An analysis of data distribution methods for Gaussian elimination in distributed-memory multicomputers

In multicomputers, an appropriate data distribution is crucial for reducing communication overhead and therefore the overall performance. For this reason, data parallel languages provide programmers with primitives, such as BLOCK and CYCLIC that can be used to distribute data across the distributed memory. However, the languages do not aid the programmer as to how the distribution should be performed to maximize the performance. Therefore, this paper presents an analysis of data distribution methods for overlapping computation and communication in the Gaussian elimination algorithm. The analysis indicates that both BLOCK and CYCLIC distributions have their own merit; however, BLOCK-CYCLIC with its hybrid characteristic consistently out performs its counterparts