Demand-Aware Power Management for Power-Constrained HPC Systems

As limited power budget is becoming one of the most crucial challenges in developing supercomputer systems, hardware overprovisioning which installs larger number of nodes beyond the limitations of the power constraint determined by Thermal Design Power is an attractive way to design extreme-scale supercomputers. In this design, power consumption of each node should be controlled by power-knobs equipped in the hardware such as dynamic voltage and frequency scaling (DVFS) or power capping mechanisms. Traditionally, in supercomputer systems, schedulers determine when and where to allocate jobs. In overprovisioned systems, the schedulers also need to care about power allocation to each job. An easy way is to set a fixed power cap for each job so that the total power consumption is within the power constraint of the system. This fixed power capping does not necessarily provide good performance since the effective power usage of jobs changes throughout their execution. Moreover, because each job has its own performance requirement, fixed power cap may not work well for all the jobs. In this paper, we propose a demand-aware power management framework for overprovisioned and power-constrained high-performance computing (HPC) systems. The job scheduler selects a job to run based on available hardware and power resources. The power manager continuously monitors power usage, predicts performance of executing jobs and optimizes power cap of each CPU so that the required performance level of each job is satisfied while improving system throughput by making good use of available power budget. Experiments on a real HPC system and with simulation for a large scale system show that the power manager can successfully control power consumption of executing jobs while achieving 1.17x improvement in system throughput.