Communication Optimizations for Parallel Computing Using Data Access Information

Given the large communication overheads characteristic of modern parallel machines, optimizations that eliminate, hide or parallelize communication may improve the performance of parallel computations. This paper describes our experience automatically applying communication optimizations in the context of Jade, a portable, implicitly parallel programming language designed for exploiting task-level concurrency. Jade programmers start with a program written in a standard serial, imperative language, then use Jade constructs to declare how parts of the program access data. The Jade implementation uses this data access information to automatically extract the concurrency and apply communication optimizations. Jade implementations exist for both shared memory and message passing machines; each Jade implementation applies communication optimizations appropriate for the machine on which it runs. We present performance results for several Jade applications running on both a shared memory machine (the Stanford DASH machine) and a message passing machine (the Intel iPSC/860). We use these results to characterize the overall performance impact of the communication optimizations. For our application set replicating data for concurrent read access and improving the locality of the computation by placing tasks close to the data that they access are the most important optimizations. Broadcasting widely accessed data has a significant performance impact on one application; other optimizations such as concurrently fetching remote data and overlapping computation with communication have no effect.