search

conda-forge / fenics-feedstock

A conda-smithy repository for fenics.
BSD 3-Clause "New" or "Revised" License
13 stars 12 forks source link

openmpi package with xdmf patch? #170

Closed sblauth closed 1 year ago

sblauth commented 1 year ago

Comment:

I've got a question regarding the MPI build of this package. I noticed that I had some issues writing .xdmf files when running on a cluster with MPI. I noticed that there might have been an issue and that a new build was generated (see #169). However, I still have trouble using fenics from the cluster. The current build (34) apparently only support mpich (trying to install this build with openmpi fails - at least when I want to use python3.11). But I still got issues writing XDMF files from the cluster with that build (no .h5 files are created).

I installed the package with openmpi and the previous build number (33), but that does not seem to have the fix included, so I still run into issues there.

Is there a build of the package with number 34 available for py3.11 for both openmpi and mpich? Am I missing something?

Thanks a lot in advance for your help!

minrk commented 1 year ago

The short answer is you should be able to workaround it if you install fenics-dolfin directly instead of fenics.

Looks like we're running into https://github.com/conda/conda-build/issues/3308 (or a related issue) where the hash contents don't propagate to dependencies within the multiple-output recipes, so the different mpi providers are not producing different build strings for the fenics metapackage. This results in only uploading 4 builds of fenics instead of the 8 it builds.

sblauth commented 1 year ago

Okay, thanks a lot for your help and for rebuilding the packages. With this, I should be able to install the openmpi version too (or just using the fenics-dolfin package).

I have also resolved my issue and for the sake of completeness: Writing .xdmf files in parallel does only work, when the target directory is created before the files are saved to it. In serial (and also perhaps when using a single compute node), the creation of the appropriate directory is done automatically, but when using multiple nodes, the directory has to be created first.

minrk commented 1 year ago

Yeah, the fenics package doesn't have anything in it - it used to be one actual conda package with everything in it, but it's been split up since then, leaving the empty metapackage around for backward-compatibility. We haven't done that with fenicsx on conda-forge - there is only fenics-dolfinx, no fenicsx metapackage.

This should be fixed by #171, at least.