Unless otherwise noted benchmarks run for 20 ps (10000 steps * with dt = 2 fs), Langevin thermostat, Berendsen barostat.
This benchmark is actually 800 ps long.
OpenMPI, pmemd.MPI, 800 ps, 2015-03-31
intel, IntelMPI, pmemd.MPI, 200 ps, 2015-10-15
intel, mvapich2, pmemd.MPI, 200 ps, 2015-10-15
intel, mvapich2, pmemd.MPI, 20 ps, 2015-03-30
MPI: /uufs/ember.arches/sys/pkg/mvapich2/std_intel/etc/mvapich2.sh (X)
Note the performance drop for small systems (< 30k atoms). Using ember-specific mvapich2 appears to help somewhat at higher processor counts. Of course, this could also be because the benchmark is too short to be very accurate. May need to re-run these numbers.