Step 1: Run basic benchmark using azhpc #1

Activity

1. mloubout commented on Mar 11, 2020

   

   mloubout

   on Mar 11, 2020

   More actions

   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).
2. ggorman commented on Mar 11, 2020

   

   ggorman

   on Mar 11, 2020

   ContributorAuthor

   More actions

   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.
3. ggorman

   changed the title [-]Run basic benchmark using azhpc[/-] [+]Step 1: Run basic benchmark using azhpc[/+] on Mar 11, 2020