Scheduling many-body short range MD simulations on a cluster of workstations and custom VLSI hardware

Abstract

Molecular dynamics is a powerful technique used to obtain static or dynamic properties of liquids and solids. The sheer computational intensity of many of these simulations demands more computational power than what any uniprocessor system can provide. Fortunately, these simulations can be parallelized, allowing faster execution times on a cluster of workstations. Of late, custom VLSI chips have been designed to provide an alternative to parallel techniques. The MD-GRAPE 2 is one such solution, offering a peak performance of 64 Gflops. We evaluate the performance and cost-effectiveness of various methods used in sequential and parallel molecular dynamics and the MD-GRAPE 2. We then illustrate how MD simulations involving more complex potential functions can be scheduled on parallel machines and the MD-GRAPE 2 simultaneously. © Springer-Verlag 2004.