(687g) Parallelization of Grand Canonical Ensemble Monte Carlo Using Prefetching and Windowing of Flat Histogram Simulations

While simulation techniques such as Molecular Dynamics (MD) efficiently utilize multi-core processors using domain decomposition, Monte Carlo (MC) simulations with single particle perturbations are difficult to domain decompose especially when interaction ranges are not significantly shorter than the domain. Thus, although serial MC simulations may be more efficient than serial MD for computing phase equilibrium properties, for example, parallel MD is often used when parallel MC software and algorithms are less abundant. However, parallel prefetching Monte Carlo simulations in the canonical ensemble (CE) can speed up a simulation by a factor of 3 when using 4 cores [1]. In this work, prefetching in the grand canonical ensemble (GCE) is shown to be about as efficient than in the CE, despite counteracting non-trivial efficiency differences in acceptance probability and processor load balancing. In addition, serial efficiency is also improved by a systematic study of simulation parameters. Finally, in addition to parallelization via prefetching, parallelization efficiency in transition-matrix MC simulations is also investigated by decomposing a range of macrostates into windows. These parallelization techniques, made available in the open-source software FEASST, reduce the time required to compute a vapor-liquid phase diagram of Lennard-Jones or SPC/E water by over an order of magnitude.

[1] H. W. Hatch, “Parallel Prefetching for Canonical Ensemble Monte Carlo Simulations”, J. Phys. Chem. B, 2020, 124, 7191-7198.