Data prefetching for linear algebra operations on high performance workstations

In a previous work it was shown that the performance of linear algebra computations , which access large amounts of data, is dependent on the behavior of the memory hierarchy. This research is aimed to use the multilevel orthogonal blocking approach in conjuntion with other software techniques to further improve the performance of linear algebra computations. The performance of the dense matrix by matrix multiplication executed on a super-scalar high performance workstation is improved using binding and nonbinding prefetching to hide the memory latency together with the well known technique of blocking. In this report our main goal is to improve the performance of the matrix by matrix multiplication on a high performance workstation. This widely applied operation has already been improved by means of techniques such as blocking and software pipelining. In this paper we show the improvements yielded by applying software prefetch in addition to those techniques aforementioned. Prefetching is a technique used to hide memory latency by avoiding the processor to stall because of data dependencies when a miss occurs. We rst present an algorithm where only blocking and software pipelining are used, followed by two new algorithms in which binding and nonbinding prefetching respectively have been added. Performance results start with 15 MFlops for jki form. Blocking and software pipelin-ing techniques get a performance of 128 MFlops for the best algorithm without prefetching, and nally, it is improved up to 166 MFlops when prefetching is applied in a computer with 200 MFlops of peak performance.