PSM2 is the only valid implementation for OmniPath (OPA). UCX also supports OmniPath, but it does so via Verbs, thus getting much lower performance (typically 3 GiB/s instead of 10 GiB/s). To work around that, give the mtl_psm2 component a higher priority than the pml_ucx component. See <https://www.mail-archive.com/users@lists.open-mpi.org/msg33581.html>. --- openmpi-4.0.2/ompi/mca/mtl/psm2/mtl_psm2_component.c 2019-11-15 17:06:15.142076840 +0100 +++ openmpi-4.0.2/ompi/mca/mtl/psm2/mtl_psm2_component.c 2019-11-15 17:06:22.242009379 +0100 @@ -126,7 +126,7 @@ ompi_mtl_psm2_component_register(void) setenv("PSM2_DEVICES", "self,shm", 0); } - param_priority = 40; + param_priority = 55; (void) mca_base_component_var_register (&mca_mtl_psm2_component.super.mtl_version, "priority", "Priority of the PSM2 MTL component", MCA_BASE_VAR_TYPE_INT, NULL, 0, 0, Likewise for PSM: we want InfiniPath to be taken care of by PSM, not by UCX, since UCX achieves worse performance. --- openmpi-4.0.2/ompi/mca/mtl/psm/mtl_psm_component.c +++ openmpi-4.0.2/ompi/mca/mtl/psm/mtl_psm_component.c @@ -92,7 +92,7 @@ ompi_mtl_psm_component_register(void) /* set priority high enough to beat ob1's default */ - param_priority = 30; + param_priority = 54; (void) mca_base_component_var_register (&mca_mtl_psm_component.super.mtl_version, "priority", "Priority of the PSM MTL component", MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,