Towards energy-aware VM scheduling in IaaS clouds through empirical studies

Energy-efficient computing has become increasingly important to modern HPC systems such as clouds. In this thesis we explore the ’green’ opportunities with virtualization technologies in clouds through systemlevel optimizations, and specifically focus on energy-savings by energyaware scheduling of virtual machines. A system-level approach of optimization for green cloud computing requires in-depth understanding of the power characteristics of virtual machines with respect to patterns of workloads running on them. The first step we took in this direction is to deploy a private cloud system with facilities provided by DAS-4 clusters and thoroughly characterize its power behavior. We broadly identified three power metrics, i.e. power, power efficiency and energy. by executing several types of high performance computing workloads on both the VM and host, we compare their performance with respect to these three metrics. In addition to that, we also analyze the composition of the total power consumption of a single work node, and evaluate the contributions of individual components, i.e., CPU, memory and HDD. As a result of these profiling experiments, a linear power model is derived to represent the power behavior of a single work node, which is further extended to describe an entire cloud. With the help of this model, an novel energy-savvy scheduler is proposed to make use of the monitoring system to support system-level optimization with on-the-fly VM scheduling and dynamic resource adaptation.