Feasibility analysis of using the maui scheduler for job simulation of large-scale pbs based clusters

Abstract

For large-scale High Performance Computing centers with a wide range of different projects and heterogeneous infrastructures, efficiency is an important consideration. Understanding how compute jobs are scheduled is necessary for improving the job scheduling strategies in order to optimize cluster utilization and job wait times. This increases the importance of a reliable simulation capability, which in turn requires accuracy and comparability with historic workloads from the cluster. Not all job schedulers have a simulation capability, including the Portable Batch System (PBS) resource manager. Hence, PBS based centers have no direct way to simulate changes and optimizations before they are applied to the production system. We propose and discuss how to run job simulations for large-scale PBS based clusters with the Maui Scheduler. This also includes awareness of node downtimes, scheduled and unexpected. For validation purposes, we use historic workloads collected at the IT4lnnovations supercomputing center. The viability of our approach is demonstrated by measuring the accuracy of the simulation results compared to the real workloads. In addition, we discuss how the change of the simulator's time step resolution affects the accuracy as well as simulation times. We are confident that our approach is also transferable to enable job simulations for other computing centers using PBS.