Proactive Energy Management for Smart Building and Compute Server Architectures
The growing demand for improved operational performance, coupled with the increasing scarcity of energy resources, calls for new approaches to improving the energy efficiency of smart buildings and computer systems. Traditional energy management techniques have been either reactive or locally predictive at best. These schemes often underperform, by either failing to meet the desired performance target or consuming excess energy. Moreover, different applications and environments create a diverse set of challenges. Therefore, there is a dire need to develop new techniques that approach energy-performance optimality under stringent and diverse application and environmental conditions. In this dissertation, we propose proactive management techniques for Heating, Ventilation, and Air-Conditioning (HVAC) in smart buildings, and for dynamic power management of heterogeneous processors. We show how the lack of future visibility and adaptivity of traditional energy management techniques proposed in these two domains degrades energy and performance. We develop proactive techniques to improve energy efficiency in buildings and processors. We first focus on building energy management. We propose to automatically assign meetings to rooms to reduce overall energy consumption. We derive an HVAC energy model for meeting assignment by characterizing building energy behavior. Using this energy model, we propose several assignment algorithms, and analyze their optimality and scalability. We also characterize how different factors impact energy savings, when it is worthwhile to use complex assignment algorithms, and when simpler methods suffice. We further extend this model to include weather factors, and develop a methodology for assigning meetings to the most appropriate room given the expected weather. We then propose to apply Model Predictive Control (MPC) to dynamically balance HVAC energy consumption and occupant comfort. Traditional energy management techniques ignore past discomfort behavior and therefore poorly trade off energy for comfort. Our MPC framework uses a probabilistic model to predict the upcoming occupancy and discomfort history, in order to adaptively balance energy consumption and occupant comfort. Our approach achieves high energy efficiency while operating within a specified discomfort target. For heterogeneous processors, we propose MPC-based dynamic General Purpose Graphics Processing Unit (GPGPU) power management. Traditional schemes ignore future kernel behavior, and may degrade performance and energy efficiency due to their inability to plan for the performance and energy characteristics of future phases. Our approach proactively configures hardware states based on recent execution history, the pattern of upcoming kernels, and the predicted behavior of those kernels, and adaptively varies its future visibility in order to achieve high energy savings with negligible performance impact and overhead. We extend this framework for workloads that use the CPU and the GPU concurrently. Our MPC approach simultaneously optimizes the CPU and GPU across adaptively-managed time windows consisting of multiple phases of CPU and GPU applications. We explore several alternatives that trade off future visibility for computational overhead, and demonstrate significant energy savings over current state-of-the-art approaches.
Computer engineering; Energy/Power Performance Management; Algorithms; Heterogeneous Processors; Smart Buildings; Computer science; machine learning; Mechanical engineering; Optimal control
Albonesi, David H.
Hencey, Brandon M.; Selman, Bart; Zhang, Zhiru; Paul, Indrani
Electrical and Computer Engineering
Ph. D., Electrical and Computer Engineering
Doctor of Philosophy
dissertation or thesis