Reservation-Based Overbooking for HPC Clusters

Abstract

HPC environments are not only used in research and academia anymore, but are also becoming commercially available and successful. HPC resource providers, which offer HPC services over the Internet, have to ensure a very high utilization rate to be competitive and profitable. One solution to improve utilization is the use of overbooking. This paper shows an improved overbooking approach for HPC providers that serves this purpose. Resources are not assigned to a job until it actually starts. This enhances the scheduler's degree of freedom and therefore improves the overbooking performance. We evaluated the potential of this approach using real-world job traces. Given a sufficient expected demand, overbooking is applicable and provides additional profit.