∇ · (κ(u)∇v).
Usually this linearization is simpler to implement than Newton and the linear problems are some-
what easier to solve. In addition to using -snes_mf_operator with this approximation to the
Jacobian, the Picard iterative procedure can be performed by defining J(x) to be A(x). Some-
times this iteration exhibits better global convergence than Newton linearization.
During successive calls to FormJacobian(), the user can either insert new matrix contexts
or reuse old ones, depending on the application requirements. For many sparse matrix formats,
reusing the old space (and merely changing the matrix elements) is more efficient; however, if the
matrix structure completely changes, creating an entirely new matrix context may be preferable.
Upon subsequent calls to the FormJacobian() routine, the user may wish to reinitialize the matrix
entries to zero by calling MatZeroEntries(). See Section 3.4 for details on the reuse of the matrix
context.
The directory ${PETSC_DIR}/src/snes/examples/tutorials provides a variety of examples.
5.2 The Nonlinear Solvers
As summarized in Table 6,SNES includes several Newton-like nonlinear solvers based on line search
techniques and trust region methods. Also provided are several nonlinear Krylov methods, as well
as nonlinear methods involving decompositions of the problem.
Each solver may have associated with it a set of options, which can be set with routines and
options database commands provided for this purpose. A complete list can be found by consulting
the manual pages or by running a program with the -help option; we discuss just a few in the
sections below.
5.2.1 Line Search Newton
The method SNESNEWTONLS (-snes_type newtonls) provides a line search Newton method for
solving systems of nonlinear equations. By default, this technique employs cubic backtracking [9].
Alternative line search techniques are listed in Table 7.
Every SNES has a line search context of type SNESLineSearch that may be retrieved using
SNESGetLineSearch(SNES snes,SNESLineSearch *ls);.
There are several default options for the line searches. The order of polynomial approximation
may be set with -snes_linesearch_order or
SNESLineSearchSetOrder(SNESLineSearch ls, PetscInt order);
for instance, 2 for quadratic or 3 for cubic. Sometimes, it may not be necessary to monitor the
progress of the nonlinear iteration. In this case, -snes_linesearch_norms or
114
![](asset-73.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
Method SNESType Options Name Default Line Search
Line search Newton SNESNEWTONLS newtonls SNESLINESEARCHBT
Trust region Newton SNESNEWTONTR newtontr –
Test Jacobian SNESTEST test –
Nonlinear Richardson SNESNRICHARDSON nrichardson SNESLINESEARCHL2
Nonlinear CG SNESNCG ncg SNESLINESEARCHCP
Nonlinear GMRES SNESNGMRES ngmres SNESLINESEARCHL2
Quasi-Newton SNESQN qn see Table 8
FAS SNESFAS fas –
Nonlinear ASM SNESNASM nasm –
ASPIN SNESASPIN aspin SNESLINESEARCHBT
Nonlinear Gauss-Seidel SNESNGS ngs –
Anderson Mixing SNESANDERSON anderson –
Newton with constraints (1) SNESVINEWTONRSLS vinewtonrsls SNESLINESEARCHBT
Newton with constraints (2) SNESVINEWTONSSLS vinewtonssls SNESLINESEARCHBT
Multi-stage Smoothers SNESMS ms –
Composite SNESCOMPOSITE composite –
KSP only (linear solver) SNESKSPONLY ksponly –
Python Shell SNESPYTHON python –
Shell SNESSHELL shell –
Table 6: PETSc Nonlinear Solvers
Line Search SNESLineSearchType Options Name
Backtracking SNESLINESEARCHBT bt
(damped) step SNESLINESEARCHBASIC basic
L2-norm Minimization SNESLINESEARCHL2 l2
Critical point SNESLINESEARCHCP cp
Shell SNESLINESEARCHSHELL shell
Table 7: PETSc Line Search Methods
115
![](asset-74.png)
5.2. THE NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
SNESLineSearchSetComputeNorms(SNESLineSearch ls,PetscBool norms);
may be used to turn off function, step, and solution norm computation at the end of the linesearch.
The default line search for the line search Newton method, SNESLINESEARCHBT involves several
parameters, which are set to defaults that are reasonable for many applications. The user can
override the defaults by using the options
-snes_linesearch_alpha <alpha>,-snes_linesearch_maxstep <max>, and
-snes_linesearch_minlambda <tol>.
Besides the backtracking linesearch, there are SNESLINESEARCHL2, which uses a polynomial
secant minimization of ||F(x)||2, and SNESLINESEARCHCP, which minimizes F(x)·Ywhere Yis
the search direction. These are both potentially iterative line searches, which may be used to find
a better-fitted steplength in the case where a single secant search is not sufficient. The number
of iterations may be set with -snes_linesearch_max_it. In addition, the convergence criteria
of the iterative line searches may be set using function tolerances -snes_linesearch_rtol and
-snes_linesearch_atol, and steplength tolerance snes_linesearch_ltol.
Custom line search types may either be defined using SNESLineSearchShell, or by creating a
custom user line search type in the model of the preexisting ones and register it using
SNESLineSearchRegister(const char sname[],PetscErrorCode
(*function)(SNESLineSearch));.
5.2.2 Trust Region Methods
The trust region method in SNES for solving systems of nonlinear equations, SNESNEWTONTR
(-snes_type newtontr), is taken from the MINPACK project [24]. Several parameters can be
set to control the variation of the trust region size during the solution process. In particular, the
user can control the initial trust region radius, computed by
∆=∆0kF0k2,
by setting ∆0via the option -snes_tr_delta0 <delta0>.
5.2.3 Nonlinear Krylov Methods
A number of nonlinear Krylov methods are provided, including Nonlinear Richardson, conjugate
gradient, GMRES, and Anderson Mixing. These methods are described individually below. They
are all instrumental to PETSc’s nonlinear preconditioning.
Nonlinear Richardson. The nonlinear Richardson iteration merely takes the form of a line
search-damped fixed-point iteration of the form
xk+1 =xk−λF(xk), k = 0,1,..., (5.5)
where the default linesearch is SNESLINESEARCHL2. This simple solver is mostly useful as a nonlinear
smoother, or to provide line search stabilization to an inner method.
Nonlinear Conjugate Gradients. Nonlinear CG is equivalent to linear CG, but with the
steplength determined by line search (SNESLINESEARCHCP by default). Five variants (Fletcher-
Reed, Hestenes-Steifel, Polak-Ribiere-Polyak, Dai-Yuan, and Conjugate Descent) are implemented
in PETSc and may be chosen using
SNESNCGSetType(SNES snes, SNESNCGType btype);
116
![](asset-75.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
QN Method SNESQNType Options Name Default Line Search
L-BFGS SNES_QN_LBFGS lbfgs SNESLINESEARCHCP
“Good” Broyden SNES_QN_BROYDEN broyden SNESLINESEARCHBASIC
“Bad” Broyden SNES_QN_BADBROYEN badbroyden SNESLINESEARCHL2
Table 8: PETSc quasi-Newton solvers
Anderson Mixing and Nonlinear GMRES Methods. Nonlinear GMRES and Anderson
Mixing methods combine the last miterates, plus a new fixed-point iteration iterate, into a residual-
minimizing new iterate.
5.2.4 Quasi-Newton Methods
Quasi-Newton methods store iterative rank-one updates to the Jacobian instead of computing it
directly. Three limited-memory quasi-Newton methods are provided, L-BFGS, which are described
in Table 8. These all are encapsulated under -snes_type qn and may be changed with snes_
qn_type. The default is L-BFGS, which provides symmetric updates to an approximate Jacobian.
This iteration is similar to the line search Newton methods.
One may also control the form of the initial Jacobian approximation with
SNESQNSetScaleType(SNES snes, SNESQNScaleType stype);
and the restart type with
SNESQNSetRestartType(SNES snes, SNESQNRestartType rtype);
5.2.5 The Full Approximation Scheme
The Full Approximation Scheme is a nonlinear multigrid correction. At each level, there is a
recursive cycle control SNES instance, and either one or two nonlinear solvers as smoothers (up and
down). Problems set up using the SNES DMDA interface are automatically coarsened. FAS differs
slightly from PCMG, in that the hierarchy is constructed recursively. However, much of the interface
is a one-to-one map. We describe the “get” operations here, and it can be assumed that each has a
corresponding “set” operation. For instance, the number of levels in the hierarchy may be retrieved
using
SNESFASGetLevels(SNES snes, PetscInt *levels);
There are four SNESFAS cycle types, SNES_FAS_MULTIPLICATIVE,SNES_FAS_ADDITIVE,
SNES_FAS_FULL, and SNES_FAS_KASKADE. The type may be set with
SNESFASSetType(SNES snes,SNESFASType fastype);.
and the cycle type, 1 for V, 2 for W, may be set with
SNESFASSetCycles(SNES snes, PetscInt cycles);.
Much like the interface to PCMG described in Section 4.4.9, there are interfaces to recover the various
levels’ cycles and smoothers. The level smoothers may be accessed with
SNESFASGetSmoother(SNES snes, PetscInt level, SNES *smooth);
SNESFASGetSmootherUp(SNES snes, PetscInt level, SNES *smooth);
SNESFASGetSmootherDown(SNES snes, PetscInt level, SNES *smooth);
117
![](asset-76.png)
5.3. GENERAL OPTIONS PETSc 3.8 March 24, 2018
and the level cycles with
SNESFASGetCycleSNES(SNES snes,PetscInt level,SNES *lsnes);.
Also akin to PCMG, the restriction and prolongation at a level may be acquired with
SNESFASGetInterpolation(SNES snes, PetscInt level, Mat *mat);
SNESFASGetRestriction(SNES snes, PetscInt level, Mat *mat);
In addition, FAS requires special restriction for solution-like variables, called injection. This may
be set with
SNESFASGetInjection(SNES snes, PetscInt level, Mat *mat);.
The coarse solve context may be acquired with
SNESFASGetCoarseSolve(SNES snes, SNES *smooth);
5.2.6 Nonlinear Additive Schwarz
Nonlinear Additive Schwarz methods (NASM) take a number of local nonlinear subproblems, solves
them independently in parallel, and combines those solutions into a new approximate solution.
SNESNASMSetSubdomains(SNES snes,PetscInt n,SNES subsnes[],VecScatter
iscatter[],VecScatter oscatter[],VecScatter gscatter[]);
allows for the user to create these local subdomains. Problems set up using the SNES DMDA interface
are automatically decomposed. To begin, the type of subdomain updates to the whole solution
are limited to two types borrowed from PCASM:PC_ASM_BASIC, in which the overlapping updates
added. PC_ASM_RESTRICT updates in a nonoverlapping fashion. This may be set with
SNESNASMSetType(SNES snes,PCASMType type);.
SNESASPIN is a helper SNES type that sets up a nonlinearly preconditioned Newton’s method using
NASM as the preconditioner.
5.3 General Options
This section discusses options and routines that apply to all SNES solvers and problem classes. In
particular, we focus on convergence tests, monitoring routines, and tools for checking derivative
computations.
5.3.1 Convergence Tests
Convergence of the nonlinear solvers can be detected in a variety of ways; the user can even specify a
customized test, as discussed below. Most of the nonlinear solvers use SNESConvergenceTestDefault(),
however, SNESNEWTONTR uses a method-specific additional convergence test as well. The conver-
gence tests involves several parameters, which are set by default to values that should be reasonable
for a wide range of problems. The user can customize the parameters to the problem at hand by
using some of the following routines and options.
One method of convergence testing is to declare convergence when the norm of the change in
the solution between successive iterations is less than some tolerance, stol. Convergence can also
be determined based on the norm of the function. Such a test can use either the absolute size of
118
![](asset-77.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
the norm, atol, or its relative decrease, rtol, from an initial guess. The following routine sets
these parameters, which are used in many of the default SNES convergence tests:
SNESSetTolerances(SNES snes,PetscReal atol,PetscReal rtol,PetscReal stol,
PetscInt its,PetscInt fcts);
This routine also sets the maximum numbers of allowable nonlinear iterations, its, and function
evaluations, fcts. The corresponding options database commands for setting these parameters
are -snes_atol <atol>,-snes_rtol <rtol>,-snes_stol <stol>,-snes_max_it <its>, and
-snes_max_funcs <fcts>. A related routine is SNESGetTolerances().
Convergence tests for trust regions methods often use an additional parameter that indicates
the minimum allowable trust region radius. The user can set this parameter with the option
-snes_trtol <trtol> or with the routine
SNESSetTrustRegionTolerance(SNES snes,PetscReal trtol);
Users can set their own customized convergence tests in SNES by using the command
SNESSetConvergenceTest(SNES snes,PetscErrorCode (*test)(SNES snes,PetscInt
it,PetscReal xnorm, PetscReal gnorm,PetscReal f,SNESConvergedReason reason,
void *cctx),void *cctx,PetscErrorCode (*destroy)(void *cctx));
The final argument of the convergence test routine, cctx, denotes an optional user-defined context
for private data. When solving systems of nonlinear equations, the arguments xnorm,gnorm, and f
are the current iterate norm, current step norm, and function norm, respectively.
SNESConvergedReason should be set positive for convergence and negative for divergence. See
include/petscsnes.h for a list of values for SNESConvergedReason.
5.3.2 Convergence Monitoring
By default the SNES solvers run silently without displaying information about the iterations. The
user can initiate monitoring with the command
SNESMonitorSet(SNES snes,PetscErrorCode (*mon)(SNES,PetscInt its,PetscReal
norm,void* mctx),void *mctx,PetscErrorCode (*monitordestroy)(void**));
The routine, mon, indicates a user-defined monitoring routine, where its and mctx respectively
denote the iteration number and an optional user-defined context for private data for the monitor
routine. The argument norm is the function norm.
The routine set by SNESMonitorSet() is called once after every successful step computa-
tion within the nonlinear solver. Hence, the user can employ this routine for any application-
specific computations that should be done after the solution update. The option -snes_monitor
activates the default SNES monitor routine, SNESMonitorDefault(), while -snes_monitor_lg_
residualnorm draws a simple line graph of the residual norm’s convergence.
One can cancel hardwired monitoring routines for SNES at runtime with -snes_monitor_
cancel.
As the Newton method converges so that the residual norm is small, say 10−10, many of the
final digits printed with the -snes_monitor option are meaningless. Worse, they are different
on different machines; due to different round-off rules used by, say, the IBM RS6000 and the Sun
SPARC. This makes testing between different machines difficult. The option -snes_monitor_short
causes PETSc to print fewer of the digits of the residual norm as it gets smaller; thus on most of
the machines it will always print the same numbers making cross-process testing easier.
The routines
119
![](asset-78.png)
5.4. INEXACT NEWTON-LIKE METHODS PETSc 3.8 March 24, 2018
SNESGetSolution(SNES snes,Vec *x);
SNESGetFunction(SNES snes,Vec *r,void *ctx,int(**func)(SNES,Vec,Vec,void*));
return the solution vector and function vector from a SNES context. These routines are useful, for
instance, if the convergence test requires some property of the solution or function other than those
passed with routine arguments.
5.3.3 Checking Accuracy of Derivatives
Since hand-coding routines for Jacobian matrix evaluation can be error prone, SNES provides easy-
to-use support for checking these matrices against finite difference versions. In the simplest form
of comparison, users can employ the option -snes_type test to compare the matrices at several
points. Although not exhaustive, this test will generally catch obvious problems. One can compare
the elements of the two matrices by using the option -snes_test_display , which causes the two
matrices to be printed to the screen.
Another means for verifying the correctness of a code for Jacobian computation is running the
problem with either the finite difference or matrix-free variant, -snes_fd or -snes_mf. see Section
5.6 or Section 5.5). If a problem converges well with these matrix approximations but not with
a user-provided routine, the problem probably lies with the hand-coded matrix. See the note in
Section 5.1.2 about assembling your Jabobian in the “preconditioner” slot B.
5.4 Inexact Newton-like Methods
Since exact solution of the linear Newton systems within (5.2) at each iteration can be costly,
modifications are often introduced that significantly reduce these expenses and yet retain the rapid
convergence of Newton’s method. Inexact or truncated Newton techniques approximately solve the
linear systems using an iterative scheme. In comparison with using direct methods for solving the
Newton systems, iterative methods have the virtue of requiring little space for matrix storage and
potentially saving significant computational work. Within the class of inexact Newton methods, of
particular interest are Newton-Krylov methods, where the subsidiary iterative technique for solving
the Newton system is chosen from the class of Krylov subspace projection methods. Note that at
runtime the user can set any of the linear solver options discussed in Chapter 4, such as -ksp_
type <ksp_method> and -pc_type <pc_method>, to set the Krylov subspace and preconditioner
methods.
Two levels of iterations occur for the inexact techniques, where during each global or outer New-
ton iteration a sequence of subsidiary inner iterations of a linear solver is performed. Appropriate
control of the accuracy to which the subsidiary iterative method solves the Newton system at each
global iteration is critical, since these inner iterations determine the asymptotic convergence rate
for inexact Newton techniques. While the Newton systems must be solved well enough to retain
fast local convergence of the Newton’s iterates, use of excessive inner iterations, particularly when
kxk−x∗kis large, is neither necessary nor economical. Thus, the number of required inner itera-
tions typically increases as the Newton process progresses, so that the truncated iterates approach
the true Newton iterates.
A sequence of nonnegative numbers {ηk}can be used to indicate the variable convergence
criterion. In this case, when solving a system of nonlinear equations, the update step of the Newton
process remains unchanged, and direct solution of the linear system is replaced by iteration on the
system until the residuals
r(i)
k=F0(xk)∆xk+F(xk)
120
![](asset-79.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
satisfy
kr(i)
kk
kF(xk)k≤ηk≤η < 1.
Here x0is an initial approximation of the solution, and k·kdenotes an arbitrary norm in <n.
By default a constant relative convergence tolerance is used for solving the subsidiary linear
systems within the Newton-like methods of SNES. When solving a system of nonlinear equations,
one can instead employ the techniques of Eisenstat and Walker [11] to compute ηkat each step of
the nonlinear solver by using the option -snes_ksp_ew_conv . In addition, by adding one’s own
KSP convergence test (see Section 4.3.2), one can easily create one’s own, problem-dependent, inner
convergence tests.
5.5 Matrix-Free Methods
The SNES class fully supports matrix-free methods. The matrices specified in the Jacobian eval-
uation routine need not be conventional matrices; instead, they can point to the data required
to implement a particular matrix-free method. The matrix-free variant is allowed only when the
linear systems are solved by an iterative method in combination with no preconditioning (PCNONE
or -pc_type none), a user-provided preconditioner matrix, or a user-provided preconditioner shell
(PCSHELL, discussed in Section 4.4); that is, obviously matrix-free methods cannot be used with a
direct solver, approximate factorization, or other preconditioner which requires access to explicit
matrix entries.
The user can create a matrix-free context for use within SNES with the routine
MatCreateSNESMF(SNES snes,Mat *mat);
This routine creates the data structures needed for the matrix-vector products that arise within
Krylov space iterative methods [4] by employing the matrix type MATSHELL, discussed in Section 3.3.
The default SNES matrix-free approximations can also be invoked with the command -snes_mf. Or,
one can retain the user-provided Jacobian preconditioner, but replace the user-provided Jacobian
matrix with the default matrix free variant with the option -snes_mf_operator.
See also
MatCreateMFFD(Vec x, Mat *mat);
for users who need a matrix-free matrix but are not using SNES.
The user can set one parameter to control the Jacobian-vector product approximation with the
command
MatMFFDSetFunctionError(Mat mat,PetscReal rerror);
The parameter rerror should be set to the square root of the relative error in the function evalua-
tions, erel; the default is the square root of machine epsilon (about 10−8in double precision), which
assumes that the functions are evaluated to full floating-point precision accuracy. This parameter
can also be set from the options database with -snes_mf_err <err>
In addition, SNES provides a way to register new routines to compute the differencing parameter
(h); see the manual page for MatMFFDSetType() and MatMFFDRegister(). We currently provide
two default routines accessible via -snes_mf_type <default or wp>. For the default approach
there is one “tuning” parameter, set with
MatMFFDDSSetUmin(Mat mat,PetscReal umin);
121
![](asset-7a.png)
5.5. MATRIX-FREE METHODS PETSc 3.8 March 24, 2018
This parameter, umin (or umin), is a bit involved; its default is 10−6. The Jacobian-vector product
is approximated via the formula
F0(u)a≈F(u+h∗a)−F(u)
h
where his computed via
h=erel ∗uTa/||a||2
2if|u0a|> umin ∗ ||a||1
=erel ∗umin ∗sign(uTa)∗ ||a||1/||a||2
2otherwise.
This approach is taken from Brown and Saad [4]. The parameter can also be set from the options
database with -snes_mf_umin <umin>
The second approach, taken from Walker and Pernice, [29], computes hvia
h=p1 + ||u||erel
||a||
This has no tunable parameters, but note that inside the nonlinear solve for the entire linear
iterative process udoes not change hence p1 + ||u|| need be computed only once. This information
may be set with the options
MatMFFDWPSetComputeNormU(Mat mat,PetscBool );
or -mat_mffd_compute_normu <true or false>. This information is used to eliminate the re-
dundant computation of these parameters, therefor reducing the number of collective operations
and improving the efficiency of the application code.
It is also possible to monitor the differencing parameters h that are computed via the routines
MatMFFDSetHHistory(Mat,PetscScalar *,int);
MatMFFDResetHHistory(Mat,PetscScalar *,int);
MatMFFDGetH(Mat,PetscScalar *);
We include an example in Figure 13 that explicitly uses a matrix-free approach. Note that
by using the option -snes_mf one can easily convert any SNES code to use a matrix-free Newton-
Krylov method without a preconditioner. As shown in this example, SNESSetFromOptions() must
be called after SNESSetJacobian() to enable runtime switching between the user-specified Jacobian
and the default SNES matrix-free form.
Table 9summarizes the various matrix situations that SNES supports. In particular, different
linear system matrices and preconditioning matrices are allowed, as well as both matrix-free and
application-provided preconditioners. If ${PETSC_DIR}/src/snes/examples/tutorials/ex3.c,
listed in Figure 13, is run with the options -snes_mf and -user_precond then it uses a matrix-free
application of the matrix-vector multiple and a user provided matrix free Jacobian.
static char help[] = "Newton methods to solve u’’ + u^{2} = f in parallel.\n\
This example employs a user-defined monitoring routine and optionally a user-
defined\n\
routine to check candidate iterates produced by line search routines.\n\
The command line options include:\n\
122
![](asset-7b.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
Matrix Use Conventional Matrix Formats Matrix-Free Versions
Jacobian Create matrix with MatCreate().∗Create matrix with MatCreateShell().
Matrix Assemble matrix with user-defined Use MatShellSetOperation() to set
routine. †various matrix actions, or use
MatCreateMFFD() or MatCreateSNESMF().
Preconditioning Create matrix with MatCreate().∗Use SNESGetKSP() and KSPGetPC()
Matrix Assemble matrix with user-defined to access the PC, then use
routine. †PCSetType(pc,PCSHELL)
followed by PCShellSetApply().
∗Use either the generic MatCreate() or a format-specific variant such as MatCreateAIJ().
†Set user-defined matrix formation routine with SNESSetJacobian().
Table 9: Jacobian Options
-pre_check_iterates : activate checking of iterates\n\
-post_check_iterates : activate checking of iterates\n\
-check_tol <tol>: set tolerance for iterate checking\n\
-user_precond : activate a (trivial) user-defined preconditioner\n\n";
/*T
Concepts: SNES^basic parallel example
Concepts: SNES^setting a user-defined monitoring routine
Processors: n
T*/
/*
Include "petscdraw.h" so that we can use distributed arrays (DMDAs).
Include "petscdraw.h" so that we can use PETSc drawing routines.
Include "petscsnes.h" so that we can use SNES solvers. Note that this
file automatically includes:
petscsys.h - base PETSc routines
petscvec.h - vectors
petscmat.h - matrices
petscis.h - index sets
petscksp.h - Krylov subspace methods
petscviewer.h - viewers
petscpc.h - preconditioners
petscksp.h - linear solvers
*/
#include <petscdm.h>
#include <petscdmda.h>
#include <petscsnes.h>
/*
User-defined routines.
*/
123
![](asset-7c.png)
5.5. MATRIX-FREE METHODS PETSc 3.8 March 24, 2018
PetscErrorCode FormJacobian(SNES,Vec,Mat,Mat,void*);
PetscErrorCode FormFunction(SNES,Vec,Vec,void*);
PetscErrorCode FormInitialGuess(Vec);
PetscErrorCode Monitor(SNES,PetscInt,PetscReal,void*);
PetscErrorCode PreCheck(SNESLineSearch,Vec,Vec,PetscBool*,void*);
PetscErrorCode PostCheck(SNESLineSearch,Vec,Vec,Vec,PetscBool*,PetscBool*,void*);
PetscErrorCode PostSetSubKSP(SNESLineSearch,Vec,Vec,Vec,PetscBool*,PetscBool*,
void*);
PetscErrorCode MatrixFreePreconditioner(PC,Vec,Vec);
/*
User-defined application context
*/
typedef struct {
DM da; /* distributed array */
Vec F; /* right-hand-side of PDE */
PetscMPIInt rank; /* rank of processor */
PetscMPIInt size; /* size of communicator */
PetscReal h; /* mesh spacing */
} ApplicationCtx;
/*
User-defined context for monitoring
*/
typedef struct {
PetscViewer viewer;
} MonitorCtx;
/*
User-defined context for checking candidate iterates that are
determined by line search methods
*/
typedef struct {
Vec last_step; /* previous iterate */
PetscReal tolerance; /* tolerance for changes between successive iterates
*/
ApplicationCtx *user;
} StepCheckCtx;
typedef struct {
PetscInt its0; /* num of prevous outer KSP iterations */
} SetSubKSPCtx;
int main(int argc,char **argv)
{
SNES snes; /* SNES context */
SNESLineSearch linesearch; /* SNESLineSearch context */
Mat J; /* Jacobian matrix */
ApplicationCtx ctx; /* user-defined context */
124
![](asset-7d.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
Vec x,r,U,F; /* vectors */
MonitorCtx monP; /* monitoring context */
StepCheckCtx checkP; /* step-checking context */
SetSubKSPCtx checkP1;
PetscBool pre_check,post_check,post_setsubksp; /* flag indicating whether
we’re checking candidate iterates */
PetscScalar xp,*FF,*UU,none = -1.0;
PetscErrorCode ierr;
PetscInt its,N = 5,i,maxit,maxf,xs,xm;
PetscReal abstol,rtol,stol,norm;
PetscBool flg;
ierr = PetscInitialize(&argc,&argv,(char*)0,help);if (ierr) return ierr;
ierr = MPI Comm rank(PETSC COMM WORLD,&ctx.rank);CHKERRQ(ierr);
ierr = MPI Comm size(PETSC COMM WORLD,&ctx.size);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-n",&N,NULL);CHKERRQ(ierr);
ctx.h = 1.0/(N-1);
/*----------------------------------
Create nonlinear solver context
----------------------------------*/
ierr = SNESCreate(PETSC COMM WORLD,&snes);CHKERRQ(ierr);
/*-----------------------------------
Create vector data structures; set function evaluation routine
-----------------------------------*/
/*
Create distributed array (DMDA) to manage parallel grid and vectors
*/
ierr = DMDACreate1d(PETSC COMM WORLD,DM BOUNDARY NONE,N,1,1,NULL,&ctx.da);
CHKERRQ(ierr);
ierr = DMSetFromOptions(ctx.da);CHKERRQ(ierr);
ierr = DMSetUp(ctx.da);CHKERRQ(ierr);
/*
Extract global and local vectors from DMDA; then duplicate for remaining
vectors that are the same types
*/
ierr = DMCreateGlobalVector(ctx.da,&x);CHKERRQ(ierr);
ierr = VecDuplicate(x,&r);CHKERRQ(ierr);
ierr = VecDuplicate(x,&F);CHKERRQ(ierr); ctx.F = F;
ierr = VecDuplicate(x,&U);CHKERRQ(ierr);
/*
Set function evaluation routine and vector. Whenever the nonlinear
solver needs to compute the nonlinear function, it will call this
routine.
125
![](asset-7e.png)
5.5. MATRIX-FREE METHODS PETSc 3.8 March 24, 2018
- Note that the final routine argument is the user-defined
context that provides application-specific data for the
function evaluation routine.
*/
ierr = SNESSetFunction(snes,r,FormFunction,&ctx);CHKERRQ(ierr);
/*-----------------------------------
Create matrix data structure; set Jacobian evaluation routine
-----------------------------------*/
ierr = MatCreate(PETSC COMM WORLD,&J);CHKERRQ(ierr);
ierr = MatSetSizes(J,PETSC DECIDE,PETSC DECIDE,N,N);CHKERRQ(ierr);
ierr = MatSetFromOptions(J);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(J,3,NULL);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(J,3,NULL,3,NULL);CHKERRQ(ierr);
/*
Set Jacobian matrix data structure and default Jacobian evaluation
routine. Whenever the nonlinear solver needs to compute the
Jacobian matrix, it will call this routine.
- Note that the final routine argument is the user-defined
context that provides application-specific data for the
Jacobian evaluation routine.
*/
ierr = SNESSetJacobian(snes,J,J,FormJacobian,&ctx);CHKERRQ(ierr);
/*
Optionally allow user-provided preconditioner
*/
ierr = PetscOptionsHasName(NULL,NULL,"-user_precond",&flg);CHKERRQ(ierr);
if (flg) {
KSP ksp;
PC pc;
ierr = SNESGetKSP(snes,&ksp);CHKERRQ(ierr);
ierr = KSPGetPC(ksp,&pc);CHKERRQ(ierr);
ierr = PCSetType(pc,PCSHELL);CHKERRQ(ierr);
ierr = PCShellSetApply(pc,MatrixFreePreconditioner);CHKERRQ(ierr);
}
/*----------------------------------
Customize nonlinear solver; set runtime options
-----------------------------------*/
/*
Set an optional user-defined monitoring routine
*/
ierr = PetscViewerDrawOpen(PETSC COMM WORLD,0,0,0,0,400,400,&monP.viewer);
CHKERRQ(ierr);
ierr = SNESMonitorSet(snes,Monitor,&monP,0);CHKERRQ(ierr);
126
![](asset-7f.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
/*
Set names for some vectors to facilitate monitoring (optional)
*/
ierr = PetscObjectSetName((PetscObject)x,"Approximate Solution");CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject)U,"Exact Solution");CHKERRQ(ierr);
/*
Set SNES/KSP/KSP/PC runtime options, e.g.,
-snes_view -snes_monitor -ksp_type <ksp> -pc_type <pc>
*/
ierr = SNESSetFromOptions(snes);CHKERRQ(ierr);
/*
Set an optional user-defined routine to check the validity of candidate
iterates that are determined by line search methods
*/
ierr = SNESGetLineSearch(snes, &linesearch);CHKERRQ(ierr);
ierr = PetscOptionsHasName(NULL,NULL,"-post_check_iterates",&post_check);
CHKERRQ(ierr);
if (post_check) {
ierr = PetscPrintf(PETSC COMM WORLD,"Activating post step checking routine\n")
;CHKERRQ(ierr);
ierr = SNESLineSearchSetPostCheck(linesearch,PostCheck,&checkP);CHKERRQ(ierr)
;
ierr = VecDuplicate(x,&(checkP.last_step));CHKERRQ(ierr);
checkP.tolerance = 1.0;
checkP.user = &ctx;
ierr = PetscOptionsGetReal(NULL,NULL,"-check_tol",&checkP.tolerance,NULL);
CHKERRQ(ierr);
}
ierr = PetscOptionsHasName(NULL,NULL,"-post_setsubksp",&post_setsubksp);
CHKERRQ(ierr);
if (post_setsubksp) {
ierr = PetscPrintf(PETSC COMM WORLD,"Activating post setsubksp\n");CHKERRQ(
ierr);
ierr = SNESLineSearchSetPostCheck(linesearch,PostSetSubKSP,&checkP1);CHKERRQ(
ierr);
}
ierr = PetscOptionsHasName(NULL,NULL,"-pre_check_iterates",&pre_check);CHKERRQ(
ierr);
if (pre_check) {
ierr = PetscPrintf(PETSC COMM WORLD,"Activating pre step checking routine\n");
CHKERRQ(ierr);
127
![](asset-80.png)
5.5. MATRIX-FREE METHODS PETSc 3.8 March 24, 2018
ierr = SNESLineSearchSetPreCheck(linesearch,PreCheck,&checkP);CHKERRQ(ierr);
}
/*
Print parameters used for convergence testing (optional) ... just
to demonstrate this routine; this information is also printed with
the option -snes_view
*/
ierr = SNESGetTolerances(snes,&abstol,&rtol,&stol,&maxit,&maxf);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC COMM WORLD,"atol=%g, rtol=%g, stol=%g, maxit=%D, maxf=%
D\n",(double)abstol,(double)rtol,(double)stol,maxit,maxf);CHKERRQ(ierr);
/*----------------------------------
Initialize application:
Store right-hand-side of PDE and exact solution
-----------------------------------*/
/*
Get local grid boundaries (for 1-dimensional DMDA):
xs, xm - starting grid index, width of local grid (no ghost points)
*/
ierr = DMDAGetCorners(ctx.da,&xs,NULL,NULL,&xm,NULL,NULL);CHKERRQ(ierr);
/*
Get pointers to vector data
*/
ierr = DMDAVecGetArray(ctx.da,F,&FF);CHKERRQ(ierr);
ierr = DMDAVecGetArray(ctx.da,U,&UU);CHKERRQ(ierr);
/*
Compute local vector entries
*/
xp = ctx.h*xs;
for (i=xs; i<xs+xm; i++) {
FF[i] = 6.0*xp + PetscPowScalar(xp+1.e-12,6.0); /* +1.e-12 is to prevent 0^6
*/
UU[i] = xp*xp*xp;
xp += ctx.h;
}
/*
Restore vectors
*/
ierr = DMDAVecRestoreArray(ctx.da,F,&FF);CHKERRQ(ierr);
ierr = DMDAVecRestoreArray(ctx.da,U,&UU);CHKERRQ(ierr);
/*----------------------------------
Evaluate initial guess; then solve nonlinear system
-----------------------------------*/
128
![](asset-81.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
/*
Note: The user should initialize the vector, x, with the initial guess
for the nonlinear solver prior to calling SNESSolve(). In particular,
to employ an initial guess of zero, the user should explicitly set
this vector to zero by calling VecSet().
*/
ierr = FormInitialGuess(x);CHKERRQ(ierr);
ierr = SNESSolve(snes,NULL,x);CHKERRQ(ierr);
ierr = SNESGetIterationNumber(snes,&its);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC COMM WORLD,"Number of SNES iterations = %D\n",its);
CHKERRQ(ierr);
/*----------------------------------
Check solution and clean up
-----------------------------------*/
/*
Check the error
*/
ierr = VecAXPY(x,none,U);CHKERRQ(ierr);
ierr = VecNorm(x,NORM 2,&norm);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC COMM WORLD,"Norm of error %g Iterations %D\n",(double)
norm,its);CHKERRQ(ierr);
/*
Free work space. All PETSc objects should be destroyed when they
are no longer needed.
*/
ierr = PetscViewerDestroy(&monP.viewer);CHKERRQ(ierr);
if (post_check) {ierr = VecDestroy(&checkP.last_step);CHKERRQ(ierr);}
ierr = VecDestroy(&x);CHKERRQ(ierr);
ierr = VecDestroy(&r);CHKERRQ(ierr);
ierr = VecDestroy(&U);CHKERRQ(ierr);
ierr = VecDestroy(&F);CHKERRQ(ierr);
ierr = MatDestroy(&J);CHKERRQ(ierr);
ierr = SNESDestroy(&snes);CHKERRQ(ierr);
ierr = DMDestroy(&ctx.da);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
/* ------------------------------------------------------------------- */
/*
FormInitialGuess - Computes initial guess.
Input/Output Parameter:
. x - the solution vector
*/
129
![](asset-82.png)
5.5. MATRIX-FREE METHODS PETSc 3.8 March 24, 2018
PetscErrorCode FormInitialGuess(Vec x)
{
PetscErrorCode ierr;
PetscScalar pfive = .50;
PetscFunctionBeginUser;
ierr = VecSet(x,pfive);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* ------------------------------------------------------------------- */
/*
FormFunction - Evaluates nonlinear function, F(x).
Input Parameters:
. snes - the SNES context
. x - input vector
. ctx - optional user-defined context, as set by SNESSetFunction()
Output Parameter:
. f - function vector
Note:
The user-defined context can contain any application-specific
data needed for the function evaluation.
*/
PetscErrorCode FormFunction(SNES snes,Vec x,Vec f,void *ctx)
{
ApplicationCtx *user = (ApplicationCtx*) ctx;
DM da = user->da;
PetscScalar *xx,*ff,*FF,d;
PetscErrorCode ierr;
PetscInt i,M,xs,xm;
Vec xlocal;
PetscFunctionBeginUser;
ierr = DMGetLocalVector(da,&xlocal);CHKERRQ(ierr);
/*
Scatter ghost points to local vector, using the 2-step process
DMGlobalToLocalBegin(), DMGlobalToLocalEnd().
By placing code between these two statements, computations can
be done while messages are in transition.
*/
ierr = DMGlobalToLocalBegin(da,x,INSERT VALUES,xlocal);CHKERRQ(ierr);
ierr = DMGlobalToLocalEnd(da,x,INSERT VALUES,xlocal);CHKERRQ(ierr);
/*
Get pointers to vector data.
- The vector xlocal includes ghost point; the vectors x and f do
130
![](asset-83.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
NOT include ghost points.
- Using DMDAVecGetArray() allows accessing the values using global
ordering
*/
ierr = DMDAVecGetArray(da,xlocal,&xx);CHKERRQ(ierr);
ierr = DMDAVecGetArray(da,f,&ff);CHKERRQ(ierr);
ierr = DMDAVecGetArray(da,user->F,&FF);CHKERRQ(ierr);
/*
Get local grid boundaries (for 1-dimensional DMDA):
xs, xm - starting grid index, width of local grid (no ghost points)
*/
ierr = DMDAGetCorners(da,&xs,NULL,NULL,&xm,NULL,NULL);CHKERRQ(ierr);
ierr = DMDAGetInfo(da,NULL,&M,NULL,NULL,NULL,NULL,NULL,NULL,NULL,NULL,NULL,NULL
,NULL);CHKERRQ(ierr);
/*
Set function values for boundary points; define local interior grid point
range:
xsi - starting interior grid index
xei - ending interior grid index
*/
if (xs == 0) { /* left boundary */
ff[0] = xx[0];
xs++;xm--;
}
if (xs+xm == M) { /* right boundary */
ff[xs+xm-1] = xx[xs+xm-1] - 1.0;
xm--;
}
/*
Compute function over locally owned part of the grid (interior points only)
*/
d = 1.0/(user->h*user->h);
for (i=xs; i<xs+xm; i++) ff[i] = d*(xx[i-1] - 2.0*xx[i] + xx[i+1]) + xx[i]*xx[i
] - FF[i];
/*
Restore vectors
*/
ierr = DMDAVecRestoreArray(da,xlocal,&xx);CHKERRQ(ierr);
ierr = DMDAVecRestoreArray(da,f,&ff);CHKERRQ(ierr);
ierr = DMDAVecRestoreArray(da,user->F,&FF);CHKERRQ(ierr);
ierr = DMRestoreLocalVector(da,&xlocal);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* ------------------------------------------------------------------- */
131
![](asset-84.png)
5.5. MATRIX-FREE METHODS PETSc 3.8 March 24, 2018
/*
FormJacobian - Evaluates Jacobian matrix.
Input Parameters:
. snes - the SNES context
. x - input vector
. dummy - optional user-defined context (not used here)
Output Parameters:
. jac - Jacobian matrix
. B - optionally different preconditioning matrix
. flag - flag indicating matrix structure
*/
PetscErrorCode FormJacobian(SNES snes,Vec x,Mat jac,Mat B,void *ctx)
{
ApplicationCtx *user = (ApplicationCtx*) ctx;
PetscScalar *xx,d,A[3];
PetscErrorCode ierr;
PetscInt i,j[3],M,xs,xm;
DM da = user->da;
PetscFunctionBeginUser;
/*
Get pointer to vector data
*/
ierr = DMDAVecGetArrayRead(da,x,&xx);CHKERRQ(ierr);
ierr = DMDAGetCorners(da,&xs,NULL,NULL,&xm,NULL,NULL);CHKERRQ(ierr);
/*
Get range of locally owned matrix
*/
ierr = DMDAGetInfo(da,NULL,&M,NULL,NULL,NULL,NULL,NULL,NULL,NULL,NULL,NULL,NULL
,NULL);CHKERRQ(ierr);
/*
Determine starting and ending local indices for interior grid points.
Set Jacobian entries for boundary points.
*/
if (xs == 0) { /* left boundary */
i = 0; A[0] = 1.0;
ierr = MatSetValues(jac,1,&i,1,&i,A,INSERT VALUES);CHKERRQ(ierr);
xs++;xm--;
}
if (xs+xm == M) { /* right boundary */
i = M-1;
A[0] = 1.0;
ierr = MatSetValues(jac,1,&i,1,&i,A,INSERT VALUES);CHKERRQ(ierr);
132
![](asset-85.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
xm--;
}
/*
Interior grid points
- Note that in this case we set all elements for a particular
row at once.
*/
d = 1.0/(user->h*user->h);
for (i=xs; i<xs+xm; i++) {
j[0] = i - 1; j[1] = i; j[2] = i + 1;
A[0] = A[2] = d; A[1] = -2.0*d + 2.0*xx[i];
ierr = MatSetValues(jac,1,&i,3,j,A,INSERT VALUES);CHKERRQ(ierr);
}
/*
Assemble matrix, using the 2-step process:
MatAssemblyBegin(), MatAssemblyEnd().
By placing code between these two statements, computations can be
done while messages are in transition.
Also, restore vector.
*/
ierr = MatAssemblyBegin(jac,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
ierr = DMDAVecRestoreArrayRead(da,x,&xx);CHKERRQ(ierr);
ierr = MatAssemblyEnd(jac,MAT FINAL ASSEMBLY);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* ------------------------------------------------------------------- */
/*
Monitor - Optional user-defined monitoring routine that views the
current iterate with an x-window plot. Set by SNESMonitorSet().
Input Parameters:
snes - the SNES context
its - iteration number
norm - 2-norm function value (may be estimated)
ctx - optional user-defined context for private data for the
monitor routine, as set by SNESMonitorSet()
Note:
See the manpage for PetscViewerDrawOpen() for useful runtime options,
such as -nox to deactivate all x-window output.
*/
PetscErrorCode Monitor(SNES snes,PetscInt its,PetscReal fnorm,void *ctx)
{
133
![](asset-86.png)
5.5. MATRIX-FREE METHODS PETSc 3.8 March 24, 2018
PetscErrorCode ierr;
MonitorCtx *monP = (MonitorCtx*) ctx;
Vec x;
PetscFunctionBeginUser;
ierr = PetscPrintf(PETSC COMM WORLD,"iter = %D,SNES Function norm %g\n",its,(
double)fnorm);CHKERRQ(ierr);
ierr = SNESGetSolution(snes,&x);CHKERRQ(ierr);
ierr = VecView(x,monP->viewer);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* ------------------------------------------------------------------- */
/*
PreCheck - Optional user-defined routine that checks the validity of
candidate steps of a line search method. Set by SNESLineSearchSetPreCheck().
Input Parameters:
snes - the SNES context
xcurrent - current solution
y - search direction and length
Output Parameters:
y - proposed step (search direction and length) (possibly changed)
changed_y - tells if the step has changed or not
*/
PetscErrorCode PreCheck(SNESLineSearch linesearch,Vec xcurrent,Vec y, PetscBool *
changed_y, void * ctx)
{
PetscFunctionBeginUser;
*changed_y = PETSC FALSE;
PetscFunctionReturn(0);
}
/* ------------------------------------------------------------------- */
/*
PostCheck - Optional user-defined routine that checks the validity of
candidate steps of a line search method. Set by SNESLineSearchSetPostCheck().
Input Parameters:
snes - the SNES context
ctx - optional user-defined context for private data for the
monitor routine, as set by SNESLineSearchSetPostCheck()
xcurrent - current solution
y - search direction and length
x - the new candidate iterate
Output Parameters:
y - proposed step (search direction and length) (possibly changed)
134
![](asset-87.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
x - current iterate (possibly modified)
*/
PetscErrorCode PostCheck(SNESLineSearch linesearch,Vec xcurrent,Vec y,Vec x,
PetscBool *changed_y,PetscBool *changed_x, void * ctx)
{
PetscErrorCode ierr;
PetscInt i,iter,xs,xm;
StepCheckCtx *check;
ApplicationCtx *user;
PetscScalar *xa,*xa_last,tmp;
PetscReal rdiff;
DM da;
SNES snes;
PetscFunctionBeginUser;
*changed_x = PETSC FALSE;
*changed_y = PETSC FALSE;
ierr = SNESLineSearchGetSNES(linesearch, &snes);CHKERRQ(ierr);
check = (StepCheckCtx*)ctx;
user = check->user;
ierr = SNESGetIterationNumber(snes,&iter);CHKERRQ(ierr);
/* iteration 1 indicates we are working on the second iteration */
if (iter > 0) {
da = user->da;
ierr = PetscPrintf(PETSC COMM WORLD,"Checking candidate step at iteration %D
with tolerance %g\n",iter,(double)check->tolerance);CHKERRQ(ierr);
/* Access local array data */
ierr = DMDAVecGetArray(da,check->last_step,&xa_last);CHKERRQ(ierr);
ierr = DMDAVecGetArray(da,x,&xa);CHKERRQ(ierr);
ierr = DMDAGetCorners(da,&xs,NULL,NULL,&xm,NULL,NULL);CHKERRQ(ierr);
/*
If we fail the user-defined check for validity of the candidate iterate,
then modify the iterate as we like. (Note that the particular
modification
below is intended simply to demonstrate how to manipulate this data, not
as a meaningful or appropriate choice.)
*/
for (i=xs; i<xs+xm; i++) {
if (!PetscAbsScalar(xa[i])) rdiff = 2*check->tolerance;
else rdiff = PetscAbsScalar((xa[i] - xa_last[i])/xa[i]);
if (rdiff > check->tolerance) {
tmp = xa[i];
xa[i] = .5*(xa[i] + xa_last[i]);
*changed_x = PETSC TRUE;
135
![](asset-88.png)
5.5. MATRIX-FREE METHODS PETSc 3.8 March 24, 2018
ierr = PetscPrintf(PETSC COMM WORLD," Altering entry %D: x=%g,
x_last=%g, diff=%g, x_new=%g\n",i,(double)PetscAbsScalar(tmp),(double)
PetscAbsScalar(xa_last[i]),(double)rdiff,(double)PetscAbsScalar(xa[i])
);CHKERRQ(ierr);
}
}
ierr = DMDAVecRestoreArray(da,check->last_step,&xa_last);CHKERRQ(ierr);
ierr = DMDAVecRestoreArray(da,x,&xa);CHKERRQ(ierr);
}
ierr = VecCopy(x,check->last_step);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* ------------------------------------------------------------------- */
/*
PostSetSubKSP - Optional user-defined routine that reset SubKSP options when
hierarchical bjacobi PC is used
e.g,
mpiexec -n 8 ./ex3 -nox -n 10000 -ksp_type fgmres -pc_type bjacobi -
pc_bjacobi_blocks 4 -sub_ksp_type gmres -sub_ksp_max_it 3 -post_setsubksp
-sub_ksp_rtol 1.e-16
Set by SNESLineSearchSetPostCheck().
Input Parameters:
linesearch - the LineSearch context
xcurrent - current solution
y - search direction and length
x - the new candidate iterate
Output Parameters:
y - proposed step (search direction and length) (possibly changed)
x - current iterate (possibly modified)
*/
PetscErrorCode PostSetSubKSP(SNESLineSearch linesearch,Vec xcurrent,Vec y,Vec x,
PetscBool *changed_y,PetscBool *changed_x, void * ctx)
{
PetscErrorCode ierr;
SetSubKSPCtx *check;
PetscInt iter,its,sub_its,maxit;
KSP ksp,sub_ksp,*sub_ksps;
PC pc;
PetscReal ksp_ratio;
SNES snes;
PetscFunctionBeginUser;
ierr = SNESLineSearchGetSNES(linesearch, &snes);CHKERRQ(ierr);
check = (SetSubKSPCtx*)ctx;
ierr = SNESGetIterationNumber(snes,&iter);CHKERRQ(ierr);
136
![](asset-89.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
ierr = SNESGetKSP(snes,&ksp);CHKERRQ(ierr);
ierr = KSPGetPC(ksp,&pc);CHKERRQ(ierr);
ierr = PCBJacobiGetSubKSP(pc,NULL,NULL,&sub_ksps);CHKERRQ(ierr);
sub_ksp = sub_ksps[0];
ierr = KSPGetIterationNumber(ksp,&its);CHKERRQ(ierr); /* outer KSP
iteration number */
ierr = KSPGetIterationNumber(sub_ksp,&sub_its);CHKERRQ(ierr); /* inner KSP
iteration number */
if (iter) {
ierr = PetscPrintf(PETSC COMM WORLD," ...PostCheck snes iteration %D,
ksp_it %d %d, subksp_it %d\n",iter,check->its0,its,sub_its);CHKERRQ(ierr);
ksp_ratio = ((PetscReal)(its))/check->its0;
maxit = (PetscInt)(ksp_ratio*sub_its + 0.5);
if (maxit < 2) maxit = 2;
ierr = KSPSetTolerances(sub_ksp,PETSC DEFAULT,PETSC DEFAULT,PETSC DEFAULT,
maxit);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC COMM WORLD," ...ksp_ratio %g, new maxit %d\n\n",
ksp_ratio,maxit);CHKERRQ(ierr);
}
check->its0 = its; /* save current outer KSP iteration number */
PetscFunctionReturn(0);
}
/* ------------------------------------------------------------------- */
/*
MatrixFreePreconditioner - This routine demonstrates the use of a
user-provided preconditioner. This code implements just the null
preconditioner, which of course is not recommended for general use.
Input Parameters:
+ pc - preconditioner
- x - input vector
Output Parameter:
. y - preconditioned vector
*/
PetscErrorCode MatrixFreePreconditioner(PC pc,Vec x,Vec y)
{
PetscErrorCode ierr;
ierr = VecCopy(x,y);CHKERRQ(ierr);
return 0;
}
Figure 13: Example of Uniprocess SNES Code - Both Conventional and Matrix-Free Jacobians
137
![](asset-8a.png)
5.6. FINITE DIFFERENCE JACOBIAN APPROXIMATIONS PETSc 3.8 March 24, 2018
5.6 Finite Difference Jacobian Approximations
PETSc provides some tools to help approximate the Jacobian matrices efficiently via finite differ-
ences. These tools are intended for use in certain situations where one is unable to compute Jacobian
matrices analytically, and matrix-free methods do not work well without a preconditioner, due to
very poor conditioning. The approximation requires several steps:
•First, one colors the columns of the (not yet built) Jacobian matrix, so that columns of the
same color do not share any common rows.
•Next, one creates a MatFDColoring data structure that will be used later in actually com-
puting the Jacobian.
•Finally, one tells the nonlinear solvers of SNES to use the SNESComputeJacobianDefaultColor()
routine to compute the Jacobians.
A code fragment that demonstrates this process is given below.
ISColoring iscoloring;
MatFDColoring fdcoloring;
MatColoring coloring;
/*
This initializes the nonzero structure of the Jacobian. This is artificial
because clearly if we had a routine to compute the Jacobian we wouldn’t
need to use finite differences.
*/
FormJacobian(snes,x, &J, &J, &user);
/*
Color the matrix, i.e. determine groups of columns that share no common
rows. These columns in the Jacobian can all be computed simultaneously.
*/
MatColoringCreate(J, &coloring);
MatColoringSetType(coloring,MATCOLORINGSL);
MatColoringSetFromOptions(coloring);
MatColoringApply(coloring, &iscoloring);
MatColoringDestroy( &coloring);
/*
Create the data structure that SNESComputeJacobianDefaultColor() uses
to compute the actual Jacobians via finite differences.
*/
MatFDColoringCreate(J,iscoloring, &fdcoloring);
ISColoringDestroy( &iscoloring);
MatFDColoringSetFunction(fdcoloring,(PetscErrorCode (*)(void))FormFunction,
&user);
MatFDColoringSetFromOptions(fdcoloring);
/*
Tell SNES to use the routine SNESComputeJacobianDefaultColor()
to compute Jacobians.
138
![](asset-8b.png)
CHAPTER 5. SNES: NONLINEAR SOLVERS PETSc 3.8 March 24, 2018
*/
SNESSetJacobian(snes,J,J,SNESComputeJacobianDefaultColor,fdcoloring);
Of course, we are cheating a bit. If we do not have an analytic formula for computing the
Jacobian, then how do we know what its nonzero structure is so that it may be colored? Determining
the structure is problem dependent, but fortunately, for most structured grid problems (the class of
problems for which PETSc was originally designed) if one knows the stencil used for the nonlinear
function one can usually fairly easily obtain an estimate of the location of nonzeros in the matrix.
This is harder in the unstructured case, and has not yet been implemented in general.
One need not necessarily use a MatColoring object to determine a coloring. For example,
if a grid can be colored directly (without using the associated matrix), then that coloring can be
provided to MatFDColoringCreate(). Note that the user must always preset the nonzero structure
in the matrix regardless of which coloring routine is used.
For sequential matrices PETSc provides three matrix coloring routines on from the MINPACK
package [24]: smallest-last (sl), largest-first (lf), and incidence-degree (id). In addition, two
implementations of parallel colorings are in PETSc, greedy (greedy) and Jones-Plassmann (jp).
These colorings, as well as the “natural” coloring for which each column has its own unique color,
may be accessed with the command line options
-mat_coloring_type <l,id,lf,natural,greedy,jp>
Alternatively, one can set a coloring type of MATCOLORINGGREEDY or MATCOLORINGJP for parallel
algorithms, or MATCOLORINGSL,MATCOLORINGID,MATCOLORINGLF,MATCOLORINGNATURAL for se-
quential algorithms when calling MatColoringSetType().
As for the matrix-free computation of Jacobians (see Section 5.5), two parameters affect the
accuracy of the finite difference Jacobian approximation. These are set with the command
MatFDColoringSetParameters(MatFDColoring fdcoloring,PetscReal rerror,PetscReal
umin);
The parameter rerror is the square root of the relative error in the function evaluations, erel; the
default is the square root of machine epsilon (about 10−8in double precision), which assumes that
the functions are evaluated to full floating-point precision accuracy. The second parameter, umin,
is a bit more involved; its default is 10e−8. Column iof the Jacobian matrix (denoted by F:i) is
approximated by the formula
F0
:i≈F(u+h∗dxi)−F(u)
h
where his computed via
h=erel ∗uiif |ui|> umin
h=erel ∗umin ∗sign(ui) otherwise.
These parameters may be set from the options database with
-mat_fd_coloring_err <err>
-mat_fd_coloring_umin <umin>
Note that MatColoring type MATCOLORINGSL,MATCOLORINGLF, and MATCOLORINGID are sequen-
tial algorithms. MATCOLORINGJP and MATCOLORINGGREEDY are parallel algorithms, although in prac-
tice they may create more colors than the sequential algorithms. If one computes the coloring
iscoloring reasonably with a parallel algorithm or by knowledge of the discretization, the routine
MatFDColoringCreate() is scalable. An example of this for 2D distributed arrays is given below
that uses the utility routine DMCreateColoring().
139
![](asset-8c.png)
5.7. VARIATIONAL INEQUALITIES PETSc 3.8 March 24, 2018
DMCreateColoring(da,IS_COLORING_GHOSTED, &iscoloring);
MatFDColoringCreate(J,iscoloring, &fdcoloring);
MatFDColoringSetFromOptions(fdcoloring);
ISColoringDestroy( &iscoloring);
Note that the routine MatFDColoringCreate() currently is only supported for the AIJ and
BAIJ matrix formats.
5.7 Variational Inequalities
SNES can also solve variational inequalities with box constraints. That are nonlinear algebraic
systems with additional inequality constraints on some or all of the variables: Lui≤ui≤Hui.
Some or all of the lower bounds may be negative infinity (indicated to PETSc with SNES_VI_NINF)
and some or all of the upper bounds may be infinity (indicated by SNES_VI_INF). The command
SNESVISetVariableBounds(SNES,Vec Lu,Vec Hu);
is used to indicate that one is solving a variational inequality. The option -snes_vi_monitor
turns on extra monitoring of the active set associated with the bounds and -snes_vi_type allows
selecting from several VI solvers, the default is preferred.
5.8 Nonlinear Preconditioning
Nonlinear preconditioning in PETSc involves the use of an inner SNES instance to define the step
for an outer SNES instance. The inner instance may be extracted using
SNESGetNPC(SNES snes,SNES *npc);
and passed run-time options using the -npc_ prefix. Nonlinear preconditioning comes in two flavors:
left and right. The side may be changed using -snes_npc_side or SNESSetNPCSide(). Left non-
linear preconditioning redefines the nonlinear function as the action of the nonlinear preconditioner
M;
FM(x) = M(x,b)−x.(5.6)
Right nonlinear preconditioning redefines the nonlinear function as the function on the action of
the nonlinear preconditioner;
F(M(x,b)) = b,(5.7)
which can be interpreted as putting the preconditioner into “striking distance” of the solution by
outer acceleration.
In addition, basic patterns of solver composition are available with the SNESType SNESCOMPOSITE.
This allows for two or more SNES instances to be combined additively or multiplicatively. By com-
mand line, a set of SNES types may be given by comma separated list argument to
-snes_composite_sneses. There are additive (SNES_COMPOSITE_ADDITIVE), additive with optimal
damping
(SNES_COMPOSITE_ADDITIVEOPTIMAL), and multiplicative (SNES_COMPOSITE_MULTIPLICATIVE) vari-
ants which may be set with
SNESCompositeSetType(SNES,SNESCompositeType);
New subsolvers may be added to the composite solver with
140
![](asset-8e.png)
5.8. NONLINEAR PRECONDITIONING PETSc 3.8 March 24, 2018
142
![](asset-8f.png)
Chapter 6
TS: Scalable ODE and DAE Solvers
The TS library provides a framework for the scalable solution of ODEs and DAEs arising from the
discretization of time-dependent PDEs.
Simple Example: Consider the PDE
ut=uxx
discretized with centered finite differences in space yielding the semi-discrete equation
(ui)t=ui+1 −2ui+ui−1
h2,
ut=˜
Au;
or with piecewise linear finite elements approximation in space u(x, t).
=Piξi(t)φi(x) yielding the
semi-discrete equation
Bξ0(t) = Aξ(t)
Now applying the backward Euler method results in
(B−dtnA)un+1 =Bun,
in which
uni=ξi(tn).
=u(xi, tn),
ξ0(tn+1).
=un+1i−uni
dtn,
Ais the stiffness matrix, and Bis the identity for finite differences or the mass matrix for the finite
element method.
The PETSc interface for solving time dependent problems assumes the problem is written in
the form
F(t, u, ˙u) = G(t, u), u(t0) = u0.
In general, this is a differential algebraic equation (DAE) 1. For ODE with nontrivial mass matrices
such as arise in FEM, the implicit/DAE interface significantly reduces overhead to prepare the
system for algebraic solvers (SNES/KSP) by having the user assemble the correctly shifted matrix.
Therefore this interface is also useful for ODE systems.
To solve an ODE or DAE one uses:
1If the matrix F˙u(t) = ∂F/∂ ˙uis nonsingular then it is an ODE and can be transformed to the standard explicit
form, although this transformation may not lead to efficient algorithms.
143
![](asset-90.png)
PETSc 3.8 March 24, 2018
•Function F(t, u, ˙u)
TSSetIFunction(TS ts,Vec R,PetscErrorCode
(*f)(TS,PetscReal,Vec,Vec,Vec,void*),void *funP);
The vector Ris an optional location to store the residual. The arguments to the function
f() are the timestep context, current time, input state u, input time derivative ˙u, and the
(optional) user-provided context funP. If F(t, u, ˙u) = ˙uthen one need not call this function.
•Function G(t, u), if it is nonzero, is provided with the function
TSSetRHSFunction(TS ts,Vec R,PetscErrorCode
(*f)(TS,PetscReal,Vec,Vec,void*),void *funP);
•Jacobian σF ˙u(tn, un,˙un) + Fu(tn, un,˙un)
If using a fully implicit or semi-implicit (IMEX) method one also can provide an appropriate
(approximate) Jacobian matrix of F().
TSSetIJacobian(TS ts,Mat A,Mat B,PetscErrorCode
(*fjac)(TS,PetscReal,Vec,Vec,PetscReal,Mat,Mat,void*),void *jacP);
The arguments for the function fjac() are the timestep context, current time, input state
u, input derivative ˙u, input shift σ, matrix A, preconditioning matrix B, and the (optional)
user-provided context jacP.
The Jacobian needed for the nonlinear system is, by the chain rule,
dF
dun=∂F
∂˙u|un∂˙u
∂u |un+∂F
∂u |un.
For any ODE integration method the approximation of ˙uis linear in unhence ∂˙u
∂u |un=σ,
where the shift σdepends on the ODE integrator and time step but not on the function being
integrated. Thus
dF
dun=σF ˙u(tn, un,˙un) + Fu(tn, un,˙un).
This explains why the user provide Jacobian is in the given form for all integration methods.
An equivalent way to derive the formula is to note that
F(tn, un,˙un) = F(tn, un, w +σ∗un)
where wis some linear combination of previous time solutions of uso that
dF
dun=σF ˙u(tn, un,˙un) + Fu(tn, un,˙un)
again by the chain rule.
For example, consider backward Euler’s method applied to the ODE F(t, u, ˙u) = ˙u−f(t, u)
with ˙u= (un−un−1)/δt and ∂˙u
∂u |un= 1/δt resulting in
dF
dun= (1/δt)F˙u+Fu(tn, un,˙un).
But F˙u= 1, in this special case, resulting in the expected Jacobian I/δt −fu(t, un).
144
![](asset-91.png)
CHAPTER 6. TS: SCALABLE ODE AND DAE SOLVERS PETSc 3.8 March 24, 2018
TS Name Reference Class Type Order
euler forward Euler one-step explicit 1
ssp multistage SSP [23] Runge-Kutta explicit ≤4
rk* multistage Runge-Kutta explicit ≥1
beuler backward Euler one-step implicit 1
cn Crank-Nicolson one-step implicit 2
theta* theta-method one-step implicit ≤2
alpha alpha-method [21] one-step implicit 2
gl general linear [5] multistep-multistage implicit ≤3
eimex extrapolated IMEX [8] one-step ≥1, adaptive
arkimex see Tab. 12 IMEX Runge-Kutta IMEX 1 −5
rosw see Tab. 13 Rosenbrock-W linearly implicit 1 −4
glee see Tab. 14 GL with global error explicit and implicit 1 −3
Table 10: Time integration schemes.
•Jacobian Gu
If using a fully implicit method and the function G() is provided, one also can provide an
appropriate (approximate) Jacobian matrix of G().
TSSetRHSJacobian(TS ts,Mat A,Mat B,
PetscErrorCode (*fjac)(TS,PetscReal,Vec,Mat,Mat,void*),void *jacP);
The arguments for the function fjac() are the timestep context, current time, input state u,
matrix A, preconditioning matrix B, and the (optional) user-provided context jacP.
Providing appropriate F() and G() for your problem allows for the easy runtime switching
between explicit, semi-implicit (IMEX), and fully implicit method.
6.1 Basic TS Options
The user first creates a TS object with the command
int TSCreate(MPI Comm comm,TS *ts);
int TSSetProblemType(TS ts,TSProblemType problemtype);
The TSProblemType is one of TS_LINEAR or TS_NONLINEAR.
To set up TS for solving an ODE, one must set the “initial conditions” for the ODE with
TSSetSolution(TS ts, Vec initialsolution);
One can set the solution method with the routine
TSSetType(TS ts,TSType type);
Currently supported types are TSEULER,TSRK (Runge-Kutta), TSBEULER,TSCN (Crank-Nicolson),
TSTHETA,TSGLLE (generalized linear), TSPSEUDO, and TSSUNDIALS (only if the Sundials package is
installed), or the command line option
-ts_type euler,rk,beuler,cn,theta,gl,pseudo,sundials,eimex,arkimex,rosw.
A list of available methods is given in Table 10.
Set the initial time with the command
145
![](asset-92.png)
6.2. USING IMPLICIT-EXPLICIT (IMEX) METHODS PETSc 3.8 March 24, 2018
TSSetTime(TS ts,PetscReal time);
One can change the timestep with the command
TSSetTimeStep(TS ts,PetscReal dt);
can determine the current timestep with the routine
TSGetTimeStep(TS ts,PetscReal* dt);
Here, “current” refers to the timestep being used to attempt to promote the solution form unto
un+1.
One sets the total number of timesteps to run or the total time to run (whatever is first) with the
commands
TSSetMaxSteps(TS ts,PetscInt maxsteps);
TSSetMaxTime(TS ts,PetscReal maxtime);
and determines the behavior near the final time with
TSSetExactFinalTime(TS ts,TSExactFinalTimeOption eftopt);
where eftopt is one of TS_EXACTFINALTIME_STEPOVER,TS_EXACTFINALTIME_INTERPOLATE, or
TS_EXACTFINALTIME_MATCHSTEP. One performs the requested number of time steps with
TSSolve(TS ts,Vec U);
The solve call implicitly sets up the timestep context; this can be done explicitly with
TSSetUp(TS ts);
One destroys the context with
TSDestroy(TS *ts);
and views it with
TSView(TS ts,PetscViewer viewer);
In place of TSSolve(), a single step can be taken using
TSStep(TS ts);
6.2 Using Implicit-Explicit (IMEX) Methods
For “stiff” problems or those with multiple time scales F() will be treated implicitly using a
method suitable for stiff problems and G() will be treated explicitly when using an IMEX method
like TSARKIMEX.F() is typically linear or weakly nonlinear while G() may have very strong
nonlinearities such as arise in non-oscillatory methods for hyperbolic PDE. The user provides three
pieces of information, the APIs for which have been described above.
•“Slow” part G(t, u) using TSSetRHSFunction().
•“Stiff” part F(t, u, ˙u) using TSSetIFunction().
•Jacobian Fu+σF ˙uusing TSSetIJacobian().
146
![](asset-93.png)
CHAPTER 6. TS: SCALABLE ODE AND DAE SOLVERS PETSc 3.8 March 24, 2018
˙y=g(t, y) nonstiff ODE F(t, y, ˙y) = ˙y,G(t, y) = g(t, y)
˙y=f(t, y) stiff ODE F(t, y, ˙y) = ˙y−f(t, y), G(t, y)=0
M˙y=f(t, y) stiff ODE with mass ma-
trix
F(t, y, ˙y) = M˙y−f(t, y), G(t, y) = 0
M˙y=g(t, y) nonstiff ODE with mass
matrix
F(t, y, ˙y) = ˙y, G(t, y) = M−1g(t, y)
˙y=g(t, y) + f(t, y) stiff-nonstiff ODE F(t, y, ˙y) = ˙y−f(t, y), G(t, y) = g(t, y)
M˙y=g(t, y) + f(t, y) stiff-nonstiff ODE with
mass matrix
F(t, y, ˙y) = M˙y−f(t, y), G(t, y) =
M−1g(t, y)
f(t, y, ˙y) = 0 implicit ODE/DAE F(t, y, ˙y) = f(t, y, ˙y), G(t, y) = 0; the
user needs to set TSSetEquationType() to
TS_EQ_IMPLICIT or higher
Table 11: In PETSc, DAEs and ODEs are formulated as F(t, y, ˙y) = G(t, y), where F() is meant
to be integrated implicitly and G() explicitly. An IMEX formulation such as M˙y=g(t, y) + f(t, y)
requires the user to provide M−1g(t, y) or solve g(t, y)−Mx = 0 in place of G(t, u). General cases
such as F(t, y, ˙y) = G(t, y) are not amenable to IMEX Runge-Kutta, but can be solved by using
fully implicit methods.
The user needs to set TSSetEquationType() to TS_EQ_IMPLICIT or higher if the problem is
implicit; e.g., F(t, u, ˙u) = M˙u−f(t, u), where Mis not the identity matrix:
•the problem is an implicit ODE (defined implicitly through TSSetIFunction()) or
•a DAE is being solved.
An IMEX problem representation can be made implicit by setting TSARKIMEXSetFullyImplicit().
Some use-case examples for TSARKIMEX are listed in Table 11 and a list of methods with a
summary of their properties is given in Table 12.
ROSW are linearized implicit Runge-Kutta methods known as Rosenbrock W-methods. They
can accommodate inexact Jacobian matrices in their formulation. A series of methods are available
in PETSc listed in Table 13.
GLEE methods
In this section, we describe explicit and implicit time stepping methods with global error estimation
that are introduced in [7]. The solution vector for a GLEE method is either [y, ˜y] or [y,ε], where
yis the solution, ˜yis the “auxiliary solution,” and εis the error. The working vector that TSGLEE
uses is Y= [y,˜y], or [y,ε]. A GLEE method is defined by
•(p, r, s): (order, steps, and stages),
•γ: factor representing the global error ratio,
•A, U, B, V : method coefficients,
•S: starting method to compute the working vector from the solution (say at the beginning of
time integration) so that Y=Sy,
•F: finalizing method to compute the solution from the working vector,y=F Y .
•Fembed: coefficients for computing the auxiliary solution ˜yfrom the working vector (˜y=
FembedY),
147
![](asset-94.png)
6.2. USING IMPLICIT-EXPLICIT (IMEX) METHODS PETSc 3.8 March 24, 2018
Name Reference Stages Order IM SA Embed. Dense Remarks
(IM) (stage) Stab. Output
a2 based on CN 2(1) 2(2) A-Stable yes yes(1) yes(2)
l2 SSP2(2,2,2)[28] 2(2) 2(1) L-Stable yes yes(1) yes(2) SSP, SDIRK
ars122 ARS122, [1] 2(1) 3(1) A-Stable yes yes(1) yes(2)
2c [15] 3(2) 2(2) L-Stable yes yes(1) yes(2) SDIRK
2d [15] 3(2) 2(2) L-Stable yes yes(1) yes(2) SDIRK
2e [15] 3(2) 2(2) L-Stable yes yes(1) yes(2) SDIRK
prssp2 PRS(3,3,2) [28] 3(3) 3(1) L-Stable yes no no SSP, nonSDIRK
3 [22] 4(3) 3(2) L-Stable yes yes(2) yes(2) SDIRK
bpr3 [3] 5(4) 3(2) L-Stable yes no no SDIRK, DAE-1
ars443 [1] 5(4) 3(1) L-Stable yes no no SDIRK
4 [22] 6(5) 4(2) L-Stable yes yes(3) yes(2,3) SDIRK
5 [22] 8(7) 5(2) L-Stable yes yes(4) yes(3) SDIRK
Table 12: List of the currently available IMEX Runge-Kutta schemes. For each method we listed the
-ts arkimex type name, the reference, the total number of stages/implicit stages, the order/stage-
order, the implicit stability properties, stiff accuracy (SA), the existence of an embedded scheme,
and dense output.
TS Reference Stages Order IM SA Embed. Dense -W PDAE Remarks
Name (IM) (stage) Stab. Output
theta1 classical 1(1) 1(1) L-Stable - - - - - -
theta2 classical 1(1) 2(2) A-Stable - - - - - -
2m Zoltan 2(2) 2(1) L-Stable No Yes(1) Yes(2) Yes No SSP
2p Zoltan 2(2) 2(1) L-Stable No Yes(1) Yes(2) Yes No SSP
ra3pw [30] 3(3) 3(1) A-Stable No Yes Yes(2) No Yes(3) -
ra34pw2 [30] 4(4) 3(1) L-Stable Yes Yes Yes(3) Yes Yes(3) -
rodas3 [33] 4(4) 3(1) L-Stable Yes Yes No No Yes -
sandu3 [33] 3(3) 3(1) L-Stable Yes Yes Yes(2) No No -
assp3p3s1c unpublished 3(2) 3(1) A-Stable No Yes Yes(2) Yes No SSP
lassp3p4s2c unpublished 4(3) 3(1) L-Stable No Yes Yes(3) Yes No SSP
lassp3p4s2c unpublished 4(3) 3(1) L-Stable No Yes Yes(3) Yes No SSP
ark3 unpublished 4(3) 3(1) L-Stable No Yes Yes(3) Yes No IMEX-RK
Table 13: List of the currently available Rosenbrock W-schemes. For each method we listed the
reference, the total number of stages and implicit stages, the scheme order and stage-order, the
implicit stability properties, stiff accuracy (SA), the existence of an embedded scheme, dense out-
put, the capacity to use inexact Jacobian matrices (-W), and high order integration of differential
algebraic equations (PDAE).
148
![](asset-95.png)
CHAPTER 6. TS: SCALABLE ODE AND DAE SOLVERS PETSc 3.8 March 24, 2018
TS Reference IM/EX (p,r,s) γForm Notes
TSGLEEi1 BE1 IM 1,3,2 0.5 yε Based on backward Euler
TSGLEE23 23 EX 2,3,2 0 yε
TSGLEE24 24 EX 2,4,2 0 y˜y
TSGLEE25I 25i EX 2,5,2 0 y˜y
TSGLEE35 35 EX 3,5,2 0 y˜y
TSGLEEEXRK2A exrk2a EX 2,6,2 0.25 yε
TSGLEERK32G1 rk32g1 EX 3,8,2 0 yε
TSGLEERK285EX rk285ex EX 2,9,2 0.25 yε
Table 14: List of the currently available GL schemes with global error estimation introduced in [7].
•Ferror: coefficients to compute the estimated error vector from the working vector (ε=
FerrorY).
•Serror: coefficients to initialize the auxiliary solution (˜yor ε) from a specified error vector (ε).
It is currently implemented only for r= 2. We have yaux =Serror[0] ∗ε+Serror[1] ∗y, where
yaux is the 2nd component of the working vector Y.
The methods can be described in two mathematically equivalent forms: propagate two components
(“y˜yform”) and propagating the solution and its estimated error (“yε form”). The two forms
are not explicitly specified in TSGLEE; rather, the specific values of B, U, S, F, Fembed, and Ferror
characterize whether the method is in y˜yor yε form.
The API used by this TS method includes:
•TSGetSolutionComponents: Get all the solution components of the working vector
ierr = TSGetSolutionComponents(TS,int*,Vec*)
Call with NULL as the last argument to get the total number of components in the working
vector Y(this is r(not r−1)), then call to get the i-th solution component.
•TSGetAuxSolution: Returns the auxiliary solution ˜y(computed as FembedY)
ierr = TSGetAuxSolution(TS,Vec*)
•TSGetTimeError: Returns the estimated error vector ε(computed as FerrorYif n= 0 or
restores the error estimate at the end of the previous step if n=−1)
ierr = TSGetTimeError(TS,PetscInt n,Vec*)
•TSSetTimeError: Initializes the auxiliary solution (˜yor ε) for a specified initial error.
ierr = TSSetTimeError(TS,Vec)
The local error is estimated as ε(n+ 1) −ε(n). This is to be used in the error control. The
error in y˜yGLEE is ε(n) = 1
1−γ∗(˜y(n)−y(n)).
Note that yand ˜yare reported to TSAdapt basic (TSADAPTBASIC), and thus it computes the
local error as εloc = (˜y−y). However, the actual local error is εloc =εn+1 −εn=1
1−γ∗[(˜y−y)n+1 −
(˜y−y)n].
149
![](asset-96.png)
6.3. USING FULLY IMPLICIT METHODS PETSc 3.8 March 24, 2018
6.3 Using fully implicit methods
To use a fully implicit method like TSTHETA or TSGL, either provide the Jacobian of F() (and G()
if G() is provided) or use a DM that provides a coloring so the Jacobian can be computed efficiently
via finite differences.
6.4 Using the Explicit Runge-Kutta timestepper with variable
timesteps
The explicit Euler and Runge-Kutta methods require the ODE be in the form
˙u=G(u, t).
The user can either call TSSetRHSFunction() and/or they can call TSSetIFunction() (so long as
the function provided to TSSetIFunction() is equivalent to ˙u+˜
F(t, u)) but the Jacobians need
not be provided.2
The Explicit Runge-Kutta timestepper with variable timesteps is an implementation of the
standard Runge-Kutta with an embedded method. The error in each timestep is calculated using the
solutions from the Runge-Kutta method and its embedded method (the 2-norm of the difference is
used). The default method is the 3rd-order Bogacki-Shampine method with a 2nd-order embedded
method (TSRK3BS). Other available methods are the 5th-order Fehlberg RK scheme with a 4th-
order embedded method (TSRK5F), the 5th-order Dormand-Prince RK scheme with a 4th-order
embedded method (TSRK5DP), and the 5th-order Bogacki-Shampine RK scheme with a 4th-order
embedded method (TSRK5BS). Variable timesteps cannot be used with RK schemes that do not
have an embedded method (TSRK1FE - 1st-order, 1-stage forward Euler, TSRK2A - 2nd-order, 2-stage
RK scheme, TSRK3 - 3rd-order, 3-stage RK scheme, TSRK4 - 4-th order, 4-stage RK scheme).
6.5 Special Cases
•˙u=Au. First compute the matrix Athen call
TSSetProblemType(ts,TS LINEAR);
TSSetRHSFunction(ts,NULL,TSComputeRHSFunctionLinear,NULL);
TSSetRHSJacobian(ts,A,A,TSComputeRHSJacobianConstant,NULL);
or
TSSetProblemType(ts,TS LINEAR);
TSSetIFunction(ts,NULL,TSComputeIFunctionLinear,NULL);
TSSetIJacobian(ts,A,A,TSComputeIJacobianConstant,NULL);
•˙u=A(t)u. Use
TSSetProblemType(ts,TS LINEAR);
TSSetRHSFunction(ts,NULL,TSComputeRHSFunctionLinear,NULL);
TSSetRHSJacobian(ts,A,A,YourComputeRHSJacobian, &appctx);
where YourComputeRHSJacobian() is a function you provide that computes Aas a function
of time. Or use
2PETSc will automatically translate the function provided to the appropriate form.
150
![](asset-97.png)
CHAPTER 6. TS: SCALABLE ODE AND DAE SOLVERS PETSc 3.8 March 24, 2018
ID Name Notes
TSADAPTNONE none no adaptivity
TSADAPTBASIC basic the default adaptor
TSADAPTGLEE glee extension of the basic adaptor
to treat TolAand TolRas sep-
arate criteria. It can also con-
trol global erorrs if the inte-
grator (e.g., TSGLEE) provides
this information
Table 15: TSAdapt: available adaptors.
TSSetProblemType(ts,TS LINEAR);
TSSetIFunction(ts,NULL,TSComputeIFunctionLinear,NULL);
TSSetIJacobian(ts,A,A,YourComputeIJacobian, &appctx);
6.6 Monitoring and visualizing solutions
•-ts_monitor - prints the time and timestep at each iteration.
•-ts_adapt_monitor - prints information about the timestep adaption calculation at each
iteration.
•-ts_monitor_lg_timestep - plots the size of each timestep, TSMonitorLGTimeStep().
•-ts_monitor_lg_solution - for ODEs with only a few components (not arising from the
discretization of a PDE) plots the solution as a function of time, TSMonitorLGSolution().
•-ts_monitor_lg_error - for ODEs with only a few components plots the error as a function
of time, only if TSSetSolutionFunction() is provided, TSMonitorLGError().
•-ts_monitor_draw_solution - plots the solution at each iteration, TSMonitorDrawSolution().
•-ts_monitor_draw_error - plots the error at each iteration only if TSSetSolutionFunction()
is provided, TSMonitorDrawSolution().
•ts_monitor_solution binary[:filename]} saves the solution at each iteration to a binary
file, TSMonitorSolution().
•-ts_monitor_solution_vtk < filename-%%03D.vts> - saves the solution at each iteration
to a file in vtk format, TSMonitorSolutionVTK().
6.7 Error control via variable time-stepping
Most of the time stepping methods avaialable in PETSc have an error estimation and error control
mechanism. This mechanism is implemented by changing the step size in order to maintain user
specified absolute and relative tolerances. The PETSc object responsible with error control is
TSAdapt. The available TSAdapt types are listed in Table 15.
When using TSADAPTBASIC (the default), the user typically provides a desired absolute TolA
or a relative TolRerror tolerance by invoking TSSetTolerances() or at the command line with
151
![](asset-98.png)
6.8. HANDLING OF DISCONTINUITIES PETSc 3.8 March 24, 2018
options -ts_atol and -ts_rtol. The error estimate is based on the local truncation error, so for
every step the algorithm verifies that the estimated local truncation error satisfies the tolerances
provided by the user and computes a new step size to be taken. For multistage methods, the local
truncation is obtained by comparing the solution yto a lower order bp=p−1 approximation, by,
where pis the order of the method and bpthe order of by.
The adaptive controller at step ncomputes a tolerance level
T oln(i) = TolA(i) + max(yn(i),byn(i))TolR(i),
and forms the acceptable error level
wlten=1
m
m
X
i=1 skyn(i)−byn(i)k
T ol(i),
where the errors are computed componentwise, mis the dimension of yand -ts_adapt_wnormtype
is 2(default). If -ts_adapt_wnormtype is infinity (max norm), then
wlten= max
1...m
kyn(i)−byn(i)k
T ol(i).
The error tolerances are satisfied when wlte ≤1.0.
The next step size is based on this error estimate, and determined by
∆tnew(t)=∆told min(αmax,max(αmin, β(1/wlte) 1
bp+1 )) ,(6.1)
where αmin =-ts_adapt_clip[0] and αmax=-ts_adapt_clip[1] keep the change in ∆tto within
a certain factor, and β < 1 is chosen through -ts_adapt_safety so that there is some margin to
which the tolerances are satisfied and so that the probability of rejection is decreased.
This adaptive controller works in the following way. After completing step k, if wltek+1 ≤1.0,
then the step is accepted and the next step is modified according to (6.1); otherwise, the step is
rejected and retaken with the step length computed in (6.1).
TSAdapt glee. TSADAPTGLEE is an extension of the basic adaptor to treat TolAand TolRas separate
criteria. it can also control global errors if the integrator (e.g., TSGLEE) provides this information.
6.8 Handling of discontinuities
For problems that involve discontinuous right hand sides, one can set an “event” function g(t, u)
for PETSc to detect and locate the times of discontinuities (zeros of g(t, u)). Events can be defined
through the event monitoring routine
TSSetEventHandler(TS ts,PetscInt nevents,PetscInt *direction,PetscBool
*terminate,PetscErrorCode (*eventhandler)(TS,PetscReal,Vec,PetscScalar*,void*
eventP),PetscErrorCode
(*postevent)(TS,PetscInt,PetscInt[],PetscReal,Vec,PetscBool,void*
eventP),void *eventP);
Here, nevents denotes the number of events, direction sets the type of zero crossing to be detected
for an event (+1 for positive zero-crossing, -1 for negative zero-crossing, and 0 for both), terminate
conveys whether the time-stepping should continue or halt when an event is located, eventmonitor
is a user- defined routine that specifies the event description, postevent is an optional user-defined
routine to take specific actions following an event.
152
![](asset-99.png)
CHAPTER 6. TS: SCALABLE ODE AND DAE SOLVERS PETSc 3.8 March 24, 2018
The arguments to eventhandler() are the timestep context, current time, input state u, array
of event function value, and the (optional) user-provided context eventP.
The arguments to postevent() routine are the timestep context, number of events occured,
indices of events occured, current time, input state u, a boolean flag indicating forward solve (1)
or adjoint solve (0), and the (optional) user-provided context eventP.
The event monitoring functionality is only available with PETSc’s implicit time-stepping solvers
TSTHETA,TSARKIMEX, and TSROSW.
6.9 Using TChem from PETSc
TChem3is a package originally developed at Sandia National Laboratory that can read in CHEMKIN4
data files and compute the right hand side function and its Jacobian for a reaction ODE system. To
utilize PETSc’s ODE solvers for these systems, first install PETSc with the additional ./configure
option --download-tchem. We currently provide two examples of its use; one for single cell reaction
and one for an “artificial” one dimensional problem with periodic boundary conditions and diffu-
sion of all species. The self-explanatory examples are $PETSC_DIR/src/ts/examples/tutorials/
extchem.c and $PETSC_DIR/src/ts/examples/tutorials/exchemfield.c.
6.10 Using Sundials from PETSc
Sundials is a parallel ODE solver developed by Hindmarsh et al. at LLNL. The TS library provides
an interface to use the CVODE component of Sundials directly from PETSc. (To configure PETSc
to use Sundials, see the installation guide, docs/installation/index.htm.)
To use the Sundials integrators, call
TSSetType(TS ts,TSType TSSUNDIALS);
or use the command line option -ts_type sundials.
Sundials’ CVODE solver comes with two main integrator families, Adams and BDF (backward
differentiation formula). One can select these with
TSSundialsSetType(TS ts,TSSundialsLmmType [SUNDIALS_ADAMS,SUNDIALS_BDF]);
or the command line option -ts_sundials_type <adams,bdf>. BDF is the default.
Sundials does not use the SNES library within PETSc for its nonlinear solvers, so one cannot
change the nonlinear solver options via SNES. Rather, Sundials uses the preconditioners within the
PC package of PETSc, which can be accessed via
TSSundialsGetPC(TS ts,PC *pc);
The user can then directly set preconditioner options; alternatively, the usual runtime options can
be employed via -pc_xxx.
Finally, one can set the Sundials tolerances via
TSSundialsSetTolerance(TS ts,double abs,double rel);
where abs denotes the absolute tolerance and rel the relative tolerance.
Other PETSc-Sundials options include
TSSundialsSetGramSchmidtType(TS ts,TSSundialsGramSchmidtType type);
3bitbucket.org/jedbrown/tchem
4en.wikipedia.org/wiki/CHEMKIN
153
![](asset-9a.png)
6.10. USING SUNDIALS FROM PETSC PETSc 3.8 March 24, 2018
where type is either SUNDIALS_MODIFIED_GS or SUNDIALS_UNMODIFIED_GS. This may be set via
the options data base with -ts_sundials_gramschmidt_type <modifed,unmodified>.
The routine
TSSundialsSetMaxl(TS ts,PetscInt restart);
sets the number of vectors in the Krylov subpspace used by GMRES. This may be set in the options
database with -ts_sundials_maxl maxl.
154
![](asset-9b.png)
Chapter 7
Performing sensitivity analysis
The TS library provides a framework based on discrete adjoint models for sensitivity analysis for
ODEs and DAEs. The ODE/DAE solution process (henceforth called the forward run) can be
obtained by using either explicit or implicit solvers in TS, depending on the problem properties.
Currently supported method types are TSRK (Runge-Kutta) explicit methods and TSTHETA implicit
methods, which include TSBEULER and TSCN.
7.1 Using the discrete adjoint methods
Consider the ODE/DAE
F(t, y, ˙y, p)=0, y(t0) = y0(p)t0≤t≤tF
and the cost function(s)
Ψi(y0, p) = Φi(yF, p) + ZtF
t0
ri(y(t), p, t)dt i = 1, ..., ncost.
The TSAdjoint routines of PETSc provide
∂Ψi
∂y0
=λi
and ∂Ψi
∂p =µi+λi(∂y0
∂p ).
To perform the discrete adjoint sensitivity analysis one first sets up the TS object for a regular
forward run but with one extra function call
TSSetSaveTrajectory(TS ts),
then calls TSSolve() in the usual manner.
One must create two arrays of ncost vectors λandµ(if there are no parameters pthen one can
use NULL for the µarray.) The λvectors are the same dimension and parallel layout as the solution
vector for the ODE, the mu vectors are of dimension p; when pis small usually all its elements
are on the first MPI process, while the vectors have no entries on the other processes. λiand mui
should be initialized with the values dΦi/dy|t=tFand dΦi/dp|t=tFrespectively. Then one calls
TSSetCostGradients(TS ts,PetscInt numcost, Vec *lambda,Vec *mu);
155
![](asset-9c.png)
7.1. USING THE DISCRETE ADJOINT METHODS PETSc 3.8 March 24, 2018
If F() is a function of pone needs to also provide the Jacobian −Fpwith
TSAdjointSetRHSJacobian(TS ts,Mat Amat,PetscErrorCode
(*fp)(TS,PetscReal,Vec,Mat,void*),void *ctx)
The arguments for the function fp() are the timestep context, current time, y, and the (optional)
user-provided context.
If there is an integral term in the cost function, i.e. r() is nonzero, one must also call
TSSetCostIntegrand(TS ts,PetscInt numcost,Vec costintegral,
PetscErrorCode (*rf)(TS,PetscReal,Vec,Vec,void*),
PetscErrorCode (*drdyf)(TS,PetscReal,Vec,Vec*,void*),
PetscErrorCode (*drdpf)(TS,PetscReal,Vec,Vec*,void*),void
*ctx)
where drdyf = dr/dy, drdpf = dr/dp.
Lastly, one starts the backward run by calling
TSAdjointSolve(TS ts).
One can get the value of the integral term by calling
TSGetCostIntegral(TS ts,Vec *q).
If TSSetCostGradients() and TSSetCostIntegrand() are called before TSSolve(), the integral
term will be evaluated in the forward run. Otherwise, the computation will take place in the
backward run (inside TSAdjointSolve()). Note that this allows the evaluation of the integral
term in the cost function without having to run the adjoint solvers.
To provide a better understanding of the use of the adjoint solvers, we introduce a simple
example, corresponding to ${PETSC_DIR}/src/ts/examples/tutorials/power_grid/ex3adj.c.
The problem is to study dynamic security of power system when there are credible contingencies
such as short-circuits or loss of generators, transmission lines, or loads. The dynamic security
constraints are incorporated as equality constraints in the form of discretized differential equations
and inequality constraints for bounds on the trajectory. The governing ODE system is
φ0=ωB(ω−ωS)
2H/ωSω0=pm−pmaxsin(φ)−D(ω−ωS), t0≤t≤tF,
where φis the phase angle and ωis the frequency.
The initial conditions at time t0are
φ(t0) = arcsin (pm/pmax),
w(t0)=1.
pmax is a positive number when the system operates normally. At an event such as fault inci-
dence/removal, pmax will change to 0 temporarily and back to the original value after the fault is
fixed. The objective is to maximize pmsubject to the above ODE constraints and φ < φSduring
all times. To accommodate the inequality constraint, we want to compute the sensitivity of the
cost function
Ψ(pm, φ) = −pm+cZtF
t0
(max(0, φ −φS))2dt
with respect to the parameter pm.numcost is 1 since it is a scalar function.
156
![](asset-9d.png)
CHAPTER 7. PERFORMING SENSITIVITY ANALYSIS PETSc 3.8 March 24, 2018
For ODE solution, PETSc requires user-provided functions to evaluate the system F(t, y, ˙y, p)
(set by TSSetIFunction() ) and its corresponding Jacobian Fy+σF ˙y(set by TSSetIJacobian()).
Note that the solution state yis [φ ω]There. For sensitivity analysis, we need to provide a routine
to compute fp= [0 1]Tusing TSAdjointSetRHSJacobian(), and three routines corresponding
to the integrand r=c(max(0, φ −φS))2,rp= [0 0]Tand ry= [2c(max(0, φ −φS)) 0]Tusing
TSSetCostIntegrand().
In the adjoint run, λand µare initialized as [0 0]Tand [−1] at the final time tF. After
TSAdjointSolve(), the sensitivity of the cost function w.r.t. to initial conditions is given by the
sensitivity variable λ(at time t0) directly. And the sensitivity of the cost function w.r.t. to the
parameter pmcan be computed (by users) as
dΨ
dpm
=µ(t0) + λ(t0)d [φ(t0)ω(t0)]T
dpm
.
For explicit methods where one does not need to provide the Jacobian Fufor the forward solve
one still does need it for the backward solve and thus must call
TSSetRHSJacobian(TS ts,Mat Amat, Mat Pmat,PetscErrorCode
(*f)(TS,PetscReal,Vec,Mat,Mat,void*),void *fP);
Examples include:
•a discrete adjoint sensitivity using explicit time stepping methods ${PETSC_DIR}/src/ts/
examples/tutorials/ex16adj.c,
•a discrete adjoint sensitivity using implicit time stepping methods ${PETSC_DIR}/src/ts/
examples/tutorials/ex20adj.c,
•an optimization using the discrete adjoint models of ERK ${PETSC_DIR}/src/ts/examples/
tutorials/ex16opt_ic.c and ${PETSC_DIR}/src/ts/examples/tutorials/ex16opt_p.c,
•an optimization using the discrete adjoint models of Theta methods for stiff DAEs ${PETSC_
DIR}/src/ts/examples/tutorials/ex20opt_ic.c and ${PETSC_DIR}/src/ts/examples/
tutorials/ex20opt_p.c,
•an ODE-constrained optimization using the discrete adjoint models of Theta methods for cost
function with an integral term ${PETSC_DIR}/src/ts/examples/tutorials/power_grid/
ex3opt.c,
•a discrete adjoint sensitivity using TSCN (Crank-Nicolson) methods for DAEs with discontinu-
ities ${PETSC_DIR}/src/ts/examples/tutorials/power_grid/stability_9bus/ex9busadj.
c,
•a DAE-constrained optimization using the discrete adjoint models of TSCN (Crank-Nicolson)
methods for cost function with an integral term ${PETSC_DIR}/src/ts/examples/tutorials/
power_grid/stability_9bus/ex9busopt.c,
•a discrete adjoint sensitivity using TSCN methods for a PDE problem ${PETSC_DIR}/src/ts/
examples/tutorials/advection-diffusion-reaction/ex5adj.c.
157
![](asset-9e.png)
7.2. CHECKPOINTING PETSc 3.8 March 24, 2018
7.2 Checkpointing
The discrete adjoint model requires the states (and stage values in the context of multistage
timestepping methods) to evaluate the Jacobian matrices during the adjoint (backward) run. By
default, PETSc stores the whole trajectory to disk as binary files, each of which contains the infor-
mation for a single time step including state, time, and stage values (optional). One can also make
PETSc store the trajectory to memory with the option -ts_trajectory_type memory. However,
there might not be sufficient memory capacity especially for large-scale problems and long-time
integration.
A so-called checkpointing scheme is needed to solve this problem. The scheme stores check-
points at selective time steps and recomputes the missing information. The revolve library is used
by PETSc TSTrajectory to generate an optimal checkpointing schedule that minimizes the recom-
putations given a limited number of available checkpoints. One can specify the number of available
checkpoints with the option -ts_trajectory_max_cps_ram [maximum number of checkpoints
in RAM]. Note that one checkpoint corresponds to one time step.
The revolve library also provides an optimal multistage checkpointing scheme that uses both
RAM and disk for storage. This scheme is automatically chosen if one uses both the option -
ts_trajectory_max_cps_ram [maximum number of checkpoints in RAM] and the option -ts_
trajectory_max_cps_disk [maximum number of checkpoints on disk].
Some other useful options are listed below.
•-ts_trajectory_view prints the total number of recomputations,
•-ts_monitor and -ts_adjoint_monitor allow users to monitor the progress of the adjoint
work flow,
•-ts_trajectory_type visualization may be used to save the whole trajectory for visual-
ization. It stores the solution and the time, but no stage values. The binary files generated
can be read into MATLAB via the script ${PETSC_DIR}/share/petsc/matlab/PetscReadB
inaryTrajectory.m.
158
![](asset-9f.png)
Chapter 8
Solving Steady-State Problems with
Pseudo-Timestepping
Simple Example: TS provides a general code for performing pseudo timestepping with a variable
timestep at each physical node point. For example, instead of directly attacking the steady-state
problem
G(u)=0,
we can use pseudo-transient continuation by solving
ut=G(u).
Using time differencing
ut.
=un+1 −un
dtn
with the backward Euler method, we obtain nonlinear equations at a series of pseudo-timesteps
1
dtnB(un+1 −un) = G(un+1).
For this problem the user must provide G(u), the time steps dtnand the left-hand-side matrix B(or
optionally, if the timestep is position independent and Bis the identity matrix, a scalar timestep),
as well as optionally the Jacobian of G(u).
More generally, this can be applied to implicit ODE and DAE for which the transient form is
F(u, ˙u)=0.
For solving steady-state problems with pseudo-timestepping one proceeds as follows.
•Provide the function G(u) with the routine
TSSetRHSFunction(TS ts,PetscErrorCode
(*f)(TS,PetscReal,Vec,Vec,void*),void *fP);
The arguments to the function f() are the timestep context, the current time, the input for
the function, the output for the function and the (optional) user-provided context variable
fP.
•Provide the (approximate) Jacobian matrix of G(u) and a function to compute it at each
Newton iteration. This is done with the command
159
![](asset-a0.png)
PETSc 3.8 March 24, 2018
TSSetRHSJacobian(TS ts,Mat Amat, Mat Pmat,PetscErrorCode
(*f)(TS,PetscReal,Vec,Mat,Mat,void*),void *fP);
The arguments for the function f() are the timestep context, the current time, the location
where the Jacobian is to be computed, the (approximate) Jacobian matrix, an alternative
approximate Jacobian matrix used to construct the preconditioner, and the optional user-
provided context, passed in as fP. The user must provide the Jacobian as a matrix; thus, if
using a matrix-free approach, one must create a MATSHELL matrix.
In addition, the user must provide a routine that computes the pseudo-timestep. This is slightly
different depending on if one is using a constant timestep over the entire grid, or it varies with
location.
•For location-independent pseudo-timestepping, one uses the routine
TSPseudoSetTimeStep(TS ts,PetscInt(*dt)(TS,PetscReal*,void*),void* dtctx);
The function dt is a user-provided function that computes the next pseudo-timestep. As a
default one can use TSPseudoTimeStepDefault(TS,PetscReal*,void*) for dt. This routine
updates the pseudo-timestep with one of two strategies: the default
dtn=dtincrement ∗dtn−1∗||F(un−1)||
||F(un)||
or, the alternative,
dtn=dtincrement ∗dt0∗||F(u0)||
||F(un)||
which can be set with the call
TSPseudoIncrementDtFromInitialDt(TS ts);
or the option -ts_pseudo_increment_dt_from_initial_dt. The value dtincrement is by de-
fault 1.1, but can be reset with the call
TSPseudoSetTimeStepIncrement(TS ts,PetscReal inc);
or the option -ts_pseudo_increment <inc>.
•For location-dependent pseudo-timestepping, the interface function has not yet been created.
160
![](asset-a2.png)
PETSc 3.8 March 24, 2018
162
![](asset-a3.png)
Chapter 10
DMPlex: Unstructured Grids in
PETSc
This chapter introduces the DMPLEX subclass of DM, which allows the user to handle unstructured
grids using the generic DM interface for hierarchy and multi-physics. DMPlex was created to remedy
a huge problem in all current PDE simulation codes, namely that the discretization was so closely
tied to the data layout and solver that switching discretizations in the same code was not possible.
Not only does this preclude the kind of comparison that is necessary for scientific investigation, but
it makes library (as opposed to monolithic application) development impossible.
10.1 Representing Unstructured Grids
The main advantage of DMPlex in representing topology is that it treats all the different pieces of
a mesh, e.g. cells, faces, edges, and vertices, in exactly the same way. This allows the interface
to be very small and simple, while remaining flexible and general. This also allows “dimension
independent programming”, which means that the same algorithm can be used unchanged for
meshes of different shapes and dimensions.
All pieces of the mesh are treated as points, which are identified by PetscInts. A mesh is built
by relating points to other points, in particular specifying a “covering” relation among the points.
For example, an edge is defined by being covered by two vertices, and a triangle can be defined by
being covered by three edges (or even by three vertices). In fact, this structure has been known for
a long time. It is a Hasse Diagram [37], which is a Directed Acyclic Graph (DAG) representing a
cell complex using the covering relation. The graph edges represent the relation, which also encodes
a partially ordered set (poset).
For example, we can encode the doublet mesh in Figure 14, which can also be represented as
the DAG in Figure 15. To use the PETSc API, we first consecutively number the mesh pieces.
The PETSc convention in 3 dimensions is to number first cells, then vertices, then faces, and then
edges. In 2 dimensions the convention is to number faces, vertices, and then edges. The user is free
to violate these conventions. In terms of the labels in Figure 14, these numberings are
f07→ 0, f17→ 1, v07→ 2, v17→ 3, v27→ 4, v37→ 5, e07→ 6, e17→ 7, e27→ 8, e37→ 9, e47→ 10
First, we declare the set of points present in a mesh,
DMPlexSetChart(dm, 0, 11);
163
![](asset-a4.png)
10.1. REPRESENTING UNSTRUCTURED GRIDS PETSc 3.8 March 24, 2018
v0
v1
v2
v3
e0
e1
e2e3
e4
f0f1
Figure 14: A 2D doublet mesh, two triangles sharing an edge.
v0v1v2v3
e0e1e2e3e4
f0f1
Figure 15: The Hasse diagram for our 2D doublet mesh, expressed as a DAG.
Note that a chart here corresponds to a semi-closed interval (e.g [0,11) = {0,1,...,10}) specifying
the range of indices we’d like to use to define points on the current rank. We then define the
covering relation, which we call the cone, which are also the in-edges in the DAG. In order to
preallocate correctly, we first setup sizes,
DMPlexSetConeSize(dm, 0, 3);
DMPlexSetConeSize(dm, 1, 3);
DMPlexSetConeSize(dm, 6, 2);
DMPlexSetConeSize(dm, 7, 2);
DMPlexSetConeSize(dm, 8, 2);
DMPlexSetConeSize(dm, 9, 2);
DMPlexSetConeSize(dm, 10, 2);
DMSetUp(dm);
and then point values,
DMPlexSetCone(dm, 0, [6, 7, 8]);
DMPlexSetCone(dm, 1, [7, 9, 10]);
DMPlexSetCone(dm, 6, [2, 3]);
DMPlexSetCone(dm, 7, [3, 4]);
DMPlexSetCone(dm, 8, [4, 2]);
DMPlexSetCone(dm, 9, [4, 5]);
DMPlexSetCone(dm, 10, [5, 3]);
There is also an API for the dual relation, using DMPlexSetSupportSize() and
DMPlexSetSupport(), but this can be calculated automatically by calling
164
![](asset-a5.png)
CHAPTER 10. DMPLEX: UNSTRUCTURED GRIDS IN PETSC PETSc 3.8 March 24, 2018
DMPlexSymmetrize(dm);
In order to support efficient queries, we also want to construct fast search structures and indices
for the different types of points, which is done using
DMPlexStratify(dm);
10.2 Data on Unstructured Grids
The strongest links between solvers and discretizations are
•the layout of data over the mesh,
•problem partitioning, and
•ordering of unknowns.
To enable modularity, we encode the operations above in simple data structures that can be un-
derstood by the linear algebra engine in PETSc without any reference to the mesh (topology) or
discretization (analysis).
10.2.1 Data Layout
Data is associated with a mesh using the PetscSection object. A PetscSection can be thought
of as a generalization of PetscLayout, in the same way that a fiber bundle is a generalization of the
normal Euclidean basis used in linear algebra. With PetscLayout, we associate a unit vector (ei)
with every point in the space, and just divide up points between processes. Using PetscSection,
we can associate a set of dofs, a small space {ek}, with every point, and though our points must
be contiguous like PetscLayout, they can be in any range [pStart,pEnd).
The sequence for setting up any PetscSection is the following:
1. Specify the chart,
2. Specify the number of dofs per point, and
3. Set up the PetscSection.
For example, using the mesh from Figure 14, we can lay out data for a continuous Galerkin P3
finite element method,
PetscInt pStart, pEnd, cStart, cEnd, c, vStart, vEnd, v, eStart, eEnd, e;
DMPlexGetChart(dm, &pStart, &pEnd);
DMPlexGetHeightStratum(dm, 0, &cStart, &cEnd);
DMPlexGetHeightStratum(dm, 1, &eStart, &eEnd);
DMPlexGetDepthStratum(dm, 0, &vStart, &vEnd);
PetscSectionSetChart(s, pStart, pEnd);
for(c = cStart; c < cEnd; ++c)
PetscSectionSetDof(s, c, 1);
for(v = vStart; v < vEnd; ++v)
PetscSectionSetDof(s, v, 1);
for(e = eStart; e < eEnd; ++e)
165
![](asset-a6.png)
10.3. EVALUATING RESIDUALS PETSc 3.8 March 24, 2018
PetscSectionSetDof(s, e, 2);
PetscSectionSetUp(s);
Now a PETSc local vector can be created manually using this layout,
PetscSectionGetStorageSize(s, &n);
VecSetSizes(localVec, n, PETSC DETERMINE);
VecSetFromOptions(localVec);
though it is usually easier to use the DM directly, which also provides global vectors,
DMSetDefaultSection(dm, s);
DMGetLocalVector(dm, &localVec);
DMGetGlobalVector(dm, &globalVec);
10.2.2 Partitioning and Ordering
In exactly the same way as in MatPartitioning or with MatGetOrdering(), the results of a
partition or order are encoded in an IS. However, the graph is not the adjacency graph of the
problem Jacobian, but the mesh itself. Once the mesh is partitioned and reordered, the data layout
from a PetscSection can be used to automatically derive a problem partitioning/ordering.
10.3 Evaluating Residuals
The evaluation of a residual or Jacobian, for most discretizations has the following general form:
•Traverse the mesh, picking out pieces which in general overlap,
•Extract some values from the solution vector, associated with this piece,
•Calculate some values for the piece, and
•Insert these values into the residual vector
DMPlex separates these different concerns by passing sets of points, which are just PetscInts, from
mesh traversal routines to data extraction routines and back. In this way, the PetscSection which
structures the data inside a Vec does not need to know anything about the mesh inside a DMPlex.
The most common mesh traversal is the transitive closure of a point, which is exactly the
transitive closure of a point in the DAG using the covering relation. Note that this closure can be
calculated orienting the arrows in either direction. For example, in a finite element calculation, we
calculate an integral over each element, and then sum up the contributions to the basis function
coefficients. The closure of the element can be expressed discretely as the transitive closure of the
element point in the mesh DAG, where each point also has an orientation. Then we can retrieve
the data using PetscSection methods,
PetscScalar *a;
PetscInt numPoints, *points = NULL, p;
VecGetArray(u,&a);
DMPlexGetTransitiveClosure(dm,cell,PETSC TRUE,&numPoints,&points);
for (p = 0; p <= numPoints*2; p += 2) {
PetscInt dof, off, d;
166
![](asset-a7.png)
CHAPTER 10. DMPLEX: UNSTRUCTURED GRIDS IN PETSC PETSc 3.8 March 24, 2018
PetscSectionGetDof(section, points[p], &dof);
PetscSectionGetOffset(section, points[p], &off);
for=(d = 0; d <= dof; ++d) {
myfunc(a[off+d]);
}
}
DMPlexRestoreTransitiveClosure(dm, cell, PETSC TRUE, &numPoints, &points);
VecRestoreArray(u, &a);
This operation is so common that we have built a convenience method around it which returns the
values in a contiguous array, correctly taking into account the orientations of various mesh points:
const PetscScalar *values;
PetscInt csize;
DMPlexVecGetClosure(dm, section, u, cell, &csize, &values);
/* Do integral in quadrature loop */
DMPlexVecRestoreClosure(dm, section, u, cell, &csize, &values);
DMPlexVecSetClosure(dm, section, residual, cell, &r, ADD VALUES);
A simple example of this kind of calculation is in DMPlexComputeL2Diff_Plex(). Note that there
is no restriction on the type of cell or dimension of the mesh in the code above, so it will work
for polyhedral cells, hybrid meshes, and meshes of any dimension, without change. We can also
reverse the covering relation, so that the code works for finite volume methods where we want the
data from neighboring cells for each face:
PetscScalar *a;
PetscInt points[2*2], numPoints, p, dofA, offA, dofB, offB;
VecGetArray(u, &a);
DMPlexGetTransitiveClosure(dm, cell, PETSC FALSE, &numPoints, &points);
assert(numPoints == 2);
PetscSectionGetDof(section, points[0*2], &dofA);
PetscSectionGetDof(section, points[1*2], &dofB);
assert(dofA == dofB);
PetscSectionGetOffset(section, points[0*2], &offA);
PetscSectionGetOffset(section, points[1*2], &offB);
myfunc(a[offA], a[offB]);
VecRestoreArray(u, &a);
This kind of calculation is used in TS $PETSC_DIR/src/ts/examples/tutorials/ex11.c.
10.4 Networks
Built on top of DMPlex, the DMNetwork subclass provides abstractions for representing general
unstructured networks such as communication networks, power grid, computer networks, trans-
portation networks, electrical circuits, graphs, and others.
167
![](asset-a8.png)
10.4. NETWORKS PETSc 3.8 March 24, 2018
10.4.1 Application flow
The general flow of an application code using DMNetwork is as follows:
1. Create a network object and a “component” library:
DMNetworkCreate(MPI Comm,DM*);
creates an empty network object. A “component” is specific application data at a node/edge
of the network required for its residual evaluation. For example, components could be resistor,
inductor data for circuit applications, edge weights for graph problems, generator/transmis-
sion line data for power grids. Components are registered by calling
DMNetworkRegisterComponent(DM,const char* name, PetscInt size, PetscInt*
compkey);
Here, name is the component name,size is the size of component data type, and compkey is
an integer key that can be used for setting/getting the component at a node or an edge.
2. Set network size (number of nodes, edges), edge connectivity.
3. Set the bare layout (graph) of the network
DMNetworkLayoutSetUp(DM dm);
DMNetworkSetSizes(DM dm, PetscInt nnodes, PetscInt nedges, PetscInt Nnodes,
PetscInt Nedges);
DMNetworkSetEdgeList(DM dm, int edgeconns[]);
4. Set components and number of variables for nodes/edges.
DMNetworkAddComponent(DM dm, PetscInt p,PetscInt component,void* component);
Multiple components can be added at a node/edge.
DMNetworkSetNumVariables(DM dm,PetscInt p,PetscInt nvar);
DMNetworkAddNumVariables(DM dm,PetscInt p,PetscInt nvar);
5. Signal network ready to be distributed.
DMSetUp(DM dm);
6. Distribute the network (also moves components attached with nodes/edges)
DMNetworkDistribute(DM oldDM, const char partitioner[], PetscInt overlap,DM
*distDM);
7. Hook up the DM with the solver: KSPSetDM(),SNESSetDM(),TSSetDM()
168
![](asset-a9.png)
CHAPTER 10. DMPLEX: UNSTRUCTURED GRIDS IN PETSC PETSc 3.8 March 24, 2018
10.4.2 Utility functions
DMNetwork provides functions for obtaining iterators for nodes/edges, checking the “ghost” status
of a node (vertex), and retrieving local/global indices of node/edge variables for inserting elements
in vectors/matrices.
DMNetworkGetEdgeRange(DM dm,PetscInt *eStart,PetscInt *eEnd);
DMNetworkGetVertexRange(DM dm,PetscInt *vStart, PetscInt *vEnd);
DMNetworkIsGhostVertex(DM dm,PetscInt p,PetscBool *isghost);
DMNetworkGetVariableOffset(DM dm,PetscInt p,PetscInt *offset);
DMNetworkGetVariableGlobalOffset(DM dm,PetscInt p,PetscInt *offsetg);
In network applications, one frequently needs to find the supporting edges for a node or the con-
necting nodes covering an edge. These can be obtained by the following two routines.
DMNetworkGetConnectedVertices(DM dm,PetscInt edge,const PetscInt *vertices[]);
DMNetworkGetSupportingEdges(DM dm,PetscInt vertex,PetscInt *nedges,const
PetscInt *edges[]) ;
10.4.3 Retrieving components
The components set at nodes/edges are stored in a container that can be accessed by
DMNetworkGetComponentDataArray(DM dm,DMNetworkComponentGenericDataType
**componentdataarray);
Using componentdataarray, individual components set at a node/edge can be retrieved by
DMNetworkGetComponentKeyOffset(DM dm,PetscInt p, PetscInt compnum, PetscInt
*compkey, PetscInt *offset);
compkey is the key set by DMNetworkRegisterComponent. An example of accessing and retrieving
the components at nodes is:
DMNetworkComponentGenericDataType *arr;
PetscInt Start, End, numcomps,offset,key,v,compnum;
DMNetworkGetComponentDataArray(dm, &arr);
DMNetworkGetVertexRange(dm, &Start, &End);
for(v=Start; v < End; v++) {
DMNetworkGetNumComponents(dm,v, &numcomps);
for(compnum=0; compnum < numcomps;compnum++) {
DMNetworkGetComponentKeyOffset(dm,v,compnum, &key, &offset);
compdata = (UserCompDataType)(arr+offset);
}
}
169
![](asset-aa.png)
10.4. NETWORKS PETSc 3.8 March 24, 2018
170
Part III
Additional Information
171
![](asset-ad.png)
Chapter 11
PETSc for Fortran Users
Most of the functionality of PETSc can be obtained by people who program purely in Fortran.
11.1 C vs. Fortran Interfaces
Only a few differences exist between the C and Fortran PETSc interfaces, are due to Fortran
syntax differences. All Fortran routines have the same names as the corresponding C versions, and
PETSc command line options are fully supported. The routine arguments follow the usual Fortran
conventions; the user need not worry about passing pointers or values. The calling sequences for the
Fortran version are in most cases identical to the C version, except for the error checking variable
discussed in Section 11.1.2 and a few routines listed in Section 11.1.9.
11.1.1 Include Files
The Fortran include files for PETSc are located in the directory ${PETSC_DIR}/include/petsc/
finclude and should be used via statements such as the following:
#include <petsc/finclude/petscXXX.h>
for example,
#include <petsc/finclude/petscksp.h>
You must also use the appropriate Fortran module which is done with
use petscXXX
for example,
use petscksp
11.1.2 Error Checking
In the Fortran version, each PETSc routine has as its final argument an integer error variable, in
contrast to the C convention of providing the error variable as the routine’s return value. The error
code is set to be nonzero if an error has been detected; otherwise, it is zero. For example, the
Fortran and C variants of KSPSolve() are given, respectively, below, where ierr denotes the error
variable:
173
![](asset-ae.png)
11.1. C VS. FORTRAN INTERFACES PETSc 3.8 March 24, 2018
call KSPSolve(ksp,b,x,ierr) ! Fortran
ierr = KSPSolve(ksp,b,x); /* C */
Fortran programmers can check these error codes with CHKERRQ(ierr), which terminates all
processes when an error is encountered. Likewise, one can set error codes within Fortran programs
by using SETERRQ(comm,p,’ ’,ierr), which again terminates all processes upon detection of an
error. Note that complete error tracebacks with CHKERRQ() and SETERRQ(), as described in Section
1.5 for C routines, are not directly supported for Fortran routines; however, Fortran programmers
can easily use the error codes in writing their own tracebacks. For example, one could use code
such as the following:
call KSPSolve(ksp,b,x,ierr)
if (ierr .ne. 0) then
print*, ’Error in routine ...’
return
end if
11.1.3 Calling Fortran Routines from C (and C Routines from Fortran)
Different machines have different methods of naming Fortran routines called from C (or C routines
called from Fortran). Most Fortran compilers change all the capital letters in Fortran routines to
lowercase. On some machines, the Fortran compiler appends an underscore to the end of each
Fortran routine name; for example, the Fortran routine Dabsc() would be called from C with
dabsc_(). Other machines change all the letters in Fortran routine names to capitals.
PETSc provides two macros (defined in C/C++) to help write portable code that mixes C/C++
and Fortran. They are PETSC_HAVE_FORTRAN_UNDERSCORE and PETSC_HAVE_FORTRAN_CAPS , which
are defined in the file ${PETSC_DIR}/${PETSC_ARCH}/include/petscconf.h. The macros are
used, for example, as follows:
#if defined(PETSC_HAVE_FORTRAN_CAPS)
#define dabsc_ DMDABSC
#elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
#define dabsc_ dabsc
#endif
.....
dabsc_( &n,x,y); /* call the Fortran function */
11.1.4 Passing Null Pointers
In several PETSc C functions, one has the option of passing a NULL (0) argument (for example, the
fifth argument of MatCreateSeqAIJ()). From Fortran, users must pass PETSC_NULL_XXX to indicate
a null argument (where XXX is INTEGER,DOUBLE,CHARACTER, or SCALAR depending on the type of
argument required); passing 0 from Fortran will crash the code. Note that the C convention of
passing NULL (or 0) cannot be used. For example, when no options prefix is desired in the routine
PetscOptionsGetInt(), one must use the following command in Fortran:
call PetscOptionsGetInt(PETSC_NULL_OPTIONS,PETSC_NULL_CHARACTER,
PETSC_NULL_CHARACTER,’-name’,N,flg,ierr)
174
![](asset-af.png)
CHAPTER 11. PETSC FOR FORTRAN USERS PETSc 3.8 March 24, 2018
This Fortran requirement is inconsistent with C, where the user can employ NULL for all null
arguments.
11.1.5 Duplicating Multiple Vectors
The Fortran interface to VecDuplicateVecs() differs slightly from the C/C++ variant because
Fortran does not allow conventional arrays to be returned in routine arguments. To create n
vectors of the same format as an existing vector, the user must declare a vector array, v_new of
size n. Then, after VecDuplicateVecs() has been called, v_new will contain (pointers to) the new
PETSc vector objects. When finished with the vectors, the user should destroy them by calling
VecDestroyVecs(). For example, the following code fragment duplicates v_old to form two new
vectors, v_new(1) and v_new(2).
Vec v_old, v_new(2)
PetscInt ierr
PetscScalar alpha
....
call VecDuplicateVecs(v_old,2,v_new,ierr)
alpha = 4.3
call VecSet(v_new(1),alpha,ierr)
alpha = 6.0
call VecSet(v_new(2),alpha,ierr)
....
call VecDestroyVecs(2, &v_new,ierr)
11.1.6 Matrix, Vector and IS Indices
All matrices, vectors and IS in PETSc use zero-based indexing, regardless of whether C or Fortran
is being used. The interface routines, such as MatSetValues() and VecSetValues(), always use
zero indexing. See Section 3.2 for further details.
11.1.7 Setting Routines
When a function pointer is passed as an argument to a PETSc function, such as the test in
KSPSetConvergenceTest(), it is assumed that this pointer references a routine written in the same
language as the PETSc interface function that was called. For instance, if KSPSetConvergenceTest()
is called from C, the test argument is assumed to be a C function. Likewise, if it is called from
Fortran, the test is assumed to be written in Fortran.
11.1.8 Compiling and Linking Fortran Programs
Figure 25 shows a sample makefile that can be used for PETSc programs. In this makefile, one can
compile and run a debugging version of the Fortran program ex3.F90 with the actions make ex3
and make runex3, respectively. The compilation command is restated below:
ex3: ex3.o
-${FLINKER}} -o ex3 ex3.o ${PETSC_LIB}
${RM} ex3.o
175
![](asset-b0.png)
11.2. SAMPLE FORTRAN PROGRAMS PETSc 3.8 March 24, 2018
11.1.9 Routines with Different Fortran Interfaces
The following Fortran routines differ slightly from their C counterparts; see the manual pages and
previous discussion in this chapter for details:
PetscInitialize(char *filename,int ierr)
PetscError(MPI_COMM,int err,char *message,int ierr)
VecGetArray(), MatDenseGetArray()
ISGetIndices(),
VecDuplicateVecs(), VecDestroyVecs()
PetscOptionsGetString()
The following functions are not supported in Fortran:
PetscFClose(), PetscFOpen(), PetscFPrintf(), PetscPrintf()
PetscPopErrorHandler(), PetscPushErrorHandler()
PetscInfo()
PetscSetDebugger()
VecGetArrays(), VecRestoreArrays()
PetscViewerASCIIGetPointer(), PetscViewerBinaryGetDescriptor()
PetscViewerStringOpen(), PetscViewerStringSPrintf()
PetscOptionsGetStringArray()
PETSc includes some support for direct use of Fortran90 pointers. Current routines include:
VecGetArrayF90(), VecRestoreArrayF90()
VecGetArrayReadF90(), VecRestoreArrayReadF90()
VecDuplicateVecsF90(), VecDestroyVecsF90()
DMDAVecGetArrayF90(), DMDAVecGetArrayReadF90(),
ISLocalToGlobalMappingGetIndicesF90()
MatDenseGetArrayF90(), MatDenseRestoreArrayF90()
ISGetIndicesF90(), ISRestoreIndicesF90()
See the manual pages for details and pointers to example programs. To use the routines VecGetArrayF90(),
VecRestoreArrayF90(),VecGetArrayReadF90(),VecRestoreArrayReadF90(),VecDuplicateVecsF90(),
and VecDestroyVecsF90(), one must use the Fortran90 vector include file,
#include "petsc/finclude/petscvec.h90"
Analogous include files for other libraries are petscdm.h90,petscmat.h90, and petscis.h90.
11.2 Sample Fortran Programs
Sample programs that illustrate the PETSc interface for Fortran are given in Figures 16 –19, cor-
responding to ${PETSC_DIR}/src/vec/vec/examples/tests/ex19f.F,${PETSC_DIR}/src/vec/
vec/examples/tutorials/ex4f.F,
${PETSC_DIR}/src/sys/classes/draw/examples/tests/ex5f.F, and ${PETSC_DIR}/src/snes/
examples/ex1f.F90, respectively. We also refer Fortran programmers to the C examples listed
throughout the manual, since PETSc usage within the two languages differs only slightly.
!
!
176
![](asset-b1.png)
CHAPTER 11. PETSC FOR FORTRAN USERS PETSc 3.8 March 24, 2018
program main
#include <petsc/finclude/petscvec.h>
use petscvec
implicit none
!
! This example demonstrates basic use of the PETSc Fortran interface
! to vectors.
!
PetscInt n
PetscErrorCode ierr
PetscBool flg
PetscScalar one,two,three,dot
PetscReal norm,rdot
Vec x,y,w
PetscOptions options
n = 20
one = 1.0
two = 2.0
three = 3.0
call PetscInitialize(PETSC_NULL_CHARACTER,ierr)
if (ierr .ne. 0) then
print*,’Unable to initialize PETSc’
stop
endif
call PetscOptionsCreate(options,ierr)
call PetscOptionsGetInt(options,PETSC_NULL_CHARACTER, &
& ’-n’,n,flg,ierr)
call PetscOptionsDestroy(options,ierr)
! Create a vector, then duplicate it
call VecCreate(PETSC COMM WORLD,x,ierr)
call VecSetSizes(x,PETSC DECIDE,n,ierr)
call VecSetFromOptions(x,ierr)
call VecDuplicate(x,y,ierr)
call VecDuplicate(x,w,ierr)
call VecSet(x,one,ierr)
call VecSet(y,two,ierr)
call VecDot(x,y,dot,ierr)
rdot = PetscRealPart(dot)
write(6,100) rdot
100 format(’Result of inner product ’,f10.4)
call VecScale(x,two,ierr)
call VecNorm(x,NORM 2,norm,ierr)
write(6,110) norm
177
![](asset-b2.png)
11.2. SAMPLE FORTRAN PROGRAMS PETSc 3.8 March 24, 2018
110 format(’Result of scaling ’,f10.4)
call VecCopy(x,w,ierr)
call VecNorm(w,NORM 2,norm,ierr)
write(6,120) norm
120 format(’Result of copy ’,f10.4)
call VecAXPY(y,three,x,ierr)
call VecNorm(y,NORM 2,norm,ierr)
write(6,130) norm
130 format(’Result of axpy ’,f10.4)
call VecDestroy(x,ierr)
call VecDestroy(y,ierr)
call VecDestroy(w,ierr)
call PetscFinalize(ierr)
end
Figure 16: Sample Fortran Program: Using PETSc Vectors
!
!
! Description: Illustrates the use of VecSetValues() to set
! multiple values at once; demonstrates VecGetArray().
!
!/*T
! Concepts: vectors^assembling;
! Concepts: vectors^arrays of vectors;
! Processors: 1
!T*/
! -----------------------------------------------------------------------
program main
#include <petsc/finclude/petscvec.h>
use petscvec
implicit none
!--------------------------------
! Macro definitions
!--------------------------------
!
! Macros to make clearer the process of setting values in vectors and
! getting values from vectors.
!
! - The element xx_a(ib) is element ib+1 in the vector x
! - Here we add 1 to the base array index to facilitate the use of
! conventional Fortran 1-based array indexing.
178
![](asset-b3.png)
CHAPTER 11. PETSC FOR FORTRAN USERS PETSc 3.8 March 24, 2018
!
#define xx_a(ib) xx_v(xx_i + (ib))
#define yy_a(ib) yy_v(yy_i + (ib))
!------------------------------------
! Beginning of program
!------------------------------------
PetscScalar xwork(6)
PetscScalar xx_v(1),yy_v(1)
PetscInt i,n,loc(6),isix
PetscErrorCode ierr
PetscOffset xx_i,yy_i
Vec x,y
call PetscInitialize(PETSC_NULL_CHARACTER,ierr)
if (ierr .ne. 0) then
print*,’PetscInitialize failed’
stop
endif
n=6
isix = 6
! Create initial vector and duplicate it
call VecCreateSeq(PETSC COMM SELF,n,x,ierr)
call VecDuplicate(x,y,ierr)
! Fill work arrays with vector entries and locations. Note that
! the vector indices are 0-based in PETSc (for both Fortran and
! C vectors)
do 10 i=1,n
loc(i) = i-1
xwork(i) = 10.0*real(i)
10 continue
! Set vector values. Note that we set multiple entries at once.
! Of course, usually one would create a work array that is the
! natural size for a particular problem (not one that is as long
! as the full vector).
call VecSetValues(x,isix,loc,xwork,INSERT VALUES,ierr)
! Assemble vector
call VecAssemblyBegin(x,ierr)
call VecAssemblyEnd(x,ierr)
179
![](asset-b4.png)
11.2. SAMPLE FORTRAN PROGRAMS PETSc 3.8 March 24, 2018
! View vector
call PetscObjectSetName(x, ’initial vector:’,ierr)
call VecView(x,PETSC VIEWER STDOUT SELF,ierr)
call VecCopy(x,y,ierr)
! Get a pointer to vector data.
! - For default PETSc vectors, VecGetArray() returns a pointer to
! the data array. Otherwise, the routine is implementation dependent.
! - You MUST call VecRestoreArray() when you no longer need access to
! the array.
! - Note that the Fortran interface to VecGetArray() differs from the
! C version. See the users manual for details.
call VecGetArray(x,xx_v,xx_i,ierr)
call VecGetArray(y,yy_v,yy_i,ierr)
! Modify vector data
do 30 i=1,n
xx_a(i) = 100.0*real(i)
yy_a(i) = 1000.0*real(i)
30 continue
! Restore vectors
call VecRestoreArray(x,xx_v,xx_i,ierr)
call VecRestoreArray(y,yy_v,yy_i,ierr)
! View vectors
call PetscObjectSetName(x, ’new vector 1:’,ierr)
call VecView(x,PETSC VIEWER STDOUT SELF,ierr)
call PetscObjectSetName(y, ’new vector 2:’,ierr)
call VecView(y,PETSC VIEWER STDOUT SELF,ierr)
! Free work space. All PETSc objects should be destroyed when they
! are no longer needed.
call VecDestroy(x,ierr)
call VecDestroy(y,ierr)
call PetscFinalize(ierr)
end
Figure 17: Sample Fortran Program: Setting Vector Entries and Accessing A Raw Vector Array
!
!
180
![](asset-b5.png)
CHAPTER 11. PETSC FOR FORTRAN USERS PETSc 3.8 March 24, 2018
program main
#include <petsc/finclude/petscsys.h>
#include <petsc/finclude/petscdraw.h>
use petscsys
implicit none
!
! This example demonstrates basic use of the Fortran interface for
!PetscDraw routines.
!
PetscDraw draw
PetscDrawLG lg
PetscDrawAxis axis
PetscErrorCode ierr
PetscBool flg
integer x,y,width,height
PetscScalar xd,yd
PetscReal ten
PetscInt i,n,w,h
n = 15
x = 0
y = 0
w = 400
h = 300
ten = 10.0
call PetscInitialize(PETSC_NULL_CHARACTER,ierr)
if (ierr .ne. 0) then
print*,’Unable to initialize PETSc’
stop
endif
! GetInt requires a PetscInt so have to do this ugly setting
call PetscOptionsGetInt(PETSC_NULL_OPTIONS,PETSC_NULL_CHARACTER, &
& ’-width’,w, flg,ierr)
width = int(w)
call PetscOptionsGetInt(PETSC_NULL_OPTIONS,PETSC_NULL_CHARACTER, &
& ’-height’,h,flg,ierr)
height = int(h)
call PetscOptionsGetInt(PETSC_NULL_OPTIONS,PETSC_NULL_CHARACTER, &
& ’-n’,n,flg,ierr)
call PetscDrawCreate(PETSC COMM WORLD,PETSC_NULL_CHARACTER, &
& PETSC_NULL_CHARACTER,x,y,width,height,draw,ierr)
call PetscDrawSetFromOptions(draw,ierr)
call PetscDrawLGCreate(draw,1,lg,ierr)
call PetscDrawLGGetAxis(lg,axis,ierr)
call PetscDrawAxisSetColors(axis,PETSC_DRAW_BLACK,PETSC_DRAW_RED, &
181
![](asset-b6.png)
11.2. SAMPLE FORTRAN PROGRAMS PETSc 3.8 March 24, 2018
& PETSC_DRAW_BLUE,ierr)
call PetscDrawAxisSetLabels(axis,’toplabel’,’xlabel’,’ylabel’, &
& ierr)
do 10, i=0,n-1
xd = real(i) - 5.0
yd = xd*xd
call PetscDrawLGAddPoint(lg,xd,yd,ierr)
10 continue
call PetscDrawLGSetUseMarkers(lg,PETSC TRUE,ierr)
call PetscDrawLGDraw(lg,ierr)
call PetscSleep(ten,ierr)
call PetscDrawLGDestroy(lg,ierr)
call PetscDrawDestroy(draw,ierr)
call PetscFinalize(ierr)
end
Figure 18: Sample Fortran Program: Using PETSc PetscDraw Routines
!
!
! Description: Uses the Newton method to solve a two-variable system.
!
!/*T
! Concepts: SNES^basic uniprocessor example
! Processors: 1
!T*/
!
! -----------------------------------------------------------------------
program main
#include <petsc/finclude/petsc.h>
use petsc
implicit none
!------------------------------------
! Variable declarations
!------------------------------------
!
! Variables:
! snes - nonlinear solver
! ksp - linear solver
! pc - preconditioner context
182
![](asset-b7.png)
CHAPTER 11. PETSC FOR FORTRAN USERS PETSc 3.8 March 24, 2018
! ksp - Krylov subspace method context
! x, r - solution, residual vectors
! J - Jacobian matrix
! its - iterations for convergence
!
SNES snes
PC pc
KSP ksp
Vec x,r
Mat J
SNESLineSearch linesearch
PetscErrorCode ierr
PetscInt its,i2,i20
PetscMPIInt size,rank
PetscScalar pfive
PetscReal tol
PetscBool setls
! Note: Any user-defined Fortran routines (such as FormJacobian)
! MUST be declared as external.
external FormFunction, FormJacobian, MyLineSearch
!--------------------------------------
! Macro definitions
!--------------------------------------
!
! Macros to make clearer the process of setting values in vectors and
! getting values from vectors. These vectors are used in the routines
! FormFunction() and FormJacobian().
! - The element lx_a(ib) is element ib in the vector x
!
#define lx_a(ib) lx_v(lx_i + (ib))
#define lf_a(ib) lf_v(lf_i + (ib))
!
!------------------------------------
! Beginning of program
!------------------------------------
call PetscInitialize(PETSC_NULL_CHARACTER,ierr)
if (ierr .ne. 0) then
print*,’Unable to initialize PETSc’
stop
endif
call MPI Comm size(PETSC COMM WORLD,size,ierr)
call MPI Comm rank(PETSC COMM WORLD,rank,ierr)
if (size .ne. 1) then
if (rank .eq. 0) then
write(6,*) ’This is a uniprocessor example only!’
183
![](asset-b8.png)
11.2. SAMPLE FORTRAN PROGRAMS PETSc 3.8 March 24, 2018
endif
SETERRA(PETSC COMM SELF,1,’’)
endif
i2 = 2
i20 = 20
!------------------------------------
! Create nonlinear solver context
!------------------------------------
call SNESCreate(PETSC COMM WORLD,snes,ierr)
!------------------------------------
! Create matrix and vector data structures; set corresponding routines
!------------------------------------
! Create vectors for solution and nonlinear function
call VecCreateSeq(PETSC COMM SELF,i2,x,ierr)
call VecDuplicate(x,r,ierr)
! Create Jacobian matrix data structure
call MatCreate(PETSC COMM SELF,J,ierr)
call MatSetSizes(J,PETSC DECIDE,PETSC DECIDE,i2,i2,ierr)
call MatSetFromOptions(J,ierr)
call MatSetUp(J,ierr)
! Set function evaluation routine and vector
call SNESSetFunction(snes,r,FormFunction,0,ierr)
! Set Jacobian matrix data structure and Jacobian evaluation routine
call SNESSetJacobian(snes,J,J,FormJacobian,0,ierr)
!------------------------------------
! Customize nonlinear solver; set runtime options
!------------------------------------
! Set linear solver defaults for this problem. By extracting the
!KSP,KSP, and PC contexts from the SNES context, we can then
! directly call any KSP,KSP, and PC routines to set various options.
call SNESGetKSP(snes,ksp,ierr)
call KSPGetPC(ksp,pc,ierr)
call PCSetType(pc,PCNONE,ierr)
tol = 1.e-4
call KSPSetTolerances(ksp,tol,PETSC_DEFAULT_REAL, &
184
![](asset-b9.png)
CHAPTER 11. PETSC FOR FORTRAN USERS PETSc 3.8 March 24, 2018
& PETSC_DEFAULT_REAL,i20,ierr)
! Set SNES/KSP/KSP/PC runtime options, e.g.,
! -snes_view -snes_monitor -ksp_type <ksp> -pc_type <pc>
! These options will override those specified above as long as
!SNESSetFromOptions() is called _after_ any other customization
! routines.
call SNESSetFromOptions(snes,ierr)
call PetscOptionsHasName(PETSC_NULL_OPTIONS,PETSC_NULL_CHARACTER, &
& ’-setls’,setls,ierr)
if (setls) then
call SNESGetLineSearch(snes, linesearch, ierr)
call SNESLineSearchSetType(linesearch, ’shell’, ierr)
call SNESLineSearchShellSetUserFunc(linesearch, MyLineSearch, &
& 0, ierr)
endif
!------------------------------------
! Evaluate initial guess; then solve nonlinear system
!------------------------------------
! Note: The user should initialize the vector, x, with the initial guess
! for the nonlinear solver prior to calling SNESSolve(). In particular,
! to employ an initial guess of zero, the user should explicitly set
! this vector to zero by calling VecSet().
pfive = 0.5
call VecSet(x,pfive,ierr)
call SNESSolve(snes,PETSC_NULL_VEC,x,ierr)
call SNESGetIterationNumber(snes,its,ierr);
if (rank .eq. 0) then
write(6,100) its
endif
100 format(’Number of SNES iterations = ’,i5)
!------------------------------------
! Free work space. All PETSc objects should be destroyed when they
! are no longer needed.
!------------------------------------
call VecDestroy(x,ierr)
call VecDestroy(r,ierr)
call MatDestroy(J,ierr)
call SNESDestroy(snes,ierr)
call PetscFinalize(ierr)
185
![](asset-ba.png)
11.2. SAMPLE FORTRAN PROGRAMS PETSc 3.8 March 24, 2018
end
!
! ------------------------------------------------------------------------
!
! FormFunction - Evaluates nonlinear function, F(x).
!
! Input Parameters:
! snes - the SNES context
! x - input vector
! dummy - optional user-defined context (not used here)
!
! Output Parameter:
! f - function vector
!
subroutine FormFunction(snes,x,f,dummy,ierr)
use petscsnes
implicit none
SNES snes
Vec x,f
PetscErrorCode ierr
integer dummy(*)
! Declarations for use with local arrays
PetscScalar lx_v(2),lf_v(2)
PetscOffset lx_i,lf_i
! Get pointers to vector data.
! - For default PETSc vectors, VecGetArray() returns a pointer to
! the data array. Otherwise, the routine is implementation dependent.
! - You MUST call VecRestoreArray() when you no longer need access to
! the array.
! - Note that the Fortran interface to VecGetArray() differs from the
! C version. See the Fortran chapter of the users manual for details.
call VecGetArrayRead(x,lx_v,lx_i,ierr)
call VecGetArray(f,lf_v,lf_i,ierr)
! Compute function
lf_a(1) = lx_a(1)*lx_a(1) &
& + lx_a(1)*lx_a(2) - 3.0
lf_a(2) = lx_a(1)*lx_a(2) &
& + lx_a(2)*lx_a(2) - 6.0
! Restore vectors
call VecRestoreArrayRead(x,lx_v,lx_i,ierr)
186
![](asset-bb.png)
CHAPTER 11. PETSC FOR FORTRAN USERS PETSc 3.8 March 24, 2018
call VecRestoreArray(f,lf_v,lf_i,ierr)
return
end
! ---------------------------------------------------------------------
!
! FormJacobian - Evaluates Jacobian matrix.
!
! Input Parameters:
! snes - the SNES context
! x - input vector
! dummy - optional user-defined context (not used here)
!
! Output Parameters:
! A - Jacobian matrix
! B - optionally different preconditioning matrix
! flag - flag indicating matrix structure
!
subroutine FormJacobian(snes,X,jac,B,dummy,ierr)
use petscsnes
implicit none
SNES snes
Vec X
Mat jac,B
PetscScalar A(4)
PetscErrorCode ierr
PetscInt idx(2),i2
integer dummy(*)
! Declarations for use with local arrays
PetscScalar lx_v(2)
PetscOffset lx_i
! Get pointer to vector data
i2 = 2
call VecGetArrayRead(x,lx_v,lx_i,ierr)
! Compute Jacobian entries and insert into matrix.
! - Since this is such a small problem, we set all entries for
! the matrix at once.
! - Note that MatSetValues() uses 0-based row and column numbers
! in Fortran as well as in C (as set here in the array idx).
idx(1) = 0
idx(2) = 1
187
![](asset-bc.png)
11.2. SAMPLE FORTRAN PROGRAMS PETSc 3.8 March 24, 2018
A(1) = 2.0*lx_a(1) + lx_a(2)
A(2) = lx_a(1)
A(3) = lx_a(2)
A(4) = lx_a(1) + 2.0*lx_a(2)
call MatSetValues(B,i2,idx,i2,idx,A,INSERT VALUES,ierr)
! Restore vector
call VecRestoreArrayRead(x,lx_v,lx_i,ierr)
! Assemble matrix
call MatAssemblyBegin(B,MAT FINAL ASSEMBLY,ierr)
call MatAssemblyEnd(B,MAT FINAL ASSEMBLY,ierr)
if (B .ne. jac) then
call MatAssemblyBegin(jac,MAT FINAL ASSEMBLY,ierr)
call MatAssemblyEnd(jac,MAT FINAL ASSEMBLY,ierr)
endif
return
end
subroutine MyLineSearch(linesearch, lctx, ierr)
use petscsnes
implicit none
SNESLineSearch linesearch
SNES snes
integer lctx
Vec x, f,g, y, w
PetscReal ynorm,gnorm,xnorm
PetscBool flag
PetscErrorCode ierr
PetscScalar mone
mone = -1.0
call SNESLineSearchGetSNES(linesearch, snes, ierr)
call SNESLineSearchGetVecs(linesearch, x, f, y, w, g, ierr)
call VecNorm(y,NORM 2,ynorm,ierr)
call VecAXPY(x,mone,y,ierr)
call SNESComputeFunction(snes,x,f,ierr)
call VecNorm(f,NORM 2,gnorm,ierr)
call VecNorm(x,NORM 2,xnorm,ierr)
call VecNorm(y,NORM 2,ynorm,ierr)
call SNESLineSearchSetNorms(linesearch, xnorm, gnorm, ynorm, &
& ierr)
flag = PETSC FALSE
188
![](asset-bd.png)
CHAPTER 11. PETSC FOR FORTRAN USERS PETSc 3.8 March 24, 2018
return
end
Figure 19: Sample Fortran Program: Using PETSc Nonlinear Solvers
11.2.1 Array Arguments
This material is no longer relevent since one should use VecGetArrayF90() and the other routines
that utilize Fortran pointers, instead of the code below, but it is included for historical reasons and
because many of the Fortran examples still utilize the old approach.
Since Fortran 77 does not allow arrays to be returned in routine arguments, all PETSc routines
that return arrays, such as VecGetArray(),MatDenseGetArray(), and ISGetIndices(), are de-
fined slightly differently in Fortran than in C. Instead of returning the array itself, these routines
accept as input a user-specified array of dimension one and return an integer index to the actual
array used for data storage within PETSc. The Fortran interface for several routines is as follows:
PetscScalar xx_v(1), aa_v(1)
PetscErrorCode ierr
PetscInt ss_v(1), dd_v(1), nloc
PetscOffset ss_i, xx_i, aa_i, dd_i
Vec x
Mat A
IS s
DM d
call VecGetArray(x,xx_v,xx_i,ierr)
call MatDenseGetArray(A,aa_v,aa_i,ierr)
call ISGetIndices(s,ss_v,ss_i,ierr)
To access array elements directly, both the user-specified array and the integer index must
then be used together. For example, the following Fortran program fragment illustrates directly
setting the values of a vector array instead of using VecSetValues(). Note the (optional) use
of the preprocessor #define statement to enable array manipulations in the conventional Fortran
manner.
#define xx_a(ib) xx_v(xx_i + (ib))
double precision xx_v(1)
PetscOffset xx_i
PetscErrorCode ierr
PetscInt i, n
Vec x
call VecGetArray(x,xx_v,xx_i,ierr)
call VecGetLocalSize(x,n,ierr)
do 10, i=1,n
xx_a(i) = 3*i + 1
10 continue
call VecRestoreArray(x,xx_v,xx_i,ierr)
Figure 17 contains an example of using VecGetArray() within a Fortran routine.
189
![](asset-be.png)
11.2. SAMPLE FORTRAN PROGRAMS PETSc 3.8 March 24, 2018
Since in this case the array is accessed directly from Fortran, indexing begins with 1, not 0
(unless the array is declared as xx_v(0:1)). This is different from the use of VecSetValues()
where, indexing always starts with 0.
Note: If using VecGetArray(),MatDenseGetArray(), or ISGetIndices(), from Fortran, the
user must not compile the Fortran code with options to check for “array entries out of bounds”
(e.g., on the IBM RS/6000 this is done with the -C compiler option, so never use the -C option
with this).
190
![](asset-bf.png)
Chapter 12
Using MATLAB with PETSc
There are three basic ways to use MATLAB with PETSc:
1. (Section 12.1) dumping files to be read into MATLAB,
2. (Section 12.2) automatically sending data from a running PETSc program to a MATLAB
process where you may interactively type MATLAB commands (or run scripts), and
3. (Section 12.3) automatically sending data back and forth between PETSc and MATLAB
where MATLAB commands are issued not interactively but from a script or the PETSc
program (this uses the MATLAB Engine).
12.1 Dumping Data for MATLAB
12.1.1 Dumping ASCII MATLAB data
One can dump PETSc matrices and vectors to the screen in an ASCII format that MATLAB
can read in directly. This is done with the command line options -vec_view ::ascii_matlab
or -mat_view ::ascii_matlab. To write a a file, use -vec_view :filename.m:ascii_matlab or
-mat_view :filename.m:ascii_matlab.
This causes the PETSc program to print the vectors and matrices every time VecAssemblyEnd()
or MatAssemblyEnd() are called. To provide finer control over when and what vectors and matri-
ces are dumped one can use the VecView() and MatView() functions with a viewer type of ASCII
(see PetscViewerASCIIOpen(),PETSC_VIEWER_STDOUT_WORLD,PETSC_VIEWER_STDOUT_SELF, or
PETSC_VIEWER_STDOUT_(MPI_Comm)). Before calling the viewer set the output type with, for ex-
ample,
PetscViewerPushFormat(PETSC VIEWER STDOUT WORLD,PETSC VIEWER ASCII MATLAB);
VecView(A,PETSC VIEWER STDOUT WORLD);
PetscViewerPopFormat(PETSC VIEWER STDOUT WORLD);
The name of each PETSc variable printed for MATLAB may be set with
PetscObjectSetName((PetscObject)A,"name");
If no name is specified, the object is given a default name using PetscObjectName().
12.1.2 Dumping Binary Data for MATLAB
One can also read PETSc binary files (see Section 15.3) directly into MATLAB via the scripts
available in $PETSC_DIR/share/matlab. This requires less disk space and is recommended for all
but the smallest data sizes. One can also use
191
![](asset-c0.png)
12.2. SENDING DATA TO AN INTERACTIVE MATLAB SESSIONPETSc 3.8 March 24, 2018
PetscViewerPushFormat(viewer,PETSC VIEWER BINARY MATLAB}
to dump both a PETSc binary file and a corresponding .info file which PetscReadBinaryMatlab.m
will use to format the binary file in more complex cases, such as using a DMDA. For an example,
see ${PETSC_DIR}/src/dm/examples/tutorials/ex7.c. In MATLAB (R2015b), one may then
generate a useful structure. For example:
setenv(’PETSC_DIR’,’~/petsc’);
setenv(’PETSC_ARCH’,’arch-darwin-double-debug’);
addpath(’~/petsc/share/petsc/matlab’);
gridData=PetscReadBinaryMatlab(’output_file’);
12.2 Sending Data to an Interactive MATLAB Session
One creates a viewer to MATLAB via
PetscViewerSocketOpen(MPI Comm,char *machine,int port,PetscViewer *v);
(port is usually set to PETSC_DEFAULT; use NULL for the machine if the MATLAB interactive session
is running on the same machine as the PETSc program) and then sends matrices or vectors via
VecView(Vec A,v);
MatView(Mat B,v);
See Section 15.3 for more on PETSc viewers. One may start the MATLAB program manually or use
the PETSc command PetscStartMatlab(MPI_Comm,char *machine,char *script,FILE **fp);
where machine and script may be NULL. It is also possible to start your PETSc program from
MATLAB via launch().
To receive the objects in MATLAB, make sure that ${PETSC_DIR}/${PETSC_ARCH}/lib/petsc/
matlab and ${PETSC_DIR}/share/petsc/matlab are in the MATLAB path. Use
p = PetscOpenSocket(); (or p = PetscOpenSocket(portnum) if you provided a port number
in your call to PetscViewerSocketOpen()), and then a=PetscBinaryRead(p); returns the ob-
ject passed from PETSc. PetscBinaryRead() may be called any number of times. Each call
should correspond on the PETSc side with viewing a single vector or matrix. close() closes
the connection from MATLAB. On the PETSc side, one should destroy the viewer object with
PetscViewerDestroy().
For an example, which includes sending data back to PETSc, see ${PETSC_DIR}/src/vec/vec/
examples/tutorials/ex42.c and the associated .m file.
12.3 Using the MATLAB Compute Engine
One creates access to the MATLAB engine via
PetscMatlabEngineCreate(MPI Comm comm,char *machine,PetscMatlabEngine *e);
where machine is the name of the machine hosting MATLAB (NULL may be used for localhost).
One can send objects to MATLAB via
PetscMatlabEnginePut(PetscMatlabEngine e,PetscObject obj);
One can get objects via
192
![](asset-c1.png)
CHAPTER 12. USING MATLAB WITH PETSC PETSc 3.8 March 24, 2018
PetscMatlabEngineGet(PetscMatlabEngine e,PetscObject obj);
Similarly, one can send arrays via
PetscMatlabEnginePutArray(PetscMatlabEngine e,int m,int n,PetscScalar
*array,char *name);
and get them back via
PetscMatlabEngineGetArray(PetscMatlabEngine e,int m,int n,PetscScalar
*array,char *name);
One cannot use MATLAB interactively in this mode but one can send MATLAB commands via
PetscMatlabEngineEvaluate(PetscMatlabEngine,"format",...);
where format has the usual printf() format. For example,
PetscMatlabEngineEvaluate(PetscMatlabEngine,"x = \%g *y + z;",avalue);
The name of each PETSc variable passed to MATLAB may be set with
PetscObjectSetName((PetscObject)A,"name");
Text responses can be returned from MATLAB via
PetscMatlabEngineGetOutput(PetscMatlabEngine,char **);
or
PetscMatlabEnginedPrintOutput(PetscMatlabEngine,FILE*).
There is a short-cut to starting the MATLAB engine with PETSC_MATLAB_ENGINE_(MPI_Comm).
193
![](asset-c2.png)
12.3. USING THE MATLAB COMPUTE ENGINE PETSc 3.8 March 24, 2018
194
![](asset-c3.png)
Chapter 13
Profiling
PETSc includes a consistent, lightweight scheme to allow the profiling of application programs.
The PETSc routines automatically log performance data if certain options are specified at runtime.
The user can also log information about application codes for a complete picture of performance.
In addition, as described in Section 13.1.1, PETSc provides a mechanism for printing informative
messages about computations. Section 13.1 introduces the various profiling options in PETSc,
while the remainder of the chapter focuses on details such as monitoring application codes and tips
for accurate profiling.
13.1 Basic Profiling Information
If an application code and the PETSc libraries have been configured with --with-log=1, the de-
fault, then various kinds of profiling of code between calls to PetscInitialize() and PetscFinalize()
can be activated at runtime. The profiling options include the following:
•-log_view - Prints an ASCII version of performance data at program’s conclusion. These
statistics are comprehensive and concise and require little overhead; thus, -log_view is in-
tended as the primary means of monitoring the performance of PETSc codes.
•-info [infofile] - Prints verbose information about code to stdout or an optional file.
This option provides details about algorithms, data structures, etc. Since the overhead of
printing such output slows a code, this option should not be used when evaluating a program’s
performance.
•-log_trace [logfile] - Traces the beginning and ending of all PETSc events. This option,
which can be used in conjunction with -info, is useful to see where a program is hanging
without running in the debugger.
As discussed in Section 13.1.3, additional profiling can be done with MPE.
13.1.1 Interpreting -log view Output: The Basics
As shown in Figure 6(in Part I), the option -log_view activates printing of profile data to standard
output at the conclusion of a program. Profiling data can also be printed at any time within a
program by calling PetscLogView().
We print performance data for each routine, organized by PETSc libraries, followed by any
user-defined events (discussed in Section 13.2). For each routine, the output data include the
195
![](asset-c4.png)
13.1. BASIC PROFILING INFORMATION PETSc 3.8 March 24, 2018
maximum time and floating point operation (flop) rate over all processes. Information about
parallel performance is also included, as discussed in the following section.
For the purpose of PETSc floating point operation counting, we define one flop as one operation
of any of the following types: multiplication, division, addition, or subtraction. For example, one
VecAXPY() operation, which computes y=αx +yfor vectors of length N, requires 2Nflop
(consisting of Nadditions and Nmultiplications). Bear in mind that flop rates present only a
limited view of performance, since memory loads and stores are the real performance barrier.
For simplicity, the remainder of this discussion focuses on interpreting profile data for the
KSP library, which provides the linear solvers at the heart of the PETSc package. Recall the
hierarchical organization of the PETSc library, as shown in Figure 1. Each KSP solver is composed
of a PC (preconditioner) and a KSP (Krylov subspace) part, which are in turn built on top of the
Mat (matrix) and Vec (vector) modules. Thus, operations in the KSP module are composed of
lower-level operations in these packages. Note also that the nonlinear solvers library, SNES, is built
on top of the KSP module, and the timestepping library, TS, is in turn built on top of SNES.
We briefly discuss interpretation of the sample output in Figure 6, which was generated by
solving a linear system on one process using restarted GMRES and ILU preconditioning. The
linear solvers in KSP consist of two basic phases, KSPSetUp() and KSPSolve(), each of which
consists of a variety of actions, depending on the particular solution technique. For the case of
using the PCILU preconditioner and KSPGMRES Krylov subspace method, the breakdown of PETSc
routines is listed below. As indicated by the levels of indentation, the operations in KSPSetUp()
include all of the operations within PCSetUp(), which in turn include MatILUFactor(), and so on.
•KSPSetUp - Set up linear solver
•PCSetUp - Set up preconditioner
•MatILUFactor - Factor preconditioning matrix
•MatILUFactorSymbolic - Symbolic factorization phase
•MatLUFactorNumeric - Numeric factorization phase
•KSPSolve - Solve linear system
•PCApply - Apply preconditioner
•MatSolve - Forward/backward triangular solves
•KSPGMRESOrthog - Orthogonalization in GMRES
•VecDot or VecMDot - Inner products
•MatMult - Matrix-vector product
•MatMultAdd - Matrix-vector product + vector addition
•VecScale,VecNorm,VecAXPY,VecCopy, ...
The summaries printed via -log_view reflect this routine hierarchy. For example, the perfor-
mance summaries for a particular high-level routine such as KSPSolve() include all of the operations
accumulated in the lower-level components that make up the routine.
Admittedly, we do not currently present the output with -log_view so that the hierarchy
of PETSc operations is completely clear, primarily because we have not determined a clean and
uniform way to do so throughout the library. Improvements may follow. However, for a particular
problem, the user should generally have an idea of the basic operations that are required for its
implementation (e.g., which operations are performed when using GMRES and ILU, as described
above), so that interpreting the -log_view data should be relatively straightforward.
196
![](asset-c5.png)
CHAPTER 13. PROFILING PETSc 3.8 March 24, 2018
mpiexec -n 4 ./ex10 -f0 medium -f1 arco6 -ksp_gmres_classicalgramschmidt -log_view -mat_type baij \
-matload_block_size 3 -pc_type bjacobi -options_left
Number of iterations = 19
Residual norm = 7.7643e-05
Number of iterations = 55
Residual norm = 6.3633e-01
---------------------------------------------- PETSc Performance Summary: ----------------------------------------------
ex10 on a rs6000 named p039 with 4 processors, by mcinnes Wed Jul 24 16:30:22 1996
Max Min Avg Total
Time (sec): 3.289e+01 1.0 3.288e+01
Objects: 1.130e+02 1.0 1.130e+02
Flop: 2.195e+08 1.0 2.187e+08 8.749e+08
Flop/sec: 6.673e+06 1.0 2.660e+07
MPI Messages: 2.205e+02 1.4 1.928e+02 7.710e+02
MPI Message Lengths: 7.862e+06 2.5 5.098e+06 2.039e+07
MPI Reductions: 1.850e+02 1.0
Summary of Stages: ---- Time ------ ----- Flop ------- -- Messages -- -- Message-lengths -- Reductions -
Avg %Total Avg %Total counts %Total avg %Total counts %Total
0: Load System 0: 1.191e+00 3.6% 3.980e+06 0.5% 3.80e+01 4.9% 6.102e+04 0.3% 1.80e+01 9.7%
1: KSPSetup 0: 6.328e-01 2.5% 1.479e+04 0.0% 0.00e+00 0.0% 0.000e+00 0.0% 0.00e+00 0.0%
2: KSPSolve 0: 2.269e-01 0.9% 1.340e+06 0.0% 1.52e+02 19.7% 9.405e+03 0.0% 3.90e+01 21.1%
3: Load System 1: 2.680e+01 107.3% 0.000e+00 0.0% 2.10e+01 2.7% 1.799e+07 88.2% 1.60e+01 8.6%
4: KSPSetup 1: 1.867e-01 0.7% 1.088e+08 2.3% 0.00e+00 0.0% 0.000e+00 0.0% 0.00e+00 0.0%
5: KSPSolve 1: 3.831e+00 15.3% 2.217e+08 97.1% 5.60e+02 72.6% 2.333e+06 11.4% 1.12e+02 60.5%
------------------------------------------------------------------------------------------------------------------------
.... [Summary of various phases, see part II below] ...
------------------------------------------------------------------------------------------------------------------------
Memory usage is given in bytes:
Object Type Creations Destructions Memory Descendants’ Mem.
Viewer 5 5 0 0
Index set 10 10 127076 0
Vector 76 76 9152040 0
Vector Scatter 2 2 106220 0
Matrix 8 8 9611488 5.59773e+06
Krylov Solver 4 4 33960 7.5966e+06
Preconditioner 4 4 16 9.49114e+06
KSP 4 4 0 1.71217e+07
Figure 20: Profiling a PETSc Program: Part I - Overall Summary
13.1.2 Interpreting -log view Output: Parallel Performance
We next discuss performance summaries for parallel programs, as shown within Figures 20 and 21,
which present the combined output generated by the -log_view option. The program that gen-
erated this data is ${PETSC_DIR}/src/ksp/ksp/examples/ex10.c. The code loads a matrix and
right-hand-side vector from a binary file and then solves the resulting linear system; the program
then repeats this process for a second linear system. This particular case was run on four processors
of an IBM SP, using restarted GMRES and the block Jacobi preconditioner, where each block was
solved with ILU.
Figure 20 presents an overall performance summary, including times, floating-point operations,
computational rates, and message-passing activity (such as the number and size of messages sent
and collective operations). Summaries for various user-defined stages of monitoring (as discussed
in Section 13.3) are also given. Information about the various phases of computation then follow
(as shown separately here in Figure 21). Finally, a summary of memory usage and object creation
and destruction is presented.
We next focus on the summaries for the various phases of the computation, as given in the table
within Figure 21. The summary for each phase presents the maximum times and flop rates over
all processes, as well as the ratio of maximum to minimum times and flop rates for all processes.
A ratio of approximately 1 indicates that computations within a given phase are well balanced
among the processes; as the ratio increases, the balance becomes increasingly poor. Also, the total
computational rate (in units of MFlop/sec) is given for each phase in the final column of the phase
197
![](asset-c6.png)
13.1. BASIC PROFILING INFORMATION PETSc 3.8 March 24, 2018
mpiexec -n 4 ./ex10 -f0 medium -f1 arco6 -ksp_gmres_classicalgramschmidt -log_view -mat_type baij \
-matload_block_size 3 -pc_type bjacobi -options_left
---------------------------------------------- PETSc Performance Summary: ----------------------------------------------
.... [Overall summary, see part I] ...
Phase summary info:
Count: number of times phase was executed
Time and Flop/sec: Max - maximum over all processors
Ratio - ratio of maximum to minimum over all processors
Mess: number of messages sent
Avg. len: average message length
Reduct: number of global reductions
Global: entire computation
Stage: optional user-defined stages of a computation. Set stages with PLogStagePush() and PLogStagePop().
%T - percent time in this phase %F - percent flop in this phase
%M - percent messages in this phase %L - percent message lengths in this phase
%R - percent reductions in this phase
Total Mflop/s: 10^6 * (sum of flop over all processors)/(max time over all processors)
------------------------------------------------------------------------------------------------------------------------
Phase Count Time (sec) Flop/sec --- Global --- --- Stage --- Total
Max Ratio Max Ratio Mess Avg len Reduct %T %F %M %L %R %T %F %M %L %R Mflop/s
------------------------------------------------------------------------------------------------------------------------
...
--- Event Stage 4: KSPSetUp 1
MatGetReordering 1 3.491e-03 1.0 0.0e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 2 0 0 0 0 0
MatILUFctrSymbol 1 6.970e-03 1.2 0.0e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 3 0 0 0 0 0
MatLUFactorNumer 1 1.829e-01 1.1 3.2e+07 1.1 0.0e+00 0.0e+00 0.0e+00 1 2 0 0 0 90 99 0 0 0 110
KSPSetUp 2 1.989e-01 1.1 2.9e+07 1.1 0.0e+00 0.0e+00 0.0e+00 1 2 0 0 0 99 99 0 0 0 102
PCSetUp 2 1.952e-01 1.1 2.9e+07 1.1 0.0e+00 0.0e+00 0.0e+00 1 2 0 0 0 97 99 0 0 0 104
PCSetUpOnBlocks 1 1.930e-01 1.1 3.0e+07 1.1 0.0e+00 0.0e+00 0.0e+00 1 2 0 0 0 96 99 0 0 0 105
--- Event Stage 5: KSPSolve 1
MatMult 56 1.199e+00 1.1 5.3e+07 1.0 1.1e+03 4.2e+03 0.0e+00 5 28 99 23 0 30 28 99 99 0 201
MatSolve 57 1.263e+00 1.0 4.7e+07 1.0 0.0e+00 0.0e+00 0.0e+00 5 27 0 0 0 33 28 0 0 0 187
VecNorm 57 1.528e-01 1.3 2.7e+07 1.3 0.0e+00 0.0e+00 2.3e+02 1 1 0 0 31 3 1 0 0 51 81
VecScale 57 3.347e-02 1.0 4.7e+07 1.0 0.0e+00 0.0e+00 0.0e+00 0 1 0 0 0 1 1 0 0 0 184
VecCopy 2 1.703e-03 1.1 0.0e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0
VecSet 3 2.098e-03 1.0 0.0e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0
VecAXPY 3 3.247e-03 1.1 5.4e+07 1.1 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 200
VecMDot 55 5.216e-01 1.2 9.8e+07 1.2 0.0e+00 0.0e+00 2.2e+02 2 20 0 0 30 12 20 0 0 49 327
VecMAXPY 57 6.997e-01 1.1 6.9e+07 1.1 0.0e+00 0.0e+00 0.0e+00 3 21 0 0 0 18 21 0 0 0 261
VecScatterBegin 56 4.534e-02 1.8 0.0e+00 0.0 1.1e+03 4.2e+03 0.0e+00 0 0 99 23 0 1 0 99 99 0 0
VecScatterEnd 56 2.095e-01 1.2 0.0e+00 0.0 0.0e+00 0.0e+00 0.0e+00 1 0 0 0 0 5 0 0 0 0 0
KSPSolve 1 3.832e+00 1.0 5.6e+07 1.0 1.1e+03 4.2e+03 4.5e+02 15 97 99 23 61 99 99 99 99 99 222
KSPGMRESOrthog 55 1.177e+00 1.1 7.9e+07 1.1 0.0e+00 0.0e+00 2.2e+02 4 39 0 0 30 29 40 0 0 49 290
PCSetUpOnBlocks 1 1.180e-05 1.1 0.0e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0
PCApply 57 1.267e+00 1.0 4.7e+07 1.0 0.0e+00 0.0e+00 0.0e+00 5 27 0 0 0 33 28 0 0 0 186
------------------------------------------------------------------------------------------------------------------------
.... [Conclusion of overall summary, see part I] ...
Figure 21: Profiling a PETSc Program: Part II - Phase Summaries
summary table.
Total Mflop/sec = 10−6∗(sum of flop over all processors)/(max time over all processors)
Note: Total computational rates <1 MFlop are listed as 0 in this column of the phase summary
table. Additional statistics for each phase include the total number of messages sent, the average
message length, and the number of global reductions.
As discussed in the preceding section, the performance summaries for higher-level PETSc rou-
tines include the statistics for the lower levels of which they are made up. For example, the
communication within matrix-vector products MatMult() consists of vector scatter operations, as
given by the routines VecScatterBegin() and VecScatterEnd().
The final data presented are the percentages of the various statistics (time (%T), flop/sec (%F),
messages(%M), average message length (%L), and reductions (%R)) for each event relative to the total
computation and to any user-defined stages (discussed in Section 13.3). These statistics can aid
in optimizing performance, since they indicate the sections of code that could benefit from various
kinds of tuning. Chapter 14 gives suggestions about achieving good performance with PETSc
codes.
198
![](asset-c7.png)
CHAPTER 13. PROFILING PETSc 3.8 March 24, 2018
13.1.3 Using -log mpe with Jumpshot
It is also possible to use the Jumpshot package [18] to visualize PETSc events. This package comes
with the MPE software, which is part of the MPICH [16] implementation of MPI. The option
-log_mpe [logfile]
creates a logfile of events appropriate for viewing with Jumpshot. The user can either use the default
logging file or specify a name via logfile. Events can be deactivated as described in Section 13.4.
The user can also log MPI events. To do this, simply consider the PETSc application as any MPI
application, and follow the MPI implementation’s instructions for logging MPI calls. For example,
when using MPICH, this merely required adding -llmpich to the library list before -lmpich.
13.2 Profiling Application Codes
PETSc automatically logs object creation, times, and floating-point counts for the library routines.
Users can easily supplement this information by monitoring their application codes as well. The
basic steps involved in logging a user-defined portion of code, called an event, are shown in the
code fragment below:
PetscLogEvent USER_EVENT;
PetscClassId classid;
PetscLogDouble user_event_flops;
PetscClassIdRegister("class name",&classid);
PetscLogEventRegister("User event name",classid,&USER_EVENT);
PetscLogEventBegin(USER_EVENT,0,0,0,0);
/* code segment to monitor */
PetscLogFlops(user_event_flops);
PetscLogEventEnd(USER_EVENT,0,0,0,0);
One must register the event by calling PetscLogEventRegister(), which assigns a unique
integer to identify the event for profiling purposes:
PetscLogEventRegister(const char string[],PetscClassId classid,PetscLogEvent *e);
Here string is a user-defined event name, and color is an optional user-defined event color (for use
with Jumpshot logging; see Section 13.1.3); one should see the manual page for details. The argu-
ment returned in eshould then be passed to the PetscLogEventBegin() and PetscLogEventEnd()
routines.
Events are logged by using the pair
PetscLogEventBegin(int event,PetscObject o1,PetscObject o2,PetscObject
o3,PetscObject o4);
PetscLogEventEnd(int event,PetscObject o1,PetscObject o2,PetscObject
o3,PetscObject o4);
The four objects are the PETSc objects that are most closely associated with the event. For
instance, in a matrix-vector product they would be the matrix and the two vectors. These objects
can be omitted by specifying 0 for o1 -o4. The code between these two routine calls will be
automatically timed and logged as part of the specified event.
The user can log the number of floating-point operations for this segment of code by calling
199
![](asset-c8.png)
13.3. PROFILING MULTIPLE SECTIONS OF CODE PETSc 3.8 March 24, 2018
PetscLogFlops(number of flop for this code segment);
between the calls to PetscLogEventBegin() and PetscLogEventEnd(). This value will automati-
cally be added to the global flop counter for the entire program.
13.3 Profiling Multiple Sections of Code
By default, the profiling produces a single set of statistics for all code between the PetscInitialize()
and PetscFinalize() calls within a program. One can independently monitor up to ten stages of
code by switching among the various stages with the commands
PetscLogStagePush(PetscLogStage stage);
PetscLogStagePop();
where stage is an integer (0-9); see the manual pages for details. The command
PetscLogStageRegister(const char *name,PetscLogStage *stage)
allows one to associate a name with a stage; these names are printed whenever summaries are
generated with -log_view or PetscLogView(). The following code fragment uses three profiling
stages within an program.
PetscInitialize(int *argc,char ***args,0,0);
/* stage 0 of code here */
PetscLogStageRegister("Stage 0 of Code", &stagenum0);
for (i=0; i<ntimes; i++) {
PetscLogStageRegister("Stage 1 of Code", &stagenum1);
PetscLogStagePush(stagenum1);
/* stage 1 of code here */
PetscLogStagePop();
PetscLogStageRegister("Stage 2 of Code", &stagenum2);
PetscLogStagePush(stagenum2);
/* stage 2 of code here */
PetscLogStagePop();
}
PetscFinalize();
Figures 20 and 21 show output generated by -log_view for a program that employs several
profiling stages. In particular, this program is subdivided into six stages: loading a matrix and right-
hand-side vector from a binary file, setting up the preconditioner, and solving the linear system; this
sequence is then repeated for a second linear system. For simplicity, Figure 21 contains output only
for stages 4 and 5 (linear solve of the second system), which comprise the part of this computation
of most interest to us in terms of performance monitoring. This code organization (solving a small
linear system followed by a larger system) enables generation of more accurate profiling statistics
for the second system by overcoming the often considerable overhead of paging, as discussed in
Section 13.8.
13.4 Restricting Event Logging
By default, all PETSc operations are logged. To enable or disable the PETSc logging of individual
events, one uses the commands
200
![](asset-c9.png)
CHAPTER 13. PROFILING PETSc 3.8 March 24, 2018
PetscLogEventActivate(int event);
PetscLogEventDeactivate(int event);
The event may be either a predefined PETSc event (as listed in the file ${PETSC_DIR}/include/
petsclog.h) or one obtained with PetscLogEventRegister() (as described in Section 13.2).
PETSc also provides routines that deactivate (or activate) logging for entire components of the
library. Currently, the components that support such logging (de)activation are Mat (matrices),
Vec (vectors), KSP (linear solvers, including KSP and PC), and SNES (nonlinear solvers):
PetscLogEventDeactivateClass(MAT_CLASSID);
PetscLogEventDeactivateClass(KSP_CLASSID); /* includes PC and KSP */
PetscLogEventDeactivateClass(VEC_CLASSID);
PetscLogEventDeactivateClass(SNES_CLASSID);
and
PetscLogEventActivateClass(MAT_CLASSID);
PetscLogEventActivateClass(KSP_CLASSID); /* includes PC and KSP */
PetscLogEventActivateClass(VEC_CLASSID);
PetscLogEventActivateClass(SNES_CLASSID);
Recall that the option -log_all produces extensive profile data, which can be a challenge for
PETScView to handle due to the memory limitations of Tcl/Tk. Thus, one should generally use
-log_all when running programs with a relatively small number of events or when disabling some
of the events that occur many times in a code (e.g., VecSetValues(),MatSetValues()).
13.5 Interpreting -log info Output: Informative Messages
Users can activate the printing of verbose information about algorithms, data structures, etc. to
the screen by using the option -info or by calling PetscInfoAllow(PETSC_TRUE). Such logging,
which is used throughout the PETSc libraries, can aid the user in understanding algorithms and
tuning program performance. For example, as discussed in Section 3.1.1,-info activates the
printing of information about memory allocation during matrix assembly.
Application programmers can employ this logging as well, by using the routine
PetscInfo(void* obj,char *message,...)
where obj is the PETSc object associated most closely with the logging statement, message. For
example, in the line search Newton methods, we use a statement such as
PetscInfo(snes,"Cubically determined step, lambda %g\n",lambda);
One can selectively turn off informative messages about any of the basic PETSc objects (e.g.,
Mat,SNES) with the command
PetscInfoDeactivateClass(int object_classid)
where object_classid is one of MAT_CLASSID,SNES_CLASSID, etc. Messages can be reactivated
with the command
PetscInfoActivateClass(int object_classid)
Such deactivation can be useful when one wishes to view information about higher-level PETSc
libraries (e.g., TS and SNES) without seeing all lower level data as well (e.g., Mat). One can deactivate
events at runtime for matrix and linear solver libraries via -info [no_mat, no_ksp].
201
![](asset-ca.png)
13.6. TIME PETSc 3.8 March 24, 2018
13.6 Time
PETSc application programmers can access the wall clock time directly with the command
PetscLogDouble time;
PetscTime(&time);CHKERRQ(ierr);
which returns the current time in seconds since the epoch, and is commonly implemented with
MPI_Wtime. A floating point number is returned in order to express fractions of a second. In
addition, as discussed in Section 13.2, PETSc can automatically profile user-defined segments of
code.
13.7 Saving Output to a File
All output from PETSc programs (including informative messages, profiling information, and con-
vergence data) can be saved to a file by using the command line option -history [filename]. If
no file name is specified, the output is stored in the file ${HOME}/.petschistory. Note that this
option only saves output printed with the PetscPrintf() and PetscFPrintf() commands, not
the standard printf() and fprintf() statements.
13.8 Accurate Profiling and Paging Overheads
One factor that often plays a significant role in profiling a code is paging by the operating system.
Generally, when running a program, only a few pages required to start it are loaded into memory
rather than the entire executable. When the execution proceeds to code segments that are not in
memory, a pagefault occurs, prompting the required pages to be loaded from the disk (a very slow
process). This activity distorts the results significantly. (The paging effects are noticeable in the
log files generated by -log_mpe, which is described in Section 13.1.3.)
To eliminate the effects of paging when profiling the performance of a program, we have found
an effective procedure is to run the exact same code on a small dummy problem before running
it on the actual problem of interest. We thus ensure that all code required by a solver is loaded
into memory during solution of the small problem. When the code proceeds to the actual (larger)
problem of interest, all required pages have already been loaded into main memory, so that the
performance numbers are not distorted.
When this procedure is used in conjunction with the user-defined stages of profiling described in
Section 13.3, we can focus easily on the problem of interest. For example, we used this technique in
the program ${PETSC_DIR}/src/ksp/ksp/examples/tutorials/ex10.c to generate the timings
within Figures 20 and 21. In this case, the profiled code of interest (solving the linear system
for the larger problem) occurs within event stages 4 and 5. Section 13.1.2 provides details about
interpreting such profiling data.
In particular, the macros
PetscPreLoadBegin(PetscBool flag,char* stagename)
PetscPreLoadStage(char *stagename)
and
PetscPreLoadEnd()
202
![](asset-cb.png)
CHAPTER 13. PROFILING PETSc 3.8 March 24, 2018
can be used to easily convert a regular PETSc program to one that uses preloading. The command
line options -preload true and -preload false may be used to turn on and off preloading at run
time for PETSc programs that use these macros.
203
![](asset-cc.png)
13.8. ACCURATE PROFILING AND PAGING OVERHEADS PETSc 3.8 March 24, 2018
204
Chapter 14
Hints for Performance Tuning
This chapter provides hints on how to get to achieve best performance with PETSc, particularly on
distributed-memory machines with multiple CPU sockets per node. We focus on machine-related
performance optimization here; algorithmic aspects like preconditioner selection are not the focus
of this section.
14.1 Maximizing Memory Bandwidth
Most operations in PETSc deal with large datasets (typically vectors and sparse matrices) and
perform relatively few arithmetic operations for each byte loaded or stored from global memory.
Therefore, the arithmetic intensity expressed as the ratio of floating point operations to the number
of bytes loaded and stored is usually well below unity for typical PETSc operations. On the other
hand, modern CPUs are able to execute on the order of 10 floating point operations for each byte
loaded or stored. As a consequence, almost all PETSc operations are limited by the rate at which
data can be loaded or stored (memory bandwidth limited ) rather than by the rate of floating point
operations.
This section discusses ways to maximize the memory bandwidth achieved by applications based
on PETSc. Where appropriate, we include benchmark results in order to provide quantitative
results on typical performance gains one can achieve through parallelization, both on a single
compute node and across nodes. In particular, we start with the answer to the common question
of why performance generally does not increase 20-fold with a 20-core CPU.
14.1.1 Memory Bandwidth vs. Processes
Consider the addition of two large vectors, with the result written to a third vector. Because
there are no dependencies across the different entries of each vector, the operation is embarrasingly
parallel.
As Fig. 22 shows, the performance gains due to parallelization on different multi- and many-
core CPUs quickly saturates. The reason is that only a fraction of the total number of CPU cores
is required to saturate the memory channels. For example, a dual-socket system equipped with
Haswell 12-core Xeon CPUs achieves more than 80 percent of achievable peak memory bandwidth
with only four processes per socket (8 total), cf. Fig. 22. Consequently, running with more than
8 MPI ranks on such a system will not increase performance substantially. For the same reason,
PETSc-based applications usually do not benefit from hyper-threading.
PETSc provides a simple way to measure memory bandwidth for different numbers of processes
via the target make streams executed from $PETSC_DIR. The output provides an overview of the
possible speedup one can obtain on the given machine (not necessarily a shared memory system).
205
![](asset-ce.png)
14.1. MAXIMIZING MEMORY BANDWIDTH PETSc 3.8 March 24, 2018
1
10
100
1000
1 10 100
Bandwidth (GB/sec)
Processes/Threads
STREAM Benchmark Results on INTEL Hardware
E5-2670 v3 (Haswell)
E5-2650 v2 (Ivy Bridge)
E5-2620 (Sandy Bridge)
Xeon Phi 7120 (Knights Corner)
Xeon Phi 7250 (Knights Landing), DDR4
Xeon Phi 7250 (Knights Landing), MCDRAM
Figure 22: Memory bandwidth obtained on Intel hardware (dual socket except KNL) over the
number of processes used. One can get close to peak memory bandwidth with only a few processes.
For example, the following is the most relevant output obtained on a dual-socket system equipped
with two six-core-CPUs with hyperthreading:
np speedup
1 1.0
2 1.58
3 2.19
4 2.42
5 2.63
6 2.69
...
21 3.82
22 3.49
23 3.79
24 3.71
Estimation of possible speedup of MPI programs based on Streams benchmark.
It appears you have 1 node(s)
On this machine, one should expect a speed-up of typical memory bandwidth-bound PETSc appli-
cations of at most 4x when running multiple MPI ranks on the node. Most of the gains are already
obtained when running with only 4-6 ranks. Because a smaller number of MPI ranks usually im-
plies better preconditioners and better performance for smaller problems, the best performance for
PETSc applications may be obtained with fewer ranks than there are physical CPU cores available.
Following the results from the above run of make streams, we recommend to use additional
nodes instead of placing additional MPI ranks on the nodes. In particular, weak scaling (i.e. con-
stant load per process, increasing the number of processes) and strong scaling (i.e. constant total
206
![](asset-cf.png)
CHAPTER 14. HINTS FOR PERFORMANCE TUNING PETSc 3.8 March 24, 2018
CPU
socket
CPU
socket Interconnect
Main Memory Main Memory
Figure 23: Schematic of a two-socket NUMA system. Processes should be spread across both CPUs
to obtain full bandwidth.
work, increasing the number of processes) studies should keep the number of processes per node
constant.
14.1.2 Non-Uniform Memory Access (NUMA) and Process Placement
CPUs in nodes with more than one CPU socket are internally connected via a high-speed fabric,
cf. Fig. 23, to enable data exchange as well as cache coherency. Because main memory on modern
systems is connected via the integrated memory controllers on each CPU, memory is accessed in
a non-uniform way: A process running on one socket has direct access to the memory channels of
the respective CPU, whereas requests for memory attached to a different CPU socket need to go
through the high-speed fabric. Consequently, best aggregate memory bandwidth on the node is
obtained when the memory controllers on each CPU are fully saturated. However, full saturation
of memory channels is only possible if the data is distributed across the different memory channels.
Data in memory on modern machines is allocated by the operating system based on a first-touch
policy. That is, memory is not allocated at the point of issuing malloc(), but at the point when
the respective memory segment is actually touched (read or write). Upon first-touch, memory is
allocated on the memory channel associated with the respective CPU the process is running on.
Only if all memory on the respective CPU is already in use (either allocated or as IO cache),
memory available through other sockets is considered.
Maximum memory bandwidth can be achieved by ensuring that processes are spread over all
sockets in the respective node. For example, the recommended placement of a 8-way parallel run on
a four-socket machine is to assign two processes to each CPU socket. To do so, one needs to know
the enumeration of cores and pass the requested information to mpirun. Consider the hardware
topology information returned by lstopo (part of the hwloc package) for the following two-socket
machine, in which each CPU consists of six cores and supports hyperthreading:
Machine (126GB total)
NUMANode L#0 (P#0 63GB)
Package L#0 + L3 L#0 (15MB)
L2 L#0 (256KB) + L1d L#0 (32KB) + L1i L#0 (32KB) + Core L#0
PU L#0 (P#0)
PU L#1 (P#12)
L2 L#1 (256KB) + L1d L#1 (32KB) + L1i L#1 (32KB) + Core L#1
PU L#2 (P#1)
PU L#3 (P#13)
L2 L#2 (256KB) + L1d L#2 (32KB) + L1i L#2 (32KB) + Core L#2
PU L#4 (P#2)
PU L#5 (P#14)
L2 L#3 (256KB) + L1d L#3 (32KB) + L1i L#3 (32KB) + Core L#3
PU L#6 (P#3)
PU L#7 (P#15)
207
![](asset-d0.png)
14.1. MAXIMIZING MEMORY BANDWIDTH PETSc 3.8 March 24, 2018
L2 L#4 (256KB) + L1d L#4 (32KB) + L1i L#4 (32KB) + Core L#4
PU L#8 (P#4)
PU L#9 (P#16)
L2 L#5 (256KB) + L1d L#5 (32KB) + L1i L#5 (32KB) + Core L#5
PU L#10 (P#5)
PU L#11 (P#17)
NUMANode L#1 (P#1 63GB)
Package L#1 + L3 L#1 (15MB)
L2 L#6 (256KB) + L1d L#6 (32KB) + L1i L#6 (32KB) + Core L#6
PU L#12 (P#6)
PU L#13 (P#18)
L2 L#7 (256KB) + L1d L#7 (32KB) + L1i L#7 (32KB) + Core L#7
PU L#14 (P#7)
PU L#15 (P#19)
L2 L#8 (256KB) + L1d L#8 (32KB) + L1i L#8 (32KB) + Core L#8
PU L#16 (P#8)
PU L#17 (P#20)
L2 L#9 (256KB) + L1d L#9 (32KB) + L1i L#9 (32KB) + Core L#9
PU L#18 (P#9)
PU L#19 (P#21)
L2 L#10 (256KB) + L1d L#10 (32KB) + L1i L#10 (32KB) + Core L#10
PU L#20 (P#10)
PU L#21 (P#22)
L2 L#11 (256KB) + L1d L#11 (32KB) + L1i L#11 (32KB) + Core L#11
PU L#22 (P#11)
PU L#23 (P#23)
The relevant physical processor IDs are shown in parentheses prefixed by P#. Here, IDs 0 and 12
share the same physical core and have a common L2 cache. IDs 0, 12, 1, 13, 2, 14, 3, 15, 4, 16, 5,
17 share the same socket and have a common L3 cache.
A good placement for a run with six processes is to locate three processes on the first socket
and three processes on the second socket. Unfortunately, mechanisms for process placement vary
across MPI implementations, so make sure to consult the manual of your MPI implementation.
The following discussion is based on how processor placement is done with MPICH and OpenMPI,
where one needs to pass --bind-to core --map-by socket to mpirun:
$> mpirun -n 6 --bind-to core --map-by socket ./stream
process 0 binding: 100000000000100000000000
process 1 binding: 000000100000000000100000
process 2 binding: 010000000000010000000000
process 3 binding: 000000010000000000010000
process 4 binding: 001000000000001000000000
process 5 binding: 000000001000000000001000
Triad: 45403.1949 Rate (MB/s)
In this configuration, process 0 is bound to the first physical core on the first socket (with IDs 0
and 12), process 1 is bound to the first core on the second socket (IDs 6 and 18), and similarly for
the remaining processes. The achieved bandwidth of 45 GB/sec is close to the practical peak of
about 50 GB/sec available on the machine. If, however, all MPI processes are located on the same
socket, memory bandwidth drops significantly:
208
![](asset-d1.png)
CHAPTER 14. HINTS FOR PERFORMANCE TUNING PETSc 3.8 March 24, 2018
\$> mpirun -n 6 --bind-to core --map-by core ./stream
process 0 binding: 100000000000100000000000
process 1 binding: 010000000000010000000000
process 2 binding: 001000000000001000000000
process 3 binding: 000100000000000100000000
process 4 binding: 000010000000000010000000
process 5 binding: 000001000000000001000000
Triad: 25510.7507 Rate (MB/s)
All processes are now mapped to cores on the same socket. As a result, only the first memory
channel is fully saturated at 25.5 GB/sec.
One must not assume that mpirun uses good defaults. To demonstrate, compare the full output
of make streams from Section 14.1.1 on the left with the results on the right obtained by passing
--bind-to core --map-by socket:
$> make streams
np speedup
1 1.0
2 1.58
3 2.19
4 2.42
5 2.63
6 2.69
7 2.31
8 2.42
9 2.37
10 2.65
11 2.3
12 2.53
13 2.43
14 2.63
15 2.74
16 2.7
17 3.28
18 3.66
19 3.95
20 3.07
21 3.82
22 3.49
23 3.79
24 3.71
$> make streams MPI_BINDING="--bind-to core --map-by socket"
np speedup
1 1.0
2 1.59
3 2.66
4 3.5
5 3.56
6 4.23
7 3.95
8 4.39
9 4.09
10 4.46
11 4.15
12 4.42
13 3.71
14 3.83
15 4.08
16 4.22
17 4.18
18 4.31
19 4.22
20 4.28
21 4.25
22 4.23
23 4.28
24 4.22
For the non-optimized version on the left, the speedup obtained when using any number of processes
between 3 and 13 is essentially constant up to fluctuations, indicating that all processes were by
default executed on the same socket. Only with 14 or more processes, the speedup number increases
again. In contrast, the results of make streams with proper processor placement shown on the right
resulted in slightly higher overall parallel speedup (identical baselines), in smaller performance
fluctuations, and more than 90 percent of peak bandwidth with only six processes.
Machines with job submission systems such as SLURM usually provide similar mechanisms
209
![](asset-d2.png)
14.1. MAXIMIZING MEMORY BANDWIDTH PETSc 3.8 March 24, 2018
for processor placements through options specified in job submission scripts. Please consult the
respective manuals.
Additional Process Placement Considerations and Details
For a typical, memory bandwidth-limited PETSc application, the primary consideration in placing
MPI processes is ensuring that processes are evenly distributed among sockets, and hence using all
available memory channels. Increasingly complex processor designs and cache hierarchies, however,
mean that performance may also be sensitive to how processes are bound to the resources within
each socket. Performance on the two processor machine in the preceding example may be relatively
insensitive to such placement decisions, because one L3 cache is shared by all cores within a NUMA
domain, and each core has its own L2 and L1 caches. However, processors that are less “flat”, with
more complex hierarchies, may be more sensitive. In many AMD Opterons or the second-generation
“Knights Landing” Intel Xeon Phi, for instance, L2 caches are shared between two cores. On these
processors, placing consecutive MPI ranks on cores that share the same L2 cache may benefit
performance if the two ranks communicate frequently with each other, because the latency between
cores sharing an L2 cache may be roughly half that of two cores not sharing one. There may be
benefit, however, in placing consecutive ranks on cores that do not share an L2 cache, because (if
there are fewer MPI ranks than cores) this increases the total L2 cache capacity and bandwidth
available to the application. There is a trade-off to be considered between placing processes close
together (in terms of shared resources) to optimize for efficient communication and synchronization
vs. farther apart to maximize available resources (memory channels, caches, I/O channels, etc.),
and the best strategy will depend on the application and the software and hardware stack.
Different process placement strategies can affect performance at least as much as some com-
monly explored settings, such as compiler optimization levels. Unfortunately, exploration of this
space is complicated by two factors: First, processor and core numberings may be completely ar-
bitrary, changing with BIOS version, etc., and second—as already noted—there is no standard
mechanism used by MPI implementations (or job schedulers) to specify process affinity. To over-
come the first issue, we recommend using the lstopo utility of the Portable Hardware Locality
(hwloc) software package (which can be installed by configuring PETSc with --download-hwloc)
to understand the processor topology of your machine. We cannot fully address the second issue—
consult the documenation for your MPI implementation and/or job scheduler—but we offer some
general observations on understanding placement options:
•An MPI implementation may support a notion of domains in which a process may be pinned.
A domain may simply correspond to a single core; however, the MPI implementation may
allow a deal of flexibility in specifying domains that encompass multiple cores, span sockets,
etc. Some implementations, such as Intel MPI, provide means to specify whether domains
should be “compact”—composed of cores sharing resources such as caches—or “scatter”-ed,
with little resource sharing (possibly even spanning sockets).
•Separate from the specification of domains, MPI implementations often support different
orderings in which MPI ranks should be bound to these domains. Intel MPI, for instance,
supports “compact” ordering to place consecutive ranks close in terms of shared resources,
“scatter” to place them far apart, and “bunch” to map proportionally to sockets while placing
ranks as close together as possible within the sockets.
•An MPI implemenation that supports process pinning should offer some way to view the
rank assignments. Use this output in conjunction with the topology obtained via lstopo or a
similar tool to determine if the placements correspond to something you believe is reasonable
210
![](asset-d3.png)
CHAPTER 14. HINTS FOR PERFORMANCE TUNING PETSc 3.8 March 24, 2018
for your application. Do not assume that the MPI implementation is doing something sensible
by default!
14.2 Performance Pitfalls and Advice
This section looks into a potpourri of performance pitfalls encountered by users in the past. Many
of these pitfalls require a deeper understanding of the system and experience to detect. The purpose
of this section is to summarize and share our experience so that these pitfalls can be avoided in the
future.
14.2.1 Debug vs. Optimized Builds
PETSc’s ./configure defaults to building PETSc with debug mode enabled. Any code develop-
ment should be done in this mode, because it provides handy debugging facilities such as accurate
stack traces, memory leak checks, or memory corruption checks. Note that PETSc has no reliable
way of knowing whether a particular run is a production or debug run. In the case that a user
requests profiling information via -log_view, a debug build of PETSc issues the following warning:
##########################################################
# #
# WARNING!!! #
# #
# This code was compiled with a debugging option, #
# To get timing results run ./configure #
# using --with-debugging=no, the performance will #
# be generally two or three times faster. #
# #
##########################################################
Conversely, one way of checking whether a particular build of PETSc has debugging enabled is to
inspect the output of -log_view.
Debug mode will generally be most useful for code development if appropriate compiler options
are set to faciliate debugging. The compiler should be instructed to generate binaries with debug
symbols (command line option -g for most compilers), and the optimization level chosen should
either completely disable optimizations (-O0 for most compilers) or enable only optimizations that
do not interfere with debugging (GCC, for instance, supports a -Og optimization level that does
this).
Only once the new code is thoroughly tested and ready for production, one should disable
debugging facilities by passing --with-debugging=no to ./configure. One should also ensure
that an appropriate compiler optimization level is set. Note that some compilers (e.g., Intel) default
to fairly comprehensive optimization levels, while others (e.g., GCC) default to no optimization at
all. The best optimization flags will depend on your code, the compiler, and the target architecture,
but we offer a few guidelines for finding those that will offer the best performance:
•Most compilers have a number of optimization levels (with level n usually specified via -On)
that provide a quick way to enable sets of several optimization flags. We suggest trying
the higher optimization levels (the highest level is not guaranteed to produce the fastest
executable, so some experimentation may be merited). With most recent processors now
supporting some form of SIMD or vector instructions, it is important to choose a level that
enables the compiler’s auto-vectorizer; many compilers do not enable auto-vectorization at
211
![](asset-d4.png)
14.2. PERFORMANCE PITFALLS AND ADVICE PETSc 3.8 March 24, 2018
lower optimization levels (e.g., GCC does not enable it below -O3 and the Intel compiler does
not enable it below -O2).
•For processors supporting newer vector instruction sets, such as Intel AVX2 and AVX-512, it
is also important to direct the compiler to generate code that targets these processors (.e.g.,
-march=knl to generate AVX-512 binaries targeting the Intel “Knights Landing” Xeon Phi
processor); otherwise, the executables built will not utilize the newer instructions sets and
will not take advantage of the vector processing units.
•Beyond choosing the optimization levels, some value-unsafe optimizations (such as using
reciprocals of values instead of dividing by those values, or allowing re-association of operands
in a series of calculations) for floating point calculations may yield significant performance
gains. Compilers often provide flags (e.g., -ffast-math in GCC) to enable a set of these
optimizations, and they may be turned on when using options for very aggressive optimization
(-fast or -Ofast in many compilers). These are worth exploring to maximize performance,
but, if employed, it important to verify that these do not cause erroneous results with your
code, since calculations may violate the IEEE standard for floating-point arithmetic.
14.2.2 Profiling
Users should not spend time optimizing a code until after having determined where it spends the
bulk of its time on realistically sized problems. As discussed in detail in Chapter 13, the PETSc
routines automatically log performance data if certain runtime options are specified.
To obtain a summary of where and how much time is spent in different sections of the code,
use one of the following options:
•Run the code with the option -log_view to print a performance summary for various phases
of the code.
•Run the code with the option -log_mpe [logfilename], which creates a logfile of events
suitable for viewing with Jumpshot (part of MPICH).
Then, focus on the sections where most of the time is spent. If you provided your own callback rou-
tines, e.g. for residual evaluations, search the profiling output for routines such as SNESFunctionEval
or SNESJacobianEval. If their relative time is significant (say, more than 30 percent), consider
optimizing these routines first. Generic instructions on how to optimize your callback functions are
difficult; you may start by reading performance optimization guides for your system’s hardware.
14.2.3 Aggregation
Performing operations on chunks of data rather than a single element at a time can significantly
enhance performance because of cache reuse or lower data motion. Typical examples are:
•Insert several (many) elements of a matrix or vector at once, rather than looping and in-
serting a single value at a time. In order to access elements in of vector repeatedly, employ
VecGetArray() to allow direct manipulation of the vector elements.
•When possible, use VecMDot() rather than a series of calls to VecDot().
•If you require a sequence of matrix-vector products with the same matrix, consider packing
your vectors into a single matrix and use matrix-matrix multiplications.
212
![](asset-d5.png)
CHAPTER 14. HINTS FOR PERFORMANCE TUNING PETSc 3.8 March 24, 2018
•Users should employ a reasonable number of PetscMalloc() calls in their codes. Hundreds or
thousands of memory allocations may be appropriate; however, if tens of thousands are being
used, then reducing the number of PetscMalloc() calls may be warranted. For example,
reusing space or allocating large chunks and dividing it into pieces can produce a significant
savings in allocation overhead. Section 14.2.7 gives details.
Aggressive aggregation of data may result in inflexible datastructures and code that is hard to
maintain. We advise users to keep these competing goals in mind and not blindly optimize for
performance only.
14.2.4 Memory Allocation for Sparse Matrix Assembly
Since the process of dynamic memory allocation for sparse matrices is inherently very expen-
sive, accurate preallocation of memory is crucial for efficient sparse matrix assembly. One should
use the matrix creation routines for particular data structures, such as MatCreateSeqAIJ() and
MatCreateAIJ() for compressed, sparse row formats, instead of the generic MatCreate() routine.
For problems with multiple degrees of freedom per node, the block, compressed, sparse row formats,
created by MatCreateSeqBAIJ() and MatCreateBAIJ(), can significantly enhance performance.
Section 3.1.1 includes extensive details and examples regarding preallocation.
14.2.5 Memory Allocation for Sparse Matrix Factorization
When symbolically factoring an AIJ matrix, PETSc has to guess how much fill there will be. Careful
use of the fill parameter in the MatILUInfo structure when calling MatLUFactorSymbolic() or
MatILUFactorSymbolic() can reduce greatly the number of mallocs and copies required, and thus
greatly improve the performance of the factorization. One way to determine a good value for the
fill parameter is to run a program with the option -info. The symbolic factorization phase will
then print information such as
Info:MatILUFactorSymbolic_AIJ:Realloc 12 Fill ratio:given 1 needed 2.16423
This indicates that the user should have used a fill estimate factor of about 2.17 (instead of 1) to
prevent the 12 required mallocs and copies. The command line option
-pc_ilu_fill 2.17
will cause PETSc to preallocate the correct amount of space for incomplete (ILU) factorization.
The corresponding option for direct (LU) factorization is -pc_factor_fill <fill_amount>.
14.2.6 Detecting Memory Allocation Problems
PETSc provides a number of tools to aid in detection of problems with memory allocation, including
leaks and use of uninitialized space. We briefly describe these below.
•The PETSc memory allocation (which collects statistics and performs error checking), is em-
ployed by default for codes compiled in a debug-mode (configured with --with-debugging=
1). PETSc memory allocation can be activated for optimized-mode (configured with --with-
debugging=0) using the option -malloc. The option -malloc=0 forces the use of conven-
tional memory allocation when debugging is enabled. When running timing tests, one should
build libraries in optimized mode.
213
![](asset-d6.png)
14.2. PERFORMANCE PITFALLS AND ADVICE PETSc 3.8 March 24, 2018
•When the PETSc memory allocation routines are used, the option -malloc_dump will print
a list of unfreed memory at the conclusion of a program. If all memory has been freed,
only a message stating the maximum allocated space will be printed. However, if some
memory remains unfreed, this information will be printed. Note that the option -malloc_
dump merely activates a call to PetscMallocDump() during PetscFinalize() the user can
also call PetscMallocDump() elsewhere in a program.
•Another useful option for use with PETSc memory allocation routines is -malloc_log,
which activates logging of all calls to malloc and reports memory usage, including all For-
tran arrays. This option provides a more complete picture than -malloc_dump for codes
that employ Fortran with hardwired arrays. The option -malloc_log activates logging
by calling PetscMallocSetDumpLog() in PetscInitialize() and then prints the log by
calling PetscMallocDumpLog() in PetscFinalize(). The user can also call these rou-
tines elsewhere in a program. When finer granularity is desired, the user should call
PetscMallocGetCurrentUsage() and PetscMallocGetMaximumUsage() for memory allocated
by PETSc, or
PetscMemoryGetCurrentUsage() and PetscMemoryGetMaximumUsage() for the total mem-
ory used by the program. Note that PetscMemorySetGetMaximumUsage() must be called
before PetscMemoryGetMaximumUsage() (typically at the beginning of the program).
14.2.7 Data Structure Reuse
Data structures should be reused whenever possible. For example, if a code often creates new
matrices or vectors, there often may be a way to reuse some of them. Very significant performance
improvements can be achieved by reusing matrix data structures with the same nonzero pattern. If
a code creates thousands of matrix or vector objects, performance will be degraded. For example,
when solving a nonlinear problem or timestepping, reusing the matrices and their nonzero structure
for many steps when appropriate can make the code run significantly faster.
A simple technique for saving work vectors, matrices, etc. is employing a user-defined context.
In C and C++ such a context is merely a structure in which various objects can be stashed; in
Fortran a user context can be an integer array that contains both parameters and pointers to
PETSc objects. See ${PETSC_DIR}/snes/examples/tutorials/ex5.c and ${PETSC_DIR}/snes/
examples/tutorials/ex5f.F90 for examples of user-defined application contexts in C and Fortran,
respectively.
14.2.8 Numerical Experiments
PETSc users should run a variety of tests. For example, there are a large number of options for the
linear and nonlinear equation solvers in PETSc, and different choices can make a very big difference
in convergence rates and execution times. PETSc employs defaults that are generally reasonable
for a wide range of problems, but clearly these defaults cannot be best for all cases. Users should
experiment with many combinations to determine what is best for a given problem and customize
the solvers accordingly.
•Use the options -snes_view,-ksp_view, etc. (or the routines KSPView(),SNESView(), etc.)
to view the options that have been used for a particular solver.
•Run the code with the option -help for a list of the available runtime commands.
•Use the option -info to print details about the solvers’ operation.
214
![](asset-d7.png)
CHAPTER 14. HINTS FOR PERFORMANCE TUNING PETSc 3.8 March 24, 2018
•Use the PETSc monitoring discussed in Chapter 13 to evaluate the performance of various
numerical methods.
14.2.9 Tips for Efficient Use of Linear Solvers
As discussed in Chapter 4, the default linear solvers are
•uniprocess: GMRES(30) with ILU(0) preconditioning
•multiprocess: GMRES(30) with block Jacobi preconditioning, where there is 1 block per pro-
cess, and each block is solved with ILU(0)
One should experiment to determine alternatives that may be better for various applications. Recall
that one can specify the KSP methods and preconditioners at runtime via the options:
-ksp_type <ksp_name> -pc_type <pc_name>
One can also specify a variety of runtime customizations for the solvers, as discussed throughout
the manual.
In particular, note that the default restart parameter for GMRES is 30, which may be too small
for some large-scale problems. One can alter this parameter with the option -ksp_gmres_restar
<restart> or by calling KSPGMRESSetRestart(). Section 4.3 gives information on setting alterna-
tive GMRES orthogonalization routines, which may provide much better parallel performance.
For elliptic problems one often obtains good performance and scalability with multigrid solvers.
Consult Section 4.4.5 for available options. Our experience is that GAMG works particularly well
for elasticity problems, whereas hypre does well for scalar problems.
14.2.10 System-Related Problems
The performance of a code can be affected by a variety of factors, including the cache behavior,
other users on the machine, etc. Below we briefly describe some common problems and possibilities
for overcoming them.
•Problem too large for physical memory size: When timing a program, one should
always leave at least a ten percent margin between the total memory a process is using and
the physical size of the machine’s memory. One way to estimate the amount of memory
used by given process is with the UNIX getrusage system routine. Also, the PETSc option
-log_view prints the amount of memory used by the basic PETSc objects, thus providing a
lower bound on the memory used. Another useful option is -malloc_log which reports all
memory, including any Fortran arrays in an application code.
•Effects of other users: If other users are running jobs on the same physical processor nodes
on which a program is being profiled, the timing results are essentially meaningless.
•Overhead of timing routines on certain machines: On certain machines, even calling
the system clock in order to time routines is slow; this skews all of the flop rates and timing
results. The file $PETSC_DIR/src/benchmarks/PetscTime.c contains a simple test problem
that will approximate the amount of time required to get the current time in a running
program. On good systems it will on the order of 10−6seconds or less.
215
![](asset-d8.png)
14.2. PERFORMANCE PITFALLS AND ADVICE PETSc 3.8 March 24, 2018
•Problem too large for good cache performance: Certain machines with lower memory
bandwidths (slow memory access) attempt to compensate by having a very large cache. Thus,
if a significant portion of an application fits within the cache, the program will achieve very
good performance; if the code is too large, the performance can degrade markedly. To analyze
whether this situation affects a particular code, one can try plotting the total flop rate as a
function of problem size. If the flop rate decreases rapidly at some point, then the problem
may likely be too large for the cache size.
•Inconsistent timings: Inconsistent timings are likely due to other users on the machine,
thrashing (using more virtual memory than available physical memory), or paging in of the
initial executable. Section 13.8 provides information on overcoming paging overhead when
profiling a code. We have found on all systems that if you follow all the advise above your
timings will be consistent within a variation of less than five percent.
216
![](asset-d9.png)
Chapter 15
Other PETSc Features
15.1 PETSc on a process subset
Users who wish to employ PETSc routines on only a subset of processes within a larger parallel
job, or who wish to use a “master” process to coordinate the work of “slave” PETSc processes,
should specify an alternative communicator for PETSC_COMM_WORLD by directly setting its value, for
example to an existing MPI_COMM_WORLD,
PETSC COMM WORLD=MPI_COMM_WORLD; /* To use an existing MPI_COMM_WORLD */
before calling PetscInitialize(), but, obviously, after calling MPI_Init().
15.2 Runtime Options
Allowing the user to modify parameters and options easily at runtime is very desirable for many
applications. PETSc provides a simple mechanism to enable such customization. To print a list of
available options for a given program, simply specify the option -help (or -h) at runtime, e.g.,
mpiexec -n 1 ./ex1 -help
Note that all runtime options correspond to particular PETSc routines that can be explicitly
called from within a program to set compile-time defaults. For many applications it is natural to
use a combination of compile-time and runtime choices. For example, when solving a linear system,
one could explicitly specify use of the Krylov subspace technique BiCGStab by calling
KSPSetType(ksp,KSPBCGS);
One could then override this choice at runtime with the option
-ksp_type tfqmr
to select the Transpose-Free QMR algorithm. (See Chapter 4for details.)
The remainder of this section discusses details of runtime options.
15.2.1 The Options Database
Each PETSc process maintains a database of option names and values (stored as text strings). This
database is generated with the command PetscInitialize(), which is listed below in its C/C++
and Fortran variants, respectively:
217
![](asset-da.png)
15.2. RUNTIME OPTIONS PETSc 3.8 March 24, 2018
PetscInitialize(int *argc,char ***args,const char *file,const char *help); /* C
*/
call PetscInitialize(character file,integer ierr) ! Fortran
The arguments argc and args (in the C/C++ version only) are the addresses of usual command
line arguments, while the file is a name of a file that can contain additional options. By default
this file is called .petscrc in the user’s home directory. The user can also specify options via the
environmental variable PETSC_OPTIONS. The options are processed in the following order:
1. file
2. environmental variable
3. command line
Thus, the command line options supersede the environmental variable options, which in turn su-
persede the options file.
The file format for specifying options is
-optionname possible_value
-anotheroptionname possible_value
...
All of the option names must begin with a dash (-) and have no intervening spaces. Note that the
option values cannot have intervening spaces either, and tab characters cannot be used between the
option names and values. The user can employ any naming convention. For uniformity throughout
PETSc, we employ the format -[prefix_]package_option (for instance, -ksp_type,-mat_view
::info, or -mg_levels_ksp_type).
Users can specify an alias for any option name (to avoid typing the sometimes lengthy default
name) by adding an alias to the .petscrc file in the format
alias -newname -oldname
For example,
alias -kspt -ksp_type
alias -sd -start_in_debugger
Comments can be placed in the .petscrc file by using #in the first column of a line.
15.2.2 Options Prefixes
Options prefixes allow specific objects to be controlled from the options database. For instance,
PCMG gives prefixes to its nested KSP objects; one may control the coarse grid solver by adding the mg_
coarse prefix, for example -mg_coarse_ksp_type preonly. One may also use
KSPSetOptionsPrefix(),DMSetOptionsPrefix() ,SNESSetOptionsPrefix(),TSSetOptionsPrefix(),
and similar functions to assign custom prefixes, useful for applications with multiple or nested
solvers.
15.2.3 User-Defined PetscOptions
Any subroutine in a PETSc program can add entries to the database with the command
PetscOptionsSetValue(PetscOptions options,char *name,char *value);
218
![](asset-db.png)
CHAPTER 15. OTHER PETSC FEATURES PETSc 3.8 March 24, 2018
though this is rarely done. To locate options in the database, one should use the commands
PetscOptionsHasName(PetscOptions options,char *pre,char *name,PetscBool *flg);
PetscOptionsGetInt(PetscOptions options,char *pre,char *name,PetscInt
*value,PetscBool *flg);
PetscOptionsGetReal(PetscOptions options,char *pre,char *name,PetscReal
*value,PetscBool *flg);
PetscOptionsGetString(PetscOptions options,char *pre,char *name,char *value,int
maxlen,PetscBool *flg);
PetscOptionsGetStringArray(PetscOptions options,char *pre,char *name,char
**values,PetscInt *nmax,PetscBool *flg);
PetscOptionsGetIntArray(PetscOptions options,char *pre,char *name,int
*value,PetscInt *nmax,PetscBool *flg);
PetscOptionsGetRealArray(PetscOptions options,char *pre,char *name,PetscReal
*value, PetscInt *nmax,PetscBool *flg);
All of these routines set flg=PETSC_TRUE if the corresponding option was found, flg=PETSC_FALSE
if it was not found. The optional argument pre indicates that the true name of the option is the
given name (with the dash “-” removed) prepended by the prefix pre. Usually pre should be set
to NULL (or PETSC_NULL_CHARACTER for Fortran); its purpose is to allow someone to rename all
the options in a package without knowing the names of the individual options. For example, when
using block Jacobi preconditioning, the KSP and PC methods used on the individual blocks can be
controlled via the options -sub_ksp_type and -sub_pc_type.
15.2.4 Keeping Track of Options
One useful means of keeping track of user-specified runtime options is use of -options_table,
which prints to stdout during PetscFinalize() a table of all runtime options that the user has
specified. A related option is -options_left, which prints the options table and indicates any
options that have not been requested upon a call to PetscFinalize(). This feature is useful to
check whether an option has been activated for a particular PETSc object (such as a solver or
matrix format), or whether an option name may have been accidentally misspelled.
15.3 Viewers: Looking at PETSc Objects
PETSc employs a consistent scheme for examining, printing, and saving objects through commands
of the form
XXXView(XXX obj,PetscViewer viewer);
Here obj is any PETSc object of type XXX, where XXX is Mat,Vec,SNES, etc. There are several
predefined viewers.
•Passing in a zero (0) for the viewer causes the object to be printed to the screen; this is useful
when viewing an object in a debugger but should be avoided in source code.
•PETSC_VIEWER_STDOUT_SELF and PETSC_VIEWER_STDOUT_WORLD causes the object to be printed
to the screen.
•PETSC_VIEWER_DRAW_SELF PETSC_VIEWER_DRAW_WORLD causes the object to be drawn in a
default X window.
219
![](asset-dc.png)
15.3. VIEWERS: LOOKING AT PETSC OBJECTS PETSc 3.8 March 24, 2018
•Passing in a viewer obtained by PetscViewerDrawOpen() causes the object to be displayed
graphically. See Section 15.9 for more on PETSc’s graphics support.
•To save an object to a file in ASCII format, the user creates the viewer object with the com-
mand PetscViewerASCIIOpen(MPI_Comm comm, char* file, PetscViewer *viewer). This
object is analogous to PETSC_VIEWER_STDOUT_SELF (for a communicator of MPI_COMM_SELF)
and PETSC_VIEWER_STDOUT_WORLD (for a parallel communicator).
•To save an object to a file in binary format, the user creates the viewer object with the com-
mand PetscViewerBinaryOpen(MPI_Comm comm,char* file,PetscViewerBinaryType type,
PetscViewer *viewer). Details of binary I/O are discussed below.
•Vector and matrix objects can be passed to a running MATLAB process with a viewer created
by PetscViewerSocketOpen(MPI_Comm comm,char *machine,int port,PetscViewer *viewer).
For more, see Section 12.2.
The user can control the format of ASCII printed objects with viewers created by
PetscViewerASCIIOpen() by calling
PetscViewerPushFormat(PetscViewer viewer,PetscViewerFormat format);
Formats include PETSC_VIEWER_DEFAULT,PETSC_VIEWER_ASCII_MATLAB, and PETSC_VIEWER_ASCII_IMPL.
The implementation-specific format, PETSC_VIEWER_ASCII_IMPL, displays the object in the most
natural way for a particular implementation.
The routines
PetscViewerPushFormat(PetscViewer viewer,PetscViewerFormat format);
PetscViewerPopFormat(PetscViewer viewer);
allow one to temporarily change the format of a viewer.
As discussed above, one can output PETSc objects in binary format by first opening a binary
viewer with PetscViewerBinaryOpen() and then using MatView(),VecView(), etc. The corre-
sponding routines for input of a binary object have the form XXXLoad(). In particular, matrix and
vector binary input is handled by the following routines:
MatLoad(PetscViewer viewer,MatType outtype,Mat *newmat);
VecLoad(PetscViewer viewer,VecType outtype,Vec *newvec);
These routines generate parallel matrices and vectors if the viewer’s communicator has more than
one process. The particular matrix and vector formats are determined from the options database;
see the manual pages for details.
One can provide additional information about matrix data for matrices stored on disk by provid-
ing an optional file matrixfilename.info, where matrixfilename is the name of the file containing
the matrix. The format of the optional file is the same as the .petscrc file and can (currently)
contain the following:
-matload_block_size <bs>
The block size indicates the size of blocks to use if the matrix is read into a block oriented data
structure (for example, MATMPIBAIJ). The diagonal information s1,s2,s3,... indicates which
(block) diagonals in the matrix have nonzero values.
220
![](asset-dd.png)
CHAPTER 15. OTHER PETSC FEATURES PETSc 3.8 March 24, 2018
15.3.1 Viewing From Options
Command-line options provide a particularly convenient way to view PETSc objects. All options of
the form -xxx_view accept colon(:)-separated compound arguments which specify a viewer type,
format, and/or destination (e.g. file name or socket) if appropriate. For example, to quickly export
a binary file containing a matrix, one may use -mat_view binary:matrix.out, or to output to
a MATLAB-compatible ASCII file, one may use -mat_view ascii:matrix.m:ascii_matlab. See
the PetscOptionsGetViewer() man page for full details, as well as the XXXViewFromOptions()
man pages (for instance, PetscDrawSetFromOptions()) for many other convenient command-line
options.
15.3.2 Using Viewers to Check Load Imbalance
The PetscViewer format PETSC_VIEWER_LOAD_BALANCE will cause certain objects to display simple
measures of their imbalance. For example
-n 4 ./ex32 -ksp_view_mat ::load_balance
will display
Nonzeros: Min 162 avg 168 max 174
indicating that one process has 162 nonzero entries in the matrix, the average number of nonzeros
per process is 168 and the maximum number of nonzeros is 174. Similar for vectors one can see
the load balancing with, for example,
-n 4 ./ex32 -ksp_view_rhs ::load_balance
The measurements of load balancing can also be done within the program with calls to the appro-
priate object viewer with the viewer format PETSC_VIEWER_LOAD_BALANCE.
15.4 Using SAWs with PETSc
The Scientific Application Web server, SAWs1, allows one to monitor running PETSc applications
from a browser. ./configure PETSc with the additional option --download-saws. Options to
use SAWs include
•-saws_options - allows setting values in the PETSc options database via the browser (works
only on one process).
•-stack_view saws - allows monitoring the current stack frame that PETSc is in; refresh to
see the new location.
•-snes_monitor_saws, -ksp_monitor_saws - monitor the solvers’ iterations from the web
browser.
For each of these you need to point your browser to http://hostname:8080, for example http:
//localhost:8080. Options that control behavior of SAWs include
•-saws_log filename - log all SAWs actions in a file.
•-saws_https certfile - use HTTPS instead of HTTP with a certificate.
•-saws_port_auto_select - have SAWs pick a port number instead of using 8080.
•-saws_port port - use port instead of 8080.
1https://bitbucket.org/saws/saws/wiki/Home
221
![](asset-de.png)
15.5. DEBUGGING PETSc 3.8 March 24, 2018
•-saws_root rootdirectory - local directory to which the SAWs browser will have read
access.
•-saws_local - use the local file system to obtain the SAWS javascript files (they much be in
rootdirectory/js).
Also see the manual pages for PetscSAWsBlock,PetscObjectSAWsTakeAccess,PetscObjectSAWsGrantAccess,
PetscObjectSAWsSetBlock,PetscStackSAWsGrantAccess PetscStackSAWsTakeAccess,KSPMonitorSAWs,
and SNESMonitorSAWs.
15.5 Debugging
PETSc programs may be debugged using one of the two options below.
•-start_in_debugger [noxterm,dbx,xxgdb,xdb,xldb,lldb] [-display name] - start all
processes in debugger
•-on_error_attach_debugger [noxterm,dbx,xxgdb,xdb,xldb,lldb] [-display name] -
start debugger only on encountering an error
Note that, in general, debugging MPI programs cannot be done in the usual manner of starting the
programming in the debugger (because then it cannot set up the MPI communication and remote
processes).
By default the GNU debugger gdb is used when -start_in_debugger or -on_error_attach_
debugger is specified. To employ either xxgdb or the common UNIX debugger dbx, one uses
command line options as indicated above. On HP-UX machines the debugger xdb should be used
instead of dbx; on RS/6000 machines the xldb debugger is supported as well. On OS X systems with
XCode tools, lldb is available. By default, the debugger will be started in a new xterm (to enable
running separate debuggers on each process), unless the option noxterm is used. In order to handle
the MPI startup phase, the debugger command cont should be used to continue execution of the
program within the debugger. Rerunning the program through the debugger requires terminating
the first job and restarting the processor(s); the usual run option in the debugger will not correctly
handle the MPI startup and should not be used. Not all debuggers work on all machines, the user
may have to experiment to find one that works correctly.
You can select a subset of the processes to be debugged (the rest just run without the debugger)
with the option
-debugger_nodes node1,node2,...
where you simply list the nodes you want the debugger to run with.
15.6 Error Handling
Errors are handled through the routine PetscError(). This routine checks a stack of error handlers
and calls the one on the top. If the stack is empty, it selects PetscTraceBackErrorHandler(),
which tries to print a traceback. A new error handler can be put on the stack with
PetscPushErrorHandler(PetscErrorCode (*HandlerFunction)(int line,char *dir,char
*file,char *message,int number,void*),void *HandlerContext)
The arguments to HandlerFunction() are the line number where the error occurred, the file in
which the error was detected, the corresponding directory, the error message, the error integer, and
the HandlerContext. The routine
222
![](asset-df.png)
CHAPTER 15. OTHER PETSC FEATURES PETSc 3.8 March 24, 2018
PetscPopErrorHandler()
removes the last error handler and discards it.
PETSc provides two additional error handlers besides PetscTraceBackErrorHandler():
PetscAbortErrorHandler()
PetscAttachErrorHandler()
The function PetscAbortErrorHandler() calls abort on encountering an error, while
PetscAttachErrorHandler() attaches a debugger to the running process if an error is detected.
At runtime, these error handlers can be set with the options -on_error_abort or -on_error_
attach_debugger [noxterm, dbx, xxgdb, xldb] [-display DISPLAY].
All PETSc calls can be traced (useful for determining where a program is hanging without
running in the debugger) with the option
-log_trace [filename]
where filename is optional. By default the traces are printed to the screen. This can also be set
with the command PetscLogTraceBegin(FILE*).
It is also possible to trap signals by using the command
PetscPushSignalHandler(PetscErrorCode (*Handler)(int,void *),void *ctx);
The default handler PetscSignalHandlerDefault() calls PetscError() and then terminates. In
general, a signal in PETSc indicates a catastrophic failure. Any error handler that the user provides
should try to clean up only before exiting. By default all PETSc programs use the default signal
handler, although the user can turn this off at runtime with the option -no_signal_handler .
There is a separate signal handler for floating-point exceptions. The option -fp_trap turns
on the floating-point trap at runtime, and the routine
PetscSetFPTrap(PetscFPTrap flag);
can be used in-line. A flag of PETSC_FP_TRAP_ON indicates that floating-point exceptions should be
trapped, while a value of PETSC_FP_TRAP_OFF (the default) indicates that they should be ignored.
Note that on certain machines, in particular the IBM RS/6000, trapping is very expensive.
A small set of macros is used to make the error handling lightweight. These macros are used
throughout the PETSc libraries and can be employed by the application programmer as well.
When an error is first detected, one should set it by calling
SETERRQ(MPI Comm comm,PetscErrorCode flag,,char *message);
The user should check the return codes for all PETSc routines (and possibly user-defined routines
as well) with
ierr = PetscRoutine(...);CHKERRQ(PetscErrorCode ierr);
Likewise, all memory allocations should be checked with
ierr = PetscMalloc1(n, &ptr);CHKERRQ(ierr);
If this procedure is followed throughout all of the user’s libraries and codes, any error will by default
generate a clean traceback of the location of the error.
Note that the macro PETSC_FUNCTION_NAME is used to keep track of routine names during error
tracebacks. Users need not worry about this macro in their application codes; however, users can
223
![](asset-e0.png)
15.7. NUMBERS PETSc 3.8 March 24, 2018
take advantage of this feature if desired by setting this macro before each user-defined routine that
may call SETERRQ(),CHKERRQ(). A simple example of usage is given below.
PetscErrorCode MyRoutine1() {
/* Declarations Here */
PetscFunctionBeginUser;
/* code here */
PetscFunctionReturn(0);
}
15.7 Numbers
PETSc supports the use of complex numbers in application programs written in C, C++, and
Fortran. To do so, we employ either the C99 complex type or the C++ versions of the PETSc
libraries in which the basic “scalar” datatype, given in PETSc codes by PetscScalar, is defined as
complex (or complex<double> for machines using templated complex class libraries). To work with
complex numbers, the user should run ./configure with the additional option --with-scalar-
type=complex. The file ${PETSC_DIR}/src/docs/website/documentation/installation.html
provides detailed instructions for installing PETSc. You can use --with-clanguage=c (the default)
to use the C99 complex numbers or --with-clanguage=c++ to use the C++ complex type2.
Recall that each variant of the PETSc libraries is stored in a different directory, given by
${PETSC_DIR}/lib/${PETSC_ARCH} according to the architecture. Thus, the libraries for complex
numbers are maintained separately from those for real numbers. When using any of the complex
numbers versions of PETSc, all vector and matrix elements are treated as complex, even if their
imaginary components are zero. Of course, one can elect to use only the real parts of the complex
numbers when using the complex versions of the PETSc libraries; however, when working only with
real numbers in a code, one should use a version of PETSc for real numbers for best efficiency.
The program ${PETSC_DIR}/src/ksp/ksp/examples/tutorials/ex11.c solves a linear sys-
tem with a complex coefficient matrix. Its Fortran counterpart is ${PETSC_DIR}/src/ksp/ksp/
examples/tutorials/ex11f.F.
15.8 Parallel Communication
When used in a message-passing environment, all communication within PETSc is done through
MPI, the message-passing interface standard [25]. Any file that includes petscsys.h (or any other
PETSc include file) can freely use any MPI routine.
15.9 Graphics
The PETSc graphics library is not intended to compete with high-quality graphics packages. In-
stead, it is intended to be easy to use interactively with PETSc programs. We urge users to generate
their publication-quality graphics using a professional graphics package. If a user wants to hook
certain packages into PETSc, he or she should send a message to petsc-maint@mcs.anl.gov; we
will see whether it is reasonable to try to provide direct interfaces.
2Note that this option is not required to use PETSc with C++
224
![](asset-e1.png)
CHAPTER 15. OTHER PETSC FEATURES PETSc 3.8 March 24, 2018
15.9.1 Windows as PetscViewers
For drawing predefined PETSc objects such as matrices and vectors, one may first create a viewer
using the command
PetscViewerDrawOpen(MPI Comm comm,char *display,char *title,int x,int y,int
w,int h,PetscViewer *viewer);
This viewer may be passed to any of the XXXView() routines. Alternately, one may use command-
line options to quickly specify viewer formats, including PetscDraw-based ones; see Section 15.3.1.
To draw directly into the viewer, one must obtain the PetscDraw object with the command
PetscViewerDrawGetDraw(PetscViewer viewer,PetscDraw *draw);
Then one can call any of the PetscDrawXXX commands on the draw object. If one obtains the draw
object in this manner, one does not call the PetscDrawOpenX() command discussed below.
Predefined viewers, PETSC_VIEWER_DRAW_WORLD and PETSC_VIEWER_DRAW_SELF, may be used
at any time. Their initial use will cause the appropriate window to be created.
Implementations using OpenGL, TikZ, and other formats may be selected with PetscDrawSetType().
PETSc can also produce movies; see PetscDrawSetSaveMovie(), and note that command-line op-
tions can also be convenient; see the PetscDrawSetFromOptions() man page.
By default, PETSc drawing tools employ a private colormap, which remedies the problem
of poor color choices for contour plots due to an external program’s mangling of the colormap.
Unfortunately, this may cause flashing of colors as the mouse is moved between the PETSc windows
and other windows. Alternatively, a shared colormap can be used via the option -draw_x_shared_
colormap.
15.9.2 Simple PetscDrawing
With the default format, one can open a window that is not associated with a viewer directly under
the X11 Window System or OpenGL with the command
PetscDrawCreate(MPI Comm comm,char *display,char *title,int x,int y,int w,int
h,PetscDraw *win);
PetscDrawSetFromOptions(win);
All drawing routines are performed relative to the window’s coordinate system and viewport. By
default, the drawing coordinates are from (0,0) to (1,1), where (0,0) indicates the lower left
corner of the window. The application program can change the window coordinates with the
command
PetscDrawSetCoordinates(PetscDraw win,PetscReal xl,PetscReal yl,PetscReal
xr,PetscReal yr);
By default, graphics will be drawn in the entire window. To restrict the drawing to a portion of
the window, one may use the command
PetscDrawSetViewPort(PetscDraw win,PetscReal xl,PetscReal yl,PetscReal
xr,PetscReal yr);
These arguments, which indicate the fraction of the window in which the drawing should be done,
must satisfy 0 ≤xl ≤xr ≤1 and 0 ≤yl ≤yr ≤1.
To draw a line, one uses the command
225
![](asset-e2.png)
15.9. GRAPHICS PETSc 3.8 March 24, 2018
PetscDrawLine(PetscDraw win,PetscReal xl,PetscReal yl,PetscReal xr,PetscReal
yr,int cl);
The argument cl indicates the color (which is an integer between 0 and 255) of the line. A
list of predefined colors may be found in include/petscdraw.h and includes PETSC_DRAW_BLACK,
PETSC_DRAW_RED,PETSC_DRAW_BLUE etc.
To ensure that all graphics actually have been displayed, one should use the command
PetscDrawFlush(PetscDraw win);
When displaying by using double buffering, which is set with the command
PetscDrawSetDoubleBuffer(PetscDraw win);
all processes must call
PetscDrawFlush(PetscDraw win);
in order to swap the buffers. From the options database one may use -draw_pause n, which causes
the PETSc application to pause nseconds at each PetscDrawPause(). A time of -1 indicates that
the application should pause until receiving mouse input from the user.
Text can be drawn with commands
PetscDrawString(PetscDraw win,PetscReal x,PetscReal y,int color,char *text);
PetscDrawStringVertical(PetscDraw win,PetscReal x,PetscReal y,int color,const
char *text);
PetscDrawStringCentered(PetscDraw win,PetscReal x,PetscReal y,int color,const
char *text);
PetscDrawStringBoxed(PetscDraw draw,PetscReal sxl,PetscReal syl,int sc,int
bc,const char text[],PetscReal *w,PetscReal *h);
The user can set the text font size or determine it with the commands
PetscDrawStringSetSize(PetscDraw win,PetscReal width,PetscReal height);
PetscDrawStringGetSize(PetscDraw win,PetscReal *width,PetscReal *height);
15.9.3 Line Graphs
PETSc includes a set of routines for manipulating simple two-dimensional graphs. These routines,
which begin with PetscDrawAxisDraw(), are usually not used directly by the application program-
mer. Instead, the programmer employs the line graph routines to draw simple line graphs. As
shown in the program, within Figure 24, line graphs are created with the command
PetscDrawLGCreate(PetscDraw win,PetscInt ncurves,PetscDrawLG *ctx);
The argument ncurves indicates how many curves are to be drawn. Points can be added to each
of the curves with the command
PetscDrawLGAddPoint(PetscDrawLG ctx,PetscReal *x,PetscReal *y);
The arguments xand yare arrays containing the next point value for each curve. Several points
for each curve may be added with
PetscDrawLGAddPoints(PetscDrawLG ctx,PetscInt n,PetscReal **x,PetscReal **y);
226
![](asset-e3.png)
CHAPTER 15. OTHER PETSC FEATURES PETSc 3.8 March 24, 2018
The line graph is drawn (or redrawn) with the command
PetscDrawLGDraw(PetscDrawLG ctx);
A line graph that is no longer needed can be destroyed with the command
PetscDrawLGDestroy(PetscDrawLG *ctx);
To plot new curves, one can reset a linegraph with the command
PetscDrawLGReset(PetscDrawLG ctx);
The line graph automatically determines the range of values to display on the two axes. The user
can change these defaults with the command
PetscDrawLGSetLimits(PetscDrawLG ctx,PetscReal xmin,PetscReal xmax,PetscReal
ymin,PetscReal ymax);
It is also possible to change the display of the axes and to label them. This procedure is done
by first obtaining the axes context with the command
PetscDrawLGGetAxis(PetscDrawLG ctx,PetscDrawAxis *axis);
One can set the axes’ colors and labels, respectively, by using the commands
PetscDrawAxisSetColors(PetscDrawAxis axis,int axis_lines,int ticks,int text);
PetscDrawAxisSetLabels(PetscDrawAxis axis,char *top,char *x,char *y);
static char help[] = "Plots a simple line graph.\n";
#if defined(PETSC_APPLE_FRAMEWORK)
#import <PETSc/petscsys.h>
#import <PETSc/petscdraw.h>
#else
#include <petscsys.h>
#include <petscdraw.h>
#endif
int main(int argc,char **argv)
{
PetscDraw draw;
PetscDrawLG lg;
PetscDrawAxis axis;
PetscInt n = 15,i,x = 0,y = 0,width = 400,height = 300,nports = 1;
PetscBool useports,flg;
const char *xlabel,*ylabel,*toplabel,*legend;
PetscReal xd,yd;
PetscDrawViewPorts *ports = NULL;
PetscErrorCode ierr;
toplabel = "Top Label"; xlabel = "X-axis Label"; ylabel = "Y-axis Label";
legend = "Legend";
227
![](asset-e4.png)
15.9. GRAPHICS PETSc 3.8 March 24, 2018
ierr = PetscInitialize(&argc,&argv,NULL,help);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-x",&x,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-y",&y,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-width",&width,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-height",&height,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-n",&n,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-nports",&nports,&useports);CHKERRQ(ierr);
ierr = PetscOptionsHasName(NULL,NULL,"-nolegend",&flg);CHKERRQ(ierr);
if (flg) legend = NULL;
ierr = PetscOptionsHasName(NULL,NULL,"-notoplabel",&flg);CHKERRQ(ierr);
if (flg) toplabel = NULL;
ierr = PetscOptionsHasName(NULL,NULL,"-noxlabel",&flg);CHKERRQ(ierr);
if (flg) xlabel = NULL;
ierr = PetscOptionsHasName(NULL,NULL,"-noylabel",&flg);CHKERRQ(ierr);
if (flg) ylabel = NULL;
ierr = PetscOptionsHasName(NULL,NULL,"-nolabels",&flg);CHKERRQ(ierr);
if (flg) {toplabel = NULL; xlabel = NULL; ylabel = NULL;}
ierr = PetscDrawCreate(PETSC COMM WORLD,0,"Title",x,y,width,height,&draw);
CHKERRQ(ierr);
ierr = PetscDrawSetFromOptions(draw);CHKERRQ(ierr);
if (useports) {
ierr = PetscDrawViewPortsCreate(draw,nports,&ports);CHKERRQ(ierr);
ierr = PetscDrawViewPortsSet(ports,0);CHKERRQ(ierr);
}
ierr = PetscDrawLGCreate(draw,1,&lg);CHKERRQ(ierr);
ierr = PetscDrawLGSetUseMarkers(lg,PETSC TRUE);CHKERRQ(ierr);
ierr = PetscDrawLGGetAxis(lg,&axis);CHKERRQ(ierr);
ierr = PetscDrawAxisSetColors(axis,PETSC_DRAW_BLACK,PETSC_DRAW_RED,
PETSC_DRAW_BLUE);CHKERRQ(ierr);
ierr = PetscDrawAxisSetLabels(axis,toplabel,xlabel,ylabel);CHKERRQ(ierr);
ierr = PetscDrawLGSetLegend(lg,&legend);CHKERRQ(ierr);
ierr = PetscDrawLGSetFromOptions(lg);CHKERRQ(ierr);
for (i=0; i<=n; i++) {
xd = (PetscReal)(i - 5); yd = xd*xd;
ierr = PetscDrawLGAddPoint(lg,&xd,&yd);CHKERRQ(ierr);
}
ierr = PetscDrawLGDraw(lg);CHKERRQ(ierr);
ierr = PetscDrawLGSave(lg);CHKERRQ(ierr);
ierr = PetscDrawViewPortsDestroy(ports);CHKERRQ(ierr);
ierr = PetscDrawLGDestroy(&lg);CHKERRQ(ierr);
ierr = PetscDrawDestroy(&draw);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
228
![](asset-e5.png)
CHAPTER 15. OTHER PETSC FEATURES PETSc 3.8 March 24, 2018
Figure 24: Example of PetscDrawing Plots
It is possible to turn off all graphics with the option -nox. This will prevent any windows from
being opened or any drawing actions to be done. This is useful for running large jobs when the
graphics overhead is too large, or for timing.
15.9.4 Graphical Convergence Monitor
For both the linear and nonlinear solvers default routines allow one to graphically monitor conver-
gence of the iterative method. These are accessed via the command line with -ksp_monitor_lg_
residualnorm and -snes_monitor_lg_residualnorm. See also Sections 4.3.3 and 5.3.2.
The two functions used are KSPMonitorLGResidualNorm() and
KSPMonitorLGResidualNormCreate(). These can easily be modified to serve specialized needs.
15.9.5 Disabling Graphics at Compile Time
To disable all X-window-based graphics, run ./configure with the additional option --with-x=0
15.10 Emacs Users
If users develop application codes using Emacs (which we highly recommend), the etags feature
can be used to search PETSc files quickly and efficiently. To use this feature, one should first check
if the file, ${PETSC_DIR}/TAGS exists. If this file is not present, it should be generated by running
make alletags from the PETSc home directory. Once the file exists, from Emacs the user should
issue the command
M-x visit-tags-table
where “M” denotes the Emacs Meta key, and enter the name of the TAGS file. Then the command
“M-.” will cause Emacs to find the file and line number where a desired PETSc function is defined.
Any string in any of the PETSc files can be found with the command “M-x tags-search”. To find
repeated occurrences, one can simply use “M-,” to find the next occurrence.
An alternative which provides reverse lookups (e.g. find all call sites for a given function)
and is somewhat more amenable to managing many projects is GNU Global, available from https:
//www.gnu.org/s/global/. Tags for PETSc and all external packages can be generated by running
the command
find $PETSC_DIR/{include,src,tutorials,$PETSC_ARCH/include} any/other/paths \
-regex ’.*\.\(cc\|hh\|cpp\|C\|hpp\|c\|h\|cu\)$’ \
| grep -v ftn-auto | gtags -f -
from your home directory or wherever you keep source code. If you are making large changes, it is
useful to either set this up to run as a cron job or to make a convenient alias so that refreshing is
easy. Then add the following to ~/.emacs to enable gtags and replace the plain etags bindings.
(when (require ’gtags)
(global-set-key "\C-cf" ’gtags-find-file)
(global-set-key "\M-." ’gtags-find-tag)
(define-key gtags-mode-map (kbd "C-c r") ’gtags-find-rtag))
(add-hook ’c-mode-common-hook
’(lambda () (gtags-mode t))) ; Or add to existing hook
229
![](asset-e6.png)
15.11. VI AND VIM USERS PETSc 3.8 March 24, 2018
15.11 Vi and Vim Users
If users develop application codes using Vi or Vim the tags feature can be used to search PETSc files
quickly and efficiently. To use this feature, one should first check if the file, ${PETSC_DIR}/CTAGS
exists. If this file is not present, it should be generated by running make alletags from the PETSc
home directory. Once the file exists, from Vi/Vim the user should issue the command
:set tags=CTAGS
from the PETSC_DIR directory and enter the name of the CTAGS file. Then the command “tag
functionname” will cause Vi/Vim to find the file and line number where a desired PETSc function
is defined. See, for example, http://www.yolinux.com/TUTORIALS/LinuxTutorialAdvanced_vi.
html for additional Vi/Vim options that allow searches etc. It is also possible to use GNU Global
with Vim; see the description for Emacs above.
15.12 Eclipse Users
If you are interested in developing code that uses PETSc from Eclipse or developing PETSc in
Eclipse and have knowledge of how to do indexing and build libraries in Eclipse, please contact us
at petsc-dev@mcs.anl.gov.
One way to index and build PETSc in Eclipse is as follows.
1. Open “File→Import→Git→Projects from Git”. In the next two panels, you can either add
your existing local repository or download PETSc from Bitbucket by providing the URL.
Most Eclipse distributions come with Git support. If not, install the EGit plugin. When
importing the project, select the wizard “Import as general project”.
2. Right-click on the project (or the “File” menu on top) and select “New →Convert to a
C/C++ Project (Adds C/C++ Nature)”. In the setting window, choose “C Project” and
specify the project type as “Shared Library”.
3. Right-click on the C project and open the “Properties” panel. Under “C/C++ Build →
Builder Settings”, set the Build directory to PETSC_DIR and make sure “Generate Makefiles
automatically” is unselected. Under the section “C/C++ General→Paths and Symbols”, add
the PETSc paths to “Includes”.
${PETSC_DIR}/include
${PETSC_DIR}/${PETSC_ARCH}/include
Under the section “C/C++ General→index”, choose “Use active build configuration”.
4. Configure PETSc normally outside Eclipse to generate a makefile and then build the project
in Eclipse. The source code will be parsed by Eclipse.
If you launch Eclipse from the Dock on Mac OS X, .bashrc will not be loaded (a known OS
X behavior, for security reasons). This will be a problem if you set the environment variables
PETSC_DIR and PETSC_ARCH in .bashrc. A solution which involves replacing the executable can be
found at http://stackoverflow.com/questions/829749. Alternatively, you can add PETSC_DIR
and PETSC_ARCH manually under “Properties →C/C++ Build →Environment”.
To allow an Eclipse code to compile with the PETSc include files and link with the PETSc
libraries, a PETSc user has suggested the following.
230
![](asset-e7.png)
CHAPTER 15. OTHER PETSC FEATURES PETSc 3.8 March 24, 2018
1. Right-click on your C project and select “Properties →C/C++ Build →Settings”
2. A new window on the righthand side appears with various settings options. Select “Includes”
and add the required PETSc paths,
${PETSC_DIR}/include
${PETSC_DIR}/${PETSC_ARCH}/include
3. Select “Libraries” under the header Linker and set the library search path:
${PETSC_DIR}/${PETSC_ARCH}/lib
and the libraries, for example
m, petsc, stdc++, mpichxx, mpich, lapack, blas, gfortran, dl, rt,gcc_s,
pthread, X11
Another PETSc user has provided the following steps to build an Eclipse index for PETSc that
can be used with their own code, without compiling PETSc source into their project.
1. In the user project source directory, create a symlink to the PETSC src/ directory.
2. Refresh the project explorer in Eclipse, so the new symlink is followed.
3. Right-click on the project in the project explorer, and choose “Index →Rebuild”. The index
should now be built.
4. Right-click on the PETSc symlink in the project explorer, and choose “Exclude from build...”
to make sure Eclipse does not try to compile PETSc with the project.
For further examples of using Eclipse with a PETSc-based application, see the documentation for
LaMEM3.
15.13 Qt Creator Users
This information was provided by Mohammad Mirzadeh. The Qt Creator IDE is part of the
Qt SDK, developed for cross-platform GUI programming using C++. It is available under GPL
v3, LGPL v2 and a commercial license and may be obtained, either as a part of Qt SDK or as
an stand-alone software, via http://qt.nokia.com/downloads/. It supports automatic makefile
generation using cross-platform qmake and cmake build systems as well as allowing one to import
projects based on existing, possibly hand-written, makefiles. Qt Creator has a visual debugger
using GDB and LLDB (on Linux and OS X) or Microsoft’s CDB (on Windows) as backends. It
also has an interface to Valgrind’s “memcheck” and “callgrind” tools to detect memory leaks and
profile code. It has built-in support for a variety of version control systems including git, mercurial,
and subversion. Finally, Qt Creator comes fully equipped with auto-completion, function look-up,
and code refactoring tools. This enables one to easily browse source files, find relevant functions,
and refactor them across an entire project.
3doc/ at https://bitbucket.org/bkaus/lamem
231
![](asset-e8.png)
15.13. QT CREATOR USERS PETSc 3.8 March 24, 2018
Creating a Project
When using Qt Creator with qmake, one needs a .pro file. This configuration file tells Qt Creator
about all build/compile options and locations of source files. One may start with a blank .pro file
and fill in configuration options as needed. For example:
# The name of the application executable
TARGET = ex1
# There are two ways to add PETSc functionality
# 1-Manual: Set all include path and libs required by PETSc
PETSC_INCLUDE = path/to/petsc_includes # e.g. obtained via running ‘make
getincludedirs’
PETSC_LIBS = path/to/petsc_libs # e.g. obtained via running ‘make getlinklibs’
INCLUDEPATH += PETSC_INCLUDES
LIBS += PETSC_LIBS
# 2-Automatic: Use the PKGCONFIG funtionality
# NOTE: PETSc.pc must be in the pkgconfig path. You might need to adjust
PKG_CONFIG_PATH
CONFIG += link_pkgconfig
PKGCONFIG += PETSc
# Set appropriate compiler and its flags
QMAKE_CC = path/to/mpicc
QMAKE_CXX = path/to/mpicxx # if this is a cpp project
QMAKE_LINK = path/to/mpicxx # if this is a cpp project
QMAKE_CFLAGS += -O3 # add extra flags here
QMAKE_CXXFLAGS += -O3
QMAKE_LFLAGS += -O3
# Add all files that must be compiled
SOURCES += ex1.c source1.c source2.cpp
HEADERS += source1.h source2.h
# OTHER_FILES are ignored during compilation but will be shown in file panel in
Qt Creator
OTHER_FILES += \
path/to/resource_file \
path/to/another_file
In this example, keywords include:
•TARGET: The name of the application executable.
•INCLUDEPATH: Used at compile time to point to required include files. Essentially, it is used as
an -I \$\$INCLUDEPATH flag for the compiler. This should include all application-specific
header files and those related to PETSc (which may be found via make getincludedirs).
232
![](asset-e9.png)
CHAPTER 15. OTHER PETSC FEATURES PETSc 3.8 March 24, 2018
•LIBS: Defines all required external libraries to link with the application. To get PETSc’s
linking libraries, use make getlinklibs.
•CONFIG: Configuration options to be used by qmake. Here, the option link_pkgconfig
instructs qmake to internally use pkgconfig to resolve INCLUDEPATH and LIBS variables.
•PKGCONFIG: Name of the configuration file (the .pc file – here PETSc.pc) to be passed to
pkgconfig. Note that for this functionality to work, PETSc.pc must be in path which might
require adjusting the PKG_CONFIG_PATH enviroment variable. For more information see http:
//doc.qt.io/qtcreator/creator-build-settings.html#build-environment.
•QMAKE_CC and QMAKE_CXX: Define which C/C++ compilers use.
•QMAKE_LINK: Defines the proper linker to be used. Relevant if compiling C++ projects.
•QMAKE_CFLAGS,QMAKE_CXXFLAGS and QMAKE_LFLAGS: Set the corresponding compile and link-
ing flags.
•SOURCES: Source files to be compiled.
•HEADERS: Header files required by the application.
•OTHER_FILES: Other files to include (source, header, or any other extension). Note that none
of the source files placed here are compiled.
More options can be included in a .pro file; see http://doc.qt.io/qt-5/qmake-project-files.
html. Once the .pro file is generated, the user can simply open it via Qt Creator. Upon opening,
one has the option to create two different build options, debug and release, and switch between the
two. For more information on using the Qt Creator interface and other more advanced aspects of
the IDE, refer to https://www.qt.io/ide/
15.14 Visual Studio Users
To use PETSc from MS Visual Studio, one would have to compile a PETSc example with its
corresponding makefile and then transcribe all compiler and linker options used in this build into
a Visual Studio project file, in the appropriate format in Visual Studio project settings.
15.15 XCode Users (The Apple GUI Development System)
15.15.1 Mac OS X
Follow the instructions in $PETSC_DIR/systems/Apple/OSX/bin/makeall to build the PETSc
framework and documentation suitable for use in XCode.
You can then use the PETSc framework in $PETSC_DIR/arch-osx/PETSc.framework in the
usual manner for Apple frameworks. See the examples in $PETSC_DIR/systems/Apple/OSX/
examples. When working in XCode, things like function name completion should work for all
PETSc functions as well as MPI functions. You must also link against the Apple Accelerate.
framework.
15.15.2 iPhone/iPad iOS
Follow the instructions in $PETSC_DIR/systems/Apple/iOS/bin/iosbuilder.py to build the PETSc
library for use on the iPhone/iPad.
You can then use the PETSc static library in $PETSC_DIR/arch-osx/libPETSc.a in the usual
manner for Apple libraries inside your iOS XCode projects; see the examples in $PETSC_DIR
/systems/Apple/iOS/examples. You must also link against the Apple Accelerate.framework.
233
![](asset-ea.png)
15.15. XCODE USERS (THE APPLE GUI DEVELOPMENT SYSTEM)PETSc 3.8 March 24, 2018
234
![](asset-eb.png)
Chapter 16
Makefiles
This chapter describes the design of the PETSc makefiles, which are key to managing code porta-
bility across a wide variety of UNIX and Windows systems.
16.1 Makefile System
To make a program named ex1, one may use the command
make PETSC_ARCH=arch ex1
which will compile the example and automatically link the appropriate libraries. The architec-
ture, arch, is one of solaris, rs6000, IRIX, hpux, arch-darwin-double-debug, etc., as de-
termined during the PETSc configuration process. Note that when using command line options
with make (as illustrated above), one must not place spaces on either side of the “=” signs. The
variable PETSC_ARCH can also be set as an environment variable.
16.1.1 Makefile Commands
The directory ${PETSC_DIR}/lib/petsc/conf contains virtually all makefile commands and cus-
tomizations to enable portability across different architectures. Most makefile commands for main-
taining the PETSc system are defined in the file ${PETSC_DIR}/lib/petsc/conf/rules These
commands, which process all appropriate files within the directory of execution, include
•lib - Update the PETSc libraries based on the source code in the directory.
•libfast - Update the libraries faster. Since libfast recompiles all source files in the directory
at once, rather than individually, this command saves time when many files must be compiled.
•clean - Remove garbage files.
The tree command enables the user to execute a particular action within a directory and all of
its subdirectories. The action is specified by ACTION=[action], where action is one of the basic
commands listed above. For example, if the command make ACTION=lib tree were executed from
the directory ${PETSC_DIR}/src/ksp/ksp the debugging library for all Krylov subspace solvers
would be built.
235
![](asset-ec.png)
16.2. PETSC FLAGS PETSc 3.8 March 24, 2018
16.1.2 Customized Makefiles
The directory ${PETSC_DIR}/ contains a subdirectory for each architecture that contains machine-
specific information, enabling the portability of our makefile system, these are ${PETSC_DIR}/${PE
TSC_ARCH}/lib/petsc/conf Each architecture directory contains two makefiles:
•petscvariables - definitions of the compilers, linkers, etc.
•petscrules - some build rules specific to this machine.
These files are generated automatically when you run ./configure.
These are included by the architecture-independent makefiles located in ${PETSC_DIR}/lib/
petsc/conf.
16.2 PETSc Flags
PETSc has several flags that determine how the source code will be compiled. The default flags
for particular versions are specified in ${PETSC_DIR}/${PETSC_ARCH}/include/petscconf.h. The
flags include
•PETSC_USE_DEBUG - The PETSc debugging options are activated. We recommend always
using this except when doing performance analysis or production runs.
•PETSC_USE_COMPLEX - The version with scalars represented as complex numbers is used.
•PETSC_USE_LOG - Various monitoring statistics on floating-point operations, and message-
passing activity are kept.
16.2.1 Sample Makefiles
Maintaining portable PETSc makefiles is very simple.
A first example, shown in Figure 25, is a “minimum” makefile for maintaining a single program
that uses the PETSc libraries. The most important lines in this makefile are the lines starting with
include:
include ${PETSC_DIR}/lib/petsc/conf/variables
include ${PETSC_DIR}/lib/petsc/conf/rules
These lines includes other makefiles that provide the needed definitions and rules for the particular
base PETSc installation (specified by PETSC_DIR) and architecture (specified by PETSC_ARCH).
(See 1.2 for information on setting these environmental variables.) As listed in the sample makefile,
the appropriate include file is automatically completely specified; the user should not alter this
statement within the makefile.
ALL: ex2
CFLAGS =
FFLAGS =
CPPFLAGS =
FPPFLAGS =
CLEANFILES = ex2
include ${PETSC_DIR}/lib/petsc/conf/variables
236
![](asset-ed.png)
CHAPTER 16. MAKEFILES PETSc 3.8 March 24, 2018
include ${PETSC_DIR}/lib/petsc/conf/rules
ex2: ex2.o chkopts
${CLINKER} -o ex2 ex2.o ${PETSC_LIB}
${RM} ex2.o
Figure 25: Sample PETSc Makefile for a Single Program
Users who wish to manage the compile process themselves and not use the rules PETSc uses
for compiling programs should only include variables. That is, use something like the makefile
shown in Figure 26.
ALL: ex2
CFLAGS = ${PETSC_CC_INCLUDES}
FFLAGS = ${PETSC_FC_INCLUDES}
include ${PETSC_DIR}/lib/petsc/conf/variables
ex2: ex2.o
mylinkercommand -o ex2 ex2.o ${PETSC_LIB}
Figure 26: Sample PETSc Makefile that does not use PETSc’s rules for compiling
The variables PETSC_CC_INCLUDES,PETSC_FC_INCLUDES and PETSC_LIB are defined by the
included ${PETSC_DIR}/lib/petsc/conf/variables file.
If you do not wish to include any PETSc makefiles in your makefile, you can use the following
commands in the PETSc root directory to get the information needed by your makefile:
make getlinklibs getincludedirs getpetscflags
All the libraries listed need to be linked into your executable and the include directories and flags
need to be passed to the compiler. Usually this is done with CFLAGS=<list of -I and other
flags> and FFLAGS=<list of -I and other flags> in your makefile.
Note that the variable PETSC_LIB (as listed on the link line in the above makefile) specifies all
of the various PETSc libraries in the appropriate order for correct linking. For users who employ
only a specific PETSc library, can use alternative variables like PETSC_SYS_LIB,PETSC_VEC_LIB,
PETSC_MAT_LIB,PETSC_DM_LIB,PETSC_KSP_LIB,PETSC_SNES_LIB or PETSC_TS_LIB.
The second sample makefile, given in Figure 27, controls the generation of several example
programs.
CFLAGS =
FFLAGS =
CPPFLAGS =
FPPFLAGS =
include ${PETSC_DIR}/lib/petsc/conf/variables
include ${PETSC_DIR}/lib/petsc/conf/rules
237
![](asset-ee.png)
16.2. PETSC FLAGS PETSc 3.8 March 24, 2018
ex1: ex1.o
-${CLINKER} -o ex1 ex1.o ${PETSC_LIB}
${RM} ex1.o
ex2: ex2.o
-${CLINKER} -o ex2 ex2.o ${PETSC_LIB}
${RM} ex2.o
ex3: ex3.o
-${FLINKER} -o ex3 ex3.o ${PETSC_LIB}
${RM} ex3.o
ex4: ex4.o
-${CLINKER} -o ex4 ex4.o ${PETSC_LIB}
${RM} ex4.o
runex1:
-@${MPIEXEC} ex1
runex2:
-@${MPIEXEC} -n 2 ./ex2 -mat_seqdense -options_left
runex3:
-@${MPIEXEC} ex3 -v -log_view
runex4:
-@${MPIEXEC} -n 4 ./ex4 -trdump
Figure 27: Sample PETSc Makefile for Several Example Programs
Again, the most important lines in this makefile are the include lines that include the files
defining all of the macro variables. Some additional variables that can be used in the makefile are
defined as follows:
•CFLAGS, FFLAGS User specified additional options for the C compiler and Fortran compiler.
•CPPFLAGS, FPPFLAGS User specified additional flags for the C preprocessor and Fortran pre-
processor.
•CLINKER, FLINKER the C and Fortran linkers.
•RM the remove command for deleting files.
Figure 28 contains a makefile that maintains a PETSc library. Although most users do not
need to understand or deal with such makefiles, they are also easily used.
ALL: lib
CFLAGS =
SOURCEC = sp1wd.c spinver.c spnd.c spqmd.c sprcm.c
SOURCEF = degree.F fnroot.F genqmd.F qmdqt.F rcm.F fn1wd.F gen1wd.F \
genrcm.F qmdrch.F rootls.F fndsep.F gennd.F qmdmrg.F qmdupd.F
SOURCEH =
LIBBASE = libpetscmat
MANSEC = Mat
include ${PETSC_DIR}/lib/petsc/conf/variables
include ${PETSC_DIR}/lib/petsc/conf/rules
238
![](asset-ef.png)
CHAPTER 16. MAKEFILES PETSc 3.8 March 24, 2018
Figure 28: Sample PETSc Makefile for Library Maintenance
The library’s name is libpetscmat.a, and the source files being added to it are indicated by
SOURCEC (for C files) and SOURCEF (for Fortran files).
The variable MANSEC indicates that any manual pages generated from this source should be
included in the Mat section.
16.3 Limitations
This approach to portable makefiles has some minor limitations, including the following:
•Each makefile must be called “makefile”.
•Each makefile can maintain at most one archive library.
239
![](asset-f0.png)
16.3. LIMITATIONS PETSc 3.8 March 24, 2018
240
![](asset-f1.png)
Chapter 17
Unimportant and Advanced Features
of Matrices and Solvers
This chapter introduces additional features of the PETSc matrices and solvers. Since most PETSc
users should not need to use these features, we recommend skipping this chapter during an initial
reading.
17.1 Extracting Submatrices
One can extract a (parallel) submatrix from a given (parallel) using
MatCreateSubMatrix(Mat A,IS rows,IS cols,MatReuse call,Mat *B);
This extracts the rows and columns of the matrix Ainto B. If call is MAT_INITIAL_MATRIX it will
create the matrix B. If call is MAT_REUSE_MATRIX it will reuse the Bcreated with a previous call.
17.2 Matrix Factorization
Normally, PETSc users will access the matrix solvers through the KSP interface, as discussed in
Chapter 4, but the underlying factorization and triangular solve routines are also directly accessible
to the user.
The LU and Cholesky matrix factorizations are split into two or three stages depending on the
user’s needs. The first stage is to calculate an ordering for the matrix. The ordering generally is
done to reduce fill in a sparse factorization; it does not make much sense for a dense matrix.
MatGetOrdering(Mat matrix,MatOrderingType type,IS* rowperm,IS* colperm);
The currently available alternatives for the ordering type are
•MATORDERINGNATURAL - Natural
•MATORDERINGND - Nested Dissection
•MATORDERING1WD - One-way Dissection
•MATORDERINGRCM - Reverse Cuthill-McKee
241
![](asset-f2.png)
17.2. MATRIX FACTORIZATION PETSc 3.8 March 24, 2018
•MATORDERINGQMD - Quotient Minimum Degree
These orderings can also be set through the options database.
Certain matrix formats may support only a subset of these; more options may be added. Check
the manual pages for up-to-date information. All of these orderings are symmetric at the moment;
ordering routines that are not symmetric may be added. Currently we support orderings only for
sequential matrices.
Users can add their own ordering routines by providing a function with the calling sequence
int reorder(Mat A,MatOrderingType type,IS* rowperm,IS* colperm);
Here Ais the matrix for which we wish to generate a new ordering, type may be ignored and
rowperm and colperm are the row and column permutations generated by the ordering routine.
The user registers the ordering routine with the command
MatOrderingRegister(MatOrderingType inname,char *path,char *sname,PetscErrorCode
(*reorder)(Mat,MatOrderingType,IS*,IS*)));
The input argument inname is a string of the user’s choice, iname is either an ordering defined in
petscmat.h or a users string, to indicate one is introducing a new ordering, while the output See
the code in src/mat/impls/order/sorder.c and other files in that directory for examples on how
the reordering routines may be written.
Once the reordering routine has been registered, it can be selected for use at runtime with
the command line option -pc_factor_mat_ordering_type sname. If reordering directly, the user
should provide the name as the second input argument of MatGetOrdering().
The following routines perform complete, in-place, symbolic, and numerical factorizations for
symmetric and nonsymmetric matrices, respectively:
MatCholeskyFactor(Mat matrix,IS permutation,const MatFactorInfo *info);
MatLUFactor(Mat matrix,IS rowpermutation,IS columnpermutation,const
MatFactorInfo *info);
The argument info->fill > 1 is the predicted fill expected in the factored matrix, as a ratio of
the original fill. For example, info->fill=2.0 would indicate that one expects the factored matrix
to have twice as many nonzeros as the original.
For sparse matrices it is very unlikely that the factorization is actually done in-place. More
likely, new space is allocated for the factored matrix and the old space deallocated, but to the user
it appears in-place because the factored matrix replaces the unfactored matrix.
The two factorization stages can also be performed separately, by using the out-of-place mode,
first one obtains that matrix object that will hold the factor
MatGetFactor(Mat matrix,const MatSolverPackage package,MatFactorType ftype,Mat
*factor);
and then performs the factorization
MatCholeskyFactorSymbolic(Mat factor,Mat matrix,IS perm,const MatFactorInfo
*info);
MatLUFactorSymbolic(Mat factor,Mat matrix,IS rowperm,IS colperm,const
MatFactorInfo *info);
MatCholeskyFactorNumeric(Mat factor,Mat matrix,const MatFactorInfo);
MatLUFactorNumeric(Mat factor,Mat matrix,const MatFactorInfo *info);
242
![](asset-f3.png)
CHAPTER 17. UNIMPORTANT AND ADVANCED FEATURES OF MATRICES AND
SOLVERS PETSc 3.8 March 24, 2018
In this case, the contents of the matrix result is undefined between the symbolic and numeric fac-
torization stages. It is possible to reuse the symbolic factorization. For the second and succeeding
factorizations, one simply calls the numerical factorization with a new input matrix and the same
factored result matrix. It is essential that the new input matrix have exactly the same nonzero
structure as the original factored matrix. (The numerical factorization merely overwrites the nu-
merical values in the factored matrix and does not disturb the symbolic portion, thus enabling
reuse of the symbolic phase.) In general, calling XXXFactorSymbolic with a dense matrix will do
nothing except allocate the new matrix; the XXXFactorNumeric routines will do all of the work.
Why provide the plain XXXfactor routines when one could simply call the two-stage routines?
The answer is that if one desires in-place factorization of a sparse matrix, the intermediate stage
between the symbolic and numeric phases cannot be stored in a result matrix, and it does not
make sense to store the intermediate values inside the original matrix that is being transformed.
We originally made the combined factor routines do either in-place or out-of-place factorization,
but then decided that this approach was not needed and could easily lead to confusion.
We do not currently support sparse matrix factorization with pivoting for numerical sta-
bility. This is because trying to both reduce fill and do pivoting can become quite compli-
cated. Instead, we provide a poor stepchild substitute. After one has obtained a reordering,
with MatGetOrdering(Mat A,MatOrdering type,IS *row,IS *col) one may call
MatReorderForNonzeroDiagonal(Mat A,PetscReal tol,IS row, IS col);
which will try to reorder the columns to ensure that no values along the diagonal are smaller
than tol in a absolute value. If small values are detected and corrected for, a nonsymmetric
permutation of the rows and columns will result. This is not guaranteed to work, but may help if
one was simply unlucky in the original ordering. When using the KSP solver interface the option
-pc_factor_nonzeros_along_diagonal <tol> may be used. Here, tol is an optional tolerance
to decide if a value is nonzero; by default it is 1.e-10.
Once a matrix has been factored, it is natural to solve linear systems. The following four
routines enable this process:
MatSolve(Mat A,Vec x, Vec y);
MatSolveTranspose(Mat A, Vec x, Vec y);
MatSolveAdd(Mat A,Vec x, Vec y, Vec w);
MatSolveTransposeAdd(Mat A, Vec x, Vec y, Vec w);
matrix Aof these routines must have been obtained from a factorization routine; otherwise, an error
will be generated. In general, the user should use the KSP solvers introduced in the next chapter
rather than using these factorization and solve routines directly.
17.3 Unimportant Details of KSP
PetscDrawAxisDraw(), are usually not used directly by the application programmer Again, virtu-
ally all users should use KSP through the KSP interface and, thus, will not need to know the details
that follow.
It is possible to generate a Krylov subspace context with the command
KSPCreate(MPI Comm comm,KSP *kps);
Before using the Krylov context, one must set the matrix-vector multiplication routine and the
preconditioner with the commands
243
![](asset-f4.png)
17.4. UNIMPORTANT DETAILS OF PC PETSc 3.8 March 24, 2018
PCSetOperators(PC pc,Mat Amat,Mat Pmat);
KSPSetPC(KSP ksp,PC pc);
In addition, the KSP solver must be initialized with
KSPSetUp(KSP ksp);
Solving a linear system is done with the command
KSPSolve(KSP ksp,Vec b,Vec x);
Finally, the KSP context should be destroyed with
KSPDestroy(KSP *ksp);
It may seem strange to put the matrix in the preconditioner rather than directly in the KSP; this
decision was the result of much agonizing. The reason is that for SSOR with Eisenstat’s trick, and
certain other preconditioners, the preconditioner has to change the matrix-vector multiply. This
procedure could not be done cleanly if the matrix were stashed in the KSP context that PC cannot
access.
Any preconditioner can supply not only the preconditioner, but also a routine that essentially
performs a complete Richardson step. The reason for this is mainly SOR. To use SOR in the
Richardson framework, that is,
un+1 =un+B(f−Aun),
is much more expensive than just updating the values. With this addition it is reasonable to state
that all our iterative methods are obtained by combining a preconditioner from the PC package with
a Krylov method from the KSP package. This strategy makes things much simpler conceptually, so
(we hope) clean code will result. Note: We had this idea already implicitly in older versions of KSP,
but, for instance, just doing Gauss-Seidel with Richardson in old KSP was much more expensive
than it had to be. With PETSc this should not be a problem.
17.4 Unimportant Details of PC
Most users will obtain their preconditioner contexts from the KSP context with the command
KSPGetPC(). It is possible to create, manipulate, and destroy PC contexts directly, although this
capability should rarely be needed. To create a PC context, one uses the command
PCCreate(MPI Comm comm,PC *pc);
The routine
PCSetType(PC pc,PCType method);
sets the preconditioner method to be used. The routine
PCSetOperators(PC pc,Mat Amat,Mat Pmat);
set the matrices that are to be used with the preconditioner. The routine
PCGetOperators(PC pc,Mat *Amat,Mat *Pmat);
returns the values set with PCSetOperators().
The preconditioners in PETSc can be used in several ways. The two most basic routines simply
apply the preconditioner or its transpose and are given, respectively, by
244
![](asset-f5.png)
CHAPTER 17. UNIMPORTANT AND ADVANCED FEATURES OF MATRICES AND
SOLVERS PETSc 3.8 March 24, 2018
PCApply(PC pc,Vec x,Vec y);
PCApplyTranspose(PC pc,Vec x,Vec y);
In particular, for a preconditioner matrix, B, that has been set via PCSetOperators(pc,Amat,Pmat),
the routine PCApply(pc,x,y) computes y=B−1xby solving the linear system By =xwith the
specified preconditioner method.
Additional preconditioner routines are
PCApplyBAorAB(PC pc,PCSide right,Vec x,Vec y,Vec work);
PCApplyBAorABTranspose(PC pc,PCSide right,Vec x,Vec y,Vec work);
PCApplyRichardson(PC pc,Vec x,Vec y,Vec work,PetscReal rtol,PetscReal atol,
PetscReal dtol,PetscInt maxits,PetscBool zeroguess,PetscInt
*outits,PCRichardsonConvergedReason*);
The first two routines apply the action of the matrix followed by the preconditioner or the pre-
conditioner followed by the matrix depending on whether the right is PC_LEFT or PC_RIGHT. The
final routine applies its iterations of Richardson’s method. The last three routines are provided
to improve efficiency for certain Krylov subspace methods.
APC context that is no longer needed can be destroyed with the command
PCDestroy(PC *pc);
245
![](asset-f6.png)
17.4. UNIMPORTANT DETAILS OF PC PETSc 3.8 March 24, 2018
246
![](asset-f7.png)
Index
-draw pause, 226
-fp trap, 24,223
-h, 24,217
-help, 217
-history, 202
-info, 68,70,195,201
-ksp atol, 84
-ksp compute eigenvalues, 86
-ksp compute eigenvalues explicitly, 86
-ksp divtol, 84
-ksp gmres cgs refinement type, 81
-ksp gmres classicalgramschmidt, 81
-ksp gmres modifiedgramschmidt, 81
-ksp gmres restart, 81
-ksp max it, 84
-ksp monitor, 84,85
-ksp monitor cancel, 85
-ksp monitor lg residualnorm, 84,85,229
-ksp monitor short, 85
-ksp monitor singular value, 85
-ksp monitor true residual, 85
-ksp pc side, 82
-ksp plot eigenvalues, 86
-ksp plot eigenvalues explicitly, 86
-ksp richardson scale, 81
-ksp rtol, 84
-ksp type, 81
-log mpe, 199,212
-log trace, 195,223
-log view, 195,197,212
-malloc, 213
-malloc dump, 24,213
-malloc log, 214
-mat coloring type, 139
-mat fd coloring err, 139
-mat fd coloring umin, 139
-mat view ::ascii matlab, 191
-mg levels, 98
-no signal handler, 223
-nox, 229
-on error attach debugger, 24
-options left, 219
-options table, 219
-pc asm type, 90
-pc bgs blocks, 90
-pc bjacobi blocks, 90
-pc composite pcs, 96
-pc composite type, 96
-pc eisenstat omega, 88
-pc factor diagonal fill, 88
-pc factor fill, 213
-pc factor in place, 88,89
-pc factor levels, 88
-pc factor mat ordering type, 89
-pc factor nonzeros along diagonal, 88,89,243
-pc factor reuse fill, 88
-pc factor reuse ordering, 88
-pc factor shift amount, 102
-pc factor shift type, 102
-pc fieldsplit detect saddle point, 100
-pc fieldsplit type, 100
-pc gasm type, 90
-pc ilu fill, 213
-pc mg cycle type, 97
-pc mg smoothdown, 97
-pc mg smoothup, 97
-pc mg type, 97
-pc sor backward, 88
-pc sor its, 88
-pc sor local backward, 88
-pc sor local forward, 88
-pc sor local symmetric, 88
-pc sor omega, 88
-pc sor symmetric, 88
-pc type, 86
-pc use amat, 96,97
-preload, 203
-snes atol, 119
-snes ksp ew conv, 121
-snes linesearch alpha, 116
-snes linesearch maxstep, 116
-snes linesearch minlambda, 116
247
![](asset-f8.png)
INDEX PETSc 3.8 March 24, 2018
-snes max funcs, 119
-snes max it, 119
-snes mf, 121
-snes mf err, 121
-snes mf operator, 121
-snes mf umin, 122
-snes monitor, 119
-snes monitor cancel, 119
-snes monitor lg residualnorm, 119,229
-snes monitor short, 119
-snes rtol, 119
-snes stol, 119
-snes test display, 120
-snes trtol, 119
-snes type, 112
-snes vi monitor, 140
-snes vi type, 140
-start in debugger, 24
-sub ksp type, 89
-sub pc type, 89
-ts pseudo increment, 160
-ts pseudo increment dt from initial dt, 160
-ts sundials gmres restart, 154
-ts sundials gramschmidt type, 154
-ts sundials type, 153
-ts type, 145
-v, 24
-vec type, 47
-vec view ::ascii matlab, 191
.petschistory, 202
.petscrc, 218
1-norm, 49,73
2-norm, 49
Adams, 153
ADD VALUES, 36,48,56,60–63,65,66,167
additive preconditioners, 96
aggregation, 212
AIJ matrix format, 67
alias, 218
AO, 47,51–53,58,78
AOApplicationToPetsc, 52,78
AOApplicationToPetscIS, 52
AOCreateBasic, 52
AOCreateBasicIS, 52,78
AODestroy, 52
AOPetscToApplication, 52,78
AOPetscToApplicationIS, 52
AOView, 52
Arnoldi, 86
array, distributed, 55
ASM, 89
assembly, 48
axis, drawing, 227
backward Euler, 145
BDF, 153
Bi-conjugate gradient, 84
block Gauss-Seidel, 89,90
block Jacobi, 89,90,219
Bogacki-Shampine, 150
boundary conditions, 75
Cai, Xiao-Chuan, 90
CG, 81
CHKERRQ, 27–30,33,35–39,107–112,125–
137,174,202,223,224,228
CHKERRQ(), 223
CHKMEMQ, 33
Cholesky, 241
coarse grid solve, 97
collective operations, 34
coloring with SNES, 138
combining preconditioners, 95
command line arguments, 25
command line options, 217
communicator, 85,217
compiler options, 211
complex numbers, 31,224
composite, 96
convergence tests, 84,118
coordinates, 225
Crank-Nicolson, 145
CSR, compressed sparse row format, 67
damping, 102
debugger, 24
debugging, 222
direct solver, 88
distributed array, 55
DM, 21,47,55–58,91,99,124,130,132,135,
150,163,166,168,169,189
DM BOUNDARY GHOSTED, 55
DM BOUNDARY MIRROR, 55
DM BOUNDARY NONE, 55,125
DM BOUNDARY PERIODIC, 55
DM BOUNDARY TWIST, 55
DMBoundaryType, 55
DMCompositeGetLocalISs, 72
248
![](asset-f9.png)
INDEX PETSc 3.8 March 24, 2018
DMCreateColoring, 139,140
DMCreateDomainDecomposition, 91
DMCreateGlobalVector, 55,56,125
DMCreateLocalVector, 55,56
DMCreateMatrix, 99
DMDA, 9,54–58,99,117,118,125,128,131,
192
DMDA STENCIL BOX, 55,58
DMDA STENCIL STAR, 55,58
DMDACreate1d, 55,125
DMDACreate2d, 55
DMDACreate3d, 55
DMDAGetAO, 58
DMDAGetCorners, 58,128,131,132,135
DMDAGetGhostCorners, 58
DMDAGetInfo, 131,132
DMDAGetScatter, 56
DMDAStencilType, 55
DMDAVecGetArray, 57,58,128,131,135
DMDAVecGetArrayDOF, 57
DMDAVecGetArrayDOFRead, 57
DMDAVecGetArrayRead, 57,132
DMDAVecRestoreArray, 57,58,128,131,136
DMDAVecRestoreArrayDOF, 57
DMDAVecRestoreArrayDOFRead, 57
DMDAVecRestoreArrayRead, 57,133
DMDestroy, 129
DMGetGlobalVector, 166
DMGetLocalToGlobalMapping, 58
DMGetLocalVector, 57,130,166
DMGlobalToLocalBegin, 56,130
DMGlobalToLocalEnd, 56,130
DMLocalToGlobalBegin, 56
DMLocalToGlobalEnd, 56
DMNetworkAddComponent, 168
DMNetworkAddNumVariables, 168
DMNetworkCreate, 168
DMNetworkDistribute, 168
DMNetworkGetComponentDataArray, 169
DMNetworkGetComponentKeyOffset, 169
DMNetworkGetConnectedVertices, 169
DMNetworkGetEdgeRange, 169
DMNetworkGetNumComponents, 169
DMNetworkGetVariableGlobalOffset, 169
DMNetworkGetVariableOffset, 169
DMNetworkGetVertexRange, 169
DMNetworkIsGhostVertex, 169
DMNetworkLayoutSetUp, 168
DMNetworkRegisterComponent, 168,169
DMNetworkSetEdgeList, 168
DMNetworkSetNumVariables, 168
DMNetworkSetSizes, 168
DMPLEX, 163
DMPlex, 163
DMPlexGetChart, 165
DMPlexGetDepthStratum, 165
DMPlexGetHeightStratum, 165
DMPlexGetTransitiveClosure, 166,167
DMPlexRestoreTransitiveClosure, 167
DMPlexSetChart, 163
DMPlexSetCone, 164
DMPlexSetConeSize, 164
DMPlexSetSupport, 164
DMPlexSetSupportSize, 164
DMPlexStratify, 165
DMPlexSymmetrize, 165
DMPlexVecGetClosure, 167
DMPlexVecRestoreClosure, 167
DMPlexVecSetClosure, 167
DMRestoreLocalVector, 57,131
DMSetDefaultSection, 166
DMSetFromOptions, 125
DMSetOptionsPrefix, 218
DMSetUp, 125,164,168
Dormand-Prince, 150
double buffer, 226
eclipse, 230
eigenvalues, 86
Eisenstat trick, 88
Emacs, 229
errors, 222
etags, in Emacs, 229
Euler, 145
Euler, backward, 145
factorization, 241
Fehlberg, 150
floating-point exceptions, 223
flushing, graphics, 226
Frobenius norm, 73
gather, 60
generalized linear, 145
ghost points, 53,54
global numbering, 51
global representation, 53
global to local mapping, 53
GMRES, 81
249
![](asset-fa.png)
INDEX PETSc 3.8 March 24, 2018
Gram-Schmidt, 81
graphics, 224
graphics, disabling, 229
grid partitioning, 77
Hermitian matrix, 82
Hindmarsh, 153
ICC, parallel, 88
IEEE floating point, 223
ILU, parallel, 88
in-place solvers, 88
index sets, 59
inexact Newton methods, 120
infinity norm, 49,73
INSERT VALUES, 28,31,32,48,54,56,60–
63,65,66,110,112,130,132,133,
179,188
InsertMode, 56,62,63
installing PETSc, 23
IS, 21,47,52,53,59–61,72,75,76,78,90,
91,166,175,189,241–243
IS GTOLM DROP, 53
IS GTOLM MASK, 53
ISBlockGetIndices, 60
ISBlockGetLocalSize, 60
ISBlockGetSize, 60
ISColoring, 138
ISColoringDestroy, 138,140
ISCreateBlock, 60
ISCreateGeneral, 59,61
ISCreateStride, 59
ISDestroy, 60,61
ISGetBlockSize, 60
ISGetIndices, 60,176,189,190
ISGetIndicesF90, 176
ISGetSize, 60
ISGlobalToLocalMappingApplyBlock, 53
ISGlobalToLocalMappingMode, 53
ISLocalToGlobalMapping, 47,51,53,54,58
ISLocalToGlobalMappingApply, 53
ISLocalToGlobalMappingApplyIS, 53
ISLocalToGlobalMappingCreate, 53
ISLocalToGlobalMappingDestroy, 53
ISPartitioningToNumbering, 78
ISRestoreIndices, 60
ISRestoreIndicesF90, 176
ISStrideGetInfo, 60
Jacobi, 89
Jacobian, 105
Jacobian, debugging, 120
Jacobian, testing, 120
Jumpshot, 199
Krylov subspace methods, 79,81
KSP, 21,26,27,29,32,34,35,38,79–87,
89,93,96–98,100–102,106,108,112,
115,121,124,126,127,136,137,143,
183–185,196,201,215,218,219,241,
243,244
KSP CG SYMMETRIC, 82
KSP GMRES CGS REFINEMENT ALWAYS,
81
KSP GMRES CGS REFINEMENT IFNEEDED,
81
KSP GMRES CGS REFINEMENT NONE, 81
KSP NORM UNPRECONDITIONED, 82
KSPBCGS, 81,83,217
KSPBCGSL, 9,83
KSPBICG, 81,83,84
KSPBuildResidual, 86
KSPBuildSolution, 86
KSPCG, 81,83
KSPCGLS, 83
KSPCGNE, 83
KSPCGS, 81,83
KSPCGSetType, 82
KSPCGType, 82
KSPCHEBYSHEV, 81,83
KSPChebyshevSetEigenvalues, 81
KSPComputeEigenvalues, 86
KSPComputeEigenvaluesExplicitly, 86
KSPConvergedDefault, 84
KSPConvergedReason, 84
KSPCR, 81,83
KSPCreate, 28,32,34,37,79,97,243
KSPDestroy, 30,32,39,80,244
KSPDGMRES, 9,83
KSPFBCGS, 83
KSPFBCGSR, 83
KSPFCG, 83
KSPFETIDP, 83
KSPFGMRES, 9,82,83
KSPGCR, 9,83
KSPGetConvergedReason, 84
KSPGetIterationNumber, 30,38,80,137
KSPGetPC, 29,80,97,98,103,108,123,126,
137,184,244
250
![](asset-fb.png)
INDEX PETSc 3.8 March 24, 2018
KSPGetRhs, 86
KSPGetSolution, 86
KSPGMRES, 81,83,196
KSPGMRESCGSRefinementType, 81
KSPGMRESClassicalGramSchmidtOrthogonalization,
81
KSPGMRESModifiedGramSchmidtOrthogonalization,
81
KSPGMRESSetCGSRefinementType, 81
KSPGMRESSetOrthogonalization, 81
KSPGMRESSetRestart, 81,215
KSPGROPPCG, 83
KSPIBCGS, 83
KSPLGMRES, 9,83
KSPLSQR, 81–83
KSPMINRES, 9,83
KSPMonitorDefault, 85
KSPMonitorLGResidualNormCreate, 85,229
KSPMonitorSAWs, 222
KSPMonitorSet, 85
KSPMonitorSingularValue, 85
KSPMonitorTrueResidualNorm, 85
KSPPGMRES, 83
KSPPIPECG, 83
KSPPIPECGRR, 83
KSPPIPECR, 83
KSPPIPEFCG, 83
KSPPIPEFGMRES, 83
KSPPIPEGCR, 83
KSPPREONLY, 81,83,89,103
KSPQCG, 82,83
KSPRICHARDSON, 81,83
KSPRichardsonSetScale, 81
KSPSetComputeEigenvalues, 86
KSPSetConvergenceTest, 84,175
KSPSetDM, 161,168
KSPSetFromOptions, 29,32,38,80
KSPSetInitialGuessNonzero, 29,82
KSPSetNormType, 82
KSPSetOperators, 29,32,38,79,81,95,98,
102,113
KSPSetOptionsPrefix, 218
KSPSetPC, 244
KSPSetPCSide, 82
KSPSetTolerances, 29,38,84,108,137,184
KSPSetType, 81,103,217
KSPSetUp, 29,80,89,196,244
KSPSolve, 30,32,38,80–82,84,93,173,174,
196,244
KSPSYMMLQ, 83
KSPTCQMR, 81,83
KSPTFQMR, 81,83
KSPTSIRM, 83
KSPType, 81,83,89
KSPView, 30,214
Lanczos, 86
line graphs, 226
line search, 105,114
linear system solvers, 79
lines, drawing, 225
local linear solves, 89
local representation, 53
local to global mapping, 53
logging, 195,212
LU, 241
Mat, 9,21,27,31,32,34,35,58,65–67,69,
71–79,98,102,106,107,110,111,
113,118,121,122,124,132,144,145,
156,157,160,183,187,189,192,196,
201,219,220,239,241–244
MAT FINAL ASSEMBLY, 28,32,36,66,110–
112,133,188
MAT FLUSH ASSEMBLY, 66
MAT INITIAL MATRIX, 241
MAT NEW NONZERO LOCATIONS, 66,75,
76
MAT REUSE MATRIX, 241
MAT ROW ORIENTED, 66
MAT SYMMETRIC, 36
MATAIJ, 32,72
MATAIJPERM, 9
MatAssemblyBegin, 28,32,36,66,68,110–
112,133,188
MatAssemblyEnd, 28,32,36,66,68,110–112,
133,188,191
MatAXPY, 74
MatCholeskyFactor, 242
MatCholeskyFactorNumeric, 242
MatCholeskyFactorSymbolic, 242
MatColoring, 138,139
MatColoringApply, 138
MatColoringCreate, 138
MatColoringDestroy, 138
MATCOLORINGGREEDY, 139
MATCOLORINGID, 139
MATCOLORINGJP, 139
251
![](asset-fc.png)
INDEX PETSc 3.8 March 24, 2018
MATCOLORINGLF, 139
MATCOLORINGNATURAL, 139
MatColoringSetFromOptions, 138
MatColoringSetType, 138,139
MATCOLORINGSL, 139
MatConvert, 74,76,104
MatCopy, 74
MatCreate, 28,31,34–36,65,104,107,123,
126,184,213
MatCreateAIJ, 69,123,213
MatCreateBAIJ, 213
MatCreateDense, 71
MatCreateMFFD, 121,123
MatCreateMPIAdj, 77
MatCreateSeqAIJ, 67–69,174,213
MatCreateSeqBAIJ, 213
MatCreateSeqDense, 71
MatCreateShell, 74,79,123
MatCreateSNESMF, 121,123
MatCreateSubMatrix, 241
MatDenseGetArray, 176,189,190
MatDenseGetArrayF90, 176
MatDenseRestoreArrayF90, 176
MatDestroy, 30,39,78,109,129,185
MatDiagonalScale, 74
MatFactorInfo, 242
MatFactorType, 242
MatFDColoring, 138,139
MatFDColoringCreate, 138–140
MatFDColoringSetFromOptions, 138,140
MatFDColoringSetFunction, 138
MatFDColoringSetParameters, 139
MatGetDiagonal, 74
MatGetFactor, 242
MatGetLocalSubMatrix, 72
MatGetOrdering, 89,166,241–243
MatGetOwnershipRange, 35,66,70
MatGetRow, 76
MatILUFactor, 196
MatILUFactorSymbolic, 196,213
MATIS, 84,93
MATLAB, 191
MatLoad, 220
MatLUFactor, 242
MatLUFactorNumeric, 196,242
MatLUFactorSymbolic, 213,242
MatMFFDDSSetUmin, 121
MatMFFDGetH, 122
MatMFFDRegister, 121
MatMFFDResetHHistory, 122
MatMFFDSetFunctionError, 121
MatMFFDSetHHistory, 122
MatMFFDSetType, 121
MatMFFDWPSetComputeNormU, 122
MATMPIAIJ, 69,70,76,90,104
MatMPIAIJSetPreallocation, 35,104,126
MATMPIBAIJ, 66,70,76,90,220
MatMPISBAIJSetPreallocation, 35
MatMult, 28,37,72–74,196,198
MatMultAdd, 73,74,196
MatMultTranspose, 73,74
MatMultTransposeAdd, 73,74
MatMumpsSetIcntl, 104
MATNEST, 72,99
MatNorm, 73,74
MatNullSpace, 94,102
MatNullSpaceCreate, 102
MatNullSpaceCreateRigidBody, 92
MATORDERING1WD, 89
MATORDERINGNATURAL, 89
MATORDERINGND, 89
MATORDERINGQMD, 89
MATORDERINGRCM, 89
MatOrderingRegister, 242
MatOrderingType, 241,242
MatPartitioning, 77,78,166
MatPartitioningApply, 78
MatPartitioningCreate, 78
MatPartitioningDestroy, 78
MatPartitioningSetAdjacency, 78
MatPartitioningSetFromOptions, 78
MatPtAP, 93
MatReorderForNonzeroDiagonal, 243
MatRestoreRow, 76
MatReuse, 74,241
matrices, 31,65
matrix ordering, 242
matrix-free Jacobians, 121
matrix-free methods, 73,79
MatScale, 74
MATSEQAIJ, 68,75,76,88,104
MatSeqAIJSetPreallocation, 35,104,126
MATSEQBAIJ, 66,76,88
MATSEQDENSE, 76
MATSEQSBAIJ, 88
MatSeqSBAIJSetPreallocation, 35
MatSetBlockSize, 92
MatSetFromOptions, 28,35,107,126,184
252
![](asset-fd.png)
INDEX PETSc 3.8 March 24, 2018
MatSetLocalToGlobalMapping, 58
MatSetNearNullSpace, 92,94
MatSetNullSpace, 93,102
MatSetOption, 36,66,75,76
MatSetSizes, 28,31,32,35,65,107,126,184
MatSetTransposeNullSpace, 102
MatSetType, 32,104
MatSetUp, 28,107,184
MatSetValues, 28,32,36,58,65,66,68,75,
76,110,112,132,133,175,187,188,
201
MatSetValuesBlocked, 66,91
MatSetValuesBlockedLocal, 66,72
MatSetValuesLocal, 58,72
MatSetValuesStencil, 58
MATSHELL, 121,160
MatShellGetContext, 74
MatShellSetOperation, 74,123
MatShift, 74
MatSolve, 196,243
MatSolve(), 243
MatSolveAdd, 243
MATSOLVERBAS, 103
MATSOLVERCHOLMOD, 103
MATSOLVERCUSPARSE, 103
MATSOLVERESSL, 103
MATSOLVERKLU, 103
MATSOLVERLUSOL, 9,103
MATSOLVERMATLAB, 103
MATSOLVERMKL PARDISO, 103
MATSOLVERMUMPS, 103
MatSolverPackage, 103,242
MATSOLVERPASTIX, 103
MATSOLVERSUPERLU, 103
MATSOLVERSUPERLU DIST, 103
MATSOLVERUMFPACK, 103
MatSolveTranspose, 243
MatSolveTransposeAdd, 243
MatSORType, 88
MatStructure, 74
MatTranspose, 74
MatType, 32,74,76,103,220
MatView, 73,191,192,220
MatZeroEntries, 74,75,114
MatZeroRows, 75,76
MatZeroRowsColumns, 76
MatZeroRowsColumnsIS, 76
MatZeroRowsIS, 75
MatZeroRowsLocal, 76
MatZeroRowsLocalIS, 76
memory allocation, 213
memory leaks, 213
MPI, 23,224
MPI Comm, 26,31,32,34,47,49,52,53,55,
59,60,62,65,69,71,73,74,77–79,
85,97,102,112,145,168,191–193,
220,223,225,243,244
MPI Comm rank, 125,183
MPI Comm size, 27,107,125,183
MPI Finalize, 26
MPI Finalize(), 26
MPI Init, 26,217
MPI Init(), 26
MPI Wtime, 202
mpiexec, 23
multigrid, 97
multigrid, additive, 97
multigrid, full, 97
multigrid, Kaskade, 97
multigrid, multiplicative, 97
multiplicative preconditioners, 96
nested dissection, 89
Newton-like methods, 105
non-uniform memory access, 207
nonlinear equation solvers, 105
NORM 1, 49,51,73
NORM 2, 30,38,49,51,129,177,178,188
NORM FROBENIUS, 73
NORM INFINITY, 49,73
NormType, 49–51,73,74
null space, 102
ODE solvers, 143,153
one-way dissection, 89
OpenGL, 225
options, 217
ordering, 242
orderings, 51,53,87,89
overlapping Schwarz, 89
partitioning, 77
PC, 21,27,29,32,38,79,80,87–91,95–98,
106,108,112,123,124,126,127,136,
137,153,183–185,196,201,219,244,
245
PC ASM BASIC, 90,118
PC ASM INTERPOLATE, 90
PC ASM NONE, 90
253
![](asset-fe.png)
INDEX PETSc 3.8 March 24, 2018
PC ASM RESTRICT, 90,118
PC COMPOSITE ADDITIVE, 96
PC COMPOSITE MULTIPLICATIVE, 95,96
PC FIELDSPLIT SCHUR PRE USER, 101
PC GASM BASIC, 90
PC GASM INTERPOLATE, 90
PC GASM NONE, 90
PC GASM RESTRICT, 90
PC LEFT, 245
PC MG ADDITIVE, 97
PC MG CYCLE W, 97
PC MG FULL, 97
PC MG KASKADE, 97
PC MG MULTIPLICATIVE, 97
PC RIGHT, 82,245
PCApply, 196,245
PCApplyBAorAB, 245
PCApplyBAorABTranspose, 245
PCApplyRichardson, 245
PCApplyTranspose, 245
PCASM, 87,90,91,118
PCASMGetSubKSP, 89
PCASMSetLocalSubdomains, 90
PCASMSetOverlap, 90
PCASMSetTotalSubdomains, 90
PCASMSetType, 90
PCASMType, 90,118
PCBDDC, 87,93–95
PCBDDCSetChangeOfBasisMat, 94
PCBDDCSetCoarseningRatio, 94
PCBDDCSetDirichletBoundaries, 94
PCBDDCSetDirichletBoundariesLocal, 94
PCBDDCSetDofsSplitting, 94
PCBDDCSetLevels, 94
PCBDDCSetLocalAdjacencyGraph, 94
PCBDDCSetNeumannBoundaries, 94
PCBDDCSetNeumannBoundariesLocal, 94
PCBDDCSetPrimalVerticesLocalIS, 94
PCBJACOBI, 87
PCBJacobiGetSubKSP, 89,137
PCBJacobiSetLocalBlocks, 90
PCBJacobiSetTotalBlocks, 90
PCCHOLESKY, 87
PCCOMPOSITE, 87,95
PCCompositeAddPC, 95
PCCompositeGetPC, 96
PCCompositeSetType, 95
PCCompositeType, 95
PCCreate, 244
PCDestroy, 245
PCEISENSTAT, 87,88
PCEisenstatSetNoDiagonalScaling, 88
PCEisenstatSetOmega, 88
PCFactorGetMatrix, 103
PCFactorSetAllowDiagonalFill, 87
PCFactorSetFill, 96
PCFactorSetLevels, 87
PCFactorSetMatSolverPackage, 103
PCFactorSetReuseFill, 87
PCFactorSetReuseOrdering, 87
PCFactorSetUseInPlace, 80,87,89
PCFIELDSPLIT, 99
PCFieldSplitSetDiagUseAmat, 100
PCFieldSplitSetFields, 99
PCFieldSplitSetIS, 99
PCFieldSplitSetOffDiagUseAmat, 100
PCFieldSplitSetSchurFactType, 101
PCFieldSplitSetType, 100
PCGAMG, 87,91
PCGAMGSetNSmooths, 91
PCGAMGSetSquareGraph, 92
PCGAMGSetThreshold, 92
PCGAMGSetType, 91
PCGASM, 87,90,91
PCGASMCreateSubdomains2D, 91
PCGASMSetOverlap, 91
PCGASMSetSubdomains, 91
PCGASMSetTotalSubdomains, 91
PCGASMSetUseDMSubdomains, 91
PCGASMType, 90
PCGetOperators, 95,244
PCHYPRESetDiscreteCurl, 92
PCHYPRESetDiscreteGradient, 92
PCICC, 87
PCILU, 87,89,196
PCISSetSubdomainDiagonalScaling, 94
PCISSetSubdomainScalingFactor, 94
PCISSetUseStiffnessScaling, 94
PCJACOBI, 29,87
PCKSP, 87,96,97
PCKSPGetKSP, 96
PCLSC, 101
PCLU, 87,103
PCMG, 91,97,117,118,218
PCMGCycleType, 97
PCMGGetCoarseSolve, 97
PCMGGetSmoother, 97,98
PCMGGetSmootherDown, 98
254
![](asset-ff.png)
INDEX PETSc 3.8 March 24, 2018
PCMGGetSmootherUp, 98
PCMGSetCycleType, 97
PCMGSetInterpolation, 98
PCMGSetLevels, 97
PCMGSetNumberSmoothDown, 97
PCMGSetNumberSmoothUp, 97
PCMGSetR, 98
PCMGSetResidual, 98
PCMGSetRestriction, 98
PCMGSetRhs, 98
PCMGSetType, 97
PCMGSetX, 98
PCMGType, 97
PCNN, 9
PCNONE, 87,108,121,184
PCREDISTRIBUTE, 76
PCRichardsonConvergedReason, 245
PCSetOperators, 244,245
PCSetType, 29,87,91,95–98,103,108,123,
126,184,244
PCSetUp, 196
PCSetUseAmat, 96,100
PCSHELL, 87,121,123,126
PCShellGetContext, 95
PCShellSetApply, 95,123,126
PCShellSetContext, 95
PCShellSetSetUp, 95
PCSide, 82,245
PCSOR, 87,88,98
PCSORSetIterations, 88
PCSORSetOmega, 88
PCSORSetSymmetric, 88
PCType, 87,89,103,244
performance pitfalls, 211
performance tuning, 205
PETSC COMM SELF, 26,47,49,61,67,71,
72,179,184
PETSC COMM WORLD, 26–28,30,35,37,
38,52,107,109,125–129,134–137,
177,181,183,184,217,228
PETSC COPY VALUES, 59,61
PETSC DECIDE, 28,31,35,37,47,49,55,
69,71,73,107,126,177,184
PETSC DEFAULT, 29,38,84,108,137,192
PETSC DETERMINE, 32,47,166
PETSC DIR, 23
PETSC FALSE, 27,35,37,66,75,76,134,
135,167,188,219
PETSC FP TRAP OFF, 223
PETSC FP TRAP ON, 223
PETSC HAVE FORTRAN CAPS, 174
PETSC HAVE FORTRAN UNDERSCORE, 174
PETSC LIB, 237
PETSC MATLAB ENGINE , 193
PETSC NULL CHARACTER, 174
PETSC NULL DOUBLE, 174
PETSC NULL INTEGER, 174
PETSC NULL SCALAR, 174
PETSC OPTIONS, 218
PETSC TRUE, 29,36,86,100,135,166,167,
182,201,219,228
PETSC USE COMPLEX, 236
PETSC USE DEBUG, 236
PETSC USE LOG, 236
PETSC VIEWER ASCII IMPL, 220
PETSC VIEWER ASCII MATLAB, 191,220
PETSC VIEWER BINARY MATLAB, 192
PETSC VIEWER DEFAULT, 220
PETSC VIEWER DRAW SELF, 219,225
PETSC VIEWER DRAW WORLD, 49,73,219,
225
PETSC VIEWER LOAD BALANCE, 221
PETSC VIEWER STDOUT , 191
PETSC VIEWER STDOUT SELF, 180,191,
219,220
PETSC VIEWER STDOUT WORLD, 30,37,
49,73,108,191,219,220
PetscAbortErrorHandler, 223
PetscBag, 9
PetscBinaryRead, 192
PetscBool, 27,35,82,87,89,102,107,116,
122,124,125,134–136,152,169,177,
181,183,188,202,219,227,245
PetscClassId, 199
PetscClassIdRegister, 199
PetscCopyMode, 53,59,60
PetscDraw, 181,225–227
PetscDrawAxis, 85,181,227
PetscDrawAxis*(), 85
PetscDrawAxisDraw, 226,243
PetscDrawAxisSetColors, 181,227,228
PetscDrawAxisSetLabels, 182,227,228
PetscDrawCreate, 181,225,228
PetscDrawDestroy, 182,228
PetscDrawFlush, 226
PetscDrawLG, 85,181,226,227
PetscDrawLG*(), 85
PetscDrawLGAddPoint, 182,226,228
255
![](asset-100.png)
INDEX PETSc 3.8 March 24, 2018
PetscDrawLGAddPoints, 226
PetscDrawLGCreate, 181,226,228
PetscDrawLGDestroy, 85,182,227,228
PetscDrawLGDraw, 182,227,228
PetscDrawLGGetAxis, 181,227,228
PetscDrawLGReset, 227
PetscDrawLGSave, 228
PetscDrawLGSetFromOptions, 228
PetscDrawLGSetLegend, 228
PetscDrawLGSetLimits, 227
PetscDrawLGSetUseMarkers, 182,228
PetscDrawLine, 226
PetscDrawOpenX, 225
PetscDrawPause, 226
PetscDrawSetCoordinates, 225
PetscDrawSetDoubleBuffer, 226
PetscDrawSetFromOptions, 181,221,225,228
PetscDrawSetSaveMovie, 225
PetscDrawSetType, 225
PetscDrawSetViewPort, 225
PetscDrawSP, 86
PetscDrawSP*(), 86
PetscDrawString, 226
PetscDrawStringBoxed, 226
PetscDrawStringCentered, 226
PetscDrawStringGetSize, 226
PetscDrawStringSetSize, 226
PetscDrawStringVertical, 226
PetscDrawViewPorts, 227
PetscDrawViewPortsCreate, 228
PetscDrawViewPortsDestroy, 228
PetscDrawViewPortsSet, 228
PetscError, 176,222,223
PetscErrorCode, 26,27,35,74,84,85,95,
98,106,107,109–111,113,116,119,
124,125,130,132–138,144,145,152,
156,157,159,160,177,179,181,183,
186–189,222–224,227,242
PetscFClose, 176
PetscFinalize, 26,30,39,109,129,178,180,
182,185,195,200,214,219,228
PetscFOpen, 176
PetscFPrintf, 38,176,202
PetscFunctionBeginUser, 130,132,134–136,
224
PetscFunctionReturn, 130,131,133,134,136,
137,224
PetscInfo, 176,201
PetscInfoActivateClass, 201
PetscInfoAllow, 201
PetscInfoDeactivateClass, 201
PetscInitialize, 25–27,35,107,125,176,177,
179,181,183,195,200,214,217,218,
228
PetscInt, 27,31,32,35,47–55,58–62,65–67,
69,71,74–77,80,81,84–90,96–98,
102,107,110,111,114,117–119,124,
125,130,132,133,135–137,141,146,
149,152,154–156,160,163,165–169,
175,177,179,181,183,187,189,219,
226,227,245
PetscLayout, 165
PetscLogEvent, 199
PetscLogEventActivate, 201
PetscLogEventActivateClass, 201
PetscLogEventBegin, 199,200
PetscLogEventDeactivate, 201
PetscLogEventDeactivateClass, 201
PetscLogEventEnd, 199,200
PetscLogEventRegister, 199,201
PetscLogFlops, 199,200
PetscLogStage, 35,200
PetscLogStagePop, 36,200
PetscLogStagePush, 36,200
PetscLogStageRegister, 36,200
PetscLogTraceBegin, 223
PetscLogView, 195,200
PetscMalloc, 213
PetscMalloc1, 223
PetscMallocDump, 214
PetscMallocDumpLog, 214
PetscMallocGetCurrentUsage, 214
PetscMallocGetMaximumUsage, 214
PetscMallocSetDumpLog, 214
PetscMatlabEngine, 192,193
PetscMatlabEngineCreate, 192
PetscMatlabEngineEvaluate, 193
PetscMatlabEngineGet, 193
PetscMatlabEngineGetArray, 193
PetscMatlabEngineGetOutput, 193
PetscMatlabEnginePut, 192
PetscMatlabEnginePutArray, 193
PetscMemoryGetCurrentUsage, 214
PetscMemoryGetMaximumUsage, 214
PetscMemorySetGetMaximumUsage, 214
PetscMPIInt, 27,107,124,183
PetscObject, 27,85,101,127,191–193,199
PetscObjectCompose, 101
256
![](asset-101.png)
INDEX PETSc 3.8 March 24, 2018
PetscObjectGetComm, 85
PetscObjectName, 191
PetscObjectSAWsGrantAccess, 222
PetscObjectSAWsSetBlock, 222
PetscObjectSAWsTakeAccess, 222
PetscObjectSetName, 27,127,180,191,193
PetscObjectSetName(), 191
PetscOffset, 179,186,187,189
PetscOptionsCreate, 177
PetscOptionsDestroy, 177
PetscOptionsGetBool, 27,29,37
PetscOptionsGetInt, 27,31,35,125,174,177,
181,219,228
PetscOptionsGetIntArray, 219
PetscOptionsGetReal, 127,219
PetscOptionsGetRealArray, 219
PetscOptionsGetString, 176,219
PetscOptionsGetStringArray, 176,219
PetscOptionsGetViewer, 221
PetscOptionsHasName, 107,126,127,185,219,
228
PetscOptionsInsertString, 29
PetscOptionsSetValue, 218
PetscPopErrorHandler, 176,223
PetscPreLoadBegin, 202
PetscPreLoadEnd, 202
PetscPreLoadStage, 202
PetscPrintf, 30,38,109,127–129,134–137,
176,202
PetscPushErrorHandler, 176,222
PetscPushSignalHandler, 223
PetscRandom, 35
PetscRandomCreate, 37
PetscRandomDestroy, 37
PetscRandomSetFromOptions, 37
PetscReal, 27,35,50,51,74,81,84–88,119,
121,124,125,133,135–137,139,144–
146,152,156,157,159,160,177,181,
183,188,219,225–228,243,245
PetscSAWsBlock, 222
PetscScalar, 27,29,31,32,35,48–51,54,57,
61,62,65,71,74–76,107,109–111,
122,125,130,132,135,152,166,167,
175,177,179,181,183,186–189,193,
224
PetscSection, 165,166
PetscSectionGetDof, 167
PetscSectionGetOffset, 167
PetscSectionGetStorageSize, 166
PetscSectionSetChart, 165
PetscSectionSetDof, 165,166
PetscSectionSetUp, 166
PetscSetDebugger, 176
PetscSetFPTrap, 223
PetscSignalHandlerDefault, 223
PetscSleep, 182
PetscStackSAWsGrantAccess, 222
PetscStackSAWsTakeAccess, 222
PetscStartMatlab, 192
PetscTime, 202
PetscTraceBackErrorHandler, 222,223
PetscViewer, 48,52,73,124,146,192,219–
221,225
PetscViewerASCIIGetPointer, 176
PetscViewerASCIIOpen, 191,220
PetscViewerBinaryGetDescriptor, 176
PetscViewerBinaryOpen, 220
PetscViewerDestroy, 129,192
PetscViewerDrawGetDraw, 225
PetscViewerDrawOpen, 73,126,133,220,225
PetscViewerFormat, 220
PetscViewerPopFormat, 191,220
PetscViewerPushFormat, 191,192,220
PetscViewerSocketOpen, 192,220
PetscViewerStringOpen, 176
PetscViewerStringSPrintf, 176
preconditioners, 86
preconditioning, 79,82
preconditioning, right and left, 245
profiling, 195,212
providing arrays for vectors, 49
Qt Creator, 231
quotient minimum degree, 89
relaxation, 88,97
reorder, 241
Residual Evaluation, 166
restart, 81
reverse Cuthill-McKee, 89
Richardson’s method, 245
Runge-Kutta, 145,150
running PETSc programs, 23
runtime options, 217
Sarkis, Marcus, 90
saws, 221
scatter, 60
SCATTER FORWARD, 60–63
257
![](asset-102.png)
INDEX PETSc 3.8 March 24, 2018
SCATTER REVERSE, 61–63
ScatterMode, 62,63
SETERRQ, 27,107,174,223,224
SETERRQ(), 223
signals, 223
singular systems, 102
smoothing, 97
SNES, 9,21,32,105,106,108–114,116–124,
127,130,132–136,138,140,141,143,
153,182–188,196,201,219
SNES FAS ADDITIVE, 117
SNES FAS FULL, 117
SNES FAS KASKADE, 117
SNES FAS MULTIPLICATIVE, 117
SNESANDERSON, 115
SNESASPIN, 115,118
SNESCOMPOSITE, 115,140
SNESCompositeAddSNES, 141
SNESCompositeGetSNES, 141
SNESCompositeSetType, 140
SNESComputeFunction, 188
SNESComputeJacobianDefaultColor, 138,139
SNESConvergedReason, 119
SNESCreate, 107,112,125,184
SNESDestroy, 109,113,129,185
SNESFAS, 115,117
SNESFASGetCoarseSolve, 118
SNESFASGetCycleSNES, 118
SNESFASGetInjection, 118
SNESFASGetInterpolation, 118
SNESFASGetLevels, 117
SNESFASGetRestriction, 118
SNESFASGetSmoother, 117
SNESFASGetSmootherDown, 117
SNESFASGetSmootherUp, 117
SNESFASSetCycles, 117
SNESFASSetType, 117
SNESFASType, 117
SNESGetFunction, 108,120
SNESGetIterationNumber, 108,129,135,136,
185
SNESGetKSP, 108,123,126,137,184
SNESGetLineSearch, 114,127,185
SNESGetNPC, 140
SNESGetSolution, 120,134
SNESGetTolerances, 119,128
SNESKSPONLY, 115
SNESLineSearch, 114,116,124,134–136,183,
188
SNESLINESEARCHBASIC, 115,117
SNESLINESEARCHBT, 115,116
SNESLINESEARCHCP, 115–117
SNESLineSearchGetSNES, 135,136,188
SNESLineSearchGetVecs, 188
SNESLINESEARCHL2, 115–117
SNESLineSearchRegister, 116
SNESLineSearchSetComputeNorms, 116
SNESLineSearchSetNorms, 188
SNESLineSearchSetOrder, 114
SNESLineSearchSetPostCheck, 127,134,136
SNESLineSearchSetPreCheck, 128,134
SNESLineSearchSetType, 185
SNESLINESEARCHSHELL, 115
SNESLineSearchShellSetUserFunc, 185
SNESLineSearchType, 115
SNESMonitorDefault, 119
SNESMonitorSAWs, 222
SNESMonitorSet, 119,126,133
SNESMS, 115
SNESNASM, 115
SNESNASMSetSubdomains, 118
SNESNASMSetType, 118
SNESNCG, 115
SNESNCGSetType, 116
SNESNEWTONLS, 114,115
SNESNEWTONTR, 115,116,118
SNESNGMRES, 115
SNESNGS, 115
SNESNRICHARDSON, 115
SNESQN, 115
SNESQNSetRestartType, 117
SNESQNSetScaleType, 117
SNESSetConvergenceTest, 119
SNESSetDM, 161,168
SNESSetFromOptions, 108,112,122,127,185
SNESSetFunction, 107,113,126,130,184
SNESSetJacobian, 102,107,108,113,122,
123,126,139,184
SNESSetNPCSide, 140
SNESSetOptionsPrefix, 218
SNESSetTolerances, 119
SNESSetTrustRegionTolerance, 119
SNESSetType, 112
SNESSHELL, 115
SNESSolve, 108,112,113,129,185
SNESTEST, 115
SNESType, 112,115,140,141
SNESView, 214
258
![](asset-103.png)
INDEX PETSc 3.8 March 24, 2018
SNESVINEWTONRSLS, 115
SNESVINEWTONSSLS, 115
SNESVISetVariableBounds, 140
SOR, 88
SOR BACKWARD SWEEP, 88
SOR FORWARD SWEEP, 88
SOR LOCAL BACKWARD SWEEP, 88
SOR LOCAL FORWARD SWEEP, 88
SOR LOCAL SYMMETRIC SWEEP, 88
SOR SYMMETRIC SWEEP, 88
spectrum, 86
SSOR, 88
stride, 59
submatrices, 241
Sundials, 153
SUNDIALS MODIFIED GS, 154
SUNDIALS UNMODIFIED GS, 154
symbolic factorization, 242
tags, in Vi/Vim, 230
text, drawing, 226
time, 202
timing, 195,212
trust region, 105,116
TS, 21,143–146,148,149,152–157,159,160,
167,196,201
TS EQ IMPLICIT, 147
TS EXACTFINALTIME INTERPOLATE, 146
TS EXACTFINALTIME MATCHSTEP, 146
TS EXACTFINALTIME STEPOVER, 146
TS LINEAR, 145,150,151
TS NONLINEAR, 145
TSAdapt, 149,151
TSADAPTBASIC, 151
TSADAPTGLEE, 151,152
TSADAPTNONE, 151
TSAdjointSetRHSJacobian, 156,157
TSAdjointSolve, 156,157
TSARKIMEX, 146,147,153
TSARKIMEXSetFullyImplicit, 147
TSBEULER, 145,155
TSCN, 145,155,157
TSComputeIFunctionLinear, 150,151
TSComputeIJacobianConstant, 150
TSComputeRHSFunctionLinear, 150
TSComputeRHSJacobianConstant, 150
TSCreate, 145
TSDestroy, 146
TSEULER, 145
TSExactFinalTimeOption, 146
TSGetAuxSolution, 149
TSGetCostIntegral, 156
TSGetSolutionComponents, 149
TSGetTimeError, 149
TSGetTimeStep, 146
TSGLEE, 147,149,151,152
TSGLEE23, 149
TSGLEE24, 149
TSGLEE35, 149
TSGLEEEXRK2A, 149
TSGLEERK285EX, 149
TSGLEERK32G1, 149
TSGLLE, 145
TSMonitorDrawSolution, 151
TSMonitorLGError, 151
TSMonitorLGSolution, 151
TSMonitorSolution, 151
TSMonitorSolutionVTK, 151
TSProblemType, 145
TSPSEUDO, 145
TSPseudoIncrementDtFromInitialDt, 160
TSPseudoSetTimeStep, 160
TSPseudoSetTimeStepIncrement, 160
TSPseudoTimeStepDefault, 160
TSRK, 9,145,155
TSRK1FE, 150
TSRK2A, 150
TSRK3, 150
TSRK3BS, 150
TSRK4, 150
TSRK5BS, 150
TSRK5DP, 150
TSRK5F, 150
TSROSW, 153
TSSetCostGradients, 155,156
TSSetCostIntegrand, 156,157
TSSetDM, 168
TSSetEquationType, 147
TSSetEventHandler, 152
TSSetExactFinalTime, 146
TSSetIFunction, 144,146,147,150,151,157
TSSetIJacobian, 102,144,146,150,151,157
TSSetMaxSteps, 146
TSSetMaxTime, 146
TSSetOptionsPrefix, 218
TSSetProblemType, 145,150,151
TSSetRHSFunction, 144,146,150,159
TSSetRHSJacobian, 145,150,157,160
259
![](asset-104.png)
INDEX PETSc 3.8 March 24, 2018
TSSetSaveTrajectory, 155
TSSetSolution, 145
TSSetSolutionFunction, 151
TSSetTime, 146
TSSetTimeError, 149
TSSetTimeStep, 146
TSSetTolerances, 151
TSSetType, 145,153
TSSetUp, 146
TSSolve, 146,155,156
TSStep, 146
TSSUNDIALS, 9,145,153
TSSundialsGetPC, 153
TSSundialsSetGramSchmidtType, 153
TSSundialsSetMaxl, 154
TSSundialsSetTolerance, 153
TSSundialsSetType, 153
TSTHETA, 145,150,153,155
TSTrajectory, 158
TSType, 145,153
TSView, 146
V-cycle, 97
Vec, 9,21,27,31,32,34,35,47–51,54–58,
60–63,73–76,80,86,95,98,102,106,
108–113,120,121,124,125,130,132,
134–137,140,144–146,149,152,155–
157,159,160,166,175,177,179,183,
186–189,192,196,201,219,220,243–
245
VecAbs, 50
VecAssemblyBegin, 48,179
VecAssemblyEnd, 48,179,191
VecAXPBY, 50
VecAXPY, 30,38,50,129,178,188,196
VecAYPX, 50
VecCopy, 50,136,137,178,180,196
VecCreate, 27,31,34,36,37,47,73,107,177
VecCreateGhost, 62
VecCreateGhostWithArray, 62
VecCreateMPI, 47,53,62,69,73
VecCreateMPIWithArray, 49
VecCreateSeq, 47,61,179,184
VecCreateSeqWithArray, 49
VecDestroy, 30,39,49,98,109,129,178,180,
185
VecDestroyVecs, 49,175,176
VecDestroyVecsF90, 176
VecDot, 50,51,177,196,212
VecDotBegin, 51
VecDotEnd, 51
VecDuplicate, 28,31,37,49,56,61,62,107,
125,127,177,179,184
VecDuplicateVecs, 49,56,62,175,176
VecDuplicateVecsF90, 176
VecGetArray, 48,50,61,108,109,111,166,
167,176,178,180,186,189,190,212
VecGetArrayF90, 176,189
VecGetArrayRead, 50,109–111,186,187
VecGetArrayReadF90, 176
VecGetArrays, 176
VecGetLocalSize, 50,189
VecGetOwnershipRange, 49
VecGetSize, 50
VecGetValues, 48,61
VecGhostGetLocalForm, 62,63
VecGhostRestoreLocalForm, 63
VecGhostUpdateBegin, 63
VecGhostUpdateEnd, 63
VecLoad, 220
VecMax, 50
VecMAXPY, 50
VecMDot, 50,51,196,212
VecMDotBegin, 51
VecMDotEnd, 51
VecMin, 50
VecMTDot, 50
VecMTDotBegin, 51
VecMTDotEnd, 51
VecNorm, 30,38,49–51,129,177,178,188,
196
VecNormBegin, 51
VecNormEnd, 51
VecPointwiseDivide, 50
VecPointwiseMult, 50
VecReciprocal, 50
VecRestoreArray, 50,108–111,167,180,186,
187,189
VecRestoreArrayF90, 176
VecRestoreArrayRead, 50,109–112,186,188
VecRestoreArrayReadF90, 176
VecRestoreArrays, 176
Vecs, 102
VecScale, 50,177,196
VecScatter, 9,47,56,60–62,118
VecScatterBegin, 60–63,198
VecScatterCreate, 60,61
VecScatterDestroy, 60,61
260
![](asset-105.png)
INDEX PETSc 3.8 March 24, 2018
VecScatterEnd, 60–63,198
VecSet, 28,29,31,37,48,50,108,113,129,
130,175,177,185
VecSetFromOptions, 28,31,36,37,47,107,
166,177
VecSetLocalToGlobalMapping, 54,58
VecSetRandom, 37
VecSetSizes, 28,31,36,37,47,107,166,177
VecSetType, 31
VecSetValues, 31,48,61,175,178,179,189,
190,201
VecSetValuesLocal, 54,58
VecShift, 50
VecSum, 50
VecSwap, 50
VecTDot, 50
VecTDotBegin, 51
VecTDotEnd, 51
vector values, getting, 61
vector values, setting, 48
vectors, 31,47
vectors, setting values with local numbering,
54
vectors, user-supplied arrays, 49
vectors, with ghost values, 62
VecType, 220
VecView, 37,48,108,134,180,191,192,220
VecWAXPY, 50
Vi, 230
Vim, 230
visual studio, 233
W-cycle, 97
wall clock time, 202
X windows, 225
xcode, 233
zero pivot, 102
261
![](asset-106.png)
INDEX PETSc 3.8 March 24, 2018
262
Bibliography
[1] U.M. Ascher, S.J. Ruuth, and R.J. Spiteri. Implicit-explicit Runge-Kutta methods for time-
dependent partial differential equations. Applied Numerical Mathematics, 25:151–167, 1997.
[2] Satish Balay, William D. Gropp, Lois Curfman McInnes, and Barry F. Smith. Efficient manage-
ment of parallelism in object oriented numerical software libraries. In E. Arge, A. M. Bruaset,
and H. P. Langtangen, editors, Modern Software Tools in Scientific Computing, pages 163–202.
Birkhauser Press, 1997.
[3] S. Boscarino, L. Pareschi, and G. Russo. Implicit-explicit Runge-Kutta schemes for hyperbolic
systems and kinetic equations in the diffusion limit. Arxiv preprint arXiv:1110.4375, 2011.
[4] Peter N. Brown and Youcef Saad. Hybrid Krylov methods for nonlinear systems of equations.
SIAM J. Sci. Stat. Comput., 11:450–481, 1990.
[5] J.C. Butcher, Z. Jackiewicz, and W.M. Wright. Error propagation of general linear methods
for ordinary differential equations. Journal of Complexity, 23(4-6):560–580, 2007.
[6] X.-C. Cai and M. Sarkis. A restricted additive Schwarz preconditioner for general sparse
linear systems. Technical Report CU-CS 843-97, Computer Science Department, University of
Colorado-Boulder, 1997. (accepted by SIAM J. of Scientific Computing).
[7] E.M. Constantinescu. Estimating global errors in time stepping. ArXiv e-prints, March 2016.
[8] E.M. Constantinescu and A. Sandu. Extrapolated implicit-explicit time stepping. SIAM
Journal on Scientific Computing, 31(6):4452–4477, 2010.
[9] J. E. Dennis, Jr. and Robert B. Schnabel. Numerical Methods for Unconstrained Optimization
and Nonlinear Equations. Prentice-Hall, Englewood Cliffs, NJ, 1983.
[10] S. Eisenstat. Efficient implementation of a class of CG methods. SIAM J. Sci. Stat. Comput.,
2:1–4, 1981.
[11] S. C. Eisenstat and H. F. Walker. Choosing the forcing terms in an inexact Newton method.
SIAM J. Scientific Computing, 17:16–32, 1996.
[12] S.C. Eisenstat, H.C. Elman, and M.H. Schultz. Variational iterative methods for nonsymmetric
systems of linear equations. SIAM Journal on Numerical Analysis, 20(2):345–357, 1983.
[13] R. Freund, G. H. Golub, and N. Nachtigal. Iterative Solution of Linear Systems, pages 57–100.
Acta Numerica. Cambridge University Press, 1992.
[14] Roland W. Freund. A transpose-free quasi-minimal residual algorithm for non-Hermitian linear
systems. SIAM J. Sci. Stat. Comput., 14:470–482, 1993.
263
![](asset-108.png)
BIBLIOGRAPHY PETSc 3.8 March 24, 2018
[15] F.X. Giraldo, J.F. Kelly, and E.M. Constantinescu. Implicit-explicit formulations of a three-
dimensional nonhydrostatic unified model of the atmosphere (NUMA). SIAM Journal on
Scientific Computing, 35(5):B1162–B1194, 2013.
[16] William Gropp and et. al. MPICH Web page. http://www.mpich.org.
[17] William Gropp, Ewing Lusk, and Anthony Skjellum. Using MPI: Portable Parallel Program-
ming with the Message Passing Interface. MIT Press, 1994.
[18] Virginia Herrarte and Ewing Lusk. Studying parallel program behavior with Upshot. Technical
Report ANL-91/15, Argonne National Laboratory, August 1991.
[19] Magnus R. Hestenes and Eduard Steifel. Methods of conjugate gradients for solving linear
systems. J. Research of the National Bureau of Standards, 49:409–436, 1952.
[20] Tobin Isaac, Georg Stadler, and Omar Ghattas. Solution of nonlinear stokes equations dis-
cretized by high-order finite elements on nonconforming and anisotropic meshes, with appli-
cation to ice sheet dynamics. SIAM Journal on Scientific Computing, 2015.
[21] K.E. Jansen, C.H. Whiting, and G.M. Hulbert. A generalized-αmethod for integrating the
filtered Navier–Stokes equations with a stabilized finite element method. Computer Methods
in Applied Mechanics and Engineering, 190(3):305–319, 2000.
[22] C.A. Kennedy and M.H. Carpenter. Additive Runge-Kutta schemes for convection-diffusion-
reaction equations. Appl. Numer. Math., 44(1-2):139–181, 2003.
[23] D.I. Ketcheson. Highly efficient strong stability-preserving Runge–Kutta methods with low-
storage implementations. SIAM Journal on Scientific Computing, 30(4):2113–2136, 2008.
[24] Jorge J. Mor´e, Danny C. Sorenson, Burton S. Garbow, and Kenneth E. Hillstrom. The
MINPACK project. In Wayne R. Cowell, editor, Sources and Development of Mathematical
Software, pages 88–111, 1984.
[25] MPI: A message-passing interface standard. International J. Supercomputing Applications,
8(3/4), 1994.
[26] Yvan Notay. Flexible Conjugate Gradients. SIAM Journal on Scientific Computing,
22(4):1444–1460, 2000.
[27] C. C. Paige and M. A. Saunders. Solution of sparse indefinite systems of linear equations.
12:617–629, 1975.
[28] L. Pareschi and G. Russo. Implicit-explicit Runge-Kutta schemes and applications to hyper-
bolic systems with relaxation. Journal of Scientific Computing, 25(1):129–155, 2005.
[29] M. Pernice and H. F. Walker. NITSOL: A Newton iterative solver for nonlinear systems. SIAM
J. Sci. Stat. Comput., 19:302–318, 1998.
[30] J. Rang and L. Angermann. New Rosenbrock W-methods of order 3 for partial differential
algebraic equations of index 1. BIT Numerical Mathematics, 45(4):761–787, 2005.
[31] Youcef Saad. A flexible inner-outer preconditioned GMRES algorithm. SIAM Journal on
Scientific Computing, 14(2):461–469, 1993.
264
[32] Youcef Saad and Martin H. Schultz. GMRES: A generalized minimal residual algorithm for
solving nonsymmetric linear systems. SIAM J. Sci. Stat. Comput., 7:856–869, 1986.
[33] A. Sandu, J.G. Verwer, J.G. Blom, E.J. Spee, G.R. Carmichael, and F.A. Potra. Benchmarking
stiff ode solvers for atmospheric chemistry problems II: Rosenbrock solvers. Atmospheric
Environment, 31(20):3459–3472, 1997.
[34] Barry F. Smith, Petter Bjørstad, and William D. Gropp. Domain Decomposition: Parallel
Multilevel Methods for Elliptic Partial Differential Equations. Cambridge University Press,
1996.
[35] Peter Sonneveld. CGS, a fast Lanczos-type solver for nonsymmetric linear systems. SIAM J.
Sci. Stat. Comput., 10:36–52, 1989.
[36] H. A. van der Vorst. BiCGSTAB: A fast and smoothly converging variant of BiCG for the
solution of nonsymmetric linear systems. SIAM J. Sci. Stat. Comput., 13:631–644, 1992.
[37] Wikipedia. Hasse diagram, 2015. http://en.wikipedia.org/wiki/Hasse_diagram.
![](asset-10a.png)
Argonne National Laboratory is a U.S. Department of Energy
laboratory managed by UChicago Argonne, LLC
Mathematics and Computer Science Division
Argonne National Laboratory
9700 South Cass Avenue, Bldg. 240
Argonne, IL 60439
www.anl.gov
Figure 29: