APS data validity without perf_event_paranoid and VTune Profiler drivers

[psing51]

Hi ,
I had recently installed intel parallel studio xe 2018u4.
I have gathered some data for the purpose of application profiling on intel 8280 CPU using the intel APS tool (shipped with 2018u4), but i did not enable the system wide monitoring while using APS binary. For Example, I did not -

1) set the /proc/sys/kernel/perf_event_paranoid value to 0  and,
2) install the  VTune Profiler/sampling  drivers .
I skipped aforementioned steps as it was listed as optional, and i am not too worried about collection overhead.


I have gathered data using following commands - 
a) mpiexec.hydra -genvall -np $SLURM_NPROCS -ppn $SLURM_NTASKS_PER_NODE  aps -c mpi   ./wrf.exe
b) mpiexec.hydra -genvall -np $SLURM_NPROCS -ppn $SLURM_NTASKS_PER_NODE  aps   ./wrf.exe

The data gathered using aforementioned  commands is valid for the purpose of analysis? as i had noticed sometimes back that there was a significant difference in the performance metrics data with b) command   ( Linux Perf vs PMU analysis )

I hope that the mpi data gathered using "aps -c mpi" should be good (not require drivers)? - as per this document it seems the drivers help only in the data collection metrics such as HPC Performance Characterization, Memory Access, and Microarchitecture Exploration.

[Vladimir_R_Intel]

Hi,

yes, in case your perf_event_paranoid lets to gather PMU metrics at least in per process mode then it is valid for analysis.

And yes "aps -c mpi" will gather only MPI related statistics and no PMU collector will be launched in this mode at all.

[Vladimir_R_Intel]

also please switch to the latest and greatest version, currently this is VTune 2020 Gold.

>>as i had noticed sometimes back that there was a significant difference in the performance metrics data with b) command   ( Linux Perf vs PMU analysis )

Could you please write a little more details what exactly was different?