Supercomputer Emulation For Evaluating Scheduling Algorithms

Abstract

Scheduling algorithms have a significant impact on the optimal utilization of HPC facilities, yet the vast majority of the research in this area is done using simulations. In working with simulations, a great deal of factors that affect a real scheduler, such as its scheduling processing time, communication latencies and the scheduler intrinsic implementation complexity are not considered. As a result, despite theoretical improvements reported in several articles, practically no new algorithms proposed have been implemented in real schedulers, with HPC facilities still using the basic first-come-first-served (FCFS) with Backfill policy scheduling algorithm. A better approach could be, therefore, the use of real schedulers in an emulation environment to evaluate new algorithms. This thesis investigates two related challenges in emulations: computational cost and faithfulness of the results to real scheduling environments. It finds that the sampling, shrinking and shuffling of a trace must be done carefully to keep the classical metrics invariant or linear variant in relation to size and times of the original workload. This is accomplished by the careful control of the submission period and the consideration of drifts in the submission period and trace duration. This methodology can help researchers to better evaluate their scheduling algorithms and help HPC administrators to optimize the parameters of production schedulers. In order to assess the proposed methodology, we evaluated both the FCFS with Backfill and Suspend/Resume scheduling algorithms. The results strongly suggest that Suspend/Resume leads to a better utilization of a supercomputer when high priorities are given to big jobs.