iMPI 5.2.1 Omni-Path tuning

[Ronald_G_2]

For older Intel MPI versions, we were told to set this:

export I_MPI_HYDRA_PMI_CONNECT=alltoall.   Is this needed with Intel MPI 5.2.1?  And is 5.2.1 the latest version? 

we also set I_MPI_FABRICS="shm:tmi" or mpirun with -PSM2.  Is this the correct setting for Omni-Path fabric?

Do you have any other env vars or options we should explore to get optimal perf of Intel MPI on Omni-Path (Broadwell host, RHEL).  Willing to experiment a bit for best performance.

Thanks!

Ron

[James_T_Intel]

Compare with using I_MPI_FABRICS=ofi or I_MPI_FABRICS=tmi (not shm:ofi or shm:tmi).  When you run with I_MPI_FABRICS=ofi, ensure that you are using the PSM2 provider (I_MPI_OFI_PROVIDER) if you have this provider installed (use I_MPI_OFI_PROVIDER_DUMP=yes to check).

Since you're asking at a very general level, I highly recommend looking through the Developer Reference for more environment variables you can use for performance experiments.

[James_T_Intel]

Also, regarding the version, there is no 5.2.1 version.  There is 5.1, which went up to Update 3, the latest version is 2017 Update 1.