GROMACS

We should-reevaluate whether/when do we need the two-step communicators implemented in mdrun (separate intra-/inter-node). Vendor optimized MPI libraries should provide optimized implementations for collectives across SMP nodes of a cluster, so we might be able to get rid of our internal optimizations in case if we determine that all modern MPIs do the right thing.

Testing is easily done by setting GMX_NO_NODECOMM env. var.