Performance of MPI in singularity/apptainer

This is an interesting article where they attempt to benchmark containerised MPI

In some cases they actually seem to see improved performance for containerised MPI.

Chatting with @angus-g the other day, it seems containerised MPI can lead to large memory overheads.

I’d be interested to hear other peoples experiences, and if you want to provide more detail @angus-g that’d be great.