Amesos2: KLU2 re-indexing operation causes throw from Map ctor in MueLu solve

[MalachiTimothyPhillips]

Bug Report

@trilinos/amesos2 @iyamazaki

#13740 introduces a Map ctor call that ends up hitting the following error in a Sierra/Fuego test that uses MueLu:

  /sierra/build/linux_rh7/nightly/intel_trilinos_master/objs/tpls/spack/spack/.spack/stage/spack-stage-trilinos-2025.02.02-hcx7s2trdqx6ncqfewvuau5py6eebw5h/spack-src/packages/tpetra/core/src/Tpetra_Map_def.hpp:841:
  Throw number = 1
  Throw test that evaluated to true: minAllGID_ < indexBase_
  Tpetra::Map constructor (noncontiguous): Minimum global ID = 0 over all process(es) is less than the given indexBase = 2005.

Specifically, setting gather_supported=false just before the branch here fixes the issue:

Trilinos/packages/amesos2/src/Amesos2_KLU2_def.hpp

Line 462 in bd59c87

if (gather_supported) {

[github-actions]

Automatic mention of the @trilinos/muelu team

[MalachiTimothyPhillips]

@iyamazaki By the way, you can also reproduce this by using KLU2 on the actual matrix. The minimum gid is > 0, so I'm pretty sure you could add a unit test with a single row and a non-zero gid to reproduce this issue.

Here are the relevant matrix files for the error we're observing on np=4:

matrix-files-for-13774.tar.gz

[MalachiTimothyPhillips]

Maybe the error stems from here:

Trilinos/packages/amesos2/src/Amesos2_TpetraCrsMatrix_MatrixAdapter_def.hpp

Line 119 in bd59c87

auto tmpMap = rcp (new contiguous_map_type (numDoFs, nRows, indexBase, rowComm));

?

[iyamazaki]

Thank you, @MalachiTimothyPhillips. I see the issue, and will look into fixing it.

[MalachiTimothyPhillips]

Closing as #13790 has been merged.

Thank you, @iyamazaki!