NLRWindSystems / WEIS / 41911922590

64%

main: 58%

Pull #479

github

Keep CI MPI runs from deadlocking on runner hardware

Some hosted ubuntu runners have an Azure MANA NIC. Open MPI's UCX PML
fails to open it ("uct_iface_open(ud_verbs/mana_0:1) failed"), MPI
communication breaks, PETSc's first broadcast fails on rank 0 during
import, and the other rank blocks forever. Run 41862056867 sat in "Run
parallel cases" until the 150 min job timeout.

- Use Open MPI's ob1 PML with self/sm/tcp BTLs and skip UCX one-sided.
  Hosted runners are single-node, so UCX adds nothing.
- Run MPI examples with python -m mpi4py so an exception on one rank
  aborts the job instead of deadlocking.
- Cap the parallel-cases step at 30 min (normally ~3 min).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

982 of 1406 new or added lines in 29 files covered. (69.84%)

8957 of 14000 relevant lines covered (63.98%)

0.64 hits per line