Abstract

In this presentation we review software issues that are critical to the successful integration of parallel computers into mainstream scientific computing. Clearly, on most systems, a compiler is the most important

software tool available to a user. We discuss compilers from the point of view of communication compilation—their ability to generate efficient communication code automatically. We illustrate with two examples of distributed-memory computers, on which almost all communication is handled by the compiler rather than by explicit calls to communication libraries.

Closely related to compilation is the need for high-quality debuggers. While single-node debuggers are important, parallel machines have their own specialized debugging needs related to the complexity of interprocess communication and synchronization. We describe a powerful simulation tool we have developed for such systems, which has proved essential in porting large applications to distributed memory systems.

Other important software tools include high-level languages, libraries, and visualization software. We discuss various aspects of these systems briefly. Ultimately, however, general-purpose supercomputing environments are likely to include more than a single computer system. Parallel computers are often highly specialized and rarely provide all of the facilities required by a complete application. Over the coming decade we will see the development of heterogeneous environments that will connect diverse supercomputers (scalar, vector, and parallel), along with high-end graphics, disk farms, and networking hubs. The real user-interface challenge will then be to provide a unified picture of such systems to potential users.