Improving System Utilization on Wireless HPC Systems with Torus Interconnects

Recent advances in wireless technologies promote the concept of wireless HPC systems, in which the wired cable interconnect in a conventional HPC system is augmented with wireless links to conFigure the system for desired interconnection topology. In this paper, we study the problem of improving system utilization and job scheduling efficiency on HPC systems with torus interconnects. First, we design and analyze a novel wireless HPC system to improve system utilization, avoid inter-job communication interference, and achieve efficient job scheduling. Second, we present a topology-aware job scheduling algorithm, which is applicable to both of our proposed wireless HPC system and a conventional 3D torus HPC system. Third, the simulation results validate the efficiency of our proposed wireless HPC system in improving system utilization. At last, we apply machine learning to quantify the impact of job arrival rate and the number of backfilling jobs on the system utilization of HPC systems with torus interconnects.