The influence of relaxed supernode partitions on the multifrontal method

Abstract

In this paper we present an algorithm for partitioning the nodes of a graph into supernodes, which improves the performance of the multifrontal method for the factorization of large, sparse matrices on vector computers. This new algorithm first partitions the graph into fundamental supernodes. Next, using a specified relaxation parameter, the supernodes are coalesced in a careful manner to create a coarser supernode partition. Using this coarser partition in the factorization generally introduces logically zero entries into the factor. This is accompanied by a decrease in the amount of sparse vector computations and data movement and an increase in the number of dense vector operations. The amount of storage required for the factor is generally increased by a small amount. On a collection of moderately sized 3-D structures, matrices speedups of 3 to 20 percent on the Cray X-MP are observed over the fundamental supernode partition which allows no logically zero entries in the factor. Using this relaxed supernode partition, the multifrontal method now factorizes the extremely sparse electric power matrices faster than the general sparse algorithm. In addition, there is potential for considerably reducing the communication requirements for an implementation of the multifrontal method on a local memory multiprocessor.