An Integrated Hardware / Software Solution for E ectiveManagement of Local Storage in High - Performance Systems ( Extended Version )

The potential of high-performance systems, especially vector and parallel machines, is generally limited by the bandwidth between processors and memory. To achieve the performance of which these machines should be capable, greater emphasis must be placed on optimizing array accesses. We propose a practical, integrated hardware/software strategy for increasing the eeectiveness of local storage management. Our scheme provides many of the advantages of both compile-time and run-time memory management techniques. In this paper, we describe our local storage facility: the priority data cache. We also describe compile-time techniques for easily and eeectively utilizing this level of local storage.