Load balancing for spatial-grid-based parallel numeric simulations on clusters of SMPs

Load distribution is an essential factor to parallel efficiency of numerical simulations that are based on spatial grids, especially on clusters of symmetric multiprocessors (SMP). This paper presents a method of mapping spatial grid nodes to processors that combines two load balancing methodologies, graph partitioning and graph matching, to achieve maximum parallel efficiency on SMP clusters. The method has been successfully applied to load distribution in a parallel computational fluid dynamics (CFD) simulation. Test runs on a PC cluster prove the effectiveness of our method.