Initial partition and large numbers of MPI ranks at low redshift.
I've been trying to run SWIFT with larger numbers of MPI ranks than we usually attempt. One of these tests uses 672 ranks running over 24 nodes of COSMA7 with a volume equivalent to EAGLE_100. That works but only when we increase the numbers of top-level cells to 72x72x72, lower values fail to get a good initial partition (the job continues past this and then usually fails with memory exhaustion on one or more nodes).
Even with a good initial partition, which now runs for many steps, the memory balance between the ranks is not great, why?