- Infos im HLRS Wiki sind nicht rechtsverbindlich und ohne Gewähr -

Difference between revisions of "VTune"

From HLRS Platforms
Jump to navigationJump to search
 
(4 intermediate revisions by the same user not shown)
Line 1: Line 1:
 
{{Infobox software
 
{{Infobox software
| description = Intel® '''VTune™ Profiler ''' is a performance analysis tool for serial and multithreaded applications. Use VTune Profiler:
+
| description = Intel® '''VTune™ Profiler ''' is an accurate performance analysis tool with low overhead for serial and multithreaded applications. Use VTune Profiler:
 
* to determine the most time-consuming (hot) functions in your application
 
* to determine the most time-consuming (hot) functions in your application
 
* to discover on very fine-grained level which section of code (loop, source code line or even data field) do not effectively utilize available processor time
 
* to discover on very fine-grained level which section of code (loop, source code line or even data field) do not effectively utilize available processor time
* to analyse communication behavior that affects threaded performance
+
* to analyse serial and multithreaded applications  
* for accurate analysis with low overhead
 
* for serial and multithreaded applications  
 
 
** MPI (MPT, OpenMPI)
 
** MPI (MPT, OpenMPI)
 
** OpenMP, Intel® oneAPI Threading Building Blocks, native threads
 
** OpenMP, Intel® oneAPI Threading Building Blocks, native threads
* Languages:
+
* for applications developed in:
 
** C/C++/C#, Fortran, Python and others
 
** C/C++/C#, Fortran, Python and others
  
Line 22: Line 20:
 
To perform the performance analyse of your application with VTune you don’t need special compiler wrapper or libraries. Just recompile and relink your code with extra –g option in order to include debug information. VTune works well for dynamically linked binaries. [https://software.intel.com/content/www/us/en/develop/documentation/vtune-help/top/set-up-analysis-target/linux-targets/analyzing-statically-linked-binaries-on-linux-targets.html Here] you can find some tips for statically linked binaries.
 
To perform the performance analyse of your application with VTune you don’t need special compiler wrapper or libraries. Just recompile and relink your code with extra –g option in order to include debug information. VTune works well for dynamically linked binaries. [https://software.intel.com/content/www/us/en/develop/documentation/vtune-help/top/set-up-analysis-target/linux-targets/analyzing-statically-linked-binaries-on-linux-targets.html Here] you can find some tips for statically linked binaries.
  
Example:
+
Example. Modules on Hawk:
 
<pre>
 
<pre>
 
module load vtune # set up VTune environment  
 
module load vtune # set up VTune environment  
 
module load gcc mpt  
 
module load gcc mpt  
 +
</pre>
 +
 +
Example. Modules on Vulcan:
 +
<pre>
 +
module load performance/vtune/2020.1 # set up VTune environment
 +
module load mpi/openmpi/4.1.1-gnu-11.1.0
 
</pre>
 
</pre>
  
Line 33: Line 37:
 
</pre>
 
</pre>
  
Run analysis:
+
'''Run analysis'''
  
 
VTune has both a GUI and command line tool: vtune-gui and vtune.  
 
VTune has both a GUI and command line tool: vtune-gui and vtune.  
Line 41: Line 45:
 
* ''memory-consumption'' - Analyze memory consumption by your Linux application, its distinct memory objects and their allocation stacks.
 
* ''memory-consumption'' - Analyze memory consumption by your Linux application, its distinct memory objects and their allocation stacks.
  
'''NOTE:''' ''The VTune project working directory and the results directory must be placed on lustre FS.''
+
'''IMPORTANT''' ''The VTune project working directory and the results directory must be placed on lustre FS.''
  
Example:
+
Example for your job script:
 
<pre>
 
<pre>
 +
..
 
module load vtune
 
module load vtune
 
module load gcc mpt
 
module load gcc mpt
Line 52: Line 57:
 
</pre>
 
</pre>
  
Report:
+
'''Report'''
  
You can also generate the report in text form using the VTune command line tool:
+
When VЕune completes the analysis, you can open the results in the vtune-gui tool. Alternatively you can also generate the report in text form using the VTune command line tool:
 
<pre>
 
<pre>
 
vtune -help report  
 
vtune -help report  
vtune -report ''summary'' -r ${WORKDIR}/results_dir
+
vtune -report summary -r ${WORKDIR}/results_dir
 
</pre>
 
</pre>
Or you can open the results of analysis in vtune-gui tool.
 
  
 
For some use cases you might need to limit the amount of raw data to be collected. Define this limit in MB through the data-limit option:
 
For some use cases you might need to limit the amount of raw data to be collected. Define this limit in MB through the data-limit option:
Line 66: Line 70:
 
</pre>
 
</pre>
  
Some more information about VTune you can find here.
+
Some more information about VTune you can find [https://kb.hlrs.de/platforms/upload/Vtune_perf_analysis.pdf here].
  
 
== See also ==
 
== See also ==

Latest revision as of 17:18, 4 June 2021

Intel® VTune™ Profiler is an accurate performance analysis tool with low overhead for serial and multithreaded applications. Use VTune Profiler:
  • to determine the most time-consuming (hot) functions in your application
  • to discover on very fine-grained level which section of code (loop, source code line or even data field) do not effectively utilize available processor time
  • to analyse serial and multithreaded applications
    • MPI (MPT, OpenMPI)
    • OpenMP, Intel® oneAPI Threading Building Blocks, native threads
  • for applications developed in:
    • C/C++/C#, Fortran, Python and others
Intel-logo.png
Developer: Intel
Platforms: Hawk, Vulcan
Category: Performance Analyzer
License: Commercial
Website: Intel® VTune™ Profiler homepage


Using Intel VTune

To perform the performance analyse of your application with VTune you don’t need special compiler wrapper or libraries. Just recompile and relink your code with extra –g option in order to include debug information. VTune works well for dynamically linked binaries. Here you can find some tips for statically linked binaries.

Example. Modules on Hawk:

module load vtune # set up VTune environment 
module load gcc mpt 

Example. Modules on Vulcan:

module load performance/vtune/2020.1 # set up VTune environment 
module load mpi/openmpi/4.1.1-gnu-11.1.0 

Compilation example:

mpicxx –O2 -g -Wl,-Bdynamic main.cpp 

Run analysis

VTune has both a GUI and command line tool: vtune-gui and vtune. The following types of analysis are available on Hawk:

  • hotspots - Analyze application flow and identify sections of code that take a long time to execute (hotspots).
  • threading - Discover how well your application is using parallelism to take advantage of all available CPUs. Identify and locate synchronization issues causing overhead or idle wait time resulting in lost performance.
  • memory-consumption - Analyze memory consumption by your Linux application, its distinct memory objects and their allocation stacks.

IMPORTANT The VTune project working directory and the results directory must be placed on lustre FS.

Example for your job script:

..
module load vtune
module load gcc mpt
WORKDIR=/your/project/dir/on/lustre
cd ${WORKDIR}
mpirun -np 128  vtune -collect hotspots -r ${WORKDIR}/results_dir -- ./a.out your_input.file

Report

When VЕune completes the analysis, you can open the results in the vtune-gui tool. Alternatively you can also generate the report in text form using the VTune command line tool:

vtune -help report 
vtune -report summary -r ${WORKDIR}/results_dir

For some use cases you might need to limit the amount of raw data to be collected. Define this limit in MB through the data-limit option:

mpirun -np 128 vtune -collect hotspots -data-limit=200 -- ./a.out

Some more information about VTune you can find here.

See also

External links