Hardware Issues in Molecular Dynamics Algorithm Design

Abstract

The diversity of architectures on which modern supercomputers are based presents a serious obstacle for the designer of software that is to operate efficiently in a variety of environments. As case studies in how to live with processor hardware, two very different kinds of contemporary computer design are introduced, one being the familiar vector computer, the other an experimental loosely-coupled network of processors. The basic concepts of each kind of computer that are relevant to algorithm development are discussed and the manner in which the machines have been harnessed to the task of large-scale molecular dynamics simulation are described. Only by taking into account the specific features of each machine was it possible to carry out what are the most extensive simulations of this kind to date, involving as many as 200,000 particles. These techniques apply to models based on differentiable potentials; an approach very different from direct numerical integration of the equations of motion is required for models that employ step potentials. A method derived from concepts of event-driven simulation and list processing that proves to be highly effective for problems of this type is presented. New results obtained by applying these techniques to the study of fluid dynamical problems at the microscopic level are briefly mentioned.