Benchmarking GPU acceleration of a parallel molecular dynamics simulation

Abstract

The parallel speedup of component subroutines in a molecular dynamics simulation is measured after implementing MPI-based CPU parallelization and GPU acceleration. GPU acceleration was observed to yield an order of magnitude reduction in simulation times (13.9x–23.6x relative to a serial run). Slightly less speedup is achieved by CPU parallelization (7.5x) over nine processors. Performance scaling is diminished by communication operations as the number of processes launched begin to exceed the number of available physical CPU cores.