|
Yes. You can get Call Graph and Sampling profiles of MPI applications using the VTune™ analyzer. For a distributed MPI application it's possible to collect sampling data only on a particular node where some part of application is executed. Just run sampling session on that node without "application to launch." However VTune was not designed to do detailed CG analyses of distributed MPI applications.
Chameleon MPI (MPICH)
You can perform CG collection only from the node where your application is started (use mpirun as a launcher, -np N mpi_app.exe as a parameter and mpi_app.exe as a module of interest). Since instrumented version of profiled application is executed on the node expect some slowdown. In some cases it might lead to discrepancies in you synchronization mechanisms.
Intel® MPI
For the Call Graph collection there is some specific. Intel® MPI launches applications with help of a daemon. VTune* during data collection tracks the hierarchy of executed processes in a special way. It cannot collect Call Graph data if Intel MPI daemon (mpdboot) is started before you run your activity. The right way is to use a special wrapper script that runs Intel MPI daemon and then your MPI application all together. Use that script as an application to launch during activity configuration. The described way gives you an opportunity to collect Call Graph data on every cluster node other than the node from where you start your application with mpiexec. Just run Intel MPI daemon under VTune setting your distributed application as a module of interest on the node where you want to collect data. Ask if you need more details.
Please review other Intel® products to speedup and analyze your distributed and shared memory applications.
http://www.intel.com/cd/software/products/asmo-na/eng/244171.htm.
This applies to:
| 
