Hi Gilles, Gilles Gouaillardet via users <users@lists.open-mpi.org> skribis:
> in order to figure out which interconnect is used, you can > > mpirun --mca pml_base_verbose 10 --mca mtl_base_verbose 10 --mca > btl_base_verbose 10 ... > > the output might be a bit verbose, so here are a few tips on how to > get it step by step Thanks for the trick. I did find out abuot “pml_base_verbose” & co., but I was hoping for an option that wouldn’t be layer-specific. I guess what you propose is a good approximation. > PSM and PSM2 are generally faster with mtl/psm and mtl/psm2, and they > are only used by pml/cm > if pml/ucx is a match, you can either blacklist it or give it a lower > priority so pml/cm is picked > mpirun --mca pml ^ucx ... > mpirun --mca pml_ucx_priority 1 ... > or force pml/cm > mpirun --mca pml cm ... > note there might be an option to tell UCX it should not try to do > PSM/PSM2, and in that case, pml/ucx would not be > selected on an Infinipath/Omnipath network. I didn’t find a way to tell UCX to ignore TrueScale and Omni-Path, which is why I ended up increasing the default priority of PSM and PSM2 instead. Thanks for your feedback, Ludo’.
