Scheduling Grid workloads on multicore clusters to minimize energy and maximize performance

Energy is a significant and growing component of the cost of running a large computing facility. A grid workload consisting of millions of jobs running on thousands of processors may consume millions of kilowatt hours of electricity. However, because a grid workload generally consists of many independent sequential processes, we may shape its execution to satisfy energy constraints. By varying the number and frequency of processors available, a scheduler may trade off energy against performance. In this paper, we explore energy and performance tradeoffs in the scheduling of grid workloads on large clusters. We build upon previous work by showing the interaction of intelligent job assignment, automated node scaling, and frequency scaling on multicore clusters. An unexpected result is that, even though low frequency is the most efficient mode of operating a single node, the careful application of frequency scaling can actually reduce overall energy consumption even further by reducing the number of nodes powered on.