Failures on sid buildbot

Issue #597 resolved
Jan Blechta created an issue

Comments (13)

  1. Jan Blechta reporter

    The original problem 1. and 2. were somehow resolved. Now we have instead problems with bad numerical stability in MUMPS. Maybe it's new major release 5.

  2. Johannes Ring

    Yes, PETSc 3.6.2 has been uploaded to Debian unstable and I have installed it on the sid buildbot. The MUMPS package in Debian is version 4.10 and not updated since 2011.

  3. Johannes Ring

    @blechta - Yes, both PETSc and MUMPS are installed as Debian packages on the sid buildbot. Nothing is installed by PETSc in the Debian package. If you think this is a bug in this version of MUMPS, I can remove PETSc from this buildbot and we will likely end up with the original problems of this bug report.

  4. Jan Blechta reporter

    No need to do that. I was just curious if this behaviour is caused by MUMPS 5 in PETSc 3.6.

  5. Jan Blechta reporter

    Branch jan/petsc-3.7-updates will fix both sid and mpich buildbot.

    MUMPS workaround is no longer needed. Let MUMPS decides about the best reordering algorithm freely.

    Let me figure out the SuperLU_dist situation later. Possibly related (but rather not): https://bitbucket.org/petsc/petsc/commits/ceeba3afeff0c18262ed13ef92e2508ca68b0ecf, http://lists.mcs.anl.gov/pipermail/petsc-users/2015-August/026671.html.

  6. Jan Blechta reporter

    Another possibility: PCFactorSetShiftType, PCFactorSetShiftAmount in PETScLUSolver.cpp.

  7. Prof Garth Wells

    I've raised this before with PETSc devs too. Apparently SuperLU_dist does very little pivoting, so can have problems with mixed/saddle point problems.

  8. Jan Blechta reporter

    Could be. And non-determinism in parallel could come from unseeded random generators in ParMETIS. Need to check if its configurable. I think I've seen it somewhere.

  9. Log in to comment