Parallel speedup appears sometimes to be quite poor when the mesh is mixed hex-tet (hex dominant or hexa core). All the partitions seemed to have similar numbers of nodes, which is the target of the MeTiS kway algorithm, but some partitions ended up with predominantly tets whereas the others are predominantly hex. Hence there are lots more elements in some partitions compared to others. This means some partitions have a lot more work involved in the solver assembly than others, and so the efficientcy of speedup is poor.

We have an expert parameter 'part cvs weighting' which, if set to true, weights each vertex within partitioning by the number of control volume sectors it touches. It has helped mixed element cases in the past.

Show Form
No comments yet. Be the first to add a comment!