Class-Partitioning Job Scheduling for Large-Scale Parallel Systems

This paper addresses the problem of poor response time for large parallel jobs under nonpreemptive backfill scheduling policies. Four monthly workloads from the large-scale NCSA O2K system are used to evaluate new scheduling policies. Our key result is that partitioning the system based on job class, where class is defined by the number of requested processors, significantly improves overall performance and the performance of large jobs, while still providing similar performance for smaller jobs, compared to the best previous priority backfill policies. Another key advantage of the class-partitioning policies is that selecting highperformance policy parameters is simpler and more intuitive than determining the number of reservations and other parameters of the priority backfill policies.