Multi-million particle molecular dynamics: II. Design considerations for distributed processing

Abstract This paper describes recent progress in implementing large-scale molecular dynamics simulations on distributed processor systems. The computational load is spread evenly among the processors by employing a spatial subdivision - a technique that introduces only minimal communication overheads dependent on the number of particles per processor, and which therefore lends itself to essentially unlimited expansion. Issues of interprocessor communication and overall control of the computation are discussed. Feasibility studies for two-dimensional systems with 4 million particles and three-dimensional systems with 2 million particles have been carried out on a hypercube array containing 64 processors; the performance was found to be comparable to that of the latest vector supercomputers, even though vectorization was not employed in the tests.