We may not need to use separate ContextCache objects for energy computation and replica propagation anymore now that https://github.com/openmm/openmm/pull/3721 has been merged. We could potentially combine a refactor here with some restructuring in openmmtools to minimize the need for each MPI process to cycle through conformations.
cc #1058: There's a weird issue where setting up a HybridTopologyFactory and running the simulation in the same script causes a >2x slowdown over resuming the simulation in a different process instance.
We may not need to use separate
ContextCache
objects for energy computation and replica propagation anymore now that https://github.com/openmm/openmm/pull/3721 has been merged. We could potentially combine a refactor here with some restructuring inopenmmtools
to minimize the need for each MPI process to cycle through conformations.