抄録
This paper describes a vertical-block decomposed parallel LU factorization method(hereafter VB-PLU), which processes effectively on a distributed memory parallel computer. Because VBPLU method is based on the block algorithm, the message aggregation technique and the block-cyclic distribution strategy, it can reduce issue of communication and improve load balance. Further-more this method ensures to use long vectors for updating in Gauss-elimination, thus it is also advantageous with vector processors. In this paper, VBPLU method is shown and computational costs are estimated. Its evaluations are confirmed through numerical experiments on two parallel computers, SR2201 and VPP300.
