Possible tests with a real Ocean model?

[willirath]

We should also explore multi-node MPI with a real application like an ocean model.

cc: @koldunovn this repo could be a good starting point for testing FESOM with multi-node containerised MPI?

[koldunovn]

Yes, this still on our agenda. Making FESOM2 work with single node was relatively easy two years ago, as it has very few dependencies and compiles relatively easily. So can be a good candidate for tests. I will not have to do it myself, but if someone interested I will be happy to help. Maybe doing it together on telecon.

[willirath]

For the MITgcm, build instructions are here: https://mitgcm.readthedocs.io/en/latest/getting_started/getting_started.html

[kathoef]

There is #4 now, which makes the example from the MITgcm build instruction page work via an srun --mpi=pmi2 --ntasks=... on NESH. I think, it would be most interesting to check out model code performance in terms of container versus host environment next? Should we come up with a useful experiment strategy here, also maybe for FESOM? An alternative thought would be trying out MPI benchmarking tools.

[kathoef]

A quick Google search returned the OSU micro benchmarks, which also seem to have a tutorial available.

[kathoef]

A few (very preliminary! this is actually not finished...) MPI benchmarking results are now available here. Should we try something similar with FESOM configurations in terms of total runtimes next?