Scheduling of Parallel Jobs in a Heterogeneous Multi-Site Environment

Most previous research on job scheduling for heterogeneous systems considers a scenario where each job or task is mapped to a single processor. On the other hand, research on parallel job scheduling has concentrated primarily on the homogeneous context. In this paper, we address the scheduling of parallel jobs in a heterogeneous multi-site environment, where each site has a homogeneous cluster of processors, but processors at different sites have different speeds. Starting with a simple greedy scheduling strategy, we propose and evaluate several enhancements using trace driven simulations. We consider the use of multiple simultaneous reservations at different sites, use of relative job efficacy as a queuing priority, and compare the use of conservative versus aggressive backfilling. Unlike the single-site case, conservative backfilling is found to be consistently superior to aggressive backfilling for the heterogeneous multisite environment. ∗ Supported in part by NSF grant EIA-9986052