Distributed supercomputing

Several design issues for a distributed system capable of supporting high-performance scientific computing are discussed. Two systems prototypes, one based on an object-oriented implementation and a second supporting coscheduling, and their application to distributed partial differential equation (PDE) solvers are presented. The questions explored relate to the proper computation and communication paradigms and the architecture of such systems. The point of view is that in supercomputing the total execution time of an application is a critical performance measure of the system and scheduling policies need to reflect user's needs rather be concerned only with requirements pertinent solely to resource utilization.< >