In addition to what Gilles said, I usually advise users in ambiguous situations to explicitly choose the transport. For example, you might want to explicitly choose using the UCX PML:
mpirun --mca pil ucx ... This way, you are 100% sure that Open MPI chose the UCX PML (if it can't choose the UCX PML for some reason, it will fail/abort -- it won't fall back to some other transport). As a sysadmin, you can also set this in the site-wide openmpi-mca-params.conf file, if you wish. Most users won't notice / care, but if someone does want to override that value, the command line has a higher precedence than the openmpi-mca-params.conf file. > On Oct 20, 2019, at 8:26 PM, Gilles Gouaillardet via users > <users@lists.open-mpi.org> wrote: > > Raymond, > > > In the case of UCX, you can > > mpirun --mca pml_base_verbose 10 ... > > If the pml/ucx component is used, then your app will run over UCX. > > If the pml/ob1 component is used, then you can > > mpirun --mca btl_base_verbose 10 ... > > btl/self should be used for communications to itself. > > if btl/uct is used for inter node communications, it means your job is > running over UCX. > > fwiw, it seems the default is to use both btl/vader and btl/uct for intra > node communications. > > > Cheers, > > > Gilles > > > > On 10/20/2019 6:28 AM, Raymond Muno via users wrote: >> Is there a way to determine, at run time, as to what choices OpenMPI made in >> terms of transports that are being utilized? We want to verify we are >> running UCX over Infiniband. >> >> I have two users, executing the identical code, with the same mpirun >> options, getting vastly different execution times on the same cluster. >> -- Jeff Squyres jsquy...@cisco.com
