Load balancing strategies for a parallel system of particles

Abstract

This paper presents a study on the load balancing of a parallel implementation of a dynamic system of particles on distributed memory architectures. The parallel implementation simulates the temporal evolution and behavior of a general system of particles. Long range interactions with the environment as well as short range interactions between particles are fairly handled. In order to deal with these constraints, a complexe data structure has been implemented. An important study on load balancing is developed for a simple case. Difficulties to theorically calculate the relocation of datas (depending upon simulation and architecture) is shown, and an heuristic is proposed for the general cases. Some results on MIMD architectures and on a network of workstations are finally briefly presented.