MPI run failure with petsc honoured

[haldaas]

Hello,
I am trying to install firedrake on my linux machine using already installed petsc (latest version). Installation completes well. However, the installation check results in 3 failures. The simple Helmholtz problem presented in the documentation runs sequentially fine. However, a segfault occurs in parallel right in the beginning of the code (when creating the mesh).
I attach here the configure logs for petsc and firedrake.
configure.log
firedrake-install.log
make.log

[wence-]

Can you show the backtrace from the segfault? I think @ksagiyam had noticed this with a recent PETSc update too, so might be working of finding out what is going on.

[wence-]

I think I've almost figured out what's going on. I think DMPlexDistribute() now happens by default (probably after https://gitlab.com/petsc/petsc/-/merge_requests/4843) when we call plex.setFromOptions() at

firedrake/firedrake/mesh.py

Line 718 in 4f61984

plex.setFromOptions()

, after which we explicitly call plex.distribute() using Chaco by default, which is a serial partitioner. The problem goes away if we run with -petscpartitioner_type "ptscotch".

Since right now we always do the distribution manually, we should wrap DMPlexDistributeSetDefault and DMPlexDistributeGetDefault and just call dmplex.distributeSetDefault(False) in the Mesh constructor before setFromOptions

[ksagiyam]

Thanks. Found that those functions were already wrapped. Updated #2282.

[wence-]

@haldaas I think we have fixed the problem, could you run firedrake-update and try parallel tests again?

[haldaas]

All pass correctly. Thank you very much!

[wence-]

Great, thanks.