Dynamic Load Balancing with Pair Potentials

We present a new load balancing algorithm inspired by Molecular Dynamics Simulations. Our main motivation is to anticipate the rising costs of tasks-scheduling caused by the growth of the number of available cores on chips. This algorithm is based on a virtual decomposition of workload in Voronoï cells centered around computing units. The method used in this paper allows cores to virtually move in order to change their computing load. Cores displacements are result of forces computation (with pair potential): attractive or repulsive forces between cores are balanced by the cores computing load (total cost of Voronoï cell). Over-charged cores are more attractive than under-charged cores (which are then more repulsive). In this paper, we demonstrate the relevance of our approach by experimenting our algorithm with a high number of automatically-generated test cases, ranging from almost stable to quickly-evolving scenarii. In all cases, our algorithm is able to quickly converge to a distribution which maintains good locality properties.