Dynamic thermal and power control for computing systems

Thermal and power management have become increasingly important for both computing and physical systems. Computing systems from real-time embedded systems to data centers require effective thermal and power management to prevent overheating and save energy. In the mean time, as a major consumer of energy buildings face challenges to reduce the energy consumption for air conditioning while maintaining comfort of occupants. In this dissertation we investigate dynamic thermal and power management for computer systems and buildings. (1) We present thermal control under utilization bound (TCUB), a novel control-theoretic thermal management algorithm designed for single core real-time embedded systems. A salient feature of TCUB is to maintain both desired processor temperature and real-time performance. (2) To address unique challenges posed by multicore processors, we develop the real-time multicore thermal control (RT-MTC) algorithm. RT-MTC employs a feedback control loop to enforce the desired temperature and CPU utilization of the multicore platform via dynamic frequency and voltage scaling. (3) We research dynamic thermal management for real-time services running on server clusters. We develop the control-theoretic thermal balancing (CTB) to dynamically balance temperature of servers via distributing clients' service requests to servers. Next, (4) we propose CloudPowerCap, a power cap management system for virtualized cloud computing infrastructure. The novelty of CloudPowerCap lies in an integrated approach to coordinate power budget management and resource management in a cloud computing environment. Finally we expand our research to physical environment by exploring several fundamental problems of thermal and power management on buildings. We analyze spatial and temporal data acquired from an real-world auditorium instrumented by a multi-modal sensor network. We propose a data mining technique to determine the appropriate number and location of temperature sensors for estimating the spatiotemporal temperature distribution of the auditorium. Furthermore, we explore the potential energy savings that can be achieved through occupancy-based HVAC scheduling based on real occupancy data of the auditorium.