Workload Adaptive Power Management with Live Phase Monitoring and Prediction

In current computer systems, power dissipation is widely recognized as one of the primary critical constraints. Improving the power efficiency of current and emerging systems has therefore become a pressing challenge and an active research area over recent years. Dynamic, on-the-fly management techniques aim to address this challenge by adaptively responding to the changes in application execution. These application patterns, commonly referred to as “phases”, expose distinct, dynamically-varying and often repetitive characteristics of workloads. Dynamic management techniques, guided by workload phase information, can effectively tune system resources to varying workload demands for improved power-efficiency. This thesis researches new methods to characterize and predict application behavior for a dynamic power management endgoal. Specifically, this work has two major thrusts. First, it explores different approaches to characterize and predict dynamically varying workload power behavior. Second, it discusses runtime management techniques for real systems that can proactively adapt processor execution to varying application execution characteristics. This work develops a runtime, real-system power model that provides processor power consumption details in terms of the component powers of different architectural units. We show that similarity analysis methods applied to these component powers help expose power phase behavior of applications. A small set of “power signatures” can represent overall workload power characteristics within 5% of the actual behavior. We develop a “transition-guided” phase detection framework that can identify repetitive application phase patterns despite system-induced variability effects. This detection strategy can identify recurrent phase signatures with less than 5% false alarms on running systems. Last, we propose a workload-adaptive dynamic power management framework guided by runtime phase predictions. This predictive power management approach is shown to improve the energy-delay product of a deployed platform by 7% when compared to existing reactive techniques and by 27% over the baseline unmanaged system.