🚧 🚧 This library is a work in progress; some features may be incomplete, incompletely documented, or incompletely tested. This file will be updated as development progresses. 🚧 🚧
MultiOpenMM is a package that enables setting up and running parallel molecular dynamics simulations, including but not limited to replica exchange simulations, using the OpenMM molecular simulation package. MultiOpenMM supports running independent simulations of multiple, possibly topologically distinct, molecular systems, within a single OpenMM context.
This can be especially useful when running, e.g., replica exchange simulations
with many replicas but few particles per replica: ordinary use of OpenMM
with one replica per GPU, for instance, would require many available GPUs but
would waste most of their computational capacity. In this case, depending on
the interactions present in the system of interest, MultiOpenMM can be used to
saturate a single GPU by stacking many non-interacting replicas into a single
OpenMM System. Scaling of interactions to account for differences in
temperature is handled automatically by MultiOpenMM. The package even supports
more flexible use cases than replica exchange, allowing systems stacked within a
single OpenMM System to have different force field terms and parameters or
molecular topologies (as might be useful for some kinds of high-throughput
screening simulations across chemical or protein sequence space). MultiOpenMM
attempts to optimize the resulting set of OpenMM Force objects in the stack to
maximize performance.
Alternatively, in cases where each replica is large, or otherwise cannot be stacked (MultiOpenMM cannot stack systems with long-range electrostatics or barostats, for instance), it may be desired to distribute work across multiple GPUs. MultiOpenMM contains a client-server system that allows simulation tasks for any number of replicas or systems to be distributed across any (possibly non-commensurate) number of OpenMM workers. Workers can even be added and removed dynamically during the course of a simulation, and simulation tasks for workers that fail will be automatically queued for redistribution to other workers.
Python (at least 3.10), OpenMM 8, NumPy, SciPy. MDTraj is required for DCD trajectory export.
Sphinx API documentation is available and can be compiled from /doc/ with,
e.g., make html.
This software is developed by Evan Pretti in the Shell research group in the Department of Chemical Engineering at the University of California, Santa Barbara. Evan Pretti gratefully acknowledges support from the Molecular Sciences Software Institute (MolSSI) as well as helpful discussions with Levi Naden at MolSSI. The Molecular Sciences Software Institute is supported by the National Science Foundation under Grant CHE-2136142.
This software is ©2024 The Regents of the University of California. All rights reserved. This software is licensed under the MIT License. For license information, see the license file.