Matrix Factorization Using Distributed Panels on the Fujitsu Ap1000

Dense linear algebra computations such as matrix factorization require the technique of`block-partitioned algorithms' for their eecient implementation on memory-hierarchy processors. For scalar-based distributed memory multiprocessors, the register, cache and oo-processor memory levels of the memory hierarchy all aaect the optimal block-partition size for such algorithms. Most studies on matrix factorization and similar algorithms have assumed that the block-partition size or panel width for the algorithm, !, to be the same as the matrix distribution block size, r, where a rectangular block-cyclic matrix distribution is being employed. Here the choice of ! = r is essentially determined by the oo-processor memory level of the memory hierarchy, with the value of ! being a tradeoo between communication startup overhead and load balance considerations. In this paper, we reexamine this assumption in the context of LU and Cholesky factorization of block-cyclic distributed matrices on scalar-based distributed memory multiprocessors, such as the Fujitsu AP1000. Here considerations of the register and cache levels of the hierarchy require a large !. We nd that the choice of !, given ! = r, leads to a tradeoo between load balance and optimal use of register and cache levels of the hierarchy (rather than communication startup), and that this tradeoo substantially limits performance. We then brieey describèdistributed pan-els' versions of these algorithms, where generally ! > r, which eeectively diminishes this tradeoo to an O(!=N) fraction of the overall computation , where N is the matrix size. Two variants of these versions, one with single rows/columns being communicated, and one with single block rows/columns being communicated, are analyzed for their load balance properties. The results of the distributed panels versions of the algorithms on the scalar-based distributed memory multiprocessor the Fujitsu AP1000 are given, which give signiicantly superior performance for distributed panels versions over the ! = r versions, with optimum performance achieved for r 1.