Is your feature request related to a problem? Please describe.
@inducer thanks for this great work of wrapping metis. I am developer of the flood and tsunami package https://github.com/anuga-community/anuga_core and we use metis (indeed your pymetis wrapper) to partition our triangular mesh for mpi runs.
In the past for our python 2 version we used metis 4 and our own wrapper of the partMeshNodal procedure. When we moved to python 3 we decided to upgrade to using metis 5. Instead of writing our own wrapper we decided to use your wrapper, and your pymetis.part_graph procedure. Worked a charm.
But on closer examination the quality of our partitions (using pymetis.part_graph) and the scaling of our parallel code, it seems that the partitions produced by pymetis.part_graph are a little inferior to the partitions created by partMeshNodal. Our communication regions seem to be about 20% larger with pymetis.part_graph
Describe the solution you'd like
So I would like to request that you also add a wrapper to partMeshNodal
Describe alternatives you've considered
I did have a very short look at your wrapper code. I would be willing to try to code up the extension myself, but would appreciate any help from you.
Is your feature request related to a problem? Please describe. @inducer thanks for this great work of wrapping metis. I am developer of the flood and tsunami package https://github.com/anuga-community/anuga_core and we use metis (indeed your pymetis wrapper) to partition our triangular mesh for mpi runs.
In the past for our python 2 version we used metis 4 and our own wrapper of the
partMeshNodal
procedure. When we moved to python 3 we decided to upgrade to using metis 5. Instead of writing our own wrapper we decided to use your wrapper, and yourpymetis.part_graph
procedure. Worked a charm.But on closer examination the quality of our partitions (using
pymetis.part_graph
) and the scaling of our parallel code, it seems that the partitions produced bypymetis.part_graph
are a little inferior to the partitions created bypartMeshNodal
. Our communication regions seem to be about 20% larger withpymetis.part_graph
Describe the solution you'd like So I would like to request that you also add a wrapper to
partMeshNodal
Describe alternatives you've considered I did have a very short look at your wrapper code. I would be willing to try to code up the extension myself, but would appreciate any help from you.