l observed an starvation issue in intel-mpi/64/4.0.3/008 when using persistent "wild" recv request, i.e. setting up persistent recv requests with MPI_ANY_SOURCE. Examining several event traces with ITCA makes me assume that those recv requests are more likely matched with send operations from processes with lower ranks numbers. This would be suboptimal as it leads to starvation of higher ranked process in large MPI jobs.
Is this the right place to post such issues?
Yes, this is an appropriate place to discuss this. Can you provide the trace file? Or do you have a sample code showing this behavior?
Technical Consulting Engineer
Intel® Cluster Tools