The Application of Multiprocessors to Scientific Problems: The Potential and the Pitfalls

The objective of research on the MIDAS project is to demonstrate the viability of a multiprocessor approach to computing and to develop a general purpose, extensible architecture which can be used to address the growing computational requirements of the scientific community. To be successful in this endeavor, however, requires more than simply designing, or even constructing, new hardware structures. To achieve high performance on future systems will require software approaches which can exploit parallel architectures as fully as possible. A critical issue, therefore, is to understand the functional requirements of a large class of applications. The requirernents must be critically examined and, in many cases, new approaches to old algorithms investigated. For highly parallel systems to be effectively utilized, they must be flexible and adaptable to specific application requirements. In particular, they will probably need a variety of control and communication mechanisms to facilitate load balancing of a problem and to minimize bottlenecks in performance. The effective utilization of highly parallel architectures, however, raises many new questions. Traditional operating system structures must, for example, be re-examined. Fault-tolerant environments, with error recovery capability, become increasingly important as the number of components and processors increases. Language extensions to support parallel operations are obviously required, and ultimately new languages are needed to explicitly exploit parallel constructs. A new generation of development and debugging tools will be necessary in order to examine programn performance and operation in an asynchronous parallel environment.