UC San Diego

Energy management has become a key issue in the design of computing systems in both mobile as well as enterprise space. For mobile systems it is important from the perspective of battery life, while for large scale systems it directly impacts operational costs, making dynamic energy management methods a critical system component. An effective and conventional mechanism to achieve runtime energy savings is to dynamically reduce the power consumption of the system by transitioning to sleep. This thesis introduces a novel online learning based meta- policy for dynamic power management, where based on characteristics of the device and workload, the system converges to the best suited power management policy for it. We show, that this approach provides superior adaptability across diverse workloads compared to state of the art single policy solutions achieving energy savings of up to 61%. The thesis next provides analysis of effectiveness of power management based approach for energy management on mobile and server class systems. The study reveals that while the effectiveness is high for mobile systems, it is on a decline for server systems due to their non energy proportional design. This motivates energy management through two alternative means - energy proportional design and workload consolidation. For energy proportional design, the thesis focuses on memory subsystem as an example and proposes a memory hierarchy using phase change memory (PRAM) and conventional DRAM. The proposed system exploits the characteristics of workloads (in terms of their read and write intensity) to intelligently place pages across both the memories, thus reducing overall energy consumption of the VM) scheduling and resource management algorithms. We show that fine grained information on how the VMs utilize share the system resources like the CPU pipeline, memory bandwidth etc. can be exploited to perform more intelligent VM consolidation and resource management to maximize the overall energy efficiency. Real-life implementation of such VM management policies out-perform state of the art techniques that primarily use just CPU utilization for VM management by 35% on average