Abstract

representations ofparallel programs simplify pi-ogra mming and debugging by providing a single, consistent fra mework that separates a program’s sequential computations from its parallel structure. arallel architectures have clearly emerged as the future environments for high-performance computation for most applications. The barrier to their widespread use is that writing parallel programs that are both efficient and portable is quite difficult. Parallel programming is more difficult than sequential programming because parallel programs must express not only the sequential computations, but also the interactions (communication and synchronization) among those computations that define the parallelism. To achieve good performance, programmers must understand this large-scale structure. Most current text-based parallel-programming languages either implicitly define parallel structure, thus requiring advanced compilation techniques, or embed communication and synchronization with sequential computation, thus making program structure difficult to understand. Furthermore, direct use of vendor-supplied procedural primitives can preclude portability. We could alleviate these difficulties with a programming process that separates specification of sequential computations from specification of synchronization and communication, and expresses synchronization and communication directly but abstractly. Directed-graph program representations can separate these specifications by permitting a two-step programming process in which programmers first design sequential components and then compose them into a parallel structure. This provides a simplified divide-and-conquer approach to design. Such program representations have other advantages. They directly represent multiple threads of control. They also display and expose the large-scale program structure that parallel programmers must understand

Citations