Thanks Jed, I'll pass that on to the debian thread. Normalising the MPI ABI sounds like a useful goal. Drew On 2018-12-08 13:15, Jed Brown wrote:

The Open MPI API is better for developers due to better type safety (all handles in MPICH are typedef'd to int). Most major commercial vendors are organized around variants of MPICH (where there is collective ABI standardization). Open MPI is more modular so most vendor stuff goes into their plugins (for those vendors working with OMPI).

I think a good solution for Linux distros (and many others) would be to make a library that is ABI compatible with OMPI, but dispatches through to MPICH. There exists a (messy) open source demonstration.

https://github.com/cea-hpc/wi4mpi/

Drew Parsons via petsc-dev <[email protected]> writes:

...

Hi PETSc developers,

Debian (and therefore Ubuntu) currently uses openmpi as the default MPI implementation. Some developers have encouraged us to switch to mpich, partly on experience of less bugs with mpich, partly around openmpi development policies.

We're amenable to the idea but don't want to rush. The next Debian stable release is coming soon (around February) and we'd anticipate a lot of transition bugs (for instance scalapack currently builds for both openmpi and mpich, build tests pass openmpi but fail mpich).

Equally importantly, pmix is in an uncertain state for mpich (and we've just finished closing outstanding pmix bugs with openmpi3. openmpi3 is stable now). Alastair McKinstry, the lead Debian MPI developer, thinks pmix support should be in order before going ahead with a transition to mpich, especially for container support (singularity).

There's a discussion thread on Debian-Science at [1] (cc: to debian-release [2]). I'm writing to invite you to add your own experience to these discussions.

Best regards, Drew Parsons Debian Developer / Debian Science member

[1] https://lists.debian.org/debian-science/2018/12/msg00017.html [2] https://lists.debian.org/debian-release/2018/12/msg00162.html