Study of workload-driven dynamic power management for high performance computing clusters

A workload-driven dynamic power management (WDPM) strategy is proposed to improve the tradeoff between power consumption and performance in homogeneous computing clusters. Through the empirical analysis of real workload logs in production systems, the WDPM strategy integrates the load prediction-based pre-wakeup approach and the dynamic feedback-based revising mechanism based on the improvement of the timeout strategy. By using the data from real workload logs and practical systems, the extensive simulations were conducted to investigate the performance and energy consumption of the strategy. The experimental results indicate that, as compared with the timeout strategy, the WDPM strategy can effectively reduce the average wait time of jobs and the node state switching times of a system with a very little increase of power consumption. Therefore, it can alleviate the performance loss and achieve the better tradeoff of performance and power consumption.