Shift/collapse on neighbor list (SC-NBL): Fast evaluation of dynamic many-body potentials in molecular dynamics simulations

Abstract

Dynamic many-body potentials (DMBPs) significantly extend the applicability of molecular dynamics (MD) simulations to broad material properties and processes, including chemical reactions. Recently, we have designed a shift/collapse (SC) algorithm that accelerates DMBP-MD simulations over the conventional approach based on linked-list cell and neighbor list (NBL) methods. Here, we present an extension of the SC algorithm called SC-NBL, which outperforms the original SC algorithm and can be implemented seamlessly in conventional NBL-based DMBP-MD programs. Excellent performance is achieved by combining communication-lean SC and computation-lean NBL approaches. For small and large granularities, the SC-NBL approach achieves 1.33× and 3.29× speedup, respectively, over the original SC approach. In addition, SC-NBL performance surpasses the conventional NBL method for DMBP-MD at large granularity with running time speedup of 2.2 folds, which was the shortcoming of the original SC algorithm.