GFortran 15.1.0 not working with OpenCoarrays 2.10.2 and open-mpi 5.0.8

[cristiansinaloa]

Hello,

The dependencies for the Mac OS package manager, Homebrew, needs to be updated if gfortran 15.1.0 no longer supports open-mpi. As of now opencoarrays still has open-mpi as a dependency.

If you could please list an alternative MPI implementation and list it as a dependency on Homebrew I would really appreciated it.

Thanks again for the amazing work you guys do.

[vehre]

Hi,

I am so to say "responsible" for the change, that OpenMPI can not be used with OpenCoarrays from gcc-15 on. I am not on Mac so testing this is harder to me, but to my understanding is the [Brewfile](https://github.com/sourceryinstitute/OpenCoarrays/blob/main/Brewfile) used for HomeBrew on Mac OS, right? And that references mpich. Do you have any idea why is want openmpi? Again, I am not on Mac OS. It's difficult to me.

[rouson]

@zbeekman can provide the best explanation. My limited understanding is that even though Homebrew can install MPICH, the brew core satisfies MPI dependencies with OpenMPI. I think the way to override that is with a tap. It appears to me that the formula to which @vehre linked is a tap so possibly that will do the trick.

[chenrui333]

👋 I have updated the homebrew formula to use mpich instead of open-mpi, Homebrew/homebrew-core#230634

[chenrui333]

• I guess this is the reason why open-mpi was picked over mpich, Update documentation to codify MPI and BLAS/LAPACK library choices & link rubydoc Homebrew/brew#6209

[zbeekman]

@chenrui333 my recollection from my time as a Homebrew maintainer was that the packages all converged around OpenMPI as the default MPI implementation. If we can ship OpenCoarrays via Homebrew with MPICH as the dependency that would be phenomenal (even though it may cause some challenges due to other libraries that are common dependencies using OpenMPI).

[chenrui333]

yeah, I guess we just need to either revert your audit change or add a allowlist list.

[ggouaillardet]

FWIW #800 is a proof of concept to have recent gfortran support Open MPI

[jsquyres]

Also, the assumption that MPICH works with coarrays because it always provides a common global unique ID for MPI windows is not correct; this is not a guarantee that the MPI standard makes. See #801 for an example.

@ggouaillardet's POC in #800 shows how to portably support both Open MPI and MPICH at the level that MPICH is supported today. But even with this change, #801 will still fail (with both Open MPI and MPICH) because OpenCoarrays is relying on a guarantee that is not provided by the MPI standard.