I agree dropping MPI for 32-bit altogether is probably not optimal.We've already got the mpi-default mechanism so apart from general bugs it shouldn't be too much trouble to switch the 32-bit systems to mpich (option 2). In a way it could also serve as an intermediate test case towards option 3 anyway, providing feedback on how readily the release arches adapt to an mpich tool chain.
The UCX status in mpich is an interesting point. Our upstreams have regularly recommended mpich over openmpi in the past, perhaps UCX changes the calculus. I figure mpich developers would fix the problems quickly enough though.
Drew On 2021-03-11 17:44, Alastair McKinstry wrote:
See the mail below from the OpenMPI maintainers. This would mean our options are: (1) Drop MPI for 32-bit archs. (2) Use MPICH for 32-bit archs, OpenMPI for 64 bit (3) Use MPICH for everything. My experience is that OpenMPI is more mature for new fabrics, etc (UCX) - I wasn't able to get UCX to work with MPICH in time for Bullseye, so I'm not favouring (3). I'd rather not drop MPI for 32-bit systems in general - while I don't expect proper HPC to run on these, its one of the simplest ways of adding parallelism, and there will be systems with large numbers of 32-bit cores. What are peoples opinions ? Best regards Alastair -------- Forwarded Message -------- SUBJECT: Dropping Open MPI 32 bit support
...
We're seriously talking about dropping 32 bit support starting with Open MPI v5.0.0: https://www.mail-archive.com/ompi-packagers@lists.open-mpi.org/msg00055.html