On Thu, Jan 18, 2018 at 05:21:48PM -0500, Aaron M. Ucko wrote: > Mark Abraham <mark.j.abra...@gmail.com> writes: > > > GROMACS dev here - I expect that will let that test pass. It just needs > > enough processes that can resolve MPI calls without deadlocking. > > Confirmed, setting OMPI_MCA_rmaps_base_oversubscribe=1 fixes the "not > enough slots" error in MdrunUtilityMpiUnitTests. The GpuUtilsUnitTests > regression remains, though.
It's possible that the latter is a gtest issue instead; that specific failing test was previously the source of a FTBFS on Hurd, though for a different reason. I can't easily account for why it would only fail with OpenMPI but not MPICH or the basic non-MPI build, though. However, I think it's a low-importance test in this case - the build doesn't use GPU support - so I may throw in another conditional to disable the test on x32, with a note to revisit the issue the next time the gtest code is updated. The thread safety warning is on all architectures, but if you still have a build tree hanging around, could you please check that set of tests with this command to be sure it's a non-issue? build/openmpi/bin/gpu_utils-test --gtest_death_test_style=threadsafe Probably no difference, but it's fast. Thanks much for the testing help. - Nicholas
