Efficient parallel programming with algorithmic skeletons

Abstract

Algorithmic skeletons are polymorphic higher-order functions representing common parallelization patterns and implemented in parallel. They can be used as the building blocks of parallel and distributed applications by integrating them into a sequential language. In this paper, we present a new approach to programming with skeletons. We integrate the skeletons into an imperative host language enhanced with higher-order functions and currying, as well as with a polymorphic type system. We thus obtain a high-level programming language which can be implemented very efficiently. After describing a series of skeletons which work with distributed arrays, we give two examples of parallel algorithms implemented in our language, namely matrix multiplication and a statistical numerical algorithm for solving partial differential equations. Rml-time measurements show that we approach the efficiency of message-passing C up to a factor between 1 and 1.75.