Career: Resource Management for Parallel and Distributed Systems

A metasystem is an ensemble of workstations, parallel, and vector machines connected by one or more networks. Metasystems offer the promise of high performance both in reduced completion time and high job throughput. The goal of this project is to develop a set of automated scheduling tools that will produce reduced completion time and high job throughput in a wide-range of metasystem environments. This investigation will also explore fundamental issues in the relationship between completion time and job throughput. The approach taken will be based on the Prophet scheduling system developed by the PI at the University of Virginia. Prophet automatically schedules data parallel jobs across a network of heterogeneous workstations to reduce completion time. In this project, Prophet will be used to schedule data parallel, task parallel, and sequential jobs in a general metasystem that may contain multicomputers, multiprocessors, and ATM networks. Prophet will be applied to real workloads that include scientific applications to test the efficacy of the scheduling tools. For parallel jobs, the completion time obtained by Prophet will compared to the completion time obtained by the Mentat scheduler. The system throughput obtained by Prophet will be compared to the throughput obtained if resource sharing is not considered in making scheduling decisions. It is also expected that insights into scheduling policies will be gained by studying the interaction between throughput and completion time.