Feedback-Directed Management of Performance and Power for Emerging Computer Systems
Xing Fu · 2016
Emerging computing systems face the critical management challenge of both performance and power simultaneously. For example, distributed real-time embedded systems such as cyber-physical systems need to reduce power consumption while enforces CPU utilization bounds on multiple uniprocessors in order to meet end-toend deadlines. Data centers operators attempt to oversubscribe data center power delivery networks to reduce throughput penalty during a power overload. This dissertation presents latest development of a control framework which adopts novel control theoretic approaches to address emerging computing systems including multi-core real-time embedded systems, distributed real-time embedded systems, and data centers. The dissertation leverages task migration and cache partitioning mechanisms in multi-core systems to reduce power and energy consumption while control CPU utilizations. We then present a control algorithm for simultaneous temperature and utilization control for distributed real-time embedded systems. Algorithms and optimizations are presented to extend state-of-the-art model predictive control technique to overcome technical challenges such as scalability. We also adapt the parameter of a power controller widely adopted by IBM servers to improve its performance significantly. A power capping algorithm for an entire data center by shifting power between data center cooling systems and IT equipments is proposed for improving performance. Finally, we present a hierarchical heuristic to minimize energy consumption of Virtual Desktop Infrastructure without violating its performance constraints. Both theoretic analysis, hardware, and simulation