PubTeX Output 2016.02.26:1105 Linux Application Tuning Guide X86 64 007 5646 010
User Manual:
Open the PDF directly: View PDF .
Page Count: 128 [warning: Documents this large are best viewed by clicking the View PDF Link!]
- New Features
- Table of Contents
- About This Guide
- 1. The SGI Compiling Environment
- 2. Performance Analysis and Debugging
- 3. Monitoring Commands
- 4. Data Process and Placement Tools
- 5. Performance Tuning
- About Performance Tuning
- Single Processor Code Tuning
- Tuning Multiprocessor Codes
- Understanding Parallel Speedup and Amdahl's Law
- Gustafson's Law
- Floating-point Program Performance
- About MPI Application Tuning
- Using Transparent Huge Pages (THPs) in MPI and SHMEM Applications
- Enabling Huge Pages in MPI and SHMEM Applications on Systems Without THP
- 6. Flexible File I/O
- 7. I/O Tuning
- 8. Suggested Shortcuts and W orkarounds
- Index
![](asset-1.png)
Linux®Application Tuning Guide for SGI®
X86-64 Based Systems
007–5646–010
![](asset-2.png)
COPYRIGHT
© 2010–2016, SGI. All rights reserved; provided portions may be copyright in third parties, as indicated elsewhere herein. No
permission is granted to copy, distribute, or create derivative works from the contents of this electronic documentation in any manner,
in whole or in part, without the prior written permission of SGI.
LIMITED RIGHTS LEGEND
The software described in this document is "commercial computer software" provided with restricted rights (except as to included
open/free source) as specified in the FAR 52.227-19 and/or the DFAR 227.7202, or successive sections. Use beyond license provisions is
a violation of worldwide intellectual property laws, treaties and conventions. This document is provided with limited rights as defined
in 52.227-14.
TRADEMARKS AND ATTRIBUTIONS
Altix, ICE, NUMAlink, OpenMP, Performance Co-Pilot, SGI, the SGI logo, SHMEM, and UV are trademarks or registered trademarks
of Silicon Graphics International Corp. or its subsidiaries in the United States and other countries.
Cray is a registered trademark of Cray, Inc.
Dinkumware is a registered trademark of Dinkumware, Ltd.
Intel, GuideView, Itanium, KAP/Pro Toolset, Phi, VTune, and Xeon are trademarks or registered trademarks of Intel Corporation, in
the United States and other countries.
Oracle and Java are registered trademarks of Oracle and/or its affiliates.
Linux is a registered trademark of Linus Torvalds in several countries.
Red Hat and Red Hat Enterprise Linux are registered trademarks of Red Hat, Inc., in the United States and other countries.
PostScript is a trademark of Adobe Systems Incorporated.
SLES and SUSE are registered trademarks of SUSE LLC in the United States and other countries.
TotalView and TotalView Technologies are registered trademarks and TVD is a trademark of Rogue Wave Software, Inc.
Windows is a registered trademark of Microsoft Corporation in the United States and/or other countries.
All other trademarks are the property of their respective owners.
![](asset-3.png)
New Features
This revision includes the following updates:
• Miscellaneous editorial and technical corrections.
007–5646–010 iii
![](asset-5.png)
Record of Revision
Version Description
001 November 2010
Original publication.
002 February 2011
Supports the SGI Performance Suite 1.1 release.
003 November 2011
Supports the SGI Performance Suite 1.3 release.
004 May 2012
Supports the SGI Performance Suite 1.4 release.
005 November 2013
Supports the SGI Performance Suite 1.7 release.
006 November 2013
Supports the SGI Performance Suite 1.7 release and includes a
correction to the PerfSocket installation documentation.
007 May 2014
Supports the SGI Performance Suite 1.8 release.
008 May 2015
Supports the SGI Performance Suite 1.10 release.
009 November 2015
Supports the SGI Performance Suite 1.11 release.
010 May 2016
Supports the SGI Performance Suite 1.12 release.
007–5646–010 v
![](asset-7.png)
Contents
About This Guide .....................xiii
Related SGI Publications . . ................... xiii
Related Publications From Other Sources . . .............. xiv
Obtaining Publications . . . ................... xv
Conventions . ........................ xv
Reader Comments ....................... xvi
1. The SGI Compiling Environment . . . . .......... 1
About the Compiling Environment .................. 1
Compiler Overview ....................... 2
Environment Modules . . . ................... 3
Library Overview ....................... 4
Static Libraries ....................... 4
Dynamic Libraries . . . ................... 5
C/C++ Libraries ....................... 5
MPI and SHMEM Libraries ................... 6
2. Performance Analysis and Debugging . . . .......... 7
About Performance Analysis and Debugging . .............. 7
Determining System Configuration .................. 8
Sources of Performance Problems .................. 15
Profiling with perf ...................... 16
Profiling with PerfSuite .................... 16
Other Performance Analysis Tools .................. 17
About Debugging ....................... 17
007–5646–010 vii
![](asset-8.png)
Contents
Using the Intel Debugger . ................... 18
Using the GNU Data Display Debugger (GNU DDD) . . ......... 19
3. Monitoring Commands . ................. 23
About the Operating System Monitoring Commands . . . ......... 23
Operating System Monitoring Commands . . .............. 23
Using the w(1) command . ................... 24
Using the ps(1) Command . ................... 24
Using the top(1) Command ................... 25
Using the vmstat(8) Command .................. 25
Using the iostat(1) command .................. 26
Using the sar(1) command ................... 27
4. Data Process and Placement Tools . . . . .......... 29
About Nonuniform Memory Access (NUMA) Computers . . ......... 29
Distributed Shared Memory (DSM) . . . .............. 29
ccNUMA Architecture . . ................... 30
Cache Coherency . . . ................... 30
Non-uniform Memory Access (NUMA) . .............. 31
About the Data and Process Placement Tools . .............. 31
About cpusets and Control Groups (cgroups) .............. 33
Using cpusets ....................... 33
Using cgroups ....................... 34
dplace Command . . . ................... 35
omplace Command . . ................... 41
taskset Command . . ................... 42
numactl Command . . ................... 44
dlook Command . . . ................... 45
viii 007–5646–010
![](asset-9.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
5. Performance Tuning . . ................. 53
About Performance Tuning . ................... 53
Single Processor Code Tuning ................... 54
Getting the Correct Results ................... 54
Managing Heap Corruption Problems . . .............. 55
Using Tuned Code . . . ................... 55
Determining Tuning Needs ................... 56
Using Compiler Options to Optimize Performance . . . ......... 56
Tuning the Cache Performance .................. 61
Managing Memory . . . ................... 63
Memory Use Strategies . ................... 63
Memory Hierarchy Latencies .................. 63
Tuning Multiprocessor Codes . ................... 64
Data Decomposition . . . ................... 64
Measuring Parallelization and Parallelizing Your Code . . ......... 66
Using SGI MPI . . . ................... 66
Using OpenMP . . . ................... 67
Identifying OpenMP Nested Parallelism . .............. 67
Using Compiler Options ................... 68
Identifying Opportunities for Loop Parallelism in Existing Code ....... 68
Fixing False Sharing . . . ................... 69
Environment Variables for Performance Tuning ............. 69
Understanding Parallel Speedup and Amdahl’s Law . . . ......... 70
Adding CPUs to Shorten Execution Time . .............. 71
Understanding Parallel Speedup .................. 72
Understanding Superlinear Speedup . . . .............. 72
007–5646–010 ix
![](asset-a.png)
Contents
Understanding Amdahl’s Law .................. 73
Calculating the Parallel Fraction of a Program .............. 74
Predicting Execution Time with n CPUs . .............. 75
Gustafson’s Law ........................ 76
Floating-point Program Performance . . . .............. 76
About MPI Application Tuning ................... 77
MPI Application Communication on SGI Hardware . . . ......... 77
MPI Job Problems and Application Design . .............. 78
MPI Performance Tools . . ................... 79
Using Transparent Huge Pages (THPs) in MPI and SHMEM Applications . . . . . . 80
Enabling Huge Pages in MPI and SHMEM Applications on Systems Without THP . . . 82
6. Flexible File I/O ..................... 83
About FFIO . ........................ 83
Environment Variables . . . ................... 84
Simple Examples ....................... 85
Multithreading Considerations ................... 88
Application Examples . . . ................... 89
Event Tracing ........................ 90
System Information and Issues ................... 90
7. I/O Tuning . . ..................... 91
About I/O Tuning ....................... 91
Application Placement and I/O Resources . . .............. 91
Layout of Filesystems and XVM for Multiple RAIDs . . . ......... 92
8. Suggested Shortcuts and Workarounds . . .......... 95
Determining Process Placement ................... 95
x 007–5646–010
![](asset-b.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Example Using pthreads .................... 96
Example Using OpenMP . ................... 98
Resetting System Limits . . ...................100
Resetting the File Limit Resource Default . ..............101
Resetting the Default Stack Size ..................103
Avoiding Segmentation Faults ..................103
Resetting Virtual Memory Size ..................104
Linux Shared Memory Accounting ..................106
OFED Tuning Requirements for SHMEM . . ..............107
Setting Java Enviroment Variables ..................108
Index . .........................109
007–5646–010 xi
![](asset-d.png)
About This Guide
This publication explains how to tune C and Fortran application programs compiled
with an Intel compiler on SGI®UVTM series systems, SGI®ICETM clusters, and SGI®
RackableTM clusters.
This guide is written for experienced programmers who are familiar with Linux
commands and with either C or Fortran programming. The focus in this document is
on achieving the highest possible performance by exploiting the features of your SGI
system. The material assumes that you know the basics of software engineering and
that you are familiar with standard methods and data structures. If you are new to
programming or software design, this guide will not be of use to you.
Related SGI Publications
The SGI Foundation Software release notes and the SGI Performance Suite release
notes contain information about the specific software packages provided in those
products. The release notes also list SGI publications that provide information about
the products. The release notes are available in the following locations:
• Online at the SGI customer portal. After you log into the SGI customer portal, you
can access the release notes.
The SGI Foundation Software release notes are posted to the following website:
https://support1-sgi.custhelp.com/app/answers/detail/a_id/4983
The SGI Performance Suite release notes are posted to the following website:
https://support1-sgi.custhelp.com/app/answers/detail/a_id/6093
Note: You must sign into the SGI customer portal, at
https://support.sgi.com/login, in order for the preceding links to work.
• On the product media. The release notes reside in a text file in the /docs directory
on the product media. For example, /docs/SGI-MPI-1.x-readme.txt.
• On the system. After installation, the release notes and other product
documentation reside in the /usr/share/doc/packages/product directory.
007–5646–010 xiii
![](asset-e.png)
About This Guide
All SGI publications are available on the SGI support portal. The following software
publications provide information about Linux implementations on SGI systems:
•SGI Foundation Software (SFS) User Guide
This manual explains how to configure and use SFS, includes information about
basic configuration for features such as the hardware event tracker (HET), CPU
frequency scaling and partitioning.
•SGI Cpuset Software Guide
Explains how to use cpusets within your application program. Cpusets restrict
processes within a program to specific processors or memory nodes.
•SGI MPI and SGI SHMEM User Guide
Describes the industry-standard message passing protocol optimized for SGI
computers. This manual describes how to tune the run-time environment to
improve the performance of an MPI message passing application on SGI
computers. The tuning methods do not involve application code changes.
•MPInside Reference Guide
Documents the SGI MPInside MPI profiling tool.
SGI creates hardware manuals that are specific to each product line. The hardware
documentation typically includes a system architecture overview and describes the
major components. It also provides the standard procedures for powering on and
powering off the system, basic troubleshooting information, and important safety and
regulatory specifications.
Related Publications From Other Sources
Compilers and performance tool information for software that runs on SGI Linux
systems is available from a variety of sources. The following additional links might
be useful to you:
• The GNU hosts Debugging with GDB and the GDB User Manual at the GNU Project
Debugger website. This website is as follows:
http://sourceware.org/gdb/documentation/
• Intel provides documentation for all the Intel Parallel Studio XE products at the
following website:
xiv 007–5646–010
![](asset-f.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
https://software.intel.com/en-us/intel-parallel-studio-xe-support/documentation
• Intel provides detailed application tuning information in the Intel®64 and IA-32
Architectures Optimization Reference Manual, which is available from the following
website:
http://www.intel.com/content/www/us/en/architecture-and-technology/64-ia-
32-architectures-optimization-manual.html?wapkw=
• Information about the Intel Xeon Processor E5 family of processors is at the
following website:
http://www.intel.com/content/www/us/en/processors/xeon/xeon-processor-e5-
family.html
• The Intel Software Network page includes information specific to the Intel VTune
Performance Amplifier. This website is as follows:
https://software.intel.com/en-us/intel-vtune-amplifier-xe
• Information about the OpenMP Standard and the OpenMP API specification for
parallel programming can be found at the following website:
http://openmp.org/wp/
Obtaining Publications
All SGI publications are available on the SGI customer portal at
http://support.sgi.com. Select the following:
Support by Product > productname > Documentation
If you do not find what you are looking for, search for document-title keywords by
selecting Search Knowledgebase and using the category Documentation.
You can view man pages by typing man title on a command line.
Conventions
The following conventions are used in this documentation:
[ ] Brackets enclose optional portions of a command or
directive line.
007–5646–010 xv
![](asset-10.png)
About This Guide
command This fixed-space font denotes literal items such as
commands, files, routines, path names, signals,
messages, and programming language structures.
... Ellipses indicate that a preceding element can be
repeated.
user input This bold, fixed-space font denotes literal items that the
user enters in interactive sessions. (Output is shown in
nonbold, fixed-space font.)
variable Italic typeface denotes variable entries and words or
concepts being defined.
manpage(x) Man page section identifiers appear in parentheses after
man page names.
Reader Comments
If you have comments about the technical accuracy, content, or organization of this
publication, contact SGI. Be sure to include the title and document number of the
publication with your comments. (Online, the document number is located in the
front matter of the publication. In printed publications, the document number is
located at the bottom of each page.)
You can contact SGI in either of the following ways:
• Send e-mail to the following address:
techpubs@sgi.com
• Contact your customer service representative and ask that an incident be filed in
the SGI incident tracking system:
http://www.sgi.com/support/supportcenters.html
SGI values your comments and will respond to them promptly.
xvi 007–5646–010
![](asset-11.png)
Chapter 1
The SGI Compiling Environment
This chapter includes the following topics:
• "About the Compiling Environment" on page 1
• "Compiler Overview" on page 2
• "Environment Modules" on page 3
• "Library Overview" on page 4
About the Compiling Environment
This chapter provides an overview of the SGI compiling environment on the SGI
family of servers.
Tuning an application involves making your program run its fastest on the available
hardware. The first step is to make your program run as efficiently as possible on a
single processor system and then consider ways to use parallel processing.
Virtual memory, also known as virtual addressing, divides a system’s relatively small
amount of physical memory among the potentially larger amount of logical processes
in a program. It does this by dividing physical memory into pages, and then
allocating pages to processes as the pages are needed.
Apage is the smallest unit of system memory allocation. Pages are added to a process
when either a page fault occurs or an allocation request is issued. Process size is
measured in pages, and two sizes are associated with every process: the total size and
the resident set size (RSS). The number of pages being used by a process and the
process size can be determined by using the pmap(1) command or by examining the
contents of the proc/pid/status directory.
Swap space is used for temporarily saving parts of a program when there is not
enough physical memory. The swap space may be on the system drive, on an
optional drive, or allocated to a particular file in a filesystem. To avoid swapping, try
not to overburden memory. Lack of adequate memory limits the number and the size
of applications that can run simultaneously on the system, and it can limit system
performance. Access time to disk is orders of magnitude slower than access to
random access memory (RAM). Performance is seriously affected when a system runs
out of memory and uses swap-to-disk while running a program. Swapping becomes
007–5646–010 1
![](asset-12.png)
1: The SGI Compiling Environment
a major bottleneck. Be sure your system is configured with enough memory to run
your applications.
Linux is a demand paging operating system, using a least-recently-used paging
algorithm. Pages are mapped into physical memory when first referenced, and pages
are brought back into memory if swapped out. In a system that uses demand paging,
the operating system copies a disk page into physical memory only if an attempt is
made to access it. That is, a page fault occurs. A page fault handler algorithm does
the necessary action. For more information, see the mmap(2) man page.
Compiler Overview
You can obtain an Intel Fortran compiler or an Intel C/C++ compiler from Intel
Corporation or from SGI. For more information, see one of the following links:
• http://software.intel.com/en-us/intel-sdp-home
• https://software.intel.com/en-us/intel-parallel-studio-xe
In addition, the GNU Fortran and C compilers are available on SGI systems.
For example, the following is the general format for the Fortran compiler command
line:
% ifort [options]filename.extension
An appropriate filename extension is required for each compiler, according to the
programming language used (Fortran, C, C++, or FORTRAN 77).
Some common compiler options are:
•-o filename: renames the output to filename.
•-g: produces additional symbol information for debugging.
•-O[level]: invokes the compiler at different optimization levels,from0to 3.
•-ldirectory_name: looks for include files in directory_name.
•-c: compiles without invoking the linker. This option produces an a.o file only.
Many processors do not handle denormalized arithmetic (for gradual underflow) in
hardware. The support of gradual underflow is implementation-dependent. Use the
-ftz option with the Intel compilers to force-flush denormalized results to zero.
2 007–5646–010
![](asset-13.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Note that frequent gradual underflow arithmetic in a program causes the program to
run very slowly, consuming large amounts of system time. You can use the time(1)
command to determine the amount of system time consumed. In this case, it is best
to trace the source of the underflows and fix the code; gradual underflow is often a
source of reduced accuracy anyway. prctl(1) allows you to query or control certain
process behavior. In a program, prctl tracks the location of floating-point errors.
Environment Modules
Amodule modifies a user’s environment dynamically. A user can load a module,
rather than change environment variables, in order to access different versions of the
compilers, loaders, libraries, and utilities that are installed on the system.
Modules can be used in the SGI compiling environment to customize the
environment. If the use of Modules is not available on your system, its installation
and use is highly recommended.
To retrieve a list of Modules that are available on your system, use the following
command:
%module avail
To load Modules into your environment, use the following commands:
%module load intel-compilers-latest mpt/2.12
Note: The preceding commands are an example only. The actual release numbers
vary depending on the version of the software you are using. See the release notes
that are distributed with your system for the pertinent release version numbers.
The following command displays a list of all arguments accepted:
sys:~> module help
Modules Release 3.1.6 (Copyright GNU GPL v2 1991):
Available Commands and Usage:
+ add|load modulefile [modulefile ...]
+ rm|unload modulefile [modulefile ...]
+ switch|swap modulefile1 modulefile2
+ display|show modulefile [modulefile ...]
007–5646–010 3
![](asset-14.png)
1: The SGI Compiling Environment
+ avail [modulefile [modulefile ...]]
+ use [-a|--append] dir [dir ...]
+ unuse dir [dir ...]
+ update
+ purge
+ list
+ clear
+ help [modulefile [modulefile ...]]
+ whatis [modulefile [modulefile ...]]
+ apropos|keyword string
+ initadd modulefile [modulefile ...]
+ initprepend modulefile [modulefile ...]
+ initrm modulefile [modulefile ...]
+ initswitch modulefile1 modulefile2
+ initlist
+ initclear
For details about using Modules, see the module(1) man page.
Library Overview
Libraries are files that contain one or more object (.o) files. Libraries simplify local
software development by hiding compilation details. Libraries are sometimes also
called archives.
The SGI compiling environment contains several types of libraries. The following
topics provide an overview about each library:
• "Static Libraries" on page 4
• "Dynamic Libraries" on page 5
• "C/C++ Libraries" on page 5
• "MPI and SHMEM Libraries" on page 6
Static Libraries
Static libraries are used when calls to the library components are satisfied at link time
by copying text from the library into the executable. To create a static library, use
ar(1) or an archiver command.
4 007–5646–010
![](asset-15.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
To use a static library, include the library name on the compiler’s command line. If
the library is not in a standard library directory, use the -L option to specify the
directory and the -l option to specify the library filename.
To build an appplication to have all static versions of standard libraries in the
application binary, use the -static option on the compiler command line.
Dynamic Libraries
Dynamic libraries are linked into the program at run time. When you load dynamic
libraries into memory, they are available for access by multiple programs. Dynamic
libraries are formed by creating a Dynamic Shared Object (DSO).
Use the link editor command, ld(1), to create a dynamic library from a series of
object files or to create a DSO from an existing static library.
To use a dynamic library, include the library on the compiler’s command line. If the
dynamic library is not in one of the standard library directories, use the -L path and
-l library_shortname compiler options during linking. You must also set the
LD_LIBRARY_PATH environment variable to the directory where the library is stored
before running the executable.
C/C++ Libraries
The Intel compiler provides the following C/C++ libraries:
•libguide.a and libguide.so, which support OpenMP-based programs.
•libsvml.a, which is the short vector math library.
•libirc.a, which includes Intel support for Profile-Guided Optimizations (PGO)
and CPU dispatch.
•libimf.a and libimf.so, which are Intel’s math libraries.
•libcprts.a and libcprts.so, which are the Dinkumware C++ libraries.
•libunwind.a and libunwind.so, which are the Unwinder libraries.
•libcxa.a and libcxa.so, which provide Intel run-time support for C++
features.
007–5646–010 5
![](asset-16.png)
1: The SGI Compiling Environment
MPI and SHMEM Libraries
The SGI MPI software package facilitates parallel programming on large computer
systems and on computer system clusters. SGI MPI supports both the Message
Passing Interface (MPI) standard and the OpenSHMEM standard. SGI’s support for
MPI and OpenSHMEM is built on top of the SGI Message Passing Toolkit (MPT). SGI
MPT is a high-performance communications middleware software product that runs
on SGI’s shared memory and cluster supercomputers. On some of these machines,
SGI MPT uses SGI Array Services to launch applications. SGI MPT is optimized for
all SGI hardware platforms.
The SHMEM application programing interface is implemented by the libsma library
and is part of the Message Passing Toolkit (MPT) product on SGI systems. The
SHMEM programming model consists of library routines that provide low-latency,
high-bandwidth communication for use in highly parallelized, scalable programs. The
routines in the SHMEM application programming interface (API) provide a
programming model for exchanging data between cooperating parallel processes. The
resulting programs are similar in style to Message Passing Interface (MPI) programs.
You can use the SHMEM API alone or in combination with MPI routines in the same
parallel program.
For more information, see the following:
• The intro_shmem(3) man page.
• The SGI MPI and SGI SHMEM User Guide.
6 007–5646–010
![](asset-17.png)
Chapter 2
Performance Analysis and Debugging
This chapter contains the following topics:
• "About Performance Analysis and Debugging" on page 7
• "Determining System Configuration" on page 8
• "Sources of Performance Problems" on page 15
• "Other Performance Analysis Tools" on page 17
• "About Debugging" on page 17
About Performance Analysis and Debugging
Tuning an application involves determining the source of performance problems and
then rectifying those problems to make your programs run their fastest on the
available hardware. Performance gains usually fall into one of three categories of
measured time:
• User CPU time, which is the time accumulated by a user process when it is
attached to a CPU and is running.
• Elapsed (wall-clock) time, which is the amount of time that passes between the
start and the termination of a process.
• System time, which is the amount of time spent on performing kernel functions
such as system calls, sched_yield, for example, or floating point errors.
Any application tuning process involves the following steps:
1. Analyzing and identifying a problem
2. Locating the problem in the code
3. Applying an optimization technique
This topics in this chapter describe how to analyze your code to determine
performance bottlenecks. For information about how to tune your application for a
single processor system and then tune it for parallel processing, see the following:
Chapter 5, "Performance Tuning" on page 53
007–5646–010 7
![](asset-18.png)
2: Performance Analysis and Debugging
Determining System Configuration
One of the first steps in application tuning is to determine the details of the system
that you are running. Depending on your system configuration, different options
might or might not provide good results.
The topology(1) command displays general information about SGI systems, with a
focus on node information. This can include node counts for blades, node IDs,
NASIDs, memory per node, system serial number, partition number, UV Hub
versions, CPU to node mappings, and general CPU information. The topology(1)
command is part of the SGI Foundation Software package.
The following is example output from two topology(1) commands:
uv-sys:~ # topology
System type: UV2000
System name: harp34-sys
Serial number: UV2-00000034
Partition number: 0
8 Blades
256 CPUs
16 Nodes
235.82 GB Memory Total
15.00 GB Max Memory on any Node
1 BASE I/O Riser
2 Network Controllers
2 Storage Controllers
2 USB Controllers
1 VGA GPU
uv-sys:~ # topology --summary --nodes --cpus
System type: UV2000
System name: harp34-sys
Serial number: UV2-00000034
Partition number: 0
8 Blades
256 CPUs
16 Nodes
235.82 GB Memory Total
15.00 GB Max Memory on any Node
1 BASE I/O Riser
2 Network Controllers
8 007–5646–010
![](asset-19.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
2 Storage Controllers
2 USB Controllers
1 VGA GPU
Index ID NASID CPUS Memory
--------------------------------------------
0 r001i11b00h0 0 16 15316 MB
1 r001i11b00h1 2 16 15344 MB
2 r001i11b01h0 4 16 15344 MB
3 r001i11b01h1 6 16 15344 MB
4 r001i11b02h0 8 16 15344 MB
5 r001i11b02h1 10 16 15344 MB
6 r001i11b03h0 12 16 15344 MB
7 r001i11b03h1 14 16 15344 MB
8 r001i11b04h0 16 16 15344 MB
9 r001i11b04h1 18 16 15344 MB
10 r001i11b05h0 20 16 15344 MB
11 r001i11b05h1 22 16 15344 MB
12 r001i11b06h0 24 16 15344 MB
13 r001i11b06h1 26 16 15344 MB
14 r001i11b07h0 28 16 15344 MB
15 r001i11b07h1 30 16 15344 MB
CPU Blade PhysID CoreID APIC-ID Family Model Speed L1(KiB) L2(KiB) L3(KiB)
---------------------------------------------------------------------------------
0 r001i11b00h0 00 00 0 6 45 2599 32d/32i 256 20480
1 r001i11b00h0 00 01 2 6 45 2599 32d/32i 256 20480
2 r001i11b00h0 00 02 4 6 45 2599 32d/32i 256 20480
3 r001i11b00h0 00 03 6 6 45 2599 32d/32i 256 20480
4 r001i11b00h0 00 04 8 6 45 2599 32d/32i 256 20480
5 r001i11b00h0 00 05 10 6 45 2599 32d/32i 256 20480
6 r001i11b00h0 00 06 12 6 45 2599 32d/32i 256 20480
7 r001i11b00h0 00 07 14 6 45 2599 32d/32i 256 20480
8 r001i11b00h1 01 00 32 6 45 2599 32d/32i 256 20480
9 r001i11b00h1 01 01 34 6 45 2599 32d/32i 256 20480
10 r001i11b00h1 01 02 36 6 45 2599 32d/32i 256 20480
11 r001i11b00h1 01 03 38 6 45 2599 32d/32i 256 20480
...
The cpumap(1) command displays logical CPUs and shows relationships between
them in a human-readable format. Aspects displayed include hyperthread
007–5646–010 9
![](asset-1a.png)
2: Performance Analysis and Debugging
relationships, last level cache sharing, and topological placement. The cpumap(1)
command gets its information from /proc/cpuinfo, the /sys/devices/system
directory structure, and /proc/sgi_uv/topology. When creating cpusets, the
numbers reported in the output section called Processor Numbering on Node(s)
correspond to the mems argument you use to define a cpuset. The cpuset mems
argument is the list of memory nodes that tasks in the cpuset are allowed to use.
For more information, see the SGI Cpuset Software Guide.
The following is example output:
uv# cpumap
Thu Sep 19 10:17:21 CDT 2013
harp34-sys.americas.sgi.com
This is an SGI UV
model name : Genuine Intel(R) CPU @ 2.60GHz
Architecture : x86_64
cpu MHz : 2599.946
cache size : 20480 KB (Last Level)
Total Number of Sockets : 16
Total Number of Cores : 128 (8 per socket)
Hyperthreading : ON
Total Number of Physical Processors : 128
Total Number of Logical Processors : 256 (2 per Phys Processor)
UV Information
HUB Version: UVHub 3.0
Number of Hubs: 16
Number of connected Hubs: 16
Number of connected NUMAlink ports: 128
=============================================================================
Hub-Processor Mapping
Hub Location Processor Numbers -- HyperThreads in ()
--- ---------- ---------------------------------------
0 r001i11b00h0 01234567
( 128 129 130 131 132 133 134 135 )
1 r001i11b00h1 8 9 10 11 12 13 14 15
10 007–5646–010
![](asset-1b.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
( 136 137 138 139 140 141 142 143 )
2 r001i11b01h0 16 17 18 19 20 21 22 23
( 144 145 146 147 148 149 150 151 )
3 r001i11b01h1 24 25 26 27 28 29 30 31
( 152 153 154 155 156 157 158 159 )
4 r001i11b02h0 32 33 34 35 36 37 38 39
( 160 161 162 163 164 165 166 167 )
5 r001i11b02h1 40 41 42 43 44 45 46 47
( 168 169 170 171 172 173 174 175 )
6 r001i11b03h0 48 49 50 51 52 53 54 55
( 176 177 178 179 180 181 182 183 )
7 r001i11b03h1 56 57 58 59 60 61 62 63
( 184 185 186 187 188 189 190 191 )
8 r001i11b04h0 64 65 66 67 68 69 70 71
( 192 193 194 195 196 197 198 199 )
9 r001i11b04h1 72 73 74 75 76 77 78 79
( 200 201 202 203 204 205 206 207 )
10 r001i11b05h0 80 81 82 83 84 85 86 87
( 208 209 210 211 212 213 214 215 )
11 r001i11b05h1 88 89 90 91 92 93 94 95
( 216 217 218 219 220 221 222 223 )
12 r001i11b06h0 96 97 98 99 100 101 102 103
( 224 225 226 227 228 229 230 231 )
13 r001i11b06h1 104 105 106 107 108 109 110 111
( 232 233 234 235 236 237 238 239 )
14 r001i11b07h0 112 113 114 115 116 117 118 119
( 240 241 242 243 244 245 246 247 )
15 r001i11b07h1 120 121 122 123 124 125 126 127
( 248 249 250 251 252 253 254 255 )
=============================================================================
Processor Numbering on Node(s)
Node (Logical) Processors
------ -------------------------
0 01234567128129130131132133134135
1 8 9 10 11 12 13 14 15 136 137 138 139 140 141 142 143
2 16 17 18 19 20 21 22 23 144 145 146 147 148 149 150 151
3 24 25 26 27 28 29 30 31 152 153 154 155 156 157 158 159
4 32 33 34 35 36 37 38 39 160 161 162 163 164 165 166 167
007–5646–010 11
![](asset-1c.png)
2: Performance Analysis and Debugging
5 40 41 42 43 44 45 46 47 168 169 170 171 172 173 174 175
6 48 49 50 51 52 53 54 55 176 177 178 179 180 181 182 183
7 56 57 58 59 60 61 62 63 184 185 186 187 188 189 190 191
8 64 65 66 67 68 69 70 71 192 193 194 195 196 197 198 199
9 72 73 74 75 76 77 78 79 200 201 202 203 204 205 206 207
10 80 81 82 83 84 85 86 87 208 209 210 211 212 213 214 215
11 88 89 90 91 92 93 94 95 216 217 218 219 220 221 222 223
12 96 97 98 99 100 101 102 103 224 225 226 227 228 229 230 231
13 104 105 106 107 108 109 110 111 232 233 234 235 236 237 238 239
14 112 113 114 115 116 117 118 119 240 241 242 243 244 245 246 247
15 120 121 122 123 124 125 126 127 248 249 250 251 252 253 254 255
=============================================================================
Sharing of Last Level (3) Caches
Socket (Logical) Processors
------ -------------------------
0 01234567128129130131132133134135
1 8 9 10 11 12 13 14 15 136 137 138 139 140 141 142 143
2 16 17 18 19 20 21 22 23 144 145 146 147 148 149 150 151
3 24 25 26 27 28 29 30 31 152 153 154 155 156 157 158 159
4 32 33 34 35 36 37 38 39 160 161 162 163 164 165 166 167
5 40 41 42 43 44 45 46 47 168 169 170 171 172 173 174 175
6 48 49 50 51 52 53 54 55 176 177 178 179 180 181 182 183
7 56 57 58 59 60 61 62 63 184 185 186 187 188 189 190 191
8 64 65 66 67 68 69 70 71 192 193 194 195 196 197 198 199
9 72 73 74 75 76 77 78 79 200 201 202 203 204 205 206 207
10 80 81 82 83 84 85 86 87 208 209 210 211 212 213 214 215
11 88 89 90 91 92 93 94 95 216 217 218 219 220 221 222 223
12 96 97 98 99 100 101 102 103 224 225 226 227 228 229 230 231
13 104 105 106 107 108 109 110 111 232 233 234 235 236 237 238 239
14 112 113 114 115 116 117 118 119 240 241 242 243 244 245 246 247
15 120 121 122 123 124 125 126 127 248 249 250 251 252 253 254 255
=============================================================================
HyperThreading
Shared Processors
-----------------
12 007–5646–010
![](asset-1d.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
( 0, 128) ( 1, 129) ( 2, 130) ( 3, 131)
( 4, 132) ( 5, 133) ( 6, 134) ( 7, 135)
( 8, 136) ( 9, 137) ( 10, 138) ( 11, 139)
( 12, 140) ( 13, 141) ( 14, 142) ( 15, 143)
( 16, 144) ( 17, 145) ( 18, 146) ( 19, 147)
( 20, 148) ( 21, 149) ( 22, 150) ( 23, 151)
( 24, 152) ( 25, 153) ( 26, 154) ( 27, 155)
( 28, 156) ( 29, 157) ( 30, 158) ( 31, 159)
( 32, 160) ( 33, 161) ( 34, 162) ( 35, 163)
( 36, 164) ( 37, 165) ( 38, 166) ( 39, 167)
( 40, 168) ( 41, 169) ( 42, 170) ( 43, 171)
( 44, 172) ( 45, 173) ( 46, 174) ( 47, 175)
( 48, 176) ( 49, 177) ( 50, 178) ( 51, 179)
( 52, 180) ( 53, 181) ( 54, 182) ( 55, 183)
( 56, 184) ( 57, 185) ( 58, 186) ( 59, 187)
( 60, 188) ( 61, 189) ( 62, 190) ( 63, 191)
( 64, 192) ( 65, 193) ( 66, 194) ( 67, 195)
( 68, 196) ( 69, 197) ( 70, 198) ( 71, 199)
( 72, 200) ( 73, 201) ( 74, 202) ( 75, 203)
( 76, 204) ( 77, 205) ( 78, 206) ( 79, 207)
( 80, 208) ( 81, 209) ( 82, 210) ( 83, 211)
( 84, 212) ( 85, 213) ( 86, 214) ( 87, 215)
( 88, 216) ( 89, 217) ( 90, 218) ( 91, 219)
( 92, 220) ( 93, 221) ( 94, 222) ( 95, 223)
( 96, 224) ( 97, 225) ( 98, 226) ( 99, 227)
( 100, 228) ( 101, 229) ( 102, 230) ( 103, 231)
( 104, 232) ( 105, 233) ( 106, 234) ( 107, 235)
( 108, 236) ( 109, 237) ( 110, 238) ( 111, 239)
( 112, 240) ( 113, 241) ( 114, 242) ( 115, 243)
( 116, 244) ( 117, 245) ( 118, 246) ( 119, 247)
( 120, 248) ( 121, 249) ( 122, 250) ( 123, 251)
( 124, 252) ( 125, 253) ( 126, 254) ( 127, 255)
The x86info(1) command displays x86 CPU diagnostics information. Type one of
the following commands to load the x86info(1) command if the command is not
already installed:
• On Red Hat Enterprise Linux (RHEL) systems, type the following:
#yum install x86info.x86_64
007–5646–010 13
![](asset-1e.png)
2: Performance Analysis and Debugging
• On SLES systems, type the following:
#zypper install x86info
The following is an example of x86info(1) command output:
uv44-sys:~ # x86info
x86info v1.25. Dave Jones 2001-2009
Feedback to .
Found 64 CPUs
--------------------------------------------------------------------------
CPU #1
EFamily: 0 EModel: 2 Family: 6 Model: 46 Stepping: 6
CPU Model: Unknown model.
Processor name string: Intel(R) Xeon(R) CPU E7520 @ 1.87GHz
Type: 0 (Original OEM) Brand: 0 (Unsupported)
Number of cores per physical package=16
Number of logical processors per socket=32
Number of logical processors per core=2
APIC ID: 0x0 Package: 0 Core: 0 SMT ID 0
--------------------------------------------------------------------------
CPU #2
EFamily: 0 EModel: 2 Family: 6 Model: 46 Stepping: 6
CPU Model: Unknown model.
Processor name string: Intel(R) Xeon(R) CPU E7520 @ 1.87GHz
Type: 0 (Original OEM) Brand: 0 (Unsupported)
Number of cores per physical package=16
Number of logical processors per socket=32
Number of logical processors per core=2
APIC ID: 0x6 Package: 0 Core: 0 SMT ID 6
--------------------------------------------------------------------------
CPU #3
EFamily: 0 EModel: 2 Family: 6 Model: 46 Stepping: 6
CPU Model: Unknown model.
Processor name string: Intel(R) Xeon(R) CPU E7520 @ 1.87GHz
Type: 0 (Original OEM) Brand: 0 (Unsupported)
Number of cores per physical package=16
Number of logical processors per socket=32
Number of logical processors per core=2
APIC ID: 0x10 Package: 0 Core: 0 SMT ID 16
--------------------------------------------------------------------------
14 007–5646–010
![](asset-1f.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
...
You can also use the uname command, which returns the kernel version and other
machine information. For example:
uv44-sys:~ # uname -a
Linux uv44-sys 2.6.32.13-0.4.1.1559.0.PTF-default #1 SMP 2010-06-15 12:47:25 +0200 x86_64 x86_64 x86_64 GNU/Linux
For more system information, change to the
/sys/devices/system/node/node0/cpu0/cache directory and list the contents.
For example:
uv44-sys:/sys/devices/system/node/node0/cpu0/cache # ls
index0 index1 index2 index3
Change directory to index0 and list the contents, as follows:
uv44-sys:/sys/devices/system/node/node0/cpu0/cache/index0 # ls
coherency_line_size level number_of_sets physical_line_partition shared_cpu_list shared_cpu_map size type
Note: The preceding example output was truncated at the right for inclusion in this
manual.
Sources of Performance Problems
The following three processes types typically cause program execution performance
slowdowns:
• CPU-bound processes, whch are processes that perform slow operations, such as
sqrt or floating-point divides, or nonpipelined operations, such as switching
between add and multiply operations.
• Memory-bound processes, which consist of code that uses poor memory strides,
occurrences of page thrashing or cache misses, or poor data placement in NUMA
systems.
• I/O-bound processes, which are processes that wait on synchronous I/O or
formatted I/O. These are also processes that wait when there is library-level or
system-level buffering.
The following topics describe some of the tools that can help pinpoint performance
slowdowns:
007–5646–010 15
![](asset-20.png)
2: Performance Analysis and Debugging
• "Profiling with perf" on page 16
• "Profiling with PerfSuite" on page 16
Profiling with perf
Linux Perf Events provides a performance analysis framework. It includes
hardware-level CPU performance monitoring unit (PMU) features, software counters,
and tracepoints. The perf RPM comes with the operating system, includes man
pages, and is not an SGI product.
For more information, see the following man pages:
•perf(1)
•perf-stat(1)
•perf-top(1)
•perf-record(1)
•perf-report(1)
•perf-list(1)
Profiling with PerfSuite
PerfSuite is a set of tools, utilities, and libraries that you can use to analyze
application software performance on Linux systems. You can use the PerfSuite tools
to perform performance-related activities, ranging from assistance with compiler
optimization reports to hardware performance counting, profiling, and MPI usage
summarization. PerfSuite is Open Source software. It is approved for licensing under
the University of Illinois/NCSA Open Source License (OSI-approved).
For more information, see one of the following websites:
• http://perfsuite.ncsa.uiuc.edu/
• http://perfsuite.sourceforge.net/
• http://perfsuite.ncsa.illinois.edu
This website hosts NCSA-specific information about using PerfSuite tools.
16 007–5646–010
![](asset-21.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
PerfSuite includes the psrun utility, which gathers hardware performance
information on an unmodified executable. For more information, see
http://perfsuite.ncsa.uiuc.edu/psrun/.
Other Performance Analysis Tools
The following tools might be useful to you when you try to optimize your code:
• The Intel®VTuneTM Amplifier XE, which is a performance and thread profiler. This
tool can perform both local sampling and remote sampling experiments. In the
remote samplling case, the VTune data collector runs on the Linux system, and an
accompanying graphical user interface (GUI) runs on a Windows system, which is
used for analyzing the results. The VTune profiler allows you to perform
interactive experiments while connected to the host through its GUI.
For information about Intel VTune Amplifier XE, see the following URL:
http://software.intel.com/en-us/intel-vtune-amplifier-xe#pid-3773–760
• Intel Inspector XE, which is a memory and thread debugger. For information
about Intel Inspector XE, see the following:
http://software.intel.com/en-us/intel-inspector-xe/
• Intel Advisor XE, which is a threading design and prototyping tool. For
information about Intel Advisor XE, see the following:
http://software.intel.com/en-us/intel-advisor-xe
About Debugging
The following debuggers are available on SGI platforms:
• The Intel Debugger and GDB. You can start the Intel Debugger and GDB with the
ddd command. The ddd command starts the Data Display Debugger, a GNU
product that provides a graphical debugging interface.
• Totalview. TotalView is graphical debugger that you can use with MPI programs.
For information about TotalView, including its licensing, see the following:
http://www.roguewave.com
007–5646–010 17
![](asset-22.png)
2: Performance Analysis and Debugging
• Allinea DDT. The Allinea DDT debugger is a graphical debugger from Allinea.
For information about Allinea DDT, see the following:
www.allinea.com/products/ddt
The following topics provide more information about some of the debuggers available
on SGI systems:
• "Using the Intel Debugger" on page 18
• "Using the GNU Data Display Debugger (GNU DDD)" on page 19
Using the Intel Debugger
The Intel Debugger for Linux is the Intel symbolic debugger. The Intel Debugger is
part of Intel Parallel Studio XE Professional Edition and above. This debugger is
based on the Eclipse GUI. This debugger works with the Intel®C and C++ compilers,
the Intel®Fortran compilers, and the GNU compilers. This product is available if
your system is licensed for the Intel compilers. You are asked during the installation
if you want to install it or not. The idb command starts the GUI. The idbc
command starts the command line interface. If you specify the -gdb option on the
idb command, the shell command line provides user commands and debugger
output similar to the GNU debugger.
You can use the Intel Debugger for Linux with single-threaded applications,
multithreaded applications, serial code, and parallel code.
Figure 2-1 on page 19 shows the GUI.
18 007–5646–010
![](asset-23.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Figure 2-1 Intel®Debugger GUI
For more information, see the following:
http://software.intel.com/en-us/articles/idb-linux/
Using the GNU Data Display Debugger (GNU DDD)
GDB is the GNU debugger. The GDB debugger supports C, C++, Fortran, and
Modula-2 programs. The following information pertains to these compilers:
• When compiling with C and C++, include the -g option on the compiler command
line. The -g option produces the dwarf2 symbols database that GDB uses.
• When using GDB for Fortran debugging, include the -g and -O0 options. Do not
use gdb for Fortran debugging when compiling with -O1 or higher. The standard
GDB debugger does not support Fortran 95 programs. To debug Fortran 95
programs, download and install the gdbf95 patch from the following website:
http://sourceforge.net/project/showfiles.php?group_id=56720
To verify that you have the correct version of GDB installed, use the gdb -v
command. The output should appear similar to the following:
007–5646–010 19
![](asset-24.png)
2: Performance Analysis and Debugging
GNU gdb 5.1.1 FORTRAN95-20020628 (RC1)
Copyright 2012 Free Software Foundation, Inc.
The Data Display Debugger provides a graphical debugging interface for the GDB
debugger and other command line debuggers. To use GDB through a GUI, use the
ddd command. Specify the --debugger option to specify the debugger you want to
use. For example, specify --debugger "idb" to specify the Intel Debugger. Use the
gdb command to start GDB’s command line interface.
When the debugger loads, the Data Display Debugger screen appears divided into
panes that show the following information:
• Array inspection
• Source code
• Disassembled code
• A command line window to the debugger engine
From the View menu, you can switch these panes on and off.
Some commonly used commands can be found on the menus. In addition, the
following actions can be useful:
• To select an address in the assembly view, click the right mouse button, and select
lookup. The gdb command runs in the command pane and shows the
corresponding source line.
• Select a variable in the source pane, and click the right mouse button. The
debugger displays the current value. Arrays appear in the array inspection
window. You can print these arrays to PostScript by using the Menu>Print Graph
option.
• To view the contents of the register file, including general, floating-point, NaT,
predicate, and application registers, select Registers from the Status menu. The
Status menu also allows you to view stack traces or to switch OpenMP threads.
For a complete list of GDB commands, use the help option or see the documentation
at the following website:
http://sourceware.org/gdb/documentation/
20 007–5646–010
![](asset-25.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Note: The current instances of GDB do not report ar.ec registers correctly. If you
are debugging rotating, register-based, software-pipelined loops at the assembly code
level, try using the Intel Debugger for Linux.
007–5646–010 21
![](asset-27.png)
Chapter 3
Monitoring Commands
This chapter includes the following topics:
• "About the Operating System Monitoring Commands" on page 23
• "Operating System Monitoring Commands" on page 23
About the Operating System Monitoring Commands
You can use operating system commands to understand the usage and limits of your
system. These commands allow you to observe both overall system performance and
single-performance execution characteristics.
The topics in this chapter describe the commands that are included in your SGI
system’s operating system. The following are additional commands and utilities that
are available:
• SGI Foundation Software utilities. SFS is included by default on your SGI system.
For information about these utilities, see the following:
SGI Foundation Software (SFS) User Guide
• Performance Co-Pilot. For documentation about this open source toolset, see the
following website:
http://www.pcp.io/documentation.html
Operating System Monitoring Commands
The following topics show several operating system commands you can use to
determine user load, system usage, and active processes:
• "Using the w(1) command" on page 24
• "Using the ps(1) Command" on page 24
• "Using the top(1) Command" on page 25
• "Using the vmstat(8) Command" on page 25
007–5646–010 23
![](asset-28.png)
3: Monitoring Commands
• "Using the iostat(1) command" on page 26
• "Using the sar(1) command" on page 27
Using the w(1) command
To obtain a high-level view of system usage that includes information about who is
logged into the system, use the w(1) command, as follows:
uv44-sys:~ # w
15:47:48 up 2:49, 5 users, load average: 0.04, 0.27, 0.42
USER TTY LOGIN@ IDLE JCPU PCPU WHAT
root pts/0 13:10 1:41m 0.07s 0.07s -bash
root pts/2 13:31 0.00s 0.14s 0.02s w
boetcher pts/4 14:30 2:13 0.73s 0.73s -csh
root pts/5 14:32 1:14m 0.04s 0.04s -bash
root pts/6 15:09 31:25 0.08s 0.08s -bash
The wcommand’s output shows who is on the system, the duration of user sessions,
processor usage by user, and currently executing user commands. The output consists
of two parts:
• The first output line shows the current time, the length of time the system has
been up, the number of users on the system, and the average number of jobs in
the run queue in the last one, five, and 15 minutes.
• The rest of the output from the wcommand shows who is logged into the system,
the duration of each user session, processor usage by user, and each user’s current
process command line.
Using the ps(1) Command
To determine active processes, use the ps(1) command, which displays a snapshot of
the process table.
The ps -A r command example that follows returns all the processes currently
running on a system:
[user@profit user]# ps -A r
PID TTY STAT TIME COMMAND
211116 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211117 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
24 007–5646–010
![](asset-29.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
211118 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211119 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211120 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211121 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211122 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211123 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211124 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211125 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211126 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211127 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211128 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211129 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211130 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
211131 pts/0 R+ 4:08 /usr/diags/bin/olconft RUNTIME=5
...
Using the top(1) Command
To monitor running processes, use the top(1) command. This command displays a
sorted list of top CPU utilization processes.
Using the vmstat(8) Command
The vmstat(8) command reports virtual memory statistics. It reports information
about processes, memory, paging, block IO, traps, and CPU activity. For more
information, see the vmstat(8) man page.
In the following vmstat(8) command, the 10 specifies a 10–second delay between
updates.
uv44-sys:~ # vmstat 10
procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu-----
r b swpd free buff cache si so bi bo in cs us sy id wa st
2 0 0 235984032 418748 8649568 0 0 0 0 0 0 0 0 100 0 0
1 0 0 236054400 418748 8645216 0 0 0 4809 256729 3401 0 0 100 0 0
1 0 0 236188016 418748 8649904 0 0 0 448 256200 631 0 0 100 0 0
2 0 0 236202976 418748 8645104 0 0 0 341 256201 1117 0 0 100 0 0
1 0 0 236088720 418748 8592616 0 0 0 847 257104 6152 0 0 100 0 0
1 0 0 235990944 418748 8648460 0 0 0 240 257085 5960 0 0 100 0 0
1 0 0 236049568 418748 8645100 0 0 0 4849 256749 3604 0 0 100 0 0
007–5646–010 25
![](asset-2a.png)
3: Monitoring Commands
Without the delay parameter, which is 10 in this example, the output returns averages
since the last reboot. Additional reports give information on a sampling period of
length delay. The process and memory reports are instantaneous in either case.
Using the iostat(1) command
The iostat(1) command monitors system input/output device loading by observing
the time the devices are active, relative to their average transfer rates. You can use
information from the iostat command to change system configuration information
to better balance the input/output load between physical disks. For more
information, see the iostat(1) man page.
In the following iostat(1) command, the 10 specifies a 10–second interval between
updates:
#iostat 10
Linux 2.6.32-430.el6.x86_64 (harp34-sys) 02/21/2014 _x86_64_ (256 CPU)
avg-cpu: %user %nice %system %iowait %steal %idle
46.24 0.01 0.67 0.01 0.00 53.08
Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn
sda 53.66 23711.65 23791.93 21795308343 21869098736
sdb 0.01 0.02 0.00 17795 0
avg-cpu: %user %nice %system %iowait %steal %idle
99.96 0.00 0.04 0.00 0.00 0.00
Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn
sda 321.20 149312.00 150423.20 1493120 1504232
sdb 0.00 0.00 0.00 0 0
avg-cpu: %user %nice %system %iowait %steal %idle
99.95 0.00 0.05 0.00 0.00 0.00
Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn
sda 305.19 146746.05 148453.95 1468928 1486024
sdb 0.00 0.00 0.00 0 0
...
26 007–5646–010
![](asset-2b.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Using the sar(1) command
The sar(1) command returns the content of selected, cumulative activity counters in
the operating system. Based on the values in the count and interval parameters, the
command writes information count times spaced at the specified interval, which is in
seconds. For more information, see the sar(1) man page. The following example
shows the sar(1) command with a request for information about CPU 1, a count of
10, and an interval of 10:
uv44-sys:~ # sar -P 1 10 10
Linux 2.6.32-416.el6.x86_64 (harp34-sys) 09/19/2013 _x86_64_ (256 CPU)
11:24:54 AM CPU %user %nice %system %iowait %steal %idle
11:25:04 AM 1 0.20 0.00 0.10 0.00 0.00 99.70
11:25:14 AM 1 10.10 0.00 0.30 0.00 0.00 89.60
11:25:24 AM 1 99.70 0.00 0.30 0.00 0.00 0.00
11:25:34 AM 1 99.70 0.00 0.30 0.00 0.00 0.00
11:25:44 AM 1 8.99 0.00 0.60 0.00 0.00 90.41
11:25:54 AM 1 0.10 0.00 0.20 0.00 0.00 99.70
11:26:04 AM 1 38.70 0.00 0.10 0.00 0.00 61.20
11:26:14 AM 1 99.80 0.00 0.10 0.00 0.00 0.10
11:26:24 AM 1 80.42 0.00 0.70 0.00 0.00 18.88
11:26:34 AM 1 0.10 0.00 0.20 0.00 0.00 99.70
Average: 1 43.78 0.00 0.29 0.00 0.00 55.93
007–5646–010 27
![](asset-2d.png)
Chapter 4
Data Process and Placement Tools
This chapter contains the following topics:
• "About Nonuniform Memory Access (NUMA) Computers" on page 29
• "About the Data and Process Placement Tools" on page 31
About Nonuniform Memory Access (NUMA) Computers
On symmetric multiprocessor (SMP) computers, all data is visible from all processors.
Each processor is functionally identical and has equal time access to every memory
address. That is, all processors have equally fast (symmetric) access to memory. These
types of systems are easy to assemble but have limited scalability due to memory
access times.
In contrast, a NUMA system has a shared address space, but the access time to
memory varies over physical address ranges and between processing elements. Each
processor has its own memory and can address the memory attached to another
processor through the Quick Path Interconnect (QPI).
In both cases, there is a single shared memory space and a single operating system
instance.
There are two levels of NUMA: intranode, managed by the Intel QPI, and internode,
managed through the SGI HUB ASIC and SGI NUMAlink technology.
The following topics explain other aspects of the SGI NUMA computers:
• "Distributed Shared Memory (DSM)" on page 29
• "ccNUMA Architecture" on page 30
Distributed Shared Memory (DSM)
Scalability is the measure of how the work done on a computing system changes as
you add CPUs, memory, network bandwidth, I/O capacity, and other resources.
Many factors, for example, memory latency across the system, can affect scalability.
007–5646–010 29
![](asset-2e.png)
4: Data Process and Placement Tools
In the SGI UV series systems, memory is physically distributed both within and
among the IRU enclosures, which consist of the compute, memory, and I/O blades.
However, memory is accessible to and shared by all devices, connected by
NUMAlink, within the single-system image (SSI). In other words, all components
connected by NUMAlink share a single Linux operating system, and they operate and
share the memory fabric of the system. Memory latency is the amount of time required
for a processor to retrieve data from memory. Memory latency is lowest when a
processor accesses local memory.
The following are the terms used to refer to the types of memory within a system:
• If a processor accesses memory that is on a compute node blade, that memory is
referred to as the node’s local memory.
• If processors access memory located on other blade nodes within the IRU or
within other NUMAlink IRUs, the memory is referred to as remote memory.
• The total memory within the NUMAlink system is referred to as global memory.
ccNUMA Architecture
As the name implies, the cache-coherent non-uniform memory access (ccNUMA)
architecture has two parts, cache coherency and nonuniform memory access, which
the following topics describe:
• "Cache Coherency" on page 30
• "Non-uniform Memory Access (NUMA)" on page 31
Cache Coherency
The SGI UV systems use caches to reduce memory latency. Although data exists in
local or remote memory, copies of the data can exist in various processor caches
throughout the system. Cache coherency keeps the cached copies consistent.
To keep the copies consistent, the ccNUMA architecture uses directory-based
coherence protocol. In directory-based coherence protocol, each 64–byte block of
memory has an entry in a table that is referred to as a directory. Like the blocks of
memory that they represent, the directories are distributed among the compute and
memory blade nodes. A block of memory is also referred to as a cache line.
Each directory entry indicates the state of the memory block that it represents. For
example, when the block is not cached, it is in an unowned state. When only one
30 007–5646–010
![](asset-2f.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
processor has a copy of the memory block, it is in an exclusive state. When more than
one processor has a copy of the block, it is in a shared state. A bit vector indicates the
caches that may contain a copy.
When a processor modifies a block of data, the processors that have the same block of
data in their caches must be notified of the modification. The SGI UV systems use an
invalidation method to maintain cache coherence. The invalidation method purges all
unmodified copies of the block of data, and the processor that wants to modify the
block receives exclusive ownership of the block.
Non-uniform Memory Access (NUMA)
In DSM systems, memory is physically located at various distances from the
processors. As a result, memory access times (latencies) are different or nonuniform.
For example, it takes less time for a processor blade to reference its locally installed
memory than to reference remote memory.
About the Data and Process Placement Tools
For cc-NUMA systems, like the SGI UV systems, performance degrades when the
application accesses remote memory, versus local memory. Because the Linux
operating system has a tendency to migrate processes, SGI recommends that you use
the data and process placement tools.
Special optimization applies to SGI UV systems to exploit multiple paths to memory,
as follows:
• By default, all pages are allocated with a first touch policy.
• The initialization loop, if executed serially, gets pages from single node.
Perform initialization in parallel, such that each processor initializes data that it is
likely to access later for calculation.
• In the parallel loop, multiple processors access that one memory page.
Figure 4-1 on page 32, shows how to code to get good data placement.
007–5646–010 31
![](asset-30.png)
4: Data Process and Placement Tools
Figure 4-1 Coding to Get Good Data Placement
The dplace(1) tool, the taskset(1) command, and the cpuset tools are built upon
the cpusets API. These tools enable your applications to avoid poor data locality
caused by process or thread drift from CPU to CPU. The omplace(1) tool works like
the dplace(1) tool and is designed for use with OpenMP applications. The
differences among these tools are as follows:
• The taskset(1) command restricts execution to the listed set of CPUs when you
specify the -c or --cpu-list option. The process is free to move among the
CPUs that you specify.
• The dplace(1) tool differs from taskset(1) in that dplace(1) binds processes to
specified CPUs in round-robin fashion. After a process is pinned, it does not
migrate, so you can use this for increasing the performance and reproducibility of
parallel codes.
32 007–5646–010
![](asset-31.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
• Cpusets are named subsets of system cpus/memories and are used extensively in
batch environments. For more information about cpusets, see the SGI Cpuset
Software Guide.
The following topics provide more information about the data and process placement
utilities:
• "About cpusets and Control Groups (cgroups)" on page 33
•"dplace Command" on page 35
•"omplace Command" on page 41
•"taskset Command" on page 42
•"numactl Command" on page 44
•"dlook Command" on page 45
About cpusets and Control Groups (cgroups)
SGI systems support both cgroups and cpusets. The cpusets are a subsystem of
cgroups.
For information about cpusets and cgroups, see the following:
• "Using cpusets" on page 33
• "Using cgroups" on page 34
Using cpusets
The cpuset facility is a workload manager tool that permits a system administrator to
restrict the number of processor and memory resources that a process or set of
processes can use. A cpuset defines a list of CPUs and memory nodes. A process
contained in a cpuset can execute only on the CPUs in that cpuset and can only
allocate memory on the memory nodes in that cpuset. Essentially, a cpuset provides
you with a CPU and a memory container or a soft partition within which you can run
sets of related tasks. Using cpusets on an SGI UV system improves cache locality and
memory access times and can substantially improve an application’s performance and
run-time repeatability.
Be aware that when placed in a cpuset, certain kernel threads can exhibit undesirable
behavior. In general, kernel threads are not confined to a cpuset, but when a
007–5646–010 33
![](asset-32.png)
4: Data Process and Placement Tools
bootcpuset is created, all the kernel threads that are able to be placed in a cpuset
become confined to the bootcpuset. In the case of the khugepaged daemon, this is
undesirable because khugepaged becomes unable to allocate memory for processes
that are on nodes outside of its cpuset. As a workaround, remove khugepaged from
the bootcpuset after the machine is up and running. The following procedure
explains how to implement the workaround.
Procedure 4-1 To remove khugepaged from the bootcpuset
1. Type the following command to retrieve the process ID of the khugepaged
daemon:
ps -ef | grep khugepaged | grep -v grep
For example, in the following output, 1054 is the process ID:
#ps -ef | grep khugepaged | grep -v grep
root 1054 2 0 Mar04 ? 00:00:02 [khugepaged]
2. Use the echo(1) command, in the following format, to remove khugepaged from
the bootcpuset:
echo khugepaged_pid > /dev/cpuset/tasks
For pid, specify the process ID for the khugepaged daemon.
3. (Optional) Script the preceding lines and run the script at boot time to ensure that
the khugepaged thread is always removed from the bootcpuset.
For general information about cpusets, see the following:
•SGI Cpuset Software Guide
• https://www.kernel.org/doc/Documentation/cgroups/cpusets.txt
Using cgroups
If you use cgroups, you can exert finer control over memory than is possible with
cpusets. If you use cgroups, be aware that their use can result in a 1–5% memory
overhead penalty. If you use a batch scheduler, verify that it supports cgroups before
you configure cgroups.
For general information about cgroups, see the following:
https://www.kernel.org/doc/Documentation/cgroups/cgroups.txt
34 007–5646–010
![](asset-33.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
dplace Command
You can use the dplace(1) command to improve the performance of processes
running on your SGI nonuniform memory access (NUMA) machine.
By default, memory is allocated to a process on the node on which the process is
executing. If a process moves from node to node while it is running, a higher
percentage of memory references are made to remote nodes. Because remote accesses
typically have higher access times, performance can degrade. CPU instruction
pipelines also have to be reloaded.
The dplace(1) command specifies scheduling and memory placement policies for the
process. You can use the dplace command to bind a related set of processes to
specific CPUs or nodes to prevent process migrations. In some cases, this improves
performance because a higher percentage of memory accesses are made to local nodes.
Processes always execute within a cpuset. The cpuset specifies the CPUs on which a
process can run. By default, processes usually execute in a cpuset that contains all the
CPUs in the system. For information about cpusets, see the SGI Cpuset Software Guide.
The dplace(1) command creates a placement container that includes all the CPUs, or
a subset of CPUs, of a cpuset. The dplace process is placed in this container and, by
default, is bound to the first CPU of the cpuset associated with the container. Then
dplace invokes exec to run the command.
The command runs within this placement container and remains bound to the first
CPU of the container. As the command forks child processes, the child processes
inherit the container and are bound to the next available CPU of the container.
If you do not specify a placement file, dplace binds processes sequentially in a
round-robin fashion to CPUs of the placement container. For example, if the current
cpuset consists of physical CPUs 2, 3, 8, and 9, the first process launched by dplace
is bound to CPU 2. The first child process forked by this process is bound to CPU 3.
The next process, regardless of whether it is forked by a parent or a child, is bound to
CPU 8, and so on. If more processes are forked than there are CPUs in the cpuset,
binding starts over with the first CPU in the cpuset.
For more information about dplace(1), see the dplace(1) man page. The dplace(1)
man page also includes examples of how to use the command.
007–5646–010 35
![](asset-34.png)
4: Data Process and Placement Tools
Example 4-1 Using the dplace(1) command with MPI Programs
The following command improves the placement of MPI programs on NUMA systems
and verifies placement of certain data structures of a long-running MPI program:
%mpirun -np 64 /usr/bin/dplace -s1 -c 0-63 ./a.out
The -s1 parameter causes dplace(1) to start placing processes with the second
process, p1. The first process, p0, is not placed because it is associated with the job
launch, not with the job itself. The -c 0-63 parameter causes dplace(1) to use
processors 0–63.
You can then use the dlook(1) command to verify placement of the data structures in
another window on one of the slave thread PIDs. For more information about the
dlook command, see "dlook Command" on page 45 and the dlook(1) man page.
Example 4-2 Using the dplace(1) command with OpenMP Programs
The following command runs an OpenMP program on logical CPUs 4 through 7
within the current cpuset:
%efc -o prog -openmp -O3 program.f
%setenv OMP_NUM_THREADS 4
%dplace -c4-7 ./prog
Example 4-3 Using the dplace(1) command with OpenMP Programs
The dplace(1) command has a static load balancing feature, so you do not have to
supply a CPU list. To place prog1 on logical CPUs 0 through 3 and prog2 on logical
CPUs 4 through 7, type the following:
%setenv OMP_NUM_THREADS 4
%dplace ./prog1 &
%dplace ./prog2 &
You can use the dplace -q command to display the static load information.
Example 4-4 Using the dplace(1) command with Linux commands
The following examples assume that you run the dplace commands from a shell that
runs in a cpuset consisting of physical CPUs 8 through 15.
36 007–5646–010
![](asset-35.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Command Run Location
dplace -c2 date
Runs the date command on physical CPU 10.
dplace make linux
Runs gcc and related processes on physical CPUs 8 through 15.
dplace -c0-4,6 make linux
Runs gcc and related processes on physical CPUs 8 through 12 or 14.
taskset 4,5,6,7 dplace app
The taskset command restricts execution to physical CPUs 12
through 15. The dplace command sequentially binds processes to
CPUs 12 through 15.
Example 4-5 Using the dplace command and a debugger for verification
To use the dplace command accurately, you should know how your placed tasks are
being created in terms of the fork,exec, and pthread_create calls. Determine
whether each of these worker calls are an MPI rank task or are groups of pthreads
created by rank tasks. Here is an example of two MPI ranks, each creating three
threads:
cat <<EOF > placefile
firsttask cpu=0
exec name=mpiapp cpu=1
fork name=mpiapp cpu=4-8:4 exact
thread name=mpiapp oncpu=4 cpu=5-7 exact thread name=mpiapp oncpu=8
cpu=9-11 exact EOF
# mpirun is placed on cpu 0 in this example
# the root mpiapp is placed on cpu 1 in this example
# or, if your version of dplace supports the "cpurel=" option:
# firsttask cpu=0
# fork name=mpiapp cpu=4-8:4 exact
# thread name=mpiapp oncpu=4 cpurel=1-3 exact
# create 2 rank tasks, each will pthread_create 3 more
007–5646–010 37
![](asset-36.png)
4: Data Process and Placement Tools
# ranks will be on 4 and 8
# thread children on 5,6,7 9,10,11
dplace -p placefile mpirun -np 2 ~cpw/bin/mpiapp -P 3 -l
exit
You can use the debugger to determine if it is working. It should show two MPI rank
applications, each with three pthreads, as follows:
>> pthreads | grep mpiapp
px *(task_struct *)e00002343c528000 17769 17769 17763 0 mpiapp
member task: e000013817540000 17795 17769 17763 0 5 mpiapp
member task: e000013473aa8000 17796 17769 17763 0 6 mpiapp
member task: e000013817c68000 17798 17769 17763 0 mpiapp
px *(task_struct *)e0000234704f0000 17770 17770 17763 0 mpiapp
member task: e000023466ed8000 17794 17770 17763 0 9 mpiapp
member task: e00002384cce0000 17797 17770 17763 0 mpiapp
member task: e00002342c448000 17799 17770 17763 0 mpiapp
You can also use the debugger to see a root application, the parent of the two MPI
rank applications, as follows:
>> ps | grep mpiapp
0xe00000340b300000 1139 17763 17729 1 0xc800000 - mpiapp
0xe00002343c528000 1139 17769 17763 0 0xc800040 - mpiapp
0xe0000234704f0000 1139 17770 17763 0 0xc800040 8 mpiapp
These are placed as specified:
>> oncpus e00002343c528000 e000013817540000 e000013473aa8000
>> e000013817c68000 e0
000234704f0000 e000023466ed8000 e00002384cce0000 e00002342c448000
task: 0xe00002343c528000 mpiapp cpus_allowed: 4
task: 0xe000013817540000 mpiapp cpus_allowed: 5
task: 0xe000013473aa8000 mpiapp cpus_allowed: 6
task: 0xe000013817c68000 mpiapp cpus_allowed: 7
task: 0xe0000234704f0000 mpiapp cpus_allowed: 8
task: 0xe000023466ed8000 mpiapp cpus_allowed: 9
task: 0xe00002384cce0000 mpiapp cpus_allowed: 10
task: 0xe00002342c448000 mpiapp cpus_allowed: 11
38 007–5646–010
![](asset-37.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Example 4-6 Using the dplace(1) command for compute thread placement troubleshooting
Sometimes compute threads do not end up on unique processors when using
commands such a dplace(1) or profile.pl. For information about Perfsuite, see
the following:
"Profiling with PerfSuite" on page 16
In this example, assume that the dplace -s1 -c0-15 command bound 16
processes to run on 0-15 CPUs. However, output from the top(1) command shows
only 13 CPUs running with CPUs 13, 14, and 15 still idle, and CPUs 0, 1 and 2 are
shared with 6 processes.
263 processes: 225 sleeping, 18 running, 3 zombie, 17 stopped
CPU states: cpu user nice system irq softirq iowait idle
total 1265.6% 0.0% 28.8% 0.0% 11.2% 0.0% 291.2%
cpu00 100.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
cpu01 90.1% 0.0% 0.0% 0.0% 9.7% 0.0% 0.0%
cpu02 99.9% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
cpu03 99.9% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
cpu04 100.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
cpu05 100.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
cpu06 100.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
cpu07 88.4% 0.0% 10.6% 0.0% 0.8% 0.0% 0.0%
cpu08 100.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
cpu09 99.9% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
cpu10 99.9% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
cpu11 88.1% 0.0% 11.2% 0.0% 0.6% 0.0% 0.0%
cpu12 99.7% 0.0% 0.2% 0.0% 0.0% 0.0% 0.0%
007–5646–010 39
![](asset-38.png)
4: Data Process and Placement Tools
cpu13 0.0% 0.0% 2.5% 0.0% 0.0% 0.0% 97.4%
cpu14 0.8% 0.0% 1.6% 0.0% 0.0% 0.0% 97.5%
cpu15 0.0% 0.0% 2.4% 0.0% 0.0% 0.0% 97.5%
Mem: 60134432k av, 15746912k used, 44387520k free, 0k shrd,
672k buff
351024k active, 13594288k inactive
Swap: 2559968k av, 0k used, 2559968k free
2652128k cached
PID USER PRI NI SIZE RSS SHARE STAT %CPU %MEM TIME CPU COMMAND
7653 ccao 25 0 115G 586M 114G R 99.9 0.9 0:08 3 mocassin
7656 ccao 25 0 115G 586M 114G R 99.9 0.9 0:08 6 mocassin
7654 ccao 25 0 115G 586M 114G R 99.8 0.9 0:08 4 mocassin
7655 ccao 25 0 115G 586M 114G R 99.8 0.9 0:08 5 mocassin
7658 ccao 25 0 115G 586M 114G R 99.8 0.9 0:08 8 mocassin
7659 ccao 25 0 115G 586M 114G R 99.8 0.9 0:08 9 mocassin
7660 ccao 25 0 115G 586M 114G R 99.8 0.9 0:08 10 mocassin
7662 ccao 25 0 115G 586M 114G R 99.7 0.9 0:08 12 mocassin
7657 ccao 25 0 115G 586M 114G R 88.5 0.9 0:07 7 mocassin
7661 ccao 25 0 115G 586M 114G R 88.3 0.9 0:07 11 mocassin
7649 ccao 25 0 115G 586M 114G R 55.2 0.9 0:04 2 mocassin
7651 ccao 25 0 115G 586M 114G R 54.1 0.9 0:03 1 mocassin
7650 ccao 25 0 115G 586M 114G R 50.0 0.9 0:04 0 mocassin
40 007–5646–010
![](asset-39.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
7647 ccao 25 0 115G 586M 114G R 49.8 0.9 0:03 0 mocassin
7652 ccao 25 0 115G 586M 114G R 44.7 0.9 0:04 2 mocassin
7648 ccao 25 0 115G 586M 114G R 35.9 0.9 0:03 1 mocassin
Even if an application starts some threads executing for a very short time, the threads
still have taken a token in the CPU list. Then, when the compute threads are finally
started, the list is exhausted and restarts from the beginning. Consequently, some
threads end up sharing the same CPU. To bypass this, try to eliminate the ghost
thread creation, as follows:
• Check for a call to the system function. This is often responsible for the placement
failure due to unexpected thread creation. If all the compute processes have the
same name, you can do this by issuing a command such as the following:
%dplace -c0-15 -n compute-process-name ...
• You can also run dplace -e -c0-32 on 16 CPUs to understand the pattern of
the thread creation. If this pattern is the same from one run to the other
(unfortunately race between thread creation often occurs), you can find the right
flag to dplace. For example, if you want to run on CPUs 0-3, with dplace -e
-C0-16 and you see that threads are always placed on CPU 0, 1, 5, and 6, then
one of the following commands should place your threads correctly:
dplace -e -c0,1,x,x,x,2,3
or
dplace -x24 -c0-3 # x24 =11000, place the 2 first and skip 3 before placing
omplace Command
The omplace(1) command controls the placement of MPI processes and OpenMP
threads. This command is a wrapper script for dplace(1). Use omplace(1), rather
than dplace(1), if your application uses MPI, OpenMP, pthreads, or hybrid
MPI/OpenMP and MPI/pthreads codes. The omplace(1) command generates the
proper dplace(1) placement file syntax automatically. It also supports some unique
options, such as block-strided CPU lists.
The omplace(1) command causes the successive threads in a hybrid MPI/OpenMP
job to be placed on unique CPUs. The CPUs are assigned in order from the effective
CPU list within the containing cpuset. The CPU placement is performed by
007–5646–010 41
![](asset-3a.png)
4: Data Process and Placement Tools
dynamically generating a placement file and invoking dplace(1) with the MPI job
launch.
For example, to run two MPI processes with four threads per process, and to display
the generated placement file, type a command similar to the following:
#mpirun -np 2 omplace -nt 4 -vv ./a.out
The preceding command places the threads as follows:
rank 0 thread 0 on CPU 0
rank 0 thread 1 on CPU 1
rank 0 thread 2 on CPU 2
rank 0 thread 3 on CPU 3
rank 1 thread 0 on CPU 4
rank 1 thread 1 on CPU 5
rank 1 thread 2 on CPU 6
rank 1 thread 3 on CPU 7
For more information, see the omplace(1) man page and the SGI MPI and SGI
SHMEM User Guide.
taskset Command
You can use the taskset(1) command to perform the following tasks:
• Restricting execution to a list of CPUs. Use the -c parameter and the
--cpu-list parameter.
• Retrieving or setting the CPU affinity of a process. Use the following parameters:
taskset [options]mask command [arg]...
taskset [options]-p[mask]pid
• Launching a new command with a specified CPU affinity.
CPU affinity is a scheduler property that bonds a process to a given set of CPUs on
the system. The Linux scheduler honors the given CPU affinity and runs the process
only on the specified CPUs. The process does not run on any other CPUs. Note that
the scheduler also supports natural CPU affinity in which the scheduler attempts to
keep processes on the same CPU as long as practical for performance reasons.
Forcing a specific CPU affinity is useful only in certain applications.
42 007–5646–010
![](asset-3b.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
The CPU affinity is represented as a bitmask, with the lowest order bit corresponding
to the first logical CPU and the highest order bit corresponding to the last logical
CPU. The mask parameter can specify more CPUs than are present. In other words, it
might be true that not all CPUs specified in the mask exist on a given system. A
retrieved mask reflects only the bits that correspond to CPUs physically on the
system. If the mask does not correspond to any valid CPUs on the system, the mask
is invalid, and the system returns an error. The masks are typically specified in
hexadecimal notation. For example:
mask specification CPUs specified
0x00000001 Processor #0
0x00000003 Processors #0 and #1
0xFFFFFFFF All processors (#0 through #31)
When taskset(1) returns, it is guaranteed that the given program has been
scheduled to a valid CPU.
The taskset(1) command does not pin a task to a specific CPU. Rather, it restricts a
task so that it does not run on any CPU that is not in the CPU list. For example, if
you use taskset(1) to launch an application that forks multiple tasks, it is possible
that the scheduler initially assigns multiple tasks to the same CPU, even though there
are idle CPUs that are in the CPU list. Scheduler load balancing software eventually
distributes the tasks so that CPU-bound tasks run on different CPUs. However, the
exact placement is not predictable and can vary from run to run. After the tasks are
evenly distributed, a task can jump to a different CPU. This outcome can affect
memory latency as pages that were node-local before the jump can be remote after
the jump.
If you are running an MPI application, SGI recommends that you do not use the
taskset(1) command because the taskset(1) command can pin the MPI shepherd
process, which wastes a CPU, and then put the remaining working MPI rank on one
of the CPUs that already had some other rank running on it. Instead of taskset(1),
SGI recommends that you use the dplace(1) command or the environment variable
MPI_DSM_CPULIST. For more information, see "dplace Command" on page 35.
If you are using a batch scheduler that creates and destroys cpusets dynamically, SGI
recommends that you use the MPI_DSM_DISTRIBUTE environment variable instead
of either the MPI_DSM_CPULIST environment variable or the dplace(1) command.
Example 1. The following example shows how to run an MPI program on eight CPUs:
#mpirun -np 8 dplace -s1 -c10,11,16-21 myMPIapplication ...
007–5646–010 43
![](asset-3c.png)
4: Data Process and Placement Tools
Example 2. The following example sets the MPI_DSM_CPULIST variable:
#setenv MPI_DSM_CPULIST 10,11,16-21 mpirun -np 8 myMPIapplication ...
Example 3. The following example runs an executable on CPU 1. The mask for CPU 1
is 0x2, so type the following:
#taskset 0x2 executable_name
Example 4. The following example moves PID 14057 to CPU 0. The mask for CPU 0 is
0x1, so type the following:
#taskset -p 0x1 14057
Example 5. The following example runs an MPI Abaqus/Standard job on an SGI UV
system with eight CPUs. Standard input is redirected to /dev/null to avoid a
SIGTTIN signal for MPT applications. Type the following:
#taskset -c 8-15 ./runme < /dev/null &
Example 6. The following example uses the taskset(1) command to lock a given
process to a particular CPU, CPU5, and then uses the profile(1) command to profile
it. The second command moves the process to another CPU, CPU3. Type the
following:
#taskset -p -c 5 16269
pid 16269’s current affinity list: 0-15
pid 16269’s new affinity list: 5
#taskset -p 16269 -c 3
pid 16269’s current affinity list: 5
pid 16269’s new affinity list: 3
For more information, see the taskset(1) man page.
numactl Command
The numactl(8) command runs processes with a specific NUMA scheduling or
memory placement policy. The policy is set for an executable command and inherited
by all of its children. In addition, numactl(8) can set a persistent policy for shared
memory segments or files. For more information, see the numactl(8) man page.
44 007–5646–010
![](asset-3d.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
dlook Command
You can use the dlook(1) command to find out where, in memory, the operating
system is placing your application’s pages and how much system and user CPU time
it is consuming. The command allows you to display the memory map and CPU
usage for a specified process. For each page in the virtual address space of the
process, dlook(1) generates the following information:
• The object that owns the page, such as a file, SYSV shared memory, a device
driver, and so on.
• The type of page, such as random access memory (RAM), FETCHOP, IOSPACE,
and so on.
If the page type is RAM memory, the following information is displayed:
– Memory attributes, such as, SHARED, DIRTY, and so on
– The node on which the page is located
– The physical address of the page (optional)
Example 4-7 Using dlook(1) with a PID
To specify a PID as a parameter to the dlook(1) command, you must be the owner of
the process, or you must be logged in as the root user. The following dlook(1)
command example shows output for the sleep process, with a PID of 191155:
$dlook 191155
_______________________________________________________________________________
Peek: sleep
Pid: 191155 Fri Sep 27 17:14:01 2013
Process memory map:
00400000-00406000 r-xp 00000000 08:08 262250 /bin/sleep
[0000000000400000-0000000000401000] 1 page on node
4 MEMORY|SHARED
[0000000000401000-0000000000402000] 1 page on node
5 MEMORY|SHARED
[0000000000403000-0000000000404000] 1 page on node
7 MEMORY|SHARED
[0000000000404000-0000000000405000] 1 page on node
8 MEMORY|SHARED
007–5646–010 45
![](asset-3e.png)
4: Data Process and Placement Tools
00605000-00606000 rw-p 00005000 08:08 262250 /bin/sleep
[0000000000605000-0000000000606000] 1 page on node
2 MEMORY|RW|DIRTY
00606000-00627000 rw-p 00000000 00:00 0 [heap]
[0000000000606000-0000000000608000] 2 pages on node
2 MEMORY|RW|DIRTY
7ffff7dd8000-7ffff7ddd000 rw-p 00000000 00:00 0
[00007ffff7dd8000-00007ffff7dda000] 2 pages on node
2 MEMORY|RW|DIRTY
[00007ffff7ddc000-00007ffff7ddd000] 1 page on node
2 MEMORY|RW|DIRTY
7ffff7fde000-7ffff7fe1000 rw-p 00000000 00:00 0
[00007ffff7fde000-00007ffff7fe1000] 3 pages on node
2 MEMORY|RW|DIRTY
7ffff7ffa000-7ffff7ffb000 rw-p 00000000 00:00 0
[00007ffff7ffa000-00007ffff7ffb000] 1 page on node
2 MEMORY|RW|DIRTY
7ffff7ffb000-7ffff7ffc000 r-xp 00000000 00:00 0 [vdso]
[00007ffff7ffb000-00007ffff7ffc000] 1 page on node
7 MEMORY|SHARED
7ffff7ffe000-7ffff7fff000 rw-p 00000000 00:00 0
[00007ffff7ffe000-00007ffff7fff000] 1 page on node
2 MEMORY|RW|DIRTY
7ffffffea000-7ffffffff000 rw-p 00000000 00:00 0 [stack]
[00007fffffffd000-00007ffffffff000] 2 pages on node
2 MEMORY|RW|DIRTY
ffffffffff600000-ffffffffff601000 r-xp 00000000 00:00 0 [vsyscall]
[ffffffffff600000-ffffffffff601000] 1 page on node
0 MEMORY|DIRTY|RESERVED
_______________________________________________________________________________
46 007–5646–010
![](asset-3f.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
The dlook(1) command generates the name of the process (Peek: sleep), the
process ID, the time, and the date it was invoked. It provides total user and system
CPU time in seconds for the process.
Under the Process memory map heading, the dlook(1) command generates
information about a process from the /proc/pid/cpu and /proc/pid/maps files. On
the left, it shows the memory segment with the offsets below in decimal. In the middle
of the output, it shows the type of access, time of execution, the PID, and the object
that owns the memory, which in this example is /lib/ld-2.2.4.so. The characters
sor pindicate whether the page is mapped as sharable (s) with other processes or is
private (p). The right side of the output page shows the number of pages of memory
consumed and shows the nodes on which the pages reside. A page is 16,384 bytes.
The node numbers reported by the dlook(1) command correspond to the numbers
reported by the cpumap(1) command under the section Processor Numbering on
Node(s). For more information, see the cpumap(1) command description in
"Determining System Configuration" on page 8.
Dirty memory means that the memory has been modified by a user.
Example 4-8 Using dlook(1) with a command
When you pass a command as an argument to dlook(1), you specify the command
and optional command arguments. The dlook(1) command issues an exec call on
the command and passes the command arguments. When the process terminates,
dlook(1) prints information about the process, as shown in the following example:
$dlook date
Thu Aug 22 10:39:20 CDT 2002
_______________________________________________________________________________
Exit: date
Pid: 4680 Thu Aug 22 10:39:20 2002
Process memory map:
2000000000030000-200000000003c000 rw-p 0000000000000000 00:00 0
[2000000000030000-200000000003c000] 3 pages on node 3 MEMORY|DIRTY
20000000002dc000-20000000002e4000 rw-p 0000000000000000 00:00 0
[20000000002dc000-20000000002e4000] 2 pages on node 3 MEMORY|DIRTY
2000000000324000-2000000000334000 rw-p 0000000000000000 00:00 0
007–5646–010 47
![](asset-40.png)
4: Data Process and Placement Tools
[2000000000324000-2000000000328000] 1 page on node 3 MEMORY|DIRTY
4000000000000000-400000000000c000 r-xp 0000000000000000 04:03 9657220 /bin/date
[4000000000000000-400000000000c000] 3 pages on node 1 MEMORY|SHARED
6000000000008000-6000000000010000 rw-p 0000000000008000 04:03 9657220 /bin/date
[600000000000c000-6000000000010000] 1 page on node 3 MEMORY|DIRTY
6000000000010000-6000000000014000 rwxp 0000000000000000 00:00 0
[6000000000010000-6000000000014000] 1 page on node 3 MEMORY|DIRTY
60000fff80000000-60000fff80004000 rw-p 0000000000000000 00:00 0
[60000fff80000000-60000fff80004000] 1 page on node 3 MEMORY|DIRTY
60000fffffff4000-60000fffffffc000 rwxp ffffffffffffc000 00:00 0
[60000fffffff4000-60000fffffffc000] 2 pages on node 3 MEMORY|DIRTY
Example 4-9 Using the dlook(1) command with the -s secs option
If you use the dlook(1) command with the -s secs option, the information is
sampled at regular internals. The example command and output are as follows:
$dlook -s 5 sleep 50
Exit: sleep
Pid: 5617 Thu Aug 22 11:16:05 2002
Process memory map:
2000000000030000-200000000003c000 rw-p 0000000000000000 00:00 0
[2000000000030000-200000000003c000] 3 pages on node 3 MEMORY|DIRTY
2000000000134000-2000000000140000 rw-p 0000000000000000 00:00 0
20000000003a4000-20000000003a8000 rw-p 0000000000000000 00:00 0
[20000000003a4000-20000000003a8000] 1 page on node 3 MEMORY|DIRTY
20000000003e0000-20000000003ec000 rw-p 0000000000000000 00:00 0
[20000000003e0000-20000000003ec000] 3 pages on node 3 MEMORY|DIRTY
4000000000000000-4000000000008000 r-xp 0000000000000000 04:03 9657225 /bin/sleep
[4000000000000000-4000000000008000] 2 pages on node 3 MEMORY|SHARED
48 007–5646–010
![](asset-41.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
6000000000004000-6000000000008000 rw-p 0000000000004000 04:03 9657225 /bin/sleep
[6000000000004000-6000000000008000] 1 page on node 3 MEMORY|DIRTY
6000000000008000-600000000000c000 rwxp 0000000000000000 00:00 0
[6000000000008000-600000000000c000] 1 page on node 3 MEMORY|DIRTY
60000fff80000000-60000fff80004000 rw-p 0000000000000000 00:00 0
[60000fff80000000-60000fff80004000] 1 page on node 3 MEMORY|DIRTY
60000fffffff4000-60000fffffffc000 rwxp ffffffffffffc000 00:00 0
[60000fffffff4000-60000fffffffc000] 2 pages on node 3 MEMORY|DIRTY
Example 4-10 Using the dlook(1) command with the mpirun(1) command
You can run a Message Passing Interface (MPI) job using the mpirun(1) command
and generate the memory map for each thread, or you can redirect the ouput to a file.
In the following example, the output has been abbreviated and bold headings added
for easier reading:
$mpirun -np 8 dlook -o dlook.out ft.C.8
Contents of dlook.out:
_______________________________________________________________________________
Exit: ft.C.8
Pid: 2306 Fri Aug 30 14:33:37 2002
Process memory map:
2000000000030000-200000000003c000 rw-p 0000000000000000 00:00 0
[2000000000030000-2000000000034000] 1 page on node 21 MEMORY|DIRTY
[2000000000034000-200000000003c000] 2 pages on node 12 MEMORY|DIRTY|SHARED
2000000000044000-2000000000060000 rw-p 0000000000000000 00:00 0
[2000000000044000-2000000000050000] 3 pages on node 12 MEMORY|DIRTY|SHARED
...
_______________________________________________________________________________
_______________________________________________________________________________
Exit: ft.C.8
Pid: 2310 Fri Aug 30 14:33:37 2002
007–5646–010 49
![](asset-42.png)
4: Data Process and Placement Tools
Process memory map:
2000000000030000-200000000003c000 rw-p 0000000000000000 00:00 0
[2000000000030000-2000000000034000] 1 page on node 25 MEMORY|DIRTY
[2000000000034000-200000000003c000] 2 pages on node 12 MEMORY|DIRTY|SHARED
2000000000044000-2000000000060000 rw-p 0000000000000000 00:00 0
[2000000000044000-2000000000050000] 3 pages on node 12 MEMORY|DIRTY|SHARED
[2000000000050000-2000000000054000] 1 page on node 25 MEMORY|DIRTY
...
_______________________________________________________________________________
_______________________________________________________________________________
Exit: ft.C.8
Pid: 2307 Fri Aug 30 14:33:37 2002
Process memory map:
2000000000030000-200000000003c000 rw-p 0000000000000000 00:00 0
[2000000000030000-2000000000034000] 1 page on node 30 MEMORY|DIRTY
[2000000000034000-200000000003c000] 2 pages on node 12 MEMORY|DIRTY|SHARED
2000000000044000-2000000000060000 rw-p 0000000000000000 00:00 0
[2000000000044000-2000000000050000] 3 pages on node 12 MEMORY|DIRTY|SHARED
[2000000000050000-2000000000054000] 1 page on node 30 MEMORY|DIRTY
...
_______________________________________________________________________________
_______________________________________________________________________________
Exit: ft.C.8
Pid: 2308 Fri Aug 30 14:33:37 2002
Process memory map:
2000000000030000-200000000003c000 rw-p 0000000000000000 00:00 0
[2000000000030000-2000000000034000] 1 page on node 0 MEMORY|DIRTY
[2000000000034000-200000000003c000] 2 pages on node 12 MEMORY|DIRTY|SHARED
2000000000044000-2000000000060000 rw-p 0000000000000000 00:00 0
[2000000000044000-2000000000050000] 3 pages on node 12 MEMORY|DIRTY|SHARED
[2000000000050000-2000000000054000] 1 page on node 0 MEMORY|DIRTY
...
50 007–5646–010
![](asset-43.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
For more information about the dlook(1) command, see the dlook(1) man page.
007–5646–010 51
![](asset-45.png)
Chapter 5
Performance Tuning
This chapter includes the following topics:
• "About Performance Tuning" on page 53
• "Single Processor Code Tuning" on page 54
• "Tuning Multiprocessor Codes" on page 64
• "Understanding Parallel Speedup and Amdahl’s Law" on page 70
• "Gustafson’s Law" on page 76
• "Floating-point Program Performance" on page 76
• "About MPI Application Tuning" on page 77
• "Using Transparent Huge Pages (THPs) in MPI and SHMEM Applications" on
page 80
• "Enabling Huge Pages in MPI and SHMEM Applications on Systems Without
THP" on page 82
About Performance Tuning
After you analyze your code to determine where performance bottlenecks are
occurring, you can turn your attention to making your programs run their fastest.
One way to do this is to use multiple CPUs in parallel processing mode. However,
this should be the last step. The first step is to make your program run as efficiently
as possible on a single processor system and then consider ways to use parallel
processing.
Intel provides tuning information, including information about the Intel processors, at
the following website:
http://www.intel.com/content/www/us/en/architecture-and-technology/64-ia-32-
architectures-optimization-manual.html.
This chapter describes the process of tuning your application for a single processor
system and then tuning it for parallel processing. It also addresses how to improve
the performance of floating-point programs and MPI applications.
007–5646–010 53
![](asset-46.png)
5: Performance Tuning
Single Processor Code Tuning
Several basic steps are used to tune performance of single-processor code:
• Get the expected answers and then tune performance. For details, see "Getting the
Correct Results" on page 54.
• Use existing tuned code, such as that found in math libraries and scientific library
packages. For details, see "Using Tuned Code" on page 55.
• Determine what needs tuning. For details, see "Determining Tuning Needs" on
page 56.
• Use the compiler to do the work. For details, see "Using Compiler Options to
Optimize Performance" on page 56.
• Consider tuning cache performance. For details, see "Tuning the Cache
Performance" on page 61.
• Set environment variables to enable higher-performance memory management
mode. For details, see "Managing Memory" on page 63.
Getting the Correct Results
One of the first steps in performance tuning is to verify that the correct answers are
being obtained. After the correct answers are obtained, tuning can be done. You can
verify answers by initially disabling specific optimizations and limiting default
optimizations. This can be accomplished by using specific compiler options and by
using debugging tools.
The following compiler options emphasize tracing and porting over performance:
Option Purpose
-O Disables all optimization. The default is -O2.
-g Preserves symbols for debugging. In the past, using -g
automatically put down the optimization level. In Intel
compiler today, you can use -O3 with -g.
-fp-model Lets you specify compiler rules for the following:
• Value safety
• Floating-point (FP) expression evaluation
54 007–5646–010
![](asset-47.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
• FPU environment access
• Precise FP exceptions
• FP contractions
The default is -fp-model fast=1. Note that -mp is
an old option and is replaced by -fp-model.
-r,-i Sets default real, integer, and logical sizes to 8 bytes,
which are useful for porting codes from Cray, Inc.
systems. This option explicitly declares intrinsic and
external library functions.
For information about debugging tools that you can use to verify that correct answers
are being obtained, see the following:
"About Debugging" on page 17
Managing Heap Corruption Problems
You can use environment variables to check for heap corruption problems in
programs that use glibc malloc/free dynamic memory management routines.
Set the MALLOC_CHECK_ environment variable to 1to print diagnostic messages or to
2to abort immediately when heap corruption is detected.
Overruns and underruns are circumstances in which an access to an array is outside
the declared boundary of the array. Underruns and overruns cannot be
simultaneously detected. The default behavior is to place inaccessible pages
immediately after allocated memory.
Using Tuned Code
Where possible, use code that has already been tuned for optimum hardware
performance.
The following mathematical functions should be used where possible to help obtain
best results:
• MKL, Intel’s Math Kernel Library. This library includes BLAS, LAPACK, and FFT
routines.
007–5646–010 55
![](asset-48.png)
5: Performance Tuning
• VML, the Vector Math Library, available as part of the MKL package
(libmkl_vml_itp.so).
• Standard Math library. Standard math library functions are provided with the
Intel compiler’s libimf.a file. If the -lm option is specified, glibc libm
routines are linked in first.
Documentation is available for MKL and VML at the following website:
https://software.intel.com/en-us/intel-parallel-studio-xe-support/documentation
Determining Tuning Needs
Use the following tools to determine what points in your code might benefit from
tuning:
Tool Purpose
time(1) Obtains an overview of user, system, and elapsed time.
gprof(1) Obtains an execution profile of your program. This is a pcsamp profile.
Use the -p compiler option to enable gprof use.
VTune Monitors performance. This is an Intel performance monitoring tool.
You can run it directly on your SGI UV system. The Linux
server/Windows client is useful when you are working on a remote
system.
psrun Measures the performance of unmodified executables. This is a
PerfSuite command-line utility. psrun takes as input a configuration
XML document that describes the desired measurement. For more
information, see the following website:
http://perfsuite.ncsa.uiuc.edu/
For information about other performance analysis tools, see Chapter 2, "Performance
Analysis and Debugging" on page 7.
Using Compiler Options to Optimize Performance
This topic describes several Intel compiler options that can optimize performance. In
addition to the performance options and processor options that this topic describes,
the following options might be useful to you:
56 007–5646–010
![](asset-49.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
• The -help option displays a short summary of the ifort or icc options.
• The -dryrun option displays the driver tool commands that ifort or icc
generate. This option does not actually perform a compile.
For more information about the Intel compiler options, see the following:
https://software.intel.com/en-us/intel-parallel-studio-xe
Use the following options to help tune performance:
Option Purpose
-fno-alias
Assumes no pointer aliasing. Pointer aliasing can create uncertainty
about the possibility that two unrelated names might refer to the
identical memory. Because of this uncertainty, the compiler assumes
that any two pointers can point to the same location in memory. This
can remove optimization opportunities, particularly for loops.
Other aliasing options include -ansi_alias and -fno_fnalias.
Note that incorrect alias assertions might generate incorrect code.
-ip
Generates single file, interprocedural optimization. A related option,
-ipo generates multifile, interprocedural optimization.
Most compiler optimizations work within a single procedure, such as
a function or a subroutine, at a time. This intra-procedural focus
restricts optimization possibilities because a compiler is forced to
make worst-case assumptions about the possible effects of a
procedure. By using inter-procedural analysis, more than a single
procedure is analyzed at once and code is optimized. It performs two
passes through the code and requires more compile time.
-O3
Enables -O2 optimizations plus more aggressive optimizations,
including loop transformation and prefetching. Loop transformations
are found in a transformation file created by the compiler; you can
examine this file to see what suggested changes have been made to
loops. Prefetch instructions allow data to be moved into the cache
before their use. A prefetch instruction is similar to a load instruction.
007–5646–010 57
![](asset-4a.png)
5: Performance Tuning
Note that Level 3 optimization may not improve performance for all
programs.
-opt_report
Generates an optimization report and places it in the file specified by
the -opt_report_file option.
-prof_gen,-prof_use
Generates and uses profiling information. These options require a
three-step compilation process:
1. Compile with proper instrumentation using -prof_gen.
2. Run the program on one or more training datasets.
3. Compile with -prof_use, which uses the profile information
from the training run.
-S
Compiles and generates an assembly listing in the .s files and does
not link. The assembly listing can be used in conjunction with the
output generated by the -opt_report option to try to determine
how well the compiler is optimizing loops.
-vec-report
Controls information specific to the vectorizer. Intel Xeon series
processors support vectorization, which can provide a powerful
performance boost.
-Ofast
Takes the place of, and is equivalent to specifying, the following
options:
-ipo -O3 -no-prec-div -static -fp-model fast=2
-xHost
58 007–5646–010
![](asset-4b.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
-diag-enable
Enables the Source Checker, which provides advanced diagnostics
based on a detailed analysis of your source code. When enabled, the
compiler performs static global analysis to find errors in software that
the compiler does not typically detect. This general source code
analysis tool is an additional diagnostic to help you debug your
programs. You can use source code analysis options to detect the
following types of potential errors in your compiled code:
• Incorrect usage of OpenMP directives
• Inconsistent object declarations in different program units
• Boundary violations
• Uninitialized memory
• Memory corruptions
• Memory leaks
• Incorrect usage of pointers and allocatable arrays
• Dead code and redundant executions
• Typographical errors or uninitialized variables
• Dangerous usage of unchecked input
Source checker analysis performs a general overview check of a
program for all possible values simultaneously. This is in contrast to
run-time checking tools that run a program with a fixed set of values
for input variables; such checking tools cannot easily check all edge
effects. By not using a fixed set of input values, the source checker
can check for obscure cases. In fact, you do not need to run the
program for Source Checker because the analysis is performed at
compilation time. The only requirement is a successful compilation.
There are limitations to Source Checker analysis. Because the Source
Checker does not fully interpret the analyzed program, it can
generate so called false-positive messages. This is a fundamental
difference between compiler errors and Source Checker errors. In the
007–5646–010 59
![](asset-4c.png)
5: Performance Tuning
case of the source checker, you decide whether the generated error is
legitimate and needs to be fixed.
The Intel compilers support additional options that are specific to each processor
model. To determine the processor used in your system, examine the contents of the
/proc/cpuinfo file. For example:
#cat /proc/cpuinfo | grep "model name"
model name : Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz
model name : Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz
model name : Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz
.
.
.
Use the information in Table 5-1 on page 60 to determine the processor in your SGI
system.
Table 5-1 SGI Systems and Intel Processors
SGI System Intel Processor Model Intel Processor Code Name
SGI UV 300 Xeon E7-8800 v4 series
Xeon E7-8800 v2 series
Xeon E7-8800 v3 series
Broadwell EX
Ivy Bridge EX
Haswell EX
SGI UV 3000 Xeon E5-4600 v3 series Haswell EP 4S
SGI UV 2000 Xeon E5-4600 series
Xeon E5-4600 v2 series Sandy Bridge EP 4S
Ivy Bridge EP 4S
SGI UV 1000 Xeon 7500 series
Xeon E7-8800 series Nehalem EX
Westmere EX
SGI ICE or SGI Rackable Xeon E5-2600 series
Xeon E5-2600 v2 series
Xeon E5-2600 v3 series
Sandy Bridge EP
Ivy Bridge EP
Haswell EP
The following list shows the processor-specific compiler options:
60 007–5646–010
![](asset-4d.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Processor Option Purpose
-xAVX Generates instructions for the Sandy Bridge processors
and the Ivy Bridge processors, which support Intel
Advanced Vector Extensions (AVX) instructions.
-xCORE-AVX2 Generates instructions for Haswell processors, which
support Intel AVX2 instructions.
-xHost Generates instructions for the highest instruction set
available on the compilation host processor.
-xSSE4.2 Generates instructions for Nehalem processors and
Westmere processors, which support Intel SSE4.2
instructions.
Tuning the Cache Performance
The processor cache stores recently-used information in a place where it can be
accessed quickly. This topic uses the following terms to describe cache performance
tuning:
•Acache line is the minimum unit of transfer from next-higher cache into this one.
•Acache hit is reference to a cache line that is present in the cache.
•Acache miss is reference to a cache line that is not present in this cache level and
must be retrieved from a higher cache, from memory, or from swap space.
• The hit time is the time to access the upper level of the memory hierarchy, which
includes the time needed to determine whether the access is a hit or a miss.
•Amiss penalty is the time to replace a block in the upper level with the
corresponding block from the lower level, plus the time to deliver this block to the
processor. The time to access the next level in the hierarchy is the major
component of the miss penalty.
There are several actions you can take to help tune cache performance:
• Avoid large power-of-two (and multiples thereof) strides and dimensions that
cause cache thrashing. Cache thrashing occurs when multiple memory accesses
require use of the same cache line. This can lead to an unnecessary number of
cache misses.
007–5646–010 61
![](asset-4e.png)
5: Performance Tuning
To prevent cache thrashing, redimension your vectors so that the size is not a
power of two. Space the vectors out in memory so that concurrently accessed
elements map to different locations in the cache. When working with
two-dimensional arrays, make the leading dimension an odd number. For
multidimensional arrays, change two or more dimensions to an odd number.
For example, assume that a cache in the hierarchy has a size of 256 KB, which is
65536 four-byte words. A Fortran program contains the following loop:
real data(655360,24)
...
do i=1,23
do j=1,655360
diff=diff+data(j,i)-data(j,i+1)
enddo
enddo
The two accesses to data are separated in memory by 655360*4 bytes, which is a
simple multiple of the cache size. They consequently load to the same location in
the cache. Because both data items cannot simultaneously coexist in that cache
location, a pattern of replace on reload occurs that considerably reduces
performance.
• Use a memory stride of 1 wherever possible. A loop over an array should access
array elements from adjacent memory addresses. When the loop iterates through
memory by consecutive word addresses, it uses every word of every cache line in
sequence and does not return to a cache line after finishing it.
If memory strides other than 1 are used, cache lines could be loaded multiple
times if an array is too large to be held in memory at one time.
• Cache bank conflicts can occur if there are two accesses to the same 16-byte-wide
bank at the same time.
A maximum of four performance monitoring events can be counted
simultaneously.
• Group together data that is used at the same time and do not use vectors in your
code, if possible. If elements that are used in one loop iteration are contiguous in
memory, it can reduce traffic to the cache and fewer cache lines will be fetched for
each iteration of the loop.
• Try to avoid the use of temporary arrays and minimize data copies.
62 007–5646–010
![](asset-4f.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Managing Memory
The following topics pertain to memory management:
• "Memory Use Strategies" on page 63
• "Memory Hierarchy Latencies" on page 63
Memory Use Strategies
The following are some general memory use goals and guidelines:
• Register reuse. Do a lot of work on the same data before working on new data.
• Cache reuse. The program is much more efficient if all of the data and instructions
fit in cache. If the data and instructions do not fit in the cache, try to use what is
in cache before using anything that is not in cache.
• Data locality. Try to access data that is nearby in memory before attempting to
access data that is far away in memory.
• I/O efficiency. Perform a large amount of I/O operations all at once, rather than a
little bit at a time. Do not mix calculations and I/O.
Memory Hierarchy Latencies
Memory is not arranged as a flat, random access storage device. It is critical to
understand that memory is a hierarchy to get good performance. Memory latency
differs within the hierarchy. Performance is affected by where the data resides.
CPUs that are waiting for memory are not doing useful work. Software should be
hierarchy-aware to achieve best performance, so observe the following guidelines:
• Perform as many operations as possible on data in registers.
• Perform as many operations as possible on data in the cache(s).
• Keep data uses spatially and temporally local.
• Consider temporal locality and spatial locality.
Memory hierarchies take advantage of temporal locality by keeping more recently
accessed data items closer to the processor. Memory hierarchies take advantage of
007–5646–010 63
![](asset-50.png)
5: Performance Tuning
spatial locality by moving contiguous words in memory to upper levels of the
hierarchy.
Tuning Multiprocessor Codes
The following topics explain multiprocessor tuning, which consists of the following
major steps:
• Perform single processor tuning, which benefits multiprocessor codes also. For
information, see "Single Processor Code Tuning" on page 54.
• Determine the parts of your code that can be parallelized. For information, see
"Data Decomposition" on page 64.
• Choose the parallelization methodology for your code. For information, see
"Measuring Parallelization and Parallelizing Your Code" on page 66.
• Analyze your code to make sure it is parallelizing properly. For information, see
Chapter 2, "Performance Analysis and Debugging" on page 7.
• Determine if false sharing exists. False sharing refers to OpenMP, not MPI. For
information, see "Fixing False Sharing" on page 69.
• Tune for data placement. For information, see Chapter 4, "Data Process and
Placement Tools" on page 29.
• Use environment variables to assist with tuning. For information, see
"Environment Variables for Performance Tuning" on page 69.
Data Decomposition
In order to efficiently use multiple processors on a system, tasks have to be found
that can be performed at the same time. There are two basic methods of defining
these tasks:
• Functional parallelism
Functional parallelism is achieved when different processors perform different
functions. This is a known approach for programmers trained in modular
programming. Disadvantages to this approach include the difficulties of defining
functions as the number of processors grow and finding functions that use an
64 007–5646–010
![](asset-51.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
equivalent amount of CPU power. This approach may also require large amounts
of synchronization and data movement.
• Data parallelism
Data parallelism is achieved when different processors perform the same function
on different parts of the data. This approach takes advantage of the large
cumulative memory. One requirement of this approach, though, is that the
problem domain be decomposed. There are two steps in data parallelism:
1. Decompose the data.
Data decomposition is breaking up the data and mapping data to processors.
You, the programmer, can break up the data explicitly by using message
passing (with MPI) and data passing (using the SHMEM library routines).
Alternatively, you can employ compiler-based MP directives to find
parallelism in implicitly decomposed data.
There are advantages and disadvantages to implicit and explicit data
decomposition:
–Implicit decomposition advantages: No data resizing is needed. All
synchronization is handled by the compiler. The source code is easier to
develop and is portable to other systems with OpenMP or High
Performance Fortran (HPF) support.
–Implicit decomposition disadvantages: The data communication is hidden
by the user
–Explicit decomposition advantages: The programmer has full control over
insertion of communication and synchronization calls. The source code is
portable to other systems. Code performance can be better than implicitly
parallelized codes.
–Explicit decomposition disadvantages: Harder to program. The source
code is harder to read and the code is longer (typically 40% more).
2. Divide the work among processors.
007–5646–010 65
![](asset-52.png)
5: Performance Tuning
Measuring Parallelization and Parallelizing Your Code
When tuning for performance, first assess the amount of code that is parallelized in
your program. Use the following formula to calculate the amount of code that is
parallelized:
p=N(T(1)-T(N)) / T(1)(N-1)
In this equation, T(1) is the time the code runs on a single CPU and T(N) is the time it
runs on N CPUs. Speedup is defined as T(1)/T(N).
If speedup/N is less than 50% (that is, N>(2-p)/(1-p)), stop using more CPUs and tune
for better scalability.
You can use one of the following to display CPU activity:
• The top(1) command.
• The vmstat(8) command.
• The open source Performance Co-Pilot tools. For example, pmval(1) (pmval
kernel.percpu.cpu.user) or the visualization command pmchart(1).
Next, focus on using one of the following parallelization methodologies:
• "Using SGI MPI" on page 66
• "Using OpenMP" on page 67
• "Identifying OpenMP Nested Parallelism" on page 67
• "Using Compiler Options" on page 68
• "Identifying Opportunities for Loop Parallelism in Existing Code" on page 68
Using SGI MPI
The SGI Performance Suite includes the SGI Message Passing Toolkit (SGI MPT). SGI
MPT includes both the SGI Message Passing Interface (SGI MPI) and SGI SHMEM.
SGI MPI is optimized and more scalable for SGI UV series systems than the generic
MPI libraries. SGI MPI takes advantage of the SGI UV architecture and SGI
nonuniform memory access (NUMA) features.
Use the -lmpi compiler option to use MPI. For a list of environment variables that
are supported, see the mpi(1) man page.
66 007–5646–010
![](asset-53.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
The MPIO_DIRECT_READ and MPIO_DIRECT_WRITE environment variables are
supported under Linux for local XFS filesystems in SGI MPT version 1.6.1 and beyond.
MPI provides the MPI-2 standard MPI I/O functions that provide file read and write
capabilities. A number of environment variables are available to tune MPI I/O
performance. The mpi_io(3) man page describes these environment variables.
For information about performance tuning for MPI applications, see the following:
•SGI MPI and SGI SHMEM User Guide
•MPInside Reference Guide
Using OpenMP
OpenMP is a shared memory multiprocessing API, which standardizes existing
practice. It is scalable for fine or coarse grain parallelism with an emphasis on
performance. It exploits the strengths of shared memory and is directive-based. The
OpenMP implementation also contains library calls and environment variables.
OpenMP is included with the C, C++, and Fortran compilers.
To use OpenMP directives, specify the ifort -openmp or icc -openmp compiler
options. These options use the OpenMP front-end that is built into the Intel
compilers. The latest Intel compiler OpenMP run-time library name is libiomp5.so.
The latest Intel compiler also supports the GNU OpenMP library as an either/or
option, in other words, do not mix-and-match the GNU library with the Intel version.
For more information, see the OpenMP standard at the following website:
http://www.openmp.org/wp/openmp-specifications
Identifying OpenMP Nested Parallelism
The following Open MP nested parallelism output shows two primary threads and
four secondary threads, called master/nested:
%cat place_nested
firsttask cpu=0
thread name=a.out oncpu=0 cpu=4 noplace=1 exact onetime thread name=a.out oncpu=0
cpu=1-3 exact thread name=a.out oncpu=4 cpu=5-7 exact
%dplace -p place_nested a.out
Master thread 0 running on cpu 0
Master thread 1 running on cpu 4
007–5646–010 67
![](asset-54.png)
5: Performance Tuning
Nested thread 0 of master 0 gets task 0 on cpu 0 Nested thread 1 of master 0 gets task 1 on cpu 1
Nested thread 2 of master 0 gets task 2 on cpu 2 Nested thread 3 of master 0 gets task 3 on cpu 3
Nested thread 0 of master 1 gets task 0 on cpu 4 Nested thread 1 of master 1 gets task 1 on cpu 5
Nested thread 2 of master 1 gets task 2 on cpu 6 Nested thread 3 of master 1 gets task 3 on cpu 7
For more information, see the dplace(1) man page.
Using Compiler Options
You can use compiler options to invoke automatic parallelization. Use the
-parallel or -par_report options to the ifort or icc compiler commands.
These options show which loops were parallelized and the reasons why some loops
were not parallelized. If a source file contains many loops, it might be necessary to
add the -override_limits flag to enable automatic parallelization. The code
generated by the -parallel option is based on the OpenMP API. The standard
OpenMP environment variables and Intel extensions apply.
There are some limitations to automatic parallelization:
• For Fortran codes, only DO loops are analyzed
• For C/C++ codes, only for loops using explicit array notation or those using
pointer increment notation are analyzed. In addition, for loops using pointer
arithmetic notation are not analyzed, nor does it analyze while or do while
loops. The compiler also does not check for blocks of code that can be run in
parallel.
Identifying Opportunities for Loop Parallelism in Existing Code
Another parallelization optimization technique is to identify loops that have a
potential for parallelism, such as the following:
• Loops without data dependencies; a data dependency conflict occurs when a loop
has results from one loop pass that are needed in future passes of the same loop.
• Loops with data dependencies because of temporary variables, reductions, nested
loops, or function calls or subroutines.
Loops that do not have a potential for parallelism are those with premature exits, too
few iterations, or those where the programming effort to avoid data dependencies is
too great.
68 007–5646–010
![](asset-55.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Fixing False Sharing
If the parallel version of your program is slower than the serial version, false sharing
might be occurring. False sharing occurs when two or more data items that appear
not to be accessed by different threads in a shared memory application correspond to
the same cache line in the processor data caches. If two threads executing on different
CPUs modify the same cache line, the cache line cannot remain resident and correct
in both CPUs, and the hardware must move the cache line through the memory
subsystem to retain coherency. This causes performance degradation and reduction in
the scalability of the application. If the data items are only read, not written, the
cache line remains in a shared state on all of the CPUs concerned. False sharing can
occur when different threads modify adjacent elements in a shared array. When two
CPUs share the same cache line of an array and the cache is decomposed, the
boundaries of the chunks split at the cache line.
If you suspect false sharing, take one of the following actions:
• Use the information in the following manual to determine the appropriate
hardware performance counter names relevant to false sharing:
http://www.intel.com/content/www/us/en/processors/architectures-software-
developer-manuals.html
• On SGI UV systems, you can use the hubstats(1) command in the SGI
Foundation Software suite to verify whether false sharing is occurring.
If false sharing is a problem, try the following solutions:
• Use the hardware counter to run a profile that monitors storage to shared cache
lines. This shows the location of the problem.
• Revise data structures or algorithms.
• Check shared data, static variables, common blocks, private variables, and public
variables in shared objects.
• Use critical regions to identify the part of the code that has the problem.
Environment Variables for Performance Tuning
You can use several different environment variables to assist in performance tuning.
For details about environment variables used to control MPI behavior, see the mpi(1)
man page.
007–5646–010 69
![](asset-56.png)
5: Performance Tuning
Several OpenMP environment variables can affect the actions of the OpenMP library.
For example, some environment variables control the behavior of threads in the
application when they have no work to perform or are waiting for other threads to
arrive at a synchronization semantic. Other environment variables can specify how
the OpenMP library schedules iterations of a loop across threads. The following
environment variables are part of the OpenMP standard:
•OMP_NUM_THREADS. The default is the number of CPUs in the system.
•OMP_SCHEDULE. The default is static.
•OMP_DYNAMIC. The default is false.
•OMP_NESTED. The default is false.
In addition to the preceding environment variables, Intel provides several OpenMP
extensions, two of which are provided through the use of the KMP_LIBRARY variable.
The KMP_LIBRARY variable sets the run-time execution mode, as follows:
• If set to serial, single-processor execution is used.
• If set to throughput, CPUs yield to other processes when waiting for work. This
is the default and is intended to provide good overall system performance in a
multiuser environment.
• If set to turnaround, worker threads do not yield while waiting for work. Setting
KMP_LIBRARY to turnaround may improve the performance of benchmarks run
on dedicated systems, where multiple users are not contending for CPU resources.
If your program generates a segmentation fault immediately upon execution, you
might need to increase KMP_STACKSIZE. This is the private stack size for threads.
The default is 4 MB. You may also need to increase your shell stacksize limit.
Understanding Parallel Speedup and Amdahl’s Law
You can use multiple CPUs in the following ways:
• Take a conventional program in C, C++, or Fortran, and have the compiler find
the parallelism that is implicit in the code.
• Write your source code to use explicit parallelism. In in the source code, specify
the parts of the program that you want to execute asynchronously and how the
parts are to coordinate with each other.
70 007–5646–010
![](asset-57.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
When your program runs on more than one CPU, its total run time should be less.
But how much less? What are the limits on the speedup? That is, if you apply 16
CPUs to the program, should it finish in 1/16th the elapsed time?
This section covers the following topics:
• "Adding CPUs to Shorten Execution Time" on page 71
• "Understanding Parallel Speedup" on page 72
• "Understanding Superlinear Speedup" on page 72
• "Understanding Amdahl’s Law" on page 73
• "Calculating the Parallel Fraction of a Program" on page 74
• "Predicting Execution Time with n CPUs" on page 75
Adding CPUs to Shorten Execution Time
You can distribute the work your program does over multiple CPUs. However, there
is always some part of the program’s logic that has to be executed serially, by a single
CPU. This sets the lower limit on program run time.
Suppose there is one loop in which the program spends 50% of the execution time. If
you can divide the iterations of this loop so that half of them are done in one CPU
while the other half are done at the same time in a different CPU, the whole loop can
be finished in half the time. The result is a 25% reduction in program execution time.
The mathematical treatment of these ideas is called Amdahl’s law, for computer
pioneer Gene Amdahl, who formalized it. There are two basic limits to the speedup
you can achieve by parallel execution:
• The fraction of the program that can be run in parallel, p, is never 100%.
• Because of hardware constraints, after a certain point, there are diminishing
benefits from each added CPU.
Tuning for parallel execution comes down to doing the best that you are able to do
within these two limits. You strive to increase the parallel fraction, p, because in some
cases even a small change in p(from 0.8 to 0.85, for example) makes a dramatic
change in the effectiveness of added CPUs.
007–5646–010 71
![](asset-58.png)
5: Performance Tuning
Then you work to ensure that each added CPU does a full CPU’s work and does not
interfere with the work of other CPUs. In the SGI UV architectures this means coding
to accomplish the following:
• Spreading the workload equally among the CPUs
• Eliminating false sharing and other types of memory contention between CPUs
• Making sure that the data used by each CPU are located in a memory near that
CPU’s node
Understanding Parallel Speedup
If half the iterations of a loop are performed on one CPU, and the other half run at
the same time on a second CPU, the whole loop should complete in half the time. For
example, consider the typical C loop in Example 5-1.
Example 5-1 Typical C Loop
for (j=0; j<MAX; ++j) {
z[j] = a[j]*b[j];
}
The compiler can automatically distribute such a loop over nCPUs (with ndecided at
run time based on the available hardware), so that each CPU performs MAX/n
iterations.
The speedup gained from applying nCPUs, Speedup(n), is the ratio of the single-CPU
execution time to the n-CPU execution time: Speedup(n)=T(1) T(n). If you measure
the single-CPU execution time of a program at 100 seconds, and the program runs in
60 seconds with two CPUs, Speedup(2) = 100 60 = 1.67.
This number captures the improvement from adding hardware. T(n) ought to be less
than T(1). If it is not, adding CPUs has made the program slower, and something is
wrong. So Speedup(n) should be a number greater than 1.0, and the greater it is, the
better. Intuitively you might hope that the speedup would be equal to the number of
CPUs (twice as many CPUs, half the time) but this ideal can seldom be achieved.
Understanding Superlinear Speedup
You expect Speedup(n) to be less than n, reflecting the fact that not all parts of a
program benefit from parallel execution. However, it is possible, in rare situations, for
72 007–5646–010
![](asset-59.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Speedup(n) to be larger than n. When the program has been sped up by more than the
increase of CPUs it is known as superlinear speedup.
A superlinear speedup does not really result from parallel execution. It comes about
because each CPU is now working on a smaller set of memory. The problem data
handled by any one CPU fits better in cache, so each CPU runs faster than the single
CPU can. A superlinear speedup is welcome, but it indicates that the sequential
program was being held back by cache effects.
Understanding Amdahl’s Law
There are always parts of a program that you cannot make parallel, where code must
run serially. For example, consider the loop. Some amount of code is devoted to
setting up the loop and allocating the work between CPUs. This housekeeping must
be done serially. Then comes parallel execution of the loop body, with all CPUs
running concurrently. At the end of the loop comes more housekeeping that must be
done serially. For example, if ndoes not divide MAX evenly, one CPU must execute
the few iterations that are left over.
Concurrency cannot speed up the serial parts of the program. Let pbe the fraction of
the program’s code that can be made parallel (pis always a fraction less than 1.0.)
The remaining fraction (1–p) of the code must run serially. In practical cases, pranges
from 0.2 to 0.99.
The potential speedup for a program is proportional to pdivided by the CPUs you can
apply, plus the remaining serial part, 1-p. As an equation, this appears as Example 5-2.
Example 5-2 Amdahl’s law: Speedup(n) Given p
1
Speedup(n) = -----------
(p/n)+(1-p)
Suppose p= 0.8; then Speedup(2) = 1 / (0.4 + 0.2) = 1.67, and Speedup(4)= 1 / (0.2 +
0.2) = 2.5. The maximum possible speedup, if you could apply an infinite number of
CPUs, would be 1 / (1-p). The fraction phas a strong effect on the possible speedup.
The reward for parallelization is small unless pis substantial (at least 0.8). To put the
point another way, the reward for increasing pis great no matter how many CPUs
you have. The more CPUs you have, the more benefit you get from increasing p.
Using only four CPUs, you need only p= 0.75 to get half the ideal speedup. With
eight CPUs, you need p= 0.85 to get half the ideal speedup.
007–5646–010 73
![](asset-5a.png)
5: Performance Tuning
There is a slightly more sophisticated version of Amdahl’s law that includes
communication overhead. This version shows that if the program has no serial part
and you increase the number of cores, the following occurs:
• The amount of computations per core diminishes.
• The communication overhead increases (unless there is no communication and
there is trivial parallelization).
• The efficiency of the code and the speedup diminishes.
The equation is as follows:
Speedup(n)=n/(1+ a*(n-1) + n*(tc/ts))
The preceding equation uses the following variables:
•nis the number of processes.
•ais the fraction of the given task not dividable into concurrent subtasks.
•ts is the time to execute the task in a single processor.
•tc is the communication overhead.
If a=0 and tc=0, there is no serial part and no communication. In this case, as in a
trivial parallelization program, you see a linear speedup.
Calculating the Parallel Fraction of a Program
You do not have to guess at the value of pfor a given program. Measure the
execution times T(1) and T(2) to calculate a measured Speedup(2) = T(1) / T(2). The
Amdahl’s law equation can be rearranged to yield pwhen Speedup (2) is known, as in
Example 5-3.
Example 5-3 Amdahl’s law: pGiven Speedup(2)
2 SpeedUp(2) - 1
p = --- * --------------
1 SpeedUp(2)
Suppose you measure T(1) = 188 seconds and T(2) = 104 seconds.
SpeedUp(2) = 188/104 = 1.81
p = 2 * ((1.81-1)/1.81) = 2*(0.81/1.81) = 0.895
74 007–5646–010
![](asset-5b.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
In some cases, the Speedup(2) = T(1)/T(2) is a value greater than 2; in other words, a
superlinear speedup. When this occurs, the formula in Example 5-3 returns a value of
pgreater than 1.0, which is clearly not useful. In this case, you need to calculate p
from two other more realistic timings, for example T(2) and T(3). The general formula
for pis shown in Example 5-4, where nand mare the two CPU counts whose
speedups are known, n>m.
Example 5-4 Amdahl’s Law: pGiven Speedup(n) and Speedup(m)
Speedup(n) - Speedup(m)
p = -------------------------------------------
(1 - 1/n)*Speedup(n) - (1 - 1/m)*Speedup(m)
For more information about superlinear speedups, see the following:
"Understanding Superlinear Speedup" on page 72
Predicting Execution Time with n CPUs
You can use the calculated value of pto extrapolate the potential speedup with higher
numbers of CPUs. The following example shows the expected time with four CPUs,
if p=0.895 and T(1)=188 seconds:
Speedup(4)= 1/((0.895/4)+(1-0.895)) = 3.04
T(4)= T(1)/Speedup(4) = 188/3.04 = 61.8
The calculation can be made routine using the computer by creating a script that
automates the calculations and extrapolates run times.
These calculations are independent of most programming issues such as language,
library, or programming model. They are not independent of hardware issues because
Amdahl’s law assumes that all CPUs are equal. At some level of parallelism, adding
a CPU no longer affects run time in a linear way. For example, on some architectures,
cache-friendly codes scale closely with Amdahl’s law up to the maximum number of
CPUs, but scaling of memory intensive applications slows as the system bus
approaches saturation. When the bus bandwidth limit is reached, the actual speedup
is less than predicted.
007–5646–010 75
![](asset-5c.png)
5: Performance Tuning
Gustafson’s Law
Gustafson’s law proposes that programmers set the size of problems to use the
available equipment to solve problems within a practical fixed time. Therefore, if
faster, more parallel equipment is available, larger problems can be solved in the same
time. Amdahl’s law is based on fixed workload or fixed problem size. It implies that
the sequential part of a program does not change with respect to machine size (for
example, the number of processors). However, the parallel part is evenly distributed
by nprocessors. The effect of Gustafson’s law was to shift research goals to select or
reformulate problems so that solving a larger problem in the same amount of time
would be possible. In particular, the law redefines efficiency as a need to minimize the
sequential part of a program even if it increases the total amount of computation. The
effect is that by running larger problems, it is hoped that the bulk of the calculation
will increase faster than the serial part of the program, allowing for better scaling.
Floating-point Program Performance
Certain floating-point programs experience slowdowns due to excessive floating point
traps called Floating-Point Software Assists (FPSWAs).
These slowdowns occur when the hardware fails to complete a floating-point
operation and requests help (emulation) from software. This happens, for instance,
with denormalized numbers.
The symptoms are a slower than normal execution and an FPSWA message in the
system log. Use the dmesg(1) to display the message. The average cost of an FPSWA
fault is quite high, around 1000 cycles/fault.
By default, the kernel prints a message similar to the following in the system log:
foo(7716): floating-point assist fault at ip 40000000000200e1
isr 0000020000000008
The kernel throttles the message in order to avoid flooding the console.
It is possible to control the behavior of the kernel on FPSWA faults using the
prctl(1) command. In particular, it is possible to get a signal delivered at the first
FPSWA. It is also possible to silence the console message.
76 007–5646–010
![](asset-5d.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
About MPI Application Tuning
When you design your MPI application, make sure to include the following in your
design:
• The pinning of MPI processes to CPUs
• The isolating of multiple MPI jobs onto different sets of sockets and Hubs
You can achieve this design by configuring a batch scheduler to create a cpuset for
every MPI job. MPI pins its processes to the sequential list of logical processors
within the containing cpuset by default, but you can control and alter the pinning
pattern using MPI_DSM_CPULIST. For more information about these programming
practices, see the following:
• The MPI_DSM_CPULIST discussion in the SGI MPI and SGI SHMEM User Guide.
• The omplace(1) and dplace(1) man pages.
• The SGI Cpuset Software Guide.
MPI Application Communication on SGI Hardware
On an SGI UV system, the following two transfer methods facilitate MPI
communication between processes:
• Shared memory
• The global reference unit (GRU), which is part of the SGI UV Hub ASIC
The SGI UV series systems use a scalable nonuniform memory access (NUMA)
architecture to allow the use of thousands of processors and terabytes of RAM in a
single Linux operating system instance. As in other large, shared-memory
systems, memory is distributed to processor sockets and accesses to memory are
cache coherent. Unlike other systems, SGI UV systems use a network of Hub
ASICs connected over NUMALink to scale to more sockets than any other x86-64
system, with excellent performance out of the box for most applications.
When running on SGI UV systems with SGI’s Message Passing Toolkit (MPT),
applications can attain higher bandwidth and lower latency for MPI calls than
when running on more conventional distributed memory clusters. However,
knowing your SGI UV system’s NUMA topology and the performance constraints
that it imposes can still help you extract peak performance. For more information
007–5646–010 77
![](asset-5e.png)
5: Performance Tuning
about the SGI UV hub, SGI UV compute blades, Intel QPI, and SGI NUMALink,
see your SGI UV hardware system user guide.
The MPI library chooses the transfer method depending on internal heuristics, the
type of MPI communication that is involved, and some user-tunable variables. When
using the GRU to transfer data and messages, the MPI library uses the GRU resources
it allocates via the GRU resource allocator, which divides up the available GRU
resources. It allocates buffer space and control blocks between the logical processors
being used by the MPI job.
MPI Job Problems and Application Design
The MPI library chooses buffer sizes and communication algorithms in an attempt to
deliver the best performance automatically to a wide variety of MPI applications, but
user tuning might be needed to improve performance. The following are some
application performance problems and some ways that you might be able to improve
MPI performance:
• Primary HyperThreads are idle.
Most high-performance computing MPI programs run best when they use only
one HyperThread per core. When an SGI UV system has multiple HyperThreads
per core, logical CPUs are numbered such that primary HyperThreads are the high
half of the logical CPU numbers. Therefore, the task of scheduling only on the
additional HyperThreads may be accomplished by scheduling MPI jobs as if only
half the full number exists, leaving the high logical CPUs idle.
You can use the cpumap(1) command to determine if cores have multiple
HyperThreads on your SGI UV system. The command’s output includes the
following:
– The number of physical and logical processors
– Whether HyperThreading is enabled
– How shared processors are paired
If an MPI job uses only half of the available logical CPUs, set
GRU_RESOURCE_FACTOR to 2 so that the MPI processes can use all the available
GRU resources on a hub, rather than reserving some of them for the idle
HyperThreads. For more information about GRU resource tuning, see the
gru_resource(3) man page.
78 007–5646–010
![](asset-5f.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
• Message bandwidth is inadequate.
Use either huge pages or transparent huge pages (THP) to ensure that your
application obtains optimal message bandwidth.
To specify the use of hugepages, use the MPI_HUGEPAGE_HEAP_SPACE
environment variable. The MPI_HUGEPAGE_HEAP_SPACE environment variable
defines the minimum amount of heap space that each MPI process can allocate
using huge pages. For information about this environment variable, see the MPI(1)
man page.
To use THPs, see "Using Transparent Huge Pages (THPs) in MPI and SHMEM
Applications" on page 80.
Some programs transfer large messages via the MPI_Send function. To enable
unbuffered, single-copy transport in these cases, you can set MPI_BUFFER_MAX to
0. For information about the MPI_BUFFER_MAX environment variable, see the
MPI(1) man page.
• MPI small or near messages are very frequent.
For small fabric hop counts, shared memory message delivery is faster than GRU
messages. To deliver all messages within an SGI UV host via shared memory, set
MPI_SHARED_NEIGHBORHOOD to host. For more information, see the MPI(1) man
page.
• Memory allocations are nonlocal.
MPI application processes normally perform best if their local memory is allocated
near the socket assigned to use it. This cannot happen if memory on that socket is
exhausted by the application or by other system consumption, for example, file
buffer cache. Use the nodeinfo(1) command to view memory consumption on
the nodes assigned to your job, and use bcfree(1) to clear out excessive file
buffer cache. PBS Professional batch scheduler installations can be configured to
issue bcfree commands in the job prologue.
For information about PBS Professional, including the availablilty of scripts, see
the PBS Professional documentation and the bcfree(1) man page.
MPI Performance Tools
SGI supports several MPI performance tools. You can use the following tools to
enhance or troubleshoot MPI program performance:
007–5646–010 79
![](asset-60.png)
5: Performance Tuning
• MPInside. MPInside is an MPI profiling tool that can help you optimize your MPI
application. The tool provides information about data transferred between ranks,
both in terms of speed and quantity.
• SGI Perfboost. SGI PerfBoost uses a wrapper library to run applications compiled
against other MPI implementations under the SGI Message Passing Toolkit (MPT)
product on SGI platforms. The PerfBoost software allows you to run SGI MPT,
which is a version of MPI optimized for SGI’s large, shared-memory systems that
can take advantage of the SGI UV Hub.
• SGI PerfCatcher. SGI PerfCatcher uses a wrapper library to return MPI and
SHMEM function profiling information. The information returned includes
percent CPU time, total time spent per function, message sizes, and load
imbalances. For more information, see the following:
For more information about the MPI performance tools, see the following:
• The MPInside(1) man page.
• The perfcatch(1) man page
•MPInside Reference Guide
•SGI MPI and SGI SHMEM User Guide
Using Transparent Huge Pages (THPs) in MPI and SHMEM Applications
On SGI UV systems, THP is important because it contributes to attaining the best
GRU-based data transfer bandwidth in Message Passing Interface (MPI) and SHMEM
programs. On newer kernels, the THP feature is enabled by default. If THP is
disabled on your SGI UV system, see "Enabling Huge Pages in MPI and SHMEM
Applications on Systems Without THP" on page 82.
On SGI ICE systems, if you use a workload manager, such as PBS Professional, your
site configuration might let you enable or disable THP on a per-job basis.
The THP feature can affect the performance of some OpenMP threaded applications
in a negative way. For certain OpenMP applications, some threads in some shared
data structures might be forced to make more nonlocal references because the
application assumes a smaller, 4–KB page size.
The THP feature affects users in the following ways:
80 007–5646–010
![](asset-61.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
• Administrators:
To activate the THP feature on a system-wide basis, write the keyword always to
the following file:
/sys/kernel/mm/transparent_hugepage/enabled
To create an environment in which individual applications can use THP if memory
is allocated accordingly within the application itself, type the following:
#echo madvise > /sys/kernel/mm/transparent_hugepage/enabled
To disable THP, type the following command:
#echo never > /sys/kernel/mm/transparent_hugepage/enabled
If the khugepaged daemon is taking a lot of time when a job is running, then
defragmentation of THP might be causing performance problems. You can type
the following command to disable defragmentation:
#echo never > /sys/kernel/mm/transparent_hugepage/defrag
If you suspect that defragmentation of THP is causing performance problems, but
you do not want to disable defragmentation, you can tune the khugepaged
daemon by editing the values in
/sys/kernel/mm/transparent_hugepage/khugepaged.
• MPI application programmers:
To determine whether THP is enabled on your system, type the following
command and note the output:
%cat /sys/kernel/mm/transparent_hugepage/enabled
The output is as follows on a system for which THP is enabled:
[always] madvise never
In the output, the bracket characters ([])appear around the keyword that is in
effect.
007–5646–010 81
![](asset-62.png)
5: Performance Tuning
Enabling Huge Pages in MPI and SHMEM Applications on Systems Without
THP
If the THP capability is disabled on your SGI UV system, you can use the
MPI_HUGEPAGE_HEAP_SPACE environment variable and the
MPT_HUGEPAGE_CONFIG command to create huge pages.
The MPT_HUGEPAGE_CONFIG command configures the system to allow huge pages to
be available to MPT’s memory allocation interceptors. The
MPI_HUGEPAGE_HEAP_SPACE environment variable enables an application to use the
huge pages reserved by the MPT_HUGEPAGE_CONFIG command.
For more information, see the MPI_HUGEPAGE_HEAP_SPACE environment variable on
the MPI(1) man page, or see the mpt_hugepage_config(1) man page.
82 007–5646–010
![](asset-63.png)
Chapter 6
Flexible File I/O
This chapter covers the following topics:
• "About FFIO" on page 83
• "Environment Variables" on page 84
• "Simple Examples" on page 85
• "Multithreading Considerations" on page 88
• "Application Examples " on page 89
• "Event Tracing " on page 90
• "System Information and Issues " on page 90
About FFIO
Flexible File I/O (FFIO) can improve the file I/O performance of existing applications
without having to resort to source code changes. The current executable remains
unchanged. Knowledge of source code is not required, but some knowledge of how
the source and the application software work can help you better interpret and
optimize FFIO results. To take advantage of FFIO, all you need to do is to set some
environment variables before running your application.
The FFIO subsystem allows you to define one or more additional I/O buffer caches
for specific files to augment the Linux kernel I/O buffer cache. The FFIO subsystem
then manages this buffer cache for you. In order to accomplish this, FFIO intercepts
standard I/O calls such as open, read, and write, and replaces them with FFIO
equivalent routines. These routines route I/O requests through the FFIO subsystem,
which uses the user-defined FFIO buffer cache.
FFIO can bypass the Linux kernel I/O buffer cache by communicating with the disk
subsystem via direct I/O. This bypass gives you precise control over cache I/O
characteristics and allows for more efficient I/O requests. For example, doing direct
I/O in large chunks (for example, 16 megabytes) allows the FFIO cache to amortize
disk access. All file buffering occurs in user space when FFIO is used with direct I/O
enabled. This differs from the Linux buffer cache mechanism, which requires a
007–5646–010 83
![](asset-64.png)
6: Flexible File I/O
context switch in order to buffer data in kernel memory. Avoiding this kind of
overhead helps FFIO to scale efficiently.
Another important distinction is that FFIO allows you to create an I/O buffer cache
dedicated to a specific application. The Linux kernel, on the other hand, has to
manage all the jobs on the entire system with a single I/O buffer cache. As a result,
FFIO typically outperforms the Linux kernel buffer cache when it comes to I/O
intensive throughput.
Environment Variables
To use FFIO, set one of the following environment variables: LD_PRELOAD or
FF_IO_OPTS.
In order to enable FFIO to trap standard I/O calls, set the LD_PRELOAD environment
variable, as follows:
#export LD_PRELOAD="/usr/lib64/libFFIO.so"
The LD_PRELOAD software is a Linux feature that instructs the linker to preload the
indicated shared libraries. In this case, libFFIO.so is preloaded and provides the
routines that replace the standard I/O calls. An application that is not dynamically
linked with the glibc library cannot work with FFIO because the standard I/O calls
cannot be intercepted. To disable FFIO, type the following:
#unset LD_PRELOAD
The FFIO buffer cache is managed by the FF_IO_OPTS environment variable. The
syntax for setting this variable can be quite complex. A simple format for defining
this variable is as follows:
export FF_IO_OPTS ’string(eie.direct.mbytes:size:num:lead:share:stride:0)’
You can use the following parameters with the FF_IO_OPTS environment variable:
string Matches the names of files that can use the buffer cache.
size Number of 4k blocks in each page of the I/O buffer cache.
num Number of pages in the I/O buffer cache.
lead The maximum number of read-ahead pages.
share A value of 1 means a shared cache, 0 means private.
84 007–5646–010
![](asset-65.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
stride Note that the number after the stride parameter is always 0.
Example 1. Assume that you want a shared buffer cache of 128 pages. Each page is to
be 16 megabytes (that is, 4096*4k). The cache has a lead of six pages and uses a stride
of one. The command is as follows:
%setenv FF_IO_OPTS ’test*(eie.direct.mbytes:4096:128:6:1:1:0)’
Each time the application opens a file, the FFIO code checks the file name to see if it
matches the string supplied by FF_IO_OPTS. The file’s path name is not considered
when checking for a match against the string. For example, file names of
/tmp/test16 and /var/tmp/testit both match.
Example 2. This more complicated usage of FF_IO_OPTS builds upon the previous
example. Multiple types of file names can share the same cache, as the following
example shows:
%setenv FF_IO_OPTS ’output* test*(eie.direct.mbytes:4096:128:6:1:1:0)’
Example 3. You can specify multiple caches with FF_IO_OPTS. In the example that
follows, files of the form output* and test* share a 128 page cache of 16 megabyte
pages. The file special42 has a 256–page private cache of 32 megabyte pages. The
command, which uses the backslash (\) continuation character, is as follows:
%setenv FF_IO_OPTS ’output* test*(eie.direct.mbytes:4096:128:6:1:1:0) \
special42(eie.direct.mbytes:8192:256:6:0:1:0)’
Additional parameters can be added to FF_IO_OPTS to create feedback that is sent to
standard output. For examples of this diagnostic output, see the following:
"Simple Examples" on page 85
Simple Examples
This topic includes some simple FFIO examples. Assume that LD_PRELOAD is set for
the correct library, and FF_IO_OPTS is defined as follows:
%setenv FF_IO_OPTS ’test*(eie.direct.mbytes:4096:128:6:1:1:0)’
It can be difficult to tell what FFIO might or might not be doing even with a simple
program. The examples in this topic use a small C program called fio that reads
007–5646–010 85
![](asset-66.png)
6: Flexible File I/O
4–megabyte chunks from a file for 100 iterations. When the program runs, it produces
the following output:
%./fio -n 100 /build/testit
Reading 4194304 bytes 100 times to /build/testit
Total time = 7.383761
Throughput = 56.804439 MB/sec
Example 1. You can direct a simple FFIO operations summary to standard output by
making the following simple addition to FF_IO_OPTS:
%setenv FF_IO_OPTS ’test*(eie.direct.mbytes:4096:128:6:1:1:0, event.summary.mbytes.notrace )’
This new setting for FF_IO_OPTS generates the following summary on standard
output when the program runs:
%./fio -n 100 /build/testit
Reading 4194304 bytes 100 times to /build/testit
Total time = 7.383761
Throughput = 56.804439 MB/sec
event_close(testit) eie <-->syscall (496 mbytes)/( 8.72 s)= 56.85 mbytes/s
oflags=0x0000000000004042=RDWR+CREAT+DIRECT
sector size =4096(bytes)
cblks =0 cbits =0x0000000000000000
current file size =512 mbytes high water file size =512 mbytes
function times wall all mbytes mbytes min max avg
called time hidden requested delivered request request request
open 1 0.00
read 2 0.61 32 32 16 16 16
reada 29 0.01 0 464 464 16 16 16
fcntl
recall
reada 29 8.11
other 5 0.00
flush 1 0.00
close 1 0.00
Two synchronous reads of 16 megabytes each were issued, for a total of 32
megabytes. In addition, there were 29 asynchronous reads (reada) issued, for a total
of 464 megabytes.
86 007–5646–010
![](asset-67.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Example 2. You can generate additional diagnostic information by specifying the
.diag modifier. The following is an example of the diagnostic output generated
when the .diag modifier is used:
%setenv FF_IO_OPTS ’test*(eie.direct.diag.mbytes:4096:128:6:1:1:0 )’
%./fio -n 100 /build/testit
Reading 4194304 bytes 100 times to /build/testit
Total time = 7.383761
Throughput = 56.804439 MB/sec
eie_close EIE final stats for file /build/testit
eie_close Used shared eie cache 1
eie_close 128 mem pages of 4096 blocks (4096 sectors), max_lead = 6 pages
eie_close advance reads used/started : 23/29 79.31% (1.78 seconds wasted)
eie_close write hits/total : 0/0 0.00%
eie_close read hits/total : 98/100 98.00%
eie_close mbytes transferred parent --> eie --> child sync async
eie_close 0 0 0 0
eie_close 400 496 2 29 (0,0)
eie_close parent <-- eie <-- child
eie_close EIE stats for Shared cache 1
eie_close 128 mem pages of 4096 blocks
eie_close advance reads used/started : 23/29 79.31% (0.00 seconds wasted)
eie_close write hits/total : 0/0 0.00%
eie_close read hits/total : 98/100 98.00%
eie_close mbytes transferred parent --> eie --> child sync async
eie_close 0 0 0
eie_close 400 496 2 29 (0,0)
The preceding output lists information for both the file and the cache. In the mbytes
transferred information, the lines in bold are for write and read operations,
respectively. Only for very simple I/O patterns can the difference between (parent
--> eie) and (eie --> child) read statistics be explained by the number of read
aheads. For random reads of a large file over a long period of time, this is not the
case. All write operations count as async.
You can generate additional diagnostic information by specifying the .diag modifier
and the .event.summary modifier. The two modifiers operate independently from
one another. The following specification uses both modifiers:
%setenv FF_IO_OPTS ’test*(eie.diag.direct.mbytes:4096:128:6:1:1:0, event.summary.mbytes.notrace )’
007–5646–010 87
![](asset-68.png)
6: Flexible File I/O
Multithreading Considerations
FFIO works with applications that use MPI for parallel processing. An MPI job assigns
each thread a number or rank. The master thread has rank 0, while the remaining
slave threads have ranks from 1 to N-l where Nis the total number of threads in the
MPI job. It is important to consider that the threads comprising an MPI job do not
necessarily have access to each others’ address space. As a result, there is no way for
the different MPI threads to share the same FFIO cache. By default, each thread
defines a separate FFIO cache based on the parameters defined by FF_IO_OPTS.
Having each MPI thread define a separate FFIO cache, based on a single environment
variable (FF_IO_OPTS), can waste a lot of memory. Fortunately, FFIO provides a
mechanism that allows you to specify a different FFIO cache for each MPI thread via
the following environment variables:
setenv FF_IO_OPTS_RANK0 ’result*(eie.direct.mbytes:4096:512:6:1:1:0)’
setenv FF_IO_OPTS_RANK1 ’output*(eie.direct.mbytes:1024:128:6:1:1:0)’
setenv FF_IO_OPTS_RANK2 ’input*(eie.direct.mbytes:2048:64:6:1:1:0)’
.
.
.
setenv FF_IO_OPTS_RANKN-1 ... (N = number of threads).
Each rank environment variable is set using the exact same syntax as FF_IO_OPTS
and each defines a distinct cache for the corresponding MPI rank. If the cache is
designated as shared, all files within the same ranking thread can use the same cache.
FFIO works with SGI MPI, HP MPI, and LAM MPI. In order to work with MPI
applications, FFIO needs to determine the rank of callers by invoking the
mpi_comm_rank_() MPI library routine. Therefore, FFIO needs to determine the
location of the MPI library used by the application. To accomplished this, set one, and
only one, of the following environment variables:
•setenv SGI_MPI /usr/lib
•setenv LAM_MPI
•setenv HP_MPI
Note: LAM MPI and HP MPI are usually distributed via a third party application.
The precise paths to the LAM and the HP MPI libraries are application dependent.
See the application installation guide to find the correct path.
88 007–5646–010
![](asset-69.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
To use the rank functionality, both the MPI and FF_IO_OPTS_RANK0 environment
variables must be set. If either variable is not set, then the MPI threads all use
FF_IO_OPTS. If both the MPI and the FF_IO_OPTS_RANK0 variables are defined but,
for example, FF_IO_OPTS_RANK2 is undefined, all rank 2 files generate a no match
with FFIO. This means that none of the rank 2 files are cached by FFIO. In this case,
the software does not default to FF_IO_OPTS.
Fortran and C/C++ applications that use the pthreads interface create threads that
share the same address space. These threads can all make use of the single FFIO
cache defined by FF_IO_OPTS.
Application Examples
FFIO has been deployed successfully with several high-performance computing
applications, such as Nastran and Abaqus. In a recent customer benchmark, an
eight-way Abaqus throughput job ran approximately twice as fast when FFIO was
used. The FFIO cache used 16–megabyte pages (that is, page_size = 4096) and the
cache size was 8.0 gigabytes. As a rule of thumb, it was determined that setting the
FFIO cache size to roughly 10-15% of the disk space required by Abaqus yielded
reasonable I/O performance. For this benchmark, the FF_IO_OPTS environment
variable was defined as follows:
%setenv FF_IO_OPTS ’*.fct *.opr* *.ord *.fil *.mdl* *.stt* *.res *.sst *.hdx *.odb* *.023
*.nck* *.sct *.lop *.ngr *.elm *.ptn* *.stp* *.eig *.lnz* *.mass *.inp* *.scn* *.ddm
*.dat* fort*(eie.direct.nodiag.mbytes:4096:512:6:1:1:0,event.summary.mbytes.notrace)’
For the MPI version of Abaqus, different caches were specified for each MPI rank, as
follows:
%setenv FF_IO_OPTS_RANK0 ’*.fct *.opr* *.ord *.fil *.mdl* *.stt* *.res *.sst *.hdx *.odb* *.023
*.nck* *.sct *.lop *.ngr *.ptn* *.stp* *.elm *.eig *.lnz* *.mass *.inp *.scn* *.ddm
*.dat* fort*(eie.direct.nodiag.mbytes:4096:512:6:1:1:0,event.summary.mbytes.notrace)’
%setenv FF_IO_OPTS_RANK1 ’*.fct *.opr* *.ord *.fil *.mdl* *.stt* *.res *.sst *.hdx *.odb* *.023
*.nck* *.sct *.lop *.ngr *.ptn* *.stp* *.elm *.eig *.lnz* *.mass *.inp *.scn* *.ddm
*.dat* fort*(eie.direct.nodiag.mbytes:4096:16:6:1:1:0,event.summary.mbytes.notrace)’
%setenv FF_IO_OPTS_RANK2 ’*.fct *.opr* *.ord *.fil *.mdl* *.stt* *.res *.sst *.hdx *.odb* *.023
*.nck* *.sct *.lop *.ngr *.ptn* *.stp* *.elm *.eig *.lnz* *.mass *.inp *.scn* *.ddm
*.dat* fort*(eie.direct.nodiag.mbytes:4096:16:6:1:1:0,event.summary.mbytes.notrace)’
007–5646–010 89
![](asset-6a.png)
6: Flexible File I/O
%setenv FF_IO_OPTS_RANK3 ’*.fct *.opr* *.ord *.fil *.mdl* *.stt* *.res *.sst *.hdx *.odb* *.023
*.nck* *.sct *.lop *.ngr *.ptn* *.stp* *.elm *.eig *.lnz* *.mass *.inp *.scn* *.ddm
*.dat* fort*(eie.direct.nodiag.mbytes:4096:16:6:1:1:0,event.summary.mbytes.notrace)’
Event Tracing
If you specify the .trace option as part of the event parameter, you can enable the
event tracing feature in FFIO.
For example:
%setenv FF_IO_OPTS ’test*(eie.direct.mbytes:4096:128:6:1:1:0, event.summary.mbytes.trace )’
This option generates files of the form ffio.events.pid for each process that is part
of the application. By default, event files are placed in /tmp. To chang this
destination, set the FFIO_TMPDIR environment variable. These files contain
time-stamped events for files using the FFIO cache and can be used to trace I/O
activity such as I/O sizes and offsets.
System Information and Issues
The FFIO subsystem supports applications written in C, C++, and Fortran. C and
C++ applications can be built with either the Intel or gcc compiler. Only Fortran
codes built with the Intel compiler work with FFIO.
The following restrictions on FFIO must also be observed:
• The FFIO implementation of pread/pwrite is not correct. The file offset
advances.
• Do not use FFIO for I/O on a socket.
• Do not link your application with the librt asynchronous I/O library.
• FFIO does not intercept calls that operate on files in /proc,/etc, and /dev.
• FFIO does not intercept calls that operate on stdin,stdout, and stderr.
• FFIO is not intended for generic I/O applications such as vi,cp,ormv, and so on.
90 007–5646–010
![](asset-6b.png)
Chapter 7
I/O Tuning
This chapter contains the following topics:
• "About I/O Tuning" on page 91
• "Application Placement and I/O Resources" on page 91
• "Layout of Filesystems and XVM for Multiple RAIDs" on page 92
About I/O Tuning
This chapter describes tuning information that you can use to improve I/O
throughput and latency.
Application Placement and I/O Resources
It is useful to place an application on the same node as its I/O resource. For graphics
applications, for example, this can improve performance up to 30 percent.
For example, assume an SGI UV system with the following devices:
#gfxtopology
Serial number: UV-00000021
Partition number: 0
8 Blades
248 CPUs
283.70 Gb Memory Total
5 I/O Risers
Blade Location NASID PCI Address X Server Display Device
----------------------------------------------------------------------
0 r001i01b08 0 0000:05:00.0 - Matrox Pilot
4 r001i01b12 8 0001:02:01.0 - SGI Scalable Graphics Capture
6 r001i01b14 12 0003:07:00.0 Layout0.0 nVidia Quadro FX 5800
0003:08:00.0 Layout0.1 nVidia Quadro FX 5800
7 r001i01b15 14 0004:03:00.0 Layout0.2 nVidia Quadro FX 5800
007–5646–010 91
![](asset-6c.png)
7: I/O Tuning
To run an OpenGL graphics program, such as glxgears(1), on the third graphics
processing unit using numactl(8), type the following command:
%numactl -N 14 -m 14 /usr/bin/glxgears -display :0.2
This example assumes the X server was started with :0 == Layout0.
The -N parameter specifies to run the command on node 14. The -m parameter
specifies to allocate memory only from node 14.
You could also use the dplace(1) command to place the application.
For information about the dplace(1) command, see the following:
"dplace Command" on page 35
Layout of Filesystems and XVM for Multiple RAIDs
There can be latency spikes in response from a RAID, and such a spikes can in effect
slow down all of the RAIDs as one I/O request completion waits for all of the striped
pieces to complete.
The latency spikes’ impact on throughput can be to stall all the I/O requests or to
delay a few I/O requests while others continue. It depends on how the I/O request is
striped across the devices. If the volumes are constructed as stripes to span all
devices and the I/O requests are sized to be full stripes, the I/O requests stall
because every I/O request has to touch every device. If the I/O requests can be
completed by touching a subset of the devices, then those that do not touch a
high-latency device can continue at full speed, while the stalled I/O requests can
complete and catch up later.
In large storage configurations, it is possible to lay out the volumes to maximize the
opportunity for the I/O requests to proceed in parallel, masking most of the effect of
a few instances of high latency.
There are at least three classes of events that cause high latency I/O operations. These
are as follows:
1. Transient disk delays - one disk pauses
2. Slow disks
3. Transient RAID controller delays
92 007–5646–010
![](asset-6d.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
The first two events affect a single logical unit number (LUN). The third event affects
all the LUNs on a controller. The first and third events appear to happen at random.
The second event is repeatable.
007–5646–010 93
![](asset-6f.png)
Chapter 8
Suggested Shortcuts and Workarounds
This chapter includes the following topics:
• "Determining Process Placement" on page 95
• "Resetting System Limits" on page 100
• "Linux Shared Memory Accounting" on page 106
• "OFED Tuning Requirements for SHMEM" on page 107
• "Setting Java Enviroment Variables" on page 108
Determining Process Placement
This topic describes methods that you can use to determine where different processes
are running. This can help you understand your application structure and help you
decide if there are obvious placement issues. Note that all examples use the C shell.
The following procedure explains how to set up the computing environment.
Procedure 8-1 To create the computing environment
1. Set up an alias as in this example, changing guest to your username:
%alias pu "ps -edaf|grep guest"
%pu
The pu command alias shows current processes.
2. Create the .toprc preferences file in your login directory to set the appropriate
top(1) options.
If you prefer to use the top(1) defaults, delete the .toprc file.
% cat <<EOF>> $HOME/.toprc
YEAbcDgHIjklMnoTP|qrsuzV{FWX
2mlt
EOF
3. Inspect all processes, determine which CPU is in use, and create an alias file for
this procedure.
007–5646–010 95
![](asset-70.png)
8: Suggested Shortcuts and Workarounds
The CPU number appears in the first column of the top(1) output:
%top -b -n 1 | sort -n | more
%alias top1 "top -b -n 1 | sort -n "
Use the following variation to produce output with column headings:
%alias top1 "top -b -n 1 | head -4 | tail -1;top -b -n 1 | sort -n"
4. View your files, replacing guest with your username:
%top -b -n 1 | sort -n | grep guest
Use the following variation to produce output with column headings:
%top -b -n 1 | head -4 | tail -1;top -b -n 1 | sort -n grep guest
The following topics present examples:
• "Example Using pthreads" on page 96
• "Example Using OpenMP" on page 98
Example Using pthreads
The following example demonstrates simple usage with a program name of th.It
sets the number of desired OpenMP threads and runs the program. Notice the
process hierarchy as shown by the PID and the PPID columns. The command usage
is as follows, where nis the number of threads:
%th n
%th 4
%pu
UID PID PPID C STIME TTY TIME CMD
root 13784 13779 0 12:41 pts/3 00:00:00 login --
guest1
guest1 13785 13784 0 12:41 pts/3 00:00:00 -csh
guest1 15062 13785 0 15:23 pts/3 00:00:00 th 4 <-- Main thread
guest1 15063 15062 0 15:23 pts/3 00:00:00 th 4 <-- daemon thread
guest1 15064 15063 99 15:23 pts/3 00:00:10 th 4 <-- worker thread 1
guest1 15065 15063 99 15:23 pts/3 00:00:10 th 4 <-- worker thread 2
guest1 15066 15063 99 15:23 pts/3 00:00:10 th 4 <-- worker thread 3
96 007–5646–010
![](asset-71.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
guest1 15067 15063 99 15:23 pts/3 00:00:10 th 4 <-- worker thread 4
guest1 15068 13857 0 15:23 pts/5 00:00:00 ps -aef
guest1 15069 13857 0 15:23 pts/5 00:00:00 grep guest1
%top -b -n 1 | sort -n | grep guest1
LC %CPU PID USER PRI NI SIZE RSS SHARE STAT %MEM TIME COMMAND
3 0.0 15072 guest1 16 0 3488 1536 3328 S 0.0 0:00 grep
5 0.0 13785 guest1 15 0 5872 3664 4592 S 0.0 0:00 csh
5 0.0 15062 guest1 16 0 15824 2080 4384 S 0.0 0:00 th
5 0.0 15063 guest1 15 0 15824 2080 4384 S 0.0 0:00 th
5 99.8 15064 guest1 25 0 15824 2080 4384 R 0.0 0:14 th
7 0.0 13826 guest1 18 0 5824 3552 5632 S 0.0 0:00 csh
10 99.9 15066 guest1 25 0 15824 2080 4384 R 0.0 0:14 th
11 99.9 15067 guest1 25 0 15824 2080 4384 R 0.0 0:14 th
13 99.9 15065 guest1 25 0 15824 2080 4384 R 0.0 0:14 th
15 0.0 13857 guest1 15 0 5840 3584 5648 S 0.0 0:00 csh
15 0.0 15071 guest1 16 0 70048 1600 69840 S 0.0 0:00 ort
15 1.5 15070 guest1 15 0 5056 2832 4288 R 0.0 0:00top
Now skip the Main and daemon processes and place the rest:
%/usr/bin/dplace -s 2 -c 4-7 th 4
%pu
UID PID PPID C STIME TTY TIME CMD
root 13784 13779 0 12:41 pts/3 00:00:00 login --
guest1
guest1 13785 13784 0 12:41 pts/3 00:00:00 -csh
guest1 15083 13785 0 15:25 pts/3 00:00:00 th 4
guest1 15084 15083 0 15:25 pts/3 00:00:00 th 4
guest1 15085 15084 99 15:25 pts/3 00:00:19 th 4
guest1 15086 15084 99 15:25 pts/3 00:00:19 th 4
guest1 15087 15084 99 15:25 pts/3 00:00:19 th 4
guest1 15088 15084 99 15:25 pts/3 00:00:19 th 4
guest1 15091 13857 0 15:25 pts/5 00:00:00 ps -aef
guest1 15092 13857 0 15:25 pts/5 00:00:00 grep guest1
%top -b -n 1 | sort -n | grep guest1
LC %CPU PID USER PRI NI SIZE RSS SHARE STAT %MEM TIME COMMAND
007–5646–010 97
![](asset-72.png)
8: Suggested Shortcuts and Workarounds
4 99.9 15085 guest1 25 0 15856 2096 6496 R 0.0 0:24 th
5 99.8 15086 guest1 25 0 15856 2096 6496 R 0.0 0:24 th
6 99.9 15087 guest1 25 0 15856 2096 6496 R 0.0 0:24 th
7 99.9 15088 guest1 25 0 15856 2096 6496 R 0.0 0:24 th
8 0.0 15095 guest1 16 0 3488 1536 3328 S 0.0 0:00 grep
12 0.0 13785 guest1 15 0 5872 3664 4592 S 0.0 0:00 csh
12 0.0 15083 guest1 16 0 15856 2096 6496 S 0.0 0:00 th
12 0.0 15084 guest1 15 0 15856 2096 6496 S 0.0 0:00 th
15 0.0 15094 guest1 16 0 70048 1600 69840 S 0.0 0:00 sort
15 1.6 15093 guest1 15 0 5056 2832 4288 R 0.0 0:00 top
Example Using OpenMP
The following example demonstrates simple OpenMP usage with a program name of
md. Set the desired number of OpenMP threads and run the program as follows:
%alias pu "ps -edaf | grep guest1
%setenv OMP_NUM_THREADS 4
%md
Use the pu alias and the top(1) command to see the output, as follows:
%pu
UID PID PPID C STIME TTY TIME CMD
root 21550 21535 0 21:48 pts/0 00:00:00 login -- guest1
guest1 21551 21550 0 21:48 pts/0 00:00:00 -csh
guest1 22183 21551 77 22:39 pts/0 00:00:03 md <-- parent / main
guest1 22184 22183 0 22:39 pts/0 00:00:00 md <-- daemon
guest1 22185 22184 0 22:39 pts/0 00:00:00 md <-- daemon helper
guest1 22186 22184 99 22:39 pts/0 00:00:03 md <-- thread 1
guest1 22187 22184 94 22:39 pts/0 00:00:03 md <-- thread 2
guest1 22188 22184 85 22:39 pts/0 00:00:03 md <-- thread 3
guest1 22189 21956 0 22:39 pts/1 00:00:00 ps -aef
guest1 22190 21956 0 22:39 pts/1 00:00:00 grep guest1
%top -b -n 1 | sort -n | grep guest1
LC %CPU PID USER PRI NI SIZE RSS SHARE STAT %MEM TIME COMMAND
2 0.0 22192 guest1 16 0 70048 1600 69840 S 0.0 0:00 sort
2 0.0 22193 guest1 16 0 3488 1536 3328 S 0.0 0:00 grep
98 007–5646–010
![](asset-73.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
2 1.6 22191 guest1 15 0 5056 2832 4288 R 0.0 0:00 top
4 98.0 22186 guest1 26 0 26432 2704 4272 R 0.0 0:11 md
8 0.0 22185 guest1 15 0 26432 2704 4272 S 0.0 0:00 md
8 87.6 22188 guest1 25 0 26432 2704 4272 R 0.0 0:10 md
9 0.0 21551 guest1 15 0 5872 3648 4560 S 0.0 0:00 csh
9 0.0 22184 guest1 15 0 26432 2704 4272 S 0.0 0:00 md
9 99.9 22183 guest1 39 0 26432 2704 4272 R 0.0 0:11 md
14 98.7 22187 guest1 39 0 26432 2704 4272 R 0.0 0:11 md
From the notation on the right of the pu list, you can see the -x 6 pattern, which is
as follows:
1. Place 1, skip 2 of them, place 3 more [011000].
2. Reverse the bit order and create the dplace(1) -x mask:
[000110] --> [0x06 ] --> decimal 6
The dplace(1) command does not currently process hexadecimal notation for
this bit mask.
The following example confirms that a simple dplace placement works correctly:
%setenv OMP_NUM_THREADS 4
%/usr/bin/dplace -x 6 -c 4-7 md
%pu
UID PID PPID C STIME TTY TIME CMD
root 21550 21535 0 21:48 pts/0 00:00:00 login -- guest1
guest1 21551 21550 0 21:48 pts/0 00:00:00 -csh
guest1 22219 21551 93 22:45 pts/0 00:00:05 md
guest1 22225 21956 0 22:45 pts/1 00:00:00 ps -aef
guest1 22226 21956 0 22:45 pts/1 00:00:00 grep guest1
guest1 22220 22219 0 22:45 pts/0 00:00:00 md
guest1 22221 22220 0 22:45 pts/0 00:00:00 md
guest1 22222 22220 93 22:45 pts/0 00:00:05 md
guest1 22223 22220 93 22:45 pts/0 00:00:05 md
guest1 22224 22220 90 22:45 pts/0 00:00:05 md
%top -b -n 1 | sort -n | grep guest1
LC %CPU PID USER PRI NI SIZE RSS SHARE STAT %MEM TIME COMMAND
2 0.0 22228 guest1 16 0 70048 1600 69840 S 0.0 0:00 sort
2 0.0 22229 guest1 16 0 3488 1536 3328 S 0.0 0:00 grep
007–5646–010 99
![](asset-74.png)
8: Suggested Shortcuts and Workarounds
2 1.6 22227 guest1 15 0 5056 2832 4288 R 0.0 0:00 top
4 0.0 22220 guest1 15 0 28496 2736 21728 S 0.0 0:00 md
4 99.9 22219 guest1 39 0 28496 2736 21728 R 0.0 0:12 md
5 99.9 22222 guest1 25 0 28496 2736 21728 R 0.0 0:11 md
6 99.9 22223 guest1 39 0 28496 2736 21728 R 0.0 0:11 md
7 99.9 22224 guest1 39 0 28496 2736 21728 R 0.0 0:11 md
9 0.0 21551 guest1 15 0 5872 3648 4560 S 0.0 0:00 csh
15 0.0 22221 guest1 15 0 28496 2736 21728 S 0.0 0:00 md
Resetting System Limits
To regulate these limits on a per-user basis for applications that do not rely on
limit.h, you can modify the limits.conf file. The system limits that you can
modify include maximum file size, maximum number of open files, maximum stack
size, and so on. To view this file, type the following:
[user@machine user]# cat /etc/security/limits.conf
# /etc/security/limits.conf
#
#Each line describes a limit for a user in the form:
#
##
#Where:
# can be:
# - an user name
# - a group name, with @group syntax
# - the wildcard *, for default entry
#
# can have the two values:
# - "soft" for enforcing the soft limits
# - "hard" for enforcing hard limits
#
# can be one of the following:
# - core - limits the core file size (KB)
# - data - max data size (KB)
# - fsize - maximum filesize (KB)
# - memlock - max locked-in-memory address space (KB)
# - nofile - max number of open files
# - rss - max resident set size (KB)
# - stack - max stack size (KB)
100 007–5646–010
![](asset-75.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
# - cpu - max CPU time (MIN)
# - nproc - max number of processes
# - as - address space limit
# - maxlogins - max number of logins for this user
# - priority - the priority to run user process with
# - locks - max number of file locks the user can hold
#
##
#* soft core 0
#* hard rss 10000
#@student hard nproc 20
#@faculty soft nproc 20
#@faculty hard nproc 50
#ftp hard nproc 0
#@student - maxlogins 4
# End of file
For information about how to change these limits, see "Resetting the File Limit
Resource Default" on page 101.
Resetting the File Limit Resource Default
Several large user applications use the value set in the limit.h file as a hard limit on
file descriptors, and that value is noted at compile time. Therefore, some applications
might need to be recompiled in order to take advantage of the SGI system hardware.
To regulate these limits on a per-user basis for applications that do not rely on
limit.h, you can modify the limits.conf file. This allows the administrator to set
the allowed number of open files per user and per group. This also requires a
one-line change to the /etc/pam.d/login file.
The following procedure explains how to change the /etc/pam.d/login file.
Procedure 8-2 To change the file limist resource default
1. Add the following line to /etc/pam.d/login:
session required /lib/security/pam_limits.so
007–5646–010 101
![](asset-76.png)
8: Suggested Shortcuts and Workarounds
2. Add the following line to /etc/security/limits.conf, where username is the
user’s login and limit is the new value for the file limit resource:
[username] hard nofile [limit]
The following command shows the new limit:
ulimit -H -n
Because of the large number of file descriptors that some applications require, such as
MPI jobs, you might need to increase the system-wide limit on the number of open
files on your SGI system. The default value for the file limit resource is 1024. The
default of 1024 file descriptors allows for approximately 199 MPI processes per host.
You can increase the file descriptor value to 8196 to allow for more than 512 MPI
processes per host by adding adding the following lines to the
/etc/security/limits.conf file:
* soft nofile 8196
* hard nofile 8196
The ulimit -a command displays all limits, as follows:
sys:~ # ulimit -a
core file size (blocks, -c) 1
data seg size (kbytes, -d) unlimited
scheduling priority (-e) 0
file size (blocks, -f) unlimited
pending signals (-i) 511876
max locked memory (kbytes, -l) 64
max memory size (kbytes, -m) 55709764
open files (-n) 1024
pipe size (512 bytes, -p) 8
POSIX message queues (bytes, -q) 819200
real-time priority (-r) 0
stack size (kbytes, -s) 8192
cpu time (seconds, -t) unlimited
max user processes (-u) 511876
virtual memory (kbytes, -v) 68057680
file locks (-x) unlimited
102 007–5646–010
![](asset-77.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
Resetting the Default Stack Size
Some applications do not run well on an SGI system with a small stack size. To set a
higher stack limit, follow the instructions in "Resetting the File Limit Resource Default"
on page 101 and add the following lines to the /etc/security/limits.conf file:
* soft stack 300000
* hard stack unlimited
These lines set a soft stack size limit of 300000 KB and an unlimited hard stack size
for all users (and all processes).
Another method that does not require root privilege relies on the fact that many MPI
implementation use ssh,rsh, or some sort of login shell to start the MPI rank
processes. If you merely need to increase the soft limit, you can modify your shell’s
startup script. For example, if your login shell is bash, add a line similar to the
following to your .bashrc file:
%ulimit -s 300000
Note that SGI MPI allows you to set your stack size limit larger. To reset the limit,
use the ulimit or limit shell command before launching an MPI program with
mpirun(1) or mpiexec_mpt(1). MPT propagates the stack limit setting to all MPI
processes in the job.
For more information on default settings, see "Resetting the File Limit Resource
Default" on page 101.
Avoiding Segmentation Faults
The default stack size in the Linux operating system is 8MB (8192 kbytes). This value
often needs to be increased to avoid segmentation fault errors. If your application
fails to run immediately, check the stack size.
You can use the ulimit -a command to view the stack size, as follows:
uv44-sys:~ # ulimit -a
core file size (blocks, -c) unlimited
data seg size (kbytes, -d) unlimited
file size (blocks, -f) unlimited
007–5646–010 103
![](asset-78.png)
8: Suggested Shortcuts and Workarounds
pending signals (-i) 204800
max locked memory (kbytes, -l) unlimited
max memory size (kbytes, -m) unlimited
open files (-n) 16384
pipe size (512 bytes, -p) 8
POSIX message queues (bytes, -q) 819200
stack size (kbytes, -s) 8192
cpu time (seconds, -t) unlimited
max user processes (-u) 204800
virtual memory (kbytes, -v) unlimited
file locks (-x) unlimited
To change the value, use a command similar to the following:
uv44-sys:~ # ulimit -s 300000
There is a similar variable for OpenMP programs. If you get a segmentation fault
right away while running a program parallelized with OpenMP, increase the
KMP_STACKSIZE to a larger size. The default size in Intel Compilers is 4MB.
For example, to increase it to 64MB, use the following commands:
• In the C shell, set the environment variable as follows:
setenv KMP_STACKSIZE 64M
• In the Bash shell, set the environment variable as follows:
export KMP_STACKSIZE=64M
Resetting Virtual Memory Size
The virtual memory parameter vmemoryuse determines the amount of virtual
memory available to your application.
104 007–5646–010
![](asset-79.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
If you are using the Bash shell, use commands such as the following when setting this
limit:
ulimit -a
ulimit -v 7128960
ulimit -v unlimited
If you are using the C shell, use commands such as the following when setting this
limit:
limit
limit vmemoryuse 7128960
limit vmemoryuse unlimited
For example. The following MPI program fails with a memory-mapping error
because of a virtual memory parameter, vmemoryuse, value that is set too low:
%limit vmemoryuse 7128960
%mpirun -v -np 4 ./program
MPI: libxmpi.so ’SGI MPI 4.9 MPT 1.14 07/18/06 08:43:15’
MPI: libmpi.so ’SGI MPI 4.9 MPT 1.14 07/18/06 08:41:05’
MPI: MPI_MSGS_MAX = 524288
MPI: MPI_BUFS_PER_PROC= 32
mmap failed (memmap_base) for 504972 pages (8273461248
bytes) Killed n
The program now succeeds when virtual memory is unlimited:
%limit vmemoryuse unlimited
%mpirun -v -np 4 ./program
MPI: libxmpi.so ’SGI MPI 4.9 MPT 1.14 07/18/06 08:43:15’
MPI: libmpi.so ’SGI MPI 4.9 MPT 1.14 07/18/06 08:41:05’
MPI: MPI_MSGS_MAX = 524288
MPI: MPI_BUFS_PER_PROC= 32
HELLO WORLD from Processor 0
HELLO WORLD from Processor 2
007–5646–010 105
![](asset-7a.png)
8: Suggested Shortcuts and Workarounds
HELLO WORLD from Processor 1
HELLO WORLD from Processor 3
Linux Shared Memory Accounting
The Linux operating system does not calculate memory utilization in a manner that is
useful for certain applications in situations where regions are shared among multiple
processes. This can lead to over-reporting of memory and to processes being killed by
schedulers that erroneously detect memory quota violations.
The get_weighted_memory_size function weighs shared memory regions by the
number of processes using the regions. Thus, if 100 processes each share a total of
10GB of memory, the weighted memory calculation shows 100MB of memory shared
per process, rather than 10GB for each process.
Because this function applies mostly to applications with large shared-memory
requirements, it is located in the SGI NUMA tools package and made available in the
libmemacct library available from a package called memacct. The library function
makes a call to the numatools kernel module, which returns the weighted sum back
to the library, and then returns back to the application.
The usage statement for the memacct call is as follows:
cc ... -lmemacct
#include <sys/types.h>
extern int get_weighted_memory_size(pid_t pid);
The syntax of the memacct call is, as follows:
int *get_weighted_memory_size(pid_t pid);
The call returns the weighted memory (RSS) size for a pid, in bytes. This call weights
the size of the shared regions by the number of processes accessing the region.
Returns -1 when an error occurs and sets errno, as follows:
ESRCH Process pid was not found.
ENOSYS The function is not implemented. Check if numatools
kernel package is up-to-date.
Normally, the following errors should not occur:
ENOENT Cannot open /proc/numatools device file.
106 007–5646–010
![](asset-7b.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
EPERM No read permission on /proc/numatools device file.
ENOTTY Inappropriate ioctl operation on /proc/numatools
device file.
EFAULT Invalid arguments. The ioctl() operation performed
by the function failed with invalid arguments.
For more information, see the memacct(3) man page.
OFED Tuning Requirements for SHMEM
You can specify the maximum number of queue pairs (QPs) for SHMEM applications
when run on large clusters over an OFED fabric, such as InfiniBand. If the
log_num_qp parameter is set to a number that is too low, the system generates the
following message:
MPT Warning: IB failed to create a QP
SHMEM codes use the InfiniBand RC protocol for communication between all pairs
of processes in the parallel job, which requires a large number of QPs. The
log_num_qp parameter defines the log2of the number of QPs. The following
procedure explains how to specify the log_num_qp parameter.
Procedure 8-3 To specify the log_num_qp parameter
1. Log into one of the hosts upon which you installed the MPT software as the root
user.
2. Use a text editor to open file /etc/modprobe.d/libmlx4.conf.
3. Add a line similar to the following to file /etc/modprobe.d/libmlx4.conf:
options mlx4_core log_num_qp=21
By default, the maximum number of queue pairs is 218 (262144). This is true
across all platforms (RHEL 7.1, RHEL 6.6, SLES 12, and SLES 11SP3).
4. Save and close the file.
5. Repeat the preceding steps on other hosts.
007–5646–010 107
![](asset-7c.png)
8: Suggested Shortcuts and Workarounds
Setting Java Enviroment Variables
When Java software starts, it checks the environment in which it is running and
configures itself to fit, assuming that it owns the entire environment. The default for
some Java implementations (for example, IBM J9 1.4.2) is to start a garbage collection
(GC) thread for every CPU it sees. Other Java implementations use other algorithms
to decide the number of GC threads to start, but the number is generally 0.5 to 1
times the number of CPUs, which is appropriate on a 1– or 2–socket system.
However, this strategy does not scale well to systems with a larger core count. Java
command line options let you control the number of GC threads that the Java virtual
machine (JVM) will use. In many cases, a single GC thread is sufficient. In other
cases, a larger number might be appropriate and can be set with the applicable
environment variable or command line option. Properly tuning the number of GC
threads for an application is an exercise in performance optimization, but a reasonable
starting point is to use one GC thread per active worker thread.
For example:
• For Oracle Java:
-XX:ParallelGCThreads
• For IBM Java:
-Xgcthreads
An example command line option:
java -XX:+UseParallelGC -XX:ParallelGCThreads=1
As an administrator, you might choose to limit the number of GC threads to a
reasonable value with an environment variable set in the global profile, for example
the /etc/profile.local file, so casual Java users can avoid difficulties. The
environment variable settings are as follows:
• For Oracle Java:
JAVA_OPTIONS="-XX:ParallelGCThreads=1"
• For IBM Java:
IBM_JAVA_OPTIONS="-Xgcthreads1"
108 007–5646–010
![](asset-7d.png)
Index
A
Amdahl’s law, 70
execution time given n and p, 75
parallel fraction p, 73
speedup(n ) given p, 73
superlinear speedup, 72
application placement and I/O resources, 91
application tuning process, 7
automatic parallelization
limitations, 68
avoiding segmentation faults, 103
C
cache bank conflicts, 61
cache coherency, 30
Cache coherent non-uniform memory access
(ccNUMA) systems, 78
cache performance, 61
ccNUMA
See also "cache coherent non-uniform memory
access", 78
ccNUMA architecture, 30
cgroups, 33
commands
dlook, 45
dplace, 35
common compiler options, 2
compiler command line, 2
compiler libaries, 6
C/C++, 5
dynamic libraries, 5
overview, 4
compiler libraries
static libraries, 4
compiler options
tracing and porting, 54
compiler options for tuning, 56
compiling environment, 1
compiler overview, 2
debugger overview, 17
libraries, 4
modules, 3
Configuring MPT
OFED, 107
CPU-bound processes, 15
cpusets, 33
D
data decomposition, 64
data dependency, 68
data parallelism, 64
data placement practices, 31
data placement tools, 29
cpusets, 31
dplace, 31
overview, 29
taskset, 31
debugger overview, 17
debuggers
gdb, 17
idb, 17
TotalView, 17
denormalized arithmetic, 2
determining parallel code amount, 66
determining tuning needs
tools used, 56
distributed shared memory (DSM), 29
dlook command, 45
dplace command, 35
007–5646–010 109
![](asset-7e.png)
Index
E
Environment variables, 69
explicit data decomposition, 64
F
False sharing, 69
file limit resources
resetting, 101
Flexible File I/O (FFIO), 88
environment variables to set, 84
operation, 83
overview, 83
simple examples, 85
floating-point programs, 76
Floating-Point Software Assist, 76
FPSWA
See "Floating-Point Software Assist", 76
functional parallelism, 64
G
Global reference unit (GRU), 77
GNU debugger, 17
Gustafson’s law, 76
I
I/O tuning
application placement, 91
layout of filesystems, 92
I/O-bound processes, 15
implicit data decomposition, 64
iostat command, 26
J
Java environment variables
setting, 108
L
layout of filesystems, 92
limits
system, 100
Linux shared memory accounting, 106
M
memory
cache coherency, 30
ccNUMA architecture, 30
distributed shared memory (DSM), 29
non-uniform memory access (NUMA), 31
memory accounting, 106
memory management, 1, 63
memory page, 1
memory strides, 61
memory-bound processes, 15
Message Passing Toolkit
for parallelization, 66
modules, 3
command examples, 3
MPI on SGI UV systems
general considerations, 77
job performance types, 78
other ccNUMA performance issues, 78
MPI on UV systems, 77
MPI profiling, 79
MPInside profiling tool, 79
110 007–5646–010
![](asset-7f.png)
Linux®Application Tuning Guide for SGI®X86-64 Based Systems
N
non-uniform memory access (NUMA), 31
NUMA Tools
command
dlook, 45
dplace, 35
O
OFED configuration for MPT, 107
OpenMP, 67
environment variables, 69
P
parallel execution
Amdahl’s law, 70
parallel fraction p, 73
parallel speedup, 72
parallelization
automatic, 68
using MPI, 66
using OpenMP, 67
perf tool, 16
performance
VTune, 17
performance analysis, 7
performance gains
types of, 7
performance problems
sources, 15
PerfSuite script, 16
process placement
determining, 95
set-up, 95
using OpenMP, 98
using pthreads, 96
profiling
MPI, 79
perf, 16
PerfSuite, 16
ps command, 24
R
resetting default system stack size, 103
resetting file limit resources, 101
resetting system limit resources, 100
resetting virtual memory size, 104
resident set size, 1
S
sar command, 27
segmentation faults, 103
setting Java environment variables, 108
SGI PerfBoost, 80
SGI PerfCatcher, 80
SHMEM, 6
shortening execution time, 71
stack size
resetting, 103
suggested shortcuts and workarounds, 95
superlinear speedup, 72
swap space, 1
system
overview, 1
system configuration, 8
system limit resources
resetting, 100
system limits
address space limit, 100
core file siz, 100
CPU time, 100
data size, 100
file locks, 100
file size, 100
locked-in-memory address space, 100
007–5646–010 111
![](asset-80.png)
Index
number of logins, 100
number of open files, 100
number of processes, 100
priority of user process, 100
resetting, 100
resident set size, 100
stack size, 100
system monitoring tools, 23
system usage commands, 23
iostat, 26
ps, 24
sar, 27
vmstat, 25
w, 24
T
taskset command, 42
tools
perf, 16
PerfSuite, 16
VTune, 17
tuning
cache performance, 61
environment variables, 69
false sharing, 69
heap corruption, 55
managing memory, 63
multiprocessor code, 64
parallelization, 66
profiling
perf, 16
PerfSuite script, 16
VTune analyzer, 17
single processor code, 54
using compiler options, 56
using math functions, 55
verifying correct results, 54
U
uname command, 15
unflow arithmetic
effects of, 2
UV Hub, 77
V
virtual addressing, 1
virtual memory, 1
vmstat command, 25
VTune performance analyzer, 17
W
w command, 24
112 007–5646–010