Re: OpenMPI 5.0 to be 32-bit only ?
On Thu, Feb 09, 2023 at 09:53:37AM +0000, Alastair McKinstry wrote:
> The push-back from upstream is that they're unconvinced anyone is actually
> using i386 for MPI.
> For example, MPI is configured to use PMIx but its thought that doesn't work
> on 32-bit, but no bugs have been reported.
> Either we increase our 32-bit testing regime, or realistically consider it
> marginal and dying.
I don't think lack of testing is the problem, we should have pretty good
coverage due to buildtime and autopkgtest tests.
There are bugs like e.g. #1003020 or #1026912 that might be due to
OpenMPI failing on 32-bit with 160 cores.
Whether spending time trying to properly fix these would be worth it,
that's a different question.
> Currently I'm favouring accepting a move to 64-bit OpenMPI as a fait
> accompli as part of code cleanups for 5.X (post Bookworm), and Debian moving
> to MPICH on at least 32-bit archs - I'd favour OpenMPI on 64-bit archs for
> better incoming-code-and-compatability support.
> I'd like to hear the case otherwise.
The case we should make is that "no one cares about 32-bit builds" from
the starting post in the GitHub issue is not true for Debian.
We do care that it *builds*, even if it might not be actually used.
 was about the benefits of switching the two architectures that were
using MPICH to OpenMPI two years ago. The mentioned "makes packages like
octave build" is due to sundials build depending on mpi-default-dev but
requiring ompi-c.pc .
m68k and sh4 are building with nocheck, whether or not there might be
additional/different test failures in packages with MPICH is unknown.
Having different MPI implementations on different architectures again
would be painful for us, especially if it would be on release architectures.
If it would be architecturally hard for upstream to continue supporting
32-bit then that's how it is, otherwise the current status quo of 32-bit
OpenMPI is good enough for us and a possible compromise might be if
upstream says "32-bit patches are welcome" and requires an
configure flag when building for 32-bit archs.
> Best regards