Power Modeling in Modern Server Systems: An Examination of Various Novel Approaches

As Cloud and other distributed networks become more complex and new virtualization technologies are increasingly adopted, there is an increasing need for versatile solutions that can effectively measure the energy consumption incurred by users of such networks and use this information to schedule user jobs with the goal of maximizing energy efficiency. In this thesis, existing work in power modeling involving performance monitoring events and the Running Average Power Limit was reviewed. While substantial work has been done in these areas, the focus has been on a narrow range of use cases. To address the lack of versatility of existing power meters, three lightweight software-based virtual power meters were developed. Utilizing simple but powerful application-agnostic power models, they offer similar performances to existing power models but have minimal overheads and are portable for use in a variety of systems. The first two power meters use performance event counters and the RAPL feature respectively to estimate power use. The third power meter is an enhanced power meter incorporating features from these two earlier power meters. Additionally, the enhanced power meter can estimate power use on a per-application basis. Furthermore, a systematic framework PowerSave was developed to analyze and evaluate energy use and performance patterns for a variety of workloads for different types of server systems under different CPU power limits. Results from PowerSave and power modeling experiments show that CPU power use accounts for most of a server's power consumption and that there is a high level of correlation between CPU power use and server power use