Improving Job Scheduling on Production Supercomputers

Job scheduling is a critical task on large-scale supercomputers, where small variety in scheduling policies can result in substantial differences in performance or resource utilization. Tremendous research has been focused on improving job scheduling theoretically. This work aims at addressing the job scheduling problem from practice. Driven by the practical motivating problems, we design and implement job scheduling schemes which can be easily deployed on production machines. All the schemes are evaluated by event-driven simulations using real workload from the production Blue Gene/P system at Argonne National Laboratory. Experimental results show our schemes can effectively improve job scheduling in terms of user satisfaction and system utilization.