Large-scale distributed programming, where processes communicate via messages, is a reality whereas large-scale shared variable programming remains an ill-understood activity. It is postulated that the distinction between these two classes is the type of communication-it is disciplined in the former case and arbitrary in the latter case. Furthermore, implementations of shared-variable systems on multiprocessors with individual caches pose several problems, including cache-coherence. This work attempts to impose certain disciplines on shared variable programming so that large-scale system design becomes feasible and the implementations avoid most of the issues associated with cache coherence. The plan of research is to identify a substantial class of shared-variable systems that enjoy some of the advantages of distributed (i.e., message-passing) systems. The systems will be constrained to a class called loosely-coupled, in which the computations of different processes are largely non-interfering even though they share variables. Loosely-coupled systems will acquire certain properties-such as absence of "race conditions" -normally associated with distributed systems. Designs of loosely-coupled systems will be accomplished without extensive analysis of scenarios based on execution sequences, process interactions or deadlock. Compositional proof theories for such systems will be developed and the efficient implementation of such a system on multiple-memory multiprocessor architectures will be indicated.
