- changed milestone to 1.7
-
assigned issue to
Enable mesh partitioning on a subset of processes
As it is currently implementd, mesh partitioners uses all the processes in an MPI communicator to perform the partitioning. For large core counts (say, more the 32K cores) this is an issue since the memory usage of most of the mesh partitioning packages depend on the number of processes (almost quadratically). Also, the quality of the partitioning is affected by the number of processes used.
Comments (4)
-
-
Issue
#637was marked as a duplicate of this issue. -
- removed milestone
Removing milestone: 1.7 (automated comment)
-
I find this issue useful for my need of partitioning mesh with optional argument of number of partitions and number of MPI processes.
Any progress happened in this issue? Seems like a old issue, but still open. Would be good if something might have already been updated in the past versions.
- Log in to comment
Should be roughly possible by giving LocalMeshData a different communicator, but will require a fair bit of cleanup to make it happen