Parallel Haskell ( Preliminary Draft )

MPP Haskell is a parallel implementation of the Haskell functional language for the Thinking Machines Inc. CM-5 large-scale distributed-memory multiprocessor. MPP Haskell is a derivative of GUM, a message-based parallel implementation of Haskell. GUM was carefully designed to minimise performance loss from low-bandwidth and high-latency communications; as an apparent consequence MPP Haskell, with the beneet of high-bandwidth and low-latency communications, achieves remarkably scalable performance. 1 Background Functional programming languages have long been touted as ideal for parallel implementation. Despite this fact, and the existence of numerous functional language implementations, very few parallel implementations have become generally available; further work in this direction seemed worthwhile. 1.1 Choice of Language Assuming the goal to be a parallel, non-strict, purely functional language implementation , and given sharply limited resources (man-hours), a practical course would be the port of an existing language implementation to a new platform. Two existing implementations were seriously considered, both high-quality, reasonably mature, and actively supported. The rst was Concurrent Clean from the University of Nigmegan 3]. One of its attractions was that it was already being ported to quite diverse hardware and software (OS) platforms, suggesting a careful design to make such ports relatively easy. While the Concurrent Clean team were agreeable to another port, their estimate of the work involved considerably exceeded the available resources. The other candidate was GUM, described by Trinder et al 5] as a freely available, portable, parallel implementation of the non-strict purely-functional language Haskell. Its use of PVM, a widely ported communications API and distributed computing infrastructure, for communication and task coordination was deliberately chosen to lessen the task of porting. GUM appeared to be a promising starting point.