Low Cost Parallelizing: A Way to be Efficient

In order to minimise the heavy cost of running parallel applications on expensive MPPs, we consider the use of workstation clusters. The main difficulty to obtain efficiency on such architectures is the fact that a slow node may dramatically decrease the overall performance of the machine. We propose a dynamic load balancing solution, based on a local partitioning scheme, consisting of moving sub-domain boundaries so as to adjust the processor load according to its available CPU capability.