A distributable algorithm for optimizing mesh partitions

Much effort has been directed towards developing mesh partitioning strategies to distribute a finite element mesh among the processors of a distributed memory parallel computer. The distribution of the mesh requires a partitioning strategy. This partitioning strategy should be designed such that the computational load is balanced and the communication is minimized. Unfortunately, many of the existing approaches for the mesh partitioning are sequential and are performed as a sequential pre-processing step on a serial machine. These approaches may not be appropriate if the mesh is very large, that is to say, this pre-processing on a serial machine may not be feasible due to memory or time constraints, or on a parallel machine if the mesh is already distributed. In this paper we propose an algorithm with a fundamentally distributed design for the optimization of mesh partitions. To assess the quality of the partitioning, the size of edge cuts is the taken as the chosen metric.