DRAS-OD: A Reinforcement Learning based Job Scheduler for On-Demand Job Scheduling in High-Performance Computing Systems

Because of the expanding system resource requests, on-demand jobs have recently been executed together with rigid jobs in high-performance computing (HPC) systems. An on-demand job usually includes a notice, which refers to the lead time or advance warning provided to the scheduling system before it needs to be executed. To make the on-demand jobs start immediately, requested system resources are usually collected according to notices before their arrival. Consequently, rigid jobs scheduled in the same HPC system will be affected because of the collected system resources of the on-demand jobs. Thus, it is important to utilize the notices of the on-demand jobs to increase the scheduling performance of the rigid and on-demand jobs in the HPC systems. In this paper, reinforcement learning-based job schedulers are built to schedule the on-demand and rigid jobs in HPC systems. The reinforcement learning-based schedulers are trained to fully utilize the idle time between notice and arrival of the on-demand job. According to the evaluation, the proposed scheduler makes more on-demand jobs starting immediately or more quickly than conventional schedulers. In the meantime, the average wait time for jobs in the HPC system has improved compared to conventional schedulers. Additionally, as the requested system resources for on-demand jobs increase, the overall scheduling performance for both on-demand and rigid jobs improves compared to the conventional schedulers.