Equilibrage et régulation de charge dans les machines parallèles à mémoire distribuée

Solving the load allocation problem is an important issue for using and operating parallel machines. We start with a state of arts of this problem for distributed memory architectures; we focus on dynamic allocation, more exactly on load balancing for regular load. An originalload balancing strategy based on a generalized prefix calculation is proposed. This strategy is correct, exact and independent on the processors interconnexion network. A load balancing kernel is implemented. We continue our study with an analysis of the execution time of this strategy. The main result is that a Gumbel distribution law modelizes the maximal execution time. Based on this results we infer a decision policy for initiating and running the load balancing process. So the proposed load balancing algorithm is efficient. An application for mechanical process simulation (large-strain deformation and recrystallization) is developed. For this dynamic application we use our load balancing kernel and the initiation and decision policies. In practice the " load balancing algorithm is always correct, stable and efficient.