gnu/packages/patches/openmpi-mtl-priorities.patch

   1 PSM2 is the only valid implementation for OmniPath (OPA).  UCX also
   2 supports OmniPath, but it does so via Verbs, thus getting much lower
   3 performance (typically 3 GiB/s instead of 10 GiB/s).
   4
   5 To work around that, give the mtl_psm2 component a higher priority
   6 than the pml_ucx component.
   7
   8 See <https://www.mail-archive.com/users@lists.open-mpi.org/msg33581.html>.
   9
  10 --- openmpi-4.0.2/ompi/mca/mtl/psm2/mtl_psm2_component.c        2019-11-15 17:06:15.142076840 +0100
  11 +++ openmpi-4.0.2/ompi/mca/mtl/psm2/mtl_psm2_component.c        2019-11-15 17:06:22.242009379 +0100
  12 @@ -126,7 +126,7 @@ ompi_mtl_psm2_component_register(void)
  13          setenv("PSM2_DEVICES", "self,shm", 0);
  14      }
  15
  16 -    param_priority = 40;
  17 +    param_priority = 55;
  18      (void) mca_base_component_var_register (&mca_mtl_psm2_component.super.mtl_version,
  19                                              "priority", "Priority of the PSM2 MTL component",
  20                                              MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
  21
  22 Likewise for PSM: we want InfiniPath to be taken care of by PSM, not by UCX,
  23 since UCX achieves worse performance.
  24
  25 --- openmpi-4.0.2/ompi/mca/mtl/psm/mtl_psm_component.c
  26 +++ openmpi-4.0.2/ompi/mca/mtl/psm/mtl_psm_component.c
  27 @@ -92,7 +92,7 @@ ompi_mtl_psm_component_register(void)
  28
  29
  30      /* set priority high enough to beat ob1's default */
  31 -    param_priority = 30;
  32 +    param_priority = 54;
  33      (void) mca_base_component_var_register (&mca_mtl_psm_component.super.mtl_version,
  34                                              "priority", "Priority of the PSM MTL component",
  35                                              MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,