This is not really a new idea, more of a discussion point but I thought I would make a post to see if anyone may have suggestions as to how to improve performance on models with unstructured meshes. I recently did a small study to see the impact of using an unstructured mesh on transport speed, and found that the performance hit is substantial (10x reduction in particles per second with the inclusion of one mesh), which is unfortunate as to resolve results on a mesh more histories are required. Naturally I would expect that the changes in performance are model and mesh dependent.
Wondering if the mesh element type impacts speed. DAGMC supports tet meshes but LibMesh supports mixed mesh types. Perhaps some hexahedral meshes can represent the shape well but require less elements.
I didn’t consider using LibMesh, I didn’t know it supported hex meshes, that’s interesting to know in the future. I’ll leave this up in case it helps someone like me in the future, but what got my problem going fast enough that it is now doable was rebuilding OpenMC as release (I had it as debug ), and increasing the number of MPI processes I was using. Thanks for the help!
@Edgar I have been looking into unstructured mesh tally performance recently and found similar behavior. One thing I discovered is that the k-d tree that we use from MOAB for accelerating spatial searches on the mesh can be optimized by explicitly choosing parameters. See the associated pull request here:
Since that has been merged in OpenMC, there are now better defaults in place but you may want to fiddle with them to see if you can eke out better performance still.
), and increasing the number of MPI processes I was using. Thanks for the help!