Test failures in test_mat_aij.py: testInvertBlockDiagonal

[James D Trotter]

I see some failing tests in "testInvertBlockDiagonal" with PETSc-3.10.3 and petsc4py-3.10.1.

I have attached the complete output, but here is a sample from one of the failing tests:

======================================================================
FAIL: testInvertBlockDiagonal (test_mat_aij.TestMatMPIAIJCRL_B_G23_B5)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test/test_mat_aij.py", line 248, in testInvertBlockDiagonal
    self.assertTrue(N.allclose(ibdiag, tmp))
AssertionError: False is not true

Is this related to the way that I have configured PETSc?

[Lisandro Dalcin]

I cannot reproduce these failures. How did you configure PETSc ? Are you using complex scalars? What architecture are you running on?

[James D Trotter]

Hi Lisandro. I am building this on an x86_64 machine with an Intel Sandy Bridge Xeon E5-2650 CPU, running Ubuntu 18.04 with Linux kernel 4.15.0.

I am not using complex scalars. The following are the configure flags I use to build PETSc:

--COPTFLAGS="-g -O2" --CXXOPTFLAGS="-g -O2" --with-debugging=yes --with-fortran-bindings=no --with-hdf5-dir=/nix/store/qm2xii1w55p4mhvazyw5zdwpii70jqn5-hdf5-1.10.3 --with-hypre-dir=/nix/store/ryzkj0sjhyank4zwnpds1gq8gxcfbwzf-hypre --with-metis-dir=/nix/store/qdn7yqbjk45dzzdn1i6pnvv3j15ayjrb-metis-5.1.0 --with-mumps-dir=/nix/store/1swxhkm9320gflpjlva9cnclpg2an57l-mumps --with-scalapack-dir=/nix/store/zqqwapc9p1zjbdgk3kck0swz7lzkd0lw-scalapack --with-ptscotch-dir=/nix/store/prl6q0z89fadhhdgjm7wr4d0k47ahwpr-scotch-6.0.4 --with-spai-dir=/nix/store/41x8qajpmfzmdyk24dvf4p0a3md26cmd-spai --with-suitesparse-dir=/nix/store/94qi9vj49ma7zrn8kc5xh0q9x6sf0yrz-suitesparse-4.4.4 --with-superlu-dir=/nix/store/2a6hqxpk1xv2i7wf523y7g4lz8y47q3c-superlu-5.2.1

In other words, I am configuring PETSc with HDF5, HYPRE, metis, mumps, ptscotch, scalapack, spai, suitesparse, and superlu.

Do you have any suggestions about how I could debug this issue?

[Lisandro Dalcin]

Mmm... maybe just add some print statements to take a look at ibdiag and tmp to see how ibdiag (computed by PETSc) and tmp (computed by numpy) differ? Is your PETSc using the same blas/lapack as numpy?

[Lisandro Dalcin]

Any news about this issue?

[James D Trotter]

Sorry, I haven’t had the time to look into this issue again. Perhaps it’s best to just close it.

Thanks for your help, anyway.