LOCALITY OPTIMIZATIONS FOR PARALLEL COMPUTING USING DATA ACCESS INFORMATION

Given the large communication overheads characteristic of modern paral lel machines, optimizations that improve locality by executing tasks c lose to data that they will access may improve the performance of para llel computations. This paper describes our experience automatically a pplying locality optimizations in the context of Jade, a portable, imp licitly parallel programming language designed for exploiting task-lev el concurrency. Jade programmers start with a program written in a sta ndard serial, imperative language, then use Jade constructs to declare how parts of the program access data. The Jade implementation uses th is data access information to automatically extract the concurrency an d apply locality optimizations. We present performance results for sev eral Jade applications running on the Stanford DASH machine. We use th ese results to characterize the overall performance impact of the loca lity optimizations. In our application set the locality optimization l evel has Little effect on the performance of two of the applications a nd a large effect on the performance of the rest of the applications. We also found that, if the locality optimization level had a significa nt effect on the performance, the maximum performance was obtained whe n the programmer explicitly placed Basks on processors rather than rel ying on the scheduling algorithm inside the Jade implementation.