Hotspot analysis of MPI application using VTune Amplifier XE 2011

[Martin_Taylor]

Hi,
I am trying to perform a HotSpot analysis on an MPI application, and I find that attempting anything other than lightweight hotspot analysis fails to give me any results.

I'm using mpiexec from within Cygwin to launch a main "Controller" executable that then spawns (using MPI_Comm_spawn) several child ("Worker") instances of itself plus a couple of other MPI processes. Trying to analyse a Worker instance results in the error: INJECTOR_ERR_FAILED_TO_CREATE_REMOTE_THREAD.
I have tried running VTune as administrator, and that makes no apparent difference.

I also read the article at http://software.intel.com/en-us/articles/why-cant-hotspots-analysis-trace-spawned-processes-of-mpi-job, and tried installing and using mpi.hydra with the -bootstrap fork option. Launching it this way I managed to get some analysis results, but only of the Controller process and of course all the interesting stuff is happening in the child Worker processes

Is there anything else I can try to get useful analysis results?

Martin

[Patrick_F_Intel1]

Hello Martin,
Just a few question first...
So you are running on windows, with the cygwin linux shell, starting an MPI process?
I assume you are using the windows VTune Amplifier version, is that right?
Pat

[Martin_Taylor]

Hi Pat
Yes, sorry - I forgot to mention I am running on Windows, and indeed I am running the Windows VTune Amplifier.
The mpi process is typically launched from a Cygwin shell with the mpiexec command and using a config file for its parameters.

Martin

[Martin_Taylor]

Hi

Do you have any ideas or feedback on this issue please?

We have some performance issues that I'm trying to investigate, and the lightweight hotspot analysis really doesn't give us enough information to diagnose them.

Thanks

Martin