https://bugzilla.suse.com/show_bug.cgi?id=1205139 https://bugzilla.suse.com/show_bug.cgi?id=1205139#c8 Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|NEW |CONFIRMED Assignee|screening-team-bugs@suse.de |nmoreychaisemartin@suse.com Flags|needinfo?(nmoreychaisemarti | |n@suse.com) | --- Comment #8 from Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com> --- There are at least 2 bugs. The 'cm' pml deadlocks when starting without any network interface available. I'm tracking the culprit down. It's calling the ofi mtl which fails to nit (as expected as there are no netdev apart from lo) but when trying to clean up, we end up in a deadlock for some reason. Not sure if it's openmpi of libfabric's fault This one explains different behaviour when building in chroot (which usually has netdevs) vs kvm (which doesn't) For python-mpi4py, there is a later deadlock which I haven't had time to look into yet. Might be a separate issue from the other packages(In reply to Atri Bhattacharya from comment #7)
(In reply to Daniel Garcia from comment #5) Will this be done for the openmpi packages so that mpiexec sets this directly, or should we do this for each package in its specific specfile?
No. The cm pml is important and ob1 usage is quite restricted (bad with RDMA hardware). This just hides the first bug above. -- You are receiving this mail because: You are on the CC list for the bug.