Implementing Context-Driven Parallel Computations

Abstract

This paper proposes a new abstract model for parallel computations called the context model, describes a novel implementation mechanism called the context-driven call, and investigates the performance and efficiency of an implementation of the model on the Helios operating system. The context model assumes that the execution order of a parallel/distributed program is determined by the location of data objects in the program. It offers two simple abstractions for describing the location of a data object, remote <somewhere> and shared, rather than using special parallel control structures normally provided by other models. Several sample applications written in an extended C language called conteXt C are given to illustrate the applicability and flexibility of using the context model. Special attention is paid to the performance and computational efficiency achieved by the implementation.