OpenMPI-related error on Rocky Linux

[popescu-v]

Description

When verifying the Khiops installation on a Rocky 8 or Rocky 9 Docker container via khiops -s or khiops -v , we obtain:

Open MPI failed an OFI Libfabric library call (fi_endpoint).  This is highly
unusual; your job may behave unpredictably (and/or abort) after this.

  Local host: ...
  Location: mtl_ofi_component.c:513
  Error: Invalid argument (22)
--------------------------------------------------------------------------
Khiops 10.2.2-b.3

See:

• https://github.com/KhiopsML/khiops/actions/runs/9468751070/job/26085981246 ; or
• https://github.com/KhiopsML/khiops-python/actions/runs/9566989588/job/26374196376#step:3:10

However, the CIs are green because khiops -s and khiops -v respectively return 0 nonetheless and only display this error as a warning.

Questions/Ideas

• This seems related to: https://github.com/ofiwg/libfabric/issues/6710
  • A potential (at least temporary) fix might be to set PSM3_MULTI_EP=1 (see https://github.com/ofiwg/libfabric/issues/6710#issuecomment-831119595)
• We need to check whether:
  • the aforementioned fix makes the error go away
  • this error makes OpenMPI unusable or yields performance loss
  • Apparently the libfabric issue has been fixed in https://github.com/ofiwg/libfabric/pull/6737 (merged as part of https://github.com/ofiwg/libfabric/pull/6877), but chances are that the libfabric versions that are available on Rocky 8 and 9 are prior to this merge.

Context

• Khiops version: 10.2.2-b.3
• OS description: Rocky 8 and Rocky 9 on the GitHub runners

[popescu-v]

Adding the aforementioned fix silences the reported error, but triggers another error:

ORTE_ERROR_LOG: Out of resource in file util/show_help.c at line 507

See https://github.com/KhiopsML/khiops-python/actions/runs/9568000054/job/26377335252#step:3:18

See also (perhaps relevant) issue: https://github.com/open-mpi/ompi/issues/5798