In recent years, collections of workstations interconnected by high- speed communication devices have come into widespread use. Though such collections often have considerable computing capacity, much of it is generally inaccessible, because only primitive means currently exist for offloading work to idle or lightly-utilized workstations when back- logs of work develop. This failure to take advantage of unused computing capacity reduces performance to a fraction of what it might be. The objective of the project is to tap the latent capacity of a network of workstations by designing and implementing distributed scheduling algorithms that are suited to the unique features of this environment. The approach, which combines measurement, analytic modeling, simulation and implementation, includes several advances over previous approaches. First, the availability of computing capacity is increased by exploiting capacity at under-utilized, as well as idle, workstations. Second, the robustness of distributed scheduling is improved by adapting the strategy pursued to changes in the system environment. Third, robustness is further promoted through the use of fully decentralized distributed scheduling algorithms. Finally, utilization of latent capacity is increased by shifting the responsibility for determining which processes can be advantageously transferred to other workstations from the user to the distributed scheduler.
