Download files
Abstract
Energy is a critical resource in all types of computing systems from servers, where energy costs dominate data centre expenses and carbon footprints, to embedded systems, where the system's battery life limits the device's functionality. In their efforts to reduce the energy use of these system's hardware manufacturers have implemented features which allow a reduced energy consumption under software control.
This thesis shows that managing these settings is a more complex problem than previously considered. Where much (but not all) of the previous academic research investigates unrealistic scenarios, this thesis presents a solution to managing the power on varying hardware.
Instead of making unrealistic assumptions, we extract a model from empirical data and characterise that model. Our models estimate the effect of different power management settings on the behaviour of the hardware platform, taking into account the workload, platform and environmental characteristics, but without any kind of a-priori knowledge of the specific workloads being run. These models encapsulate a system's knowledge of the platform.
We also developed a \emph{generalised energy-delay} policy which allows us to quickly express the instantaneous importance of both performance and energy to the system. It allows us to select a power management strategy from a number of options.
This thesis shows, by evaluation on a number of platforms, that our implementation, Koala, can accurately meet energy and performance goals. In some cases, our system saves 26\% of the system-level energy required for a task, while losing only 1\% performance. This is nearly 46\% of the dynamic energy.
Taking advantage of all energy-saving opportunities requires detailed platform, workload and environmental information. Given this knowledge, we reach the exciting conclusion that near optimal power management is possible on real operating systems, with real platforms and real workloads.