Running a sampling experiment with pat_run

Author:: Hewlett Packard Enterprise Development LP.
Copyright:: Copyright 2024-2025 Hewlett Packard Enterprise Development LP.

Setting up RajaPerf (optional)

For reference, at the time of writing these steps set up raja-perf:

> mkdir build_mpi
> cd build_mpi
> cmake -DENABLE_MPI=On ..
   ... about 40 lines of cmake output ...
> make -j
   [  1%] Building CXX object blt/thirdparty_builtin/googletest-master-2020-01-07/googletest/CMakeFiles/gtest.dir/src/gtest-all.cc.o
   [  1%] Building CXX object blt/tests/smoke/CMakeFiles/blt_mpi_smoke.dir/blt_mpi_smoke.cpp.o
     ...
   [100%] Built target test-raja-perf-suite.exe

With a bunch of compiler warnings.

Collecting the samples

`pat_run` is super simple to run, just insert “pat_run” in your launch line, we add the “-r” option to generate a report:

> srun -p bardpeak -n 2 --exclusive pat_run -r ./bin/raja-perf.exe -pftol 0.05  -k Apps_HALOEXCHANGE_FUSED
WARNING: For optimal performance analysis load the 'perftools-preload' module before linking the executable file.
CrayPat/X:  Version 24.11.0 Revision 31a512b4d sles15.5_x86_64  10/02/24 19:11:06
... output from running the application ...
Processing step 1 of 12
Suggested trace options file:   /lus/cflus02/jvogt/raja/RAJAPerf/RAJAPerf-v2022.10.0/build_mpi/raja-perf.exe+1332189-23322133s/build-options.apa

Processing step 12 of 12
CrayPat/X:  Version 24.11.0 Revision 31a512b4d sles15.5_x86_64  10/02/24 19:11:06

Number of PEs (MPI ranks):    2

Numbers of PEs per Node:      2
... The rest of the default report ...

Troubleshooting

The warning about perftools-preload can usually be ignored. It’s not that it can’t improve the results in some cases, it’s more that the improvement is not enough to justify the rebuild - especially if your build isn’t set up to relink without recompiling.

More tips for dealing with programs that do not exectte

The default report

The printed default report contains a set of tables that give a good basic view of what is going on in you program. It’s contents are described on the gettting started page. The default report

Going further

This walkthrough created the default report. `pat_run` supports more specialized data collection via the `-m mode-name` argument, which accepts these mode names:

lite-events — Event profile
lite-gpu — GPU kernel and data movement along with event profile
lite-hbm — High bandwidth memory data (for X86-64 systems only)
lite-loops — Loop estimates (CCE compiler only)

The modes are described in detail in here: perftools-lite

`pat_run` saves the experiment result in the current directory. distinguish different runs, in this example `raja-perf.exe+349876-26340827s`. You can `pat_report` or Apprentice to look at these results. Apprentice 3

You can also run without the “-r” option to skip the printed report.

Of course, much more information is available in the man page: pat_run