...
- do not use -xHost, because Intel does not "recognize" AMD (officially for security reasons . Therefore use: -xcore-avx2
- These options were used by NEC during the FESOM2 benchmark:
PT = -O3 -qopt-report5 -no-prec-div -ip -fp-model=fast=2 -implicitnone -march=core-avx2 -fPIC –qopenmp -qopt-malloc-options=2 -qopt-prefetch=5 -unroll-aggressive
These are (at least partially) quite important for good performance. However, we do not have the experience which are more or less critical. Be careful, some options might kill reproducibility (e.g., -fp-model=fast=2). - Natalja is now responsible for this: https://docs.dkrz.de/doc/levante/running-jobs/runtime-settings.html#open-mpi-4-0-0-and-lat let's stll benefit from here knowledge
- Independent on the MPI used, please try runtime setting
UCX_TLS=knem,dc_x,self
for your jobs. According to NEC for smaller Jobs it might be beneficial to replace "dc_x" with "rc_x".
Spack
On albedo we mainly use spack to install software and provide module files.
...