Scheduling Parallel Computations in Clusters of Workstations

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

Dramatic advances in the technologies, coupled with economic factors, have given rise to a new modality of parallel computing wherein workstations are connected together to act cooperatively as a single parallel computer - a Network of Workstations (NOW). Many algorithmic devices that ensured efficient interprocessor communication and coordination in a tightly coupled parallel computing environment no longer guarantee efficiency within a NOW. This project derives theoretical understanding of factors that enhance - and detract from - efficiency of parallel computing in NOWs. It also develops algorithmic techniques that translate such understanding into practically efficient scheduling strategies. The scheduling algorithms that emerge from research will provide provably predictable performance for large, significant classes of computations, such as those that arise in many scientific applications.