On Multiprocessor Scheduling

Large-scale multiprocessor systems offer the potential of cost-effective high-performance computing, but several key technologies—both hardware and software—are necessary before this potential can be realized. This paper examines the state of the art in one such software component: multiprocessor scheduling. After surveying several important papers and recent work in the field, I argue for a move away from traditional, centralized schemes limited to the scheduling of loop-derived parallelism. In order to gain insight into the complexity required for good scheduler performance in large systems, the paper analyzes the behavior of a particularly simple distributed scheduling algorithm based on work stealing. This analysis indicates that although this simple algorithm has some encouraging properties, it also has one characteristic that may prove to fundamentally limit its scalability. By clearly identifying this characteristic, this paper provides insight into the kinds of mechanisms required for scheduling algorithms to perform well on large-scale systems.