Potential of Applying kNN with Soft Walltime to Improve Scheduling Performance

Abstract

Utilizing performance is one of the most important and difficult tasks on High-Performance Computing (HPC) systems. In order to balance between simplicity and efficiency, many HPC systems are exploiting First-Come-First-Served schedulers with backfilling policies. Previous researches demonstrated the benefits of walltime approximation approaches on backfilling improvement. However, applying predicted values might affect scheduling guarantees due to the risk of early job termination caused by underestimation. The recent soft walltime feature from OpenPBS has eliminated this concern by restricting system-generated walltime only for the scheduling step.In this work, we present our simple k-nearest neighbors approach to improve the performance of the conservative backfilling algorithm. The inaccurate user estimate of a job is refined using the historic data about its most similar jobs. Then, we exploit this correction safely with the soft walltime scheme and perform simulations on real scheduling logs. Our simulation results highlight that even our simple setup can help to correct user estimates significantly. Moreover, the scheduling performance is also improved on some appropriate conditions.