search

conda-forge / openmpi-feedstock

A conda-smithy repository for openmpi.
BSD 3-Clause "New" or "Revised" License
10 stars 25 forks source link

Issue with UCC Initialization: "mca_coll_ucc_save_coll_handlers" #183

Open j34ni opened 1 month ago

j34ni commented 1 month ago

Description:

Users may encounter this kind of error related to UCC initialization in OpenMPI:

Error: coll_ucc_module.c:474 - mca_coll_ucc_module_enable() mca_coll_ucc_save_coll_handlers failed

This was documented in the OpenMPI repository (see: https://github.com/open-mpi/ompi/issues/9885).

Apparently the problem arises because the default priority for UCC is 10, which conflicts with the priorities of basic and libnbc.

Workaround: Users can bypass this issue by setting a higher priority for UCC:

export OMPI_MCA_coll_ucc_priority=100

Recommendation: Should we not inform users about this issue (and workaround), and possibly look into adjusting the default configurations to avoid such conflict?

dalcinl commented 1 month ago

If there is an upstream patch accepted, once that patch is merged, then we can add that patch here and produce a new build.