Automatic exploitation of dual level parallelism on a network of multiprocessors

Abstract

We describe a software environment for high performance distributed computing on a network of multiprocessor workstations. In designing this environment, we have used a problem oriented approach as opposed to the traditional algorithm oriented approach. This paradigm shift enables us to generate efficient programs automatically for a well defined class of problems. Thus, our system frees the users from the esoteric tasks of algorithm design and implementation. An important feature of our system is its ability to handle the large variation in granularity-we call this dual level parallelism-in a hybrid processing environment. This feature is the key to the superior efficiency delivered by the system. We give preliminary results from a case study in which our system is used to generate programs automatically for a scientific application, with a network of multiprocessors as the target platform.