Power-aware Resource Allocation via Online Simulation with Multiple-queue Backfilling

Although traditional scheduling policies for high-end parallel systems focus on minimizing average job wait time while maximizing system utilization, actual supercomputer workload traces confirm the existence of significant periods of time of low utilization. Previous work has shown that, in the context of backfilling schedulers, portions of such high-end systems can be selectively powered down to reduce power consumption and cooling costs while maintaining reasonable job performance. This is accomplished by using online simulations of a model of the actual system to automate scheduler parameter adjustment. Here we augment our power-aware scheduling policy by incorporating multiple queues into the backfilling mechanism, which provides dramatic improvement in job performance while retaining significant power savings.