Compile time transformations for sparse matrix computation used in PERMAS to improve locality

Automatic scheduling in parallel/distributed systems for coarse grained irregular problems such as sparse matrix factorization is challenging since it requires efficient run-time support to execute it. In the literature there are importants contributions about parallelization for this kind of problems. However, the potential performance gain by the parallelism can be seriously reduced by long data acces latencies. Because the large matrix of this problems are stored in disk, it is important to take advantage of data locality in programs to minimises memory page fault. In this report we present a study of the behaviour of the compile-time reestructuring techniques for this type of codes. It can be shown that it results in performance improvements when they are applied before the run-time parallelization step. This results are the basis for an ongoing research on automatic blocking at run-time. Keywords data locality, loop transformations, sparse matrix, cholesky factorization. 2 1 In..