Example for running MPI for ReplicaExchangeSampler?

[mrshirts]

Is there a good example for a script that parallelizes the runs generated by ReplicaExchangeSampler over cores via MPI (or other means)? Note that are systems are small enough (<100 particles) that we don't appear to be getting that much performance benefit of GPU over CPU. We are having a hard time getting this work, and we didn't find anything in the official documentation.

Thanks!

[andrrizzi]

ReplicaExchangeSampler should have automatic detection of MPI and it should parallelize the replicas over MPI processes accordingly. So it should be possible to just run mpirun [mpirun_options] my_script_with_rex.py.

To make it run over CPU rather than GPU, you should be able t set the platform in the ContextCache object used to generate the OpenMM Contexts. By default, ReplicaExchange uses a global context cache that you configure at the beginning of your script like this

from openmmtools.cache import global_context_cache
global_context_cache.platform = openmm.getPlatform('CPU')

Note however this bug: #449. This is giving us a hard time, and I am short on dev time right now, unfortunately.

[mrshirts]

Thanks so much for the info! I suspect we will either be running on 1 GPU at a time, or parallelized on CPU's, so we hopefully will be OK.

[andrrizzi]

Parallelizing over CPU might indeed work since I don't think we have tested the bug on CPUs. So far, however, the problem seems to be with MPI rather than the GPU platform in particular.

[mrshirts]

BTW, it's working, but we are still sorting out how to optimize what it does.