Package: libopenmpi-dev Followup-For: Bug #1100120 Control: retitle -1 libopenmpi-dev: mpi4py spawn tests fail with MPI_ERR_UNKNOWN Control: severity -1 normal
The OPAL ERROR or libucs segfault at the end of the tests can be avoided by setting OMPI_MCA_btl_tcp_if_include=lo (that's with OMPI_ not PRTE_ prefix, see Bug#1087784 for hypre) Is it expected that we should need to be setting OMPI_MCA_btl_tcp_if_include=lo in debian/tests (and debian/rules)? But that still leaves the spawn test failures with MPI_ERR_UNKNOWN (and the PMIx errors). Updating the bug title so. Looking into the history of spawn tests, spawn is known to give problems. Some problems with mpich were dealt with previously, https://github.com/pmodels/mpich/issues/7073 https://github.com/mpi4py/mpi4py/issues/541 Indeed mpi4py has assumed that openmpi spawn tests fail, and has an explicit skip checking the openmpi version in test_spawn.py This skip condition was recently changed. Previously spawn tests were skipped for all openmpi 5.0.x versions with @unittest.skipMPI('openmpi(>=5.0.0,<5.1.0)', skip_spawn()) Recently that was relaxed (https://github.com/mpi4py/mpi4py/pull/601) to @unittest.skipMPI('openmpi(>=5.0.0,<5.0.7)', skip_spawn()) And debian's openmpi has just updated to 5.0.7. So mpi4py is now running the spawn tests where previously it was skipping them, resulting in the error reported here. I'm not sure if mpi4py upstream meant the condition to say "<=5.0.7" rather than "<5.0.7", or was wrong about spawn now working properly in openmpi 5.0.7. But we can see that it is not working, so mpi4py will want to continue skipping spawn tests for now. In summary, this is a known issue in openmpi, so I'm downgrading the severity to normal.
