[petsc-dev] hang while saving non-distributed Exodus mesh

Vaclav Hapla vaclav.hapla at erdw.ethz.ch
Thu May 24 11:11:58 CDT 2018


I wanted to load an Exodus mesh and immediately save it into HDF5 (i.e. use PETSc just for the conversion) without distributing (needed for reproducing another issue). See https://bitbucket.org/petsc/petsc/branch/haplav/fix-dmplex-immediate-exodus-hdf5-save.

But then there's a problem with labels created by the reader (e.g. "Cell Sets"). Only rank 0 has them. This in turn leads to the HDF5 writer hanging in DMPlexWriteLabels_HDF5_Static.

I think this should be fixed somehow and I will gladly do that. But my question is, what should be fixed? Should DMPlexCreateExodus produce consistent labels (I mean the number of labels and their names being the same on all ranks of the DM's communicator), or DMPlexWriteLabels_HDF5_Static cope somehow with inconsistent labels?

I mean in longer term it would be nice to enforce labels being always consistent (e.g. DMSetLabelValue being collective) ...

Thanks for opinions,

More information about the petsc-dev mailing list