Maximizing Energy Budget Utilization Using Dynamic Power Cap Control

The operation cost of a High Performance Computing (HPC) system is mainly dominated by the electricity expense and is often comparable with the initial installation cost of the system. Since the electricity expense is proportional to the amount of power usage, i.e., energy consumption, future HPC systems will likely be operated under a policy so that the energy consumption does not exceed a predefined threshold. This threshold is called an energy budget. The goal of this paper is to improve the ratio of actual energy consumption to the allocated energy budget, called energy budget utilization. This paper proposes a power management method that dynamically adjusts the power cap of each job by considering the power consumption and energy budget on each node. Specifically, the power cap of each job is adjusted so that the surplus energy generated by other preceding jobs is consumed to increase the energy budget utilization while improving the execution performance by using more power. The performance evaluation is conducted by simulating a job scheduler with synthetic job traces as well as real-world job traces from the Parallel Workload Archive (PWA). For the synthetic job traces, the proposed method can improve the energy budget utilization by 17.5% on average, and also reduce the makespan by 7.6% on average. The evaluation with real-world job traces also derives the same conclusion as that with synthetic traces. The evaluation results with five PWA job sets suggest the effectiveness of the proposed method in practical use. Accordingly, the proposed method improves the utilization of the energy budget and thus the system performance.