HPC operation with time-dependent cluster-wide power capping

HPC systems have increased in size and power consumption. This has lead to a shift from a pure performance centric standpoint to power and energy aware scheduling and management considerations for HPC. This trend was further accelerated by rising energy prices and the energy crisis that began in 2022. Digital Twins have become valuable tools that enable energy and power aware scheduling of HPC clusters. This paper uses an existing Digital Twin and extends it with a node energy model that allows the prediction of the cluster power consumption. The Digital Twin is then used to simulate system-wide power capping for different energy shortages functions of varying degree. Different policies are proposed and tested towards their effectiveness in improving the job wait times and overall throughput under limiting conditions. Based on a real world HPC cluster, these policies are implemented. Depending on the pattern of the energy limitation and workload, improvements of up to 40 percent are possible compared to scheduling without policies for these conditions.

---

Thematic Sessions: Regular Papers