Bug#1004556: libmpich-dev: The simplest MPI program compiled with mpich crashes

[Drew Parsons]

On 2022-02-22 19:49, Alastair McKinstry wrote:
> I can disable pmix support in mpich.

Thanks Alastair.  My tests are now passing with mpich 4.0-3

> Pmix is working fine in openmpi, so it’s an mpich/pmix issue of some
> sort (or maybe ch4)

"Working fine" might not be quite the word for it. There's some strange 
stuff going on with openmpi, at last in multinode execution (RMA). 
https://github.com/open-mpi/ompi/issues/10026

That's actually the reason why I was rebuilding with mpich.  nwchem is 
completely useless with openmpi in a multi-node job.  (well, they hope 
the situation will be better with openmpi 5).

Drew


> 
> Regards
> Alastair
> 
> On 22/02/2022, 14:57, "debian-science-maintainers on behalf of Drew
> Parsons"
> <debian-science-maintainers-bounces+mckinstry=debian.org at alioth-lists.debian.net
> on behalf of dparsons at debian.org> wrote:
> 
>     Package: mpich
>     Followup-For: Bug #1004556
> 
>     My guess is that this bug is ongoing in 4.0-2 because of pmix 
> support.
> 
>     4.0-2 is still configured 
> --with-pmix=/usr/lib/x86_64-linux-gnu/pmix2
>     and still Depends: libpmix2 (>= 4.1.2)
> 
>     pmix support was added in 4.0~b1-2 along with ucx.
> 
>     I gather ucx was deactivated in 4.0-2 but pmix was not. Looks like 
> pmix
>     also needs to go (unless the problem is ch4, which was also added 
> in
>     4.0~b1-2). But the error message references PMIX.
> 
> 
>     Ironically, I find that an executable compiled against mpich 4.0-1
>     fails with mpiexec.mpich (as raised in this bug) but actually 
> passes
>     when run with mpiexec.openmpi.  Awkward.
> 
>     --
>     debian-science-maintainers mailing list
>     debian-science-maintainers at alioth-lists.debian.net
> 
> https://alioth-lists.debian.net/cgi-bin/mailman/listinfo/debian-science-maintainers