Bug#996204: transition: numerical library stack
On 2021-10-30 11:04, Sebastian Ramacher wrote:
On 2021-10-30 00:50:49 +0200, Drew Parsons wrote:
I've got dolfinx passing (or skipping) its own tests on 32-bit.
There's a problem with 32-bit MPI though. i386 and armhf are failing
to run
pytest over python unit tests in MPI
...
Might be something more insidious going on with 32-bit python and MPI.
mpi4py is also failing i386 tests, see
https://ci.debian.net/packages/m/mpi4py/testing/i386/
https://github.com/mpi4py/mpi4py/issues/105
Could you please file a serious bug agains mpi4py so that this issue is
documented in the BTS?
...
If this issue is solely caused by mpi4py, then it should be fixed there
and we can ignore the test results for fenics-dolfinx for this
transition.
It's not certain that mpi4py is at fault though. For one thing, not all
dolfinx MPI python tests fail on i386. And armhf is passing mpi4py
tests. mpi4py is generally careful, meaning that it's also possible that
any MPI fault is in openmpi itself. (openmpi's self tests don't seem to
be as comprehensive as mpi4py test).
It could be that both dolfinx and mpi4py test failures are a symptom of
a problem elsewhere. At the same time there have been other 32-bit
problems in dolfinx itself (Bug#995599 C++ demos).
I filed mpi4py Bug#998106 in any case to track mpi4py/issues/105,
I marked it serious but commented that I think that's overkill, since
the mpi4py test failure is in one specific place and can be avoided by
simply skipping that test on i386.
In any case, everything is blocked on the sundials upload.
True, we have a bit of time before ultimate migration.
Reply to: