Correctly profiling MPI application on the MIC (separate results for each process)

[Alastair_M_]

Dear all,

Following the instructions here https://software.intel.com/en-us/forums/topic/358774 I have been able to profile my MPI application on the MIC, however the results are not collected exactly how I would expect.

I am not seeing a separate results directory for each MPI process, rather a single result set for the whole application.

I am using the following command line with Vtune 2013 Update 15:

/opt/intel/vtune_amplifier_xe_2013/bin64/amplxe-cl -collect knc-hotspots -r res -knob enable-user-tasks=true -app-working-dir /usr/bin -source-search-dir=/home/alastair/mpi_testing/src -search-dir=/home/alastair/mpi_testing/src -- ssh mic0 mpirun -n 128 -wdir /mic0fs/ /home/alastair/mpi_testing 

What am I actually seeing in the results?  Is this an aggregation of profiles for all MPI processes?

What is required to see the results of each process individually?

Best regards,

Alastair

[TimP · Accepted solution]

As Peter indicated in that reference you have aggregated results of all ranks and threads. Filtering by cores would select individual ranks. Assuming you pin each rank to a group of cores as is normal.

[Alastair_M_]

Tim Prince wrote:

As Peter indicated in that reference you have aggregated results of all ranks and threads. Filtering by cores would select individual ranks.

Hi Tim,

Thanks for your reply.  I didn't realise that I could filter from this to point to individual MPI ranks.  I was expecting to see the result for each rank in a separate directory, as indicated in the reference.

I see now that I can filter by thread inside Vtune.

Thanks for your help Tim.

Best regards,

Alastair