Open ggorman opened 4 years ago
One MPI process per socket
Depends on the nodes, if you run on IB**
ones (the fancy HPC ones) these are weird AMD chips with 2 sockets but 30 NUMA(on the IB_120, 15 NUMA on IB_60) domains so lot better to run with one process per NUMA domain. and very little openmp (4 threads per numa domain).
From my perspective, the point is this issue to 'knowledge capture' these as options/parameters in the first instance. We can always drop then later as specific platforms.
See installation and basic benchmark in https://github.com/devitocodes/azhpc/blob/master/README.md
Requirements: