Abstract
Parallel processing is facing a software crisis. The primary reasons for this crisis are the short life span and small installation base of parallel architectures. In this article, we propose a solution to this problem in the form of an architecture-adaptable programming environment. Our method is different from high-level procedural programming languages in two ways: (1) our system automatically selects the appropriate parallel algorithm to solve the given problem efficiently on the specified architecture; (2) by using a divide-and-conquer template as the basic mechanism for achieving parallelism, we considerably simplify the implementation of the system on a new platform. There is a trade-off, however: the loss of generality. From a pragmatic point of view, this is not a major liability since our strategy will be useful in building domain-specific problem solving environments and application-oriented compilers, which can be easily and effectively ported to diverse architectures. We give preliminary results from a case study in which our method is used to adapt the parallel implementations of the conjugate gradient algorithm on a multiprocessor, a multicomputer, and a workstation network.