Reducing Energy Costs for IBM Blue Gene/P via Power-Aware Job Scheduling

Energy expense is becoming increasingly dominant in the operating costs of high-performance computing (HPC) systems. At the same time, electricity prices vary significantly at different times of the day. Furthermore, job power profiles also differ greatly, especially on HPC systems. In this paper, we propose a smart, power-aware job scheduling approach for HPC systems based on variable energy prices and job power profiles. In particular, we propose a 0-1 knapsack model and demonstrate its flexibility and effectiveness for scheduling jobs, with the goal of reducing energy cost and not degrading system utilization. We design scheduling strategies for Blue Gene/P, a typical partition-based system. Experiments with both synthetic data and real job traces from production systems show that our power-aware job scheduling approach can reduce the energy cost significantly, up to 25%, with only slight impact on system utilization.