MPI Topics on the HPE Cray EX Supercomputer

Presenter: Harvey Richardson (HPE)

Archived materials on LUMI:

• Slides: /appl/local/training/2p3day-20250303/files/LUMI-2p3day-20250303-404-Cray_MPI_on_Slingshot.pdf

• Recording: /appl/local/training/2p3day-20250303/recordings/404-Cray_MPI_on_Slingshot.mp4

These materials can only be distributed to actual users of LUMI (active user account).

Q&A

1. Based on the slide (28 maybe) should we expect very similar perfomrance having 1 GPU per MPI vs 2 GPUs per MPI comunicating with each other when the GPUS are in the same ? Since both cases use P2P

   • Could you make an example?

   Sorry. For later. Assuming GPU-GPU communication. There are two cases 1 GPU per MPI, using MPI for excahnging data or 1MPI process with 2 GPUs using direct peer to peer access

   • OK, so MPI uses P2P for that case (IPC) but there are corner cases (small message size for instance), so performance can be different. There is a range of env variables that you can use to tune P2P (check man mpi and grep for IPC). (sorry for the initial confusion, I thought P2P to be Point-to-Point MPI communications...)