Engineering large parallel functional programs

Abstract

The design and implementation of useful programming languages, whether sequential or parallel, should be driven by large, realistic applications. In constructing several medium- and large-scale programs in Glasgow Parallel Haskell, GpH, a parallel extension of Haskell, the group at Glasgow has investigated several important engineering issues: Real Application Parallelism. The programs achieve good wall-clock speedups and acceptable scale-up on both a shared-memory and a distributed memory machine. The programs typify a number of application areas and use a number of different parallel paradigms, e.g. pipelining or divide- and-conquer, often combining several paradigms in a single program. Language Issues. Although the largely implicit parallelism in GpH is a satisfactory programming model in general the base constructs for introducing and controlling parallelism tend to obfuscate the semantics of large programs. As a result we developed evaluation strategies, a more abstract, and systematic mechanism for introducing and controlling parallelism. Engineering Environment. The development and performance tuning of these programs emphasised the importance of an integrated engineering environment. In the process we have refined components of this environment like the simulator, the runtime system, and the profiling tools. KeywordsHomomorphic ImageFunctional ProgramPipeline StageRuntime SystemFunctional LanguageThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.