vtune hardware event-based sampling for mpi app

[eisenlohr__john]

I am trying to use vtune to analyze an mpi program which is a mixture of python, c and fortran. I am able to do hotspots collection for this program just fine, but when I try to do advanced hotspots or snb-general-exploration it hangs.

I run with

mpirun -np 1 -l amplxe-cl -collect snb-general-exploration --result-dir appge -- python <...args...> : -np 7 python <...args...>

I'm running on a single node with 16 cores for now but would like to do this analysis for larger cases on multiple nodes. Compiler version is 14.0.2.

I am able to run a "hello world" mpi application written in c with the snb-general-exploration or advanced-hotspots collection using a command similar to what I'm trying to use with the python/c/fortran app, so it is not an issue with sampling drivers.

Any ideas how to get this to work? Thanks in advance.

[Peter_W_Intel]

> I'm running on a single node with 16 cores for now but would like to do this analysis for larger cases on multiple node

VTune(TM) Amplifier XE doesn't work for multiple nodes directly, we can profile MPI program on single node to simulate behaviors for multiple node. If your application works on multiple profile, and you want to collect performance on these nodes, you may install VTune on these nodes separately, and system profiling can be used - it means if you MPI program works across the node, VTune on these code will do the data collection. However there is no communication on VTune between threads.    

> I am able to run a "hello world" mpi application written in c with the snb-general-exploration or advanced-hotspots collection using a command similar to what I'm trying to use with the python/c/fortran app

Can you do "amplxe-cl -collect snb-general-exploration --result-dir appge -- python python-script <...args..." ? I assume that binaries built with C/C++ & Fortran will be loaded by python script and mpirun also was used in script?