Experience of sparse matrix codes on the CRAY-1

We consider direct methods based on Gaussian elimination for solving sparse sets of linear equations. Among conventional approaches, band and frontal methods are obviously vectorizable and general sparse methods equally do not vectorize easily since they involve indirect addressing in inner loops. We illustrate these effects with actual times from runs on the Cray at Harwell. To avoid indirect addressing we have been developing code that uses a “multi-frontal” technique. This moves the reals within storage in such a way that all operations are performed on full matrices although the pivotal strategy is minimum degree. We describe how the in-core and out-of-core versions perform on theCRAY-1.