|dc.description.abstract||Power consumption has long been a concern for portable consumer electronics, but has recently become an increasing concern for larger, power-hungry systems such as servers and clusters. This concern has arisen from the associated financial cost and environmental impact, where the cost of powering and cooling a large-scale system deployment can be on the order of millions of dollars a year. Such a substantial power consumption additionally contributes significant levels of greenhouse gas emissions.
Therefore, software-based power management policies have been used to more effectively manage a system’s power consumption. However, managing power consumption requires fine-grained power values for evaluating the run-time tradeoff between power and performance. Despite hardware power meters providing a convenient and accurate source of power values, they are incapable of providing the fine-grained, per-application power measurements required in power management.
To meet this challenge, this thesis proposes a novel power modelling method called W-Classifier. In this method, a parameterised micro-benchmark is designed to reproduce a selection of representative, synthetic workloads for quantifying the relationship between key performance events and the corresponding power values. Using the micro-benchmark enables W-Classifier to be application independent, which is a novel feature of the method. To improve accuracy, W-Classifier uses run-time workload classification and derives a collection of workload-specific linear functions for power estimation, which is another novel feature for power modelling.
To further improve accuracy, W-Classifier addresses a number of common misconceptions in power modelling, which were found to impact both modelling accuracy and evaluation. These misconceptions have arisen from differences in the experimental setup and configuration, such as, execution time, handling of thermal effects and performance event selection. These differences can influence the perceived modelling accuracy, resulting in potentially misleading or erroneous conclusions if sufficient care is not taken. As a result, W-Classifier has adopted a number of additional steps to ensure good modelling practices, which were not available in previous work.
In addition to improving modelling accuracy, the workload classification used in W-Classifier can be leveraged by power management policies to provide execution context to the power values. The workload context enables more informed power policies to be implemented, improving the balance between power and performance in achieving effective power management.
According to experimental results, the modelling accuracy of W-Classifier is significantly better than previous multi-variable modelling techniques due to a collection of workload-specific power models derived through run-time workload classification. Therefore, W-Classifier can accurately estimate power consumption for a broader range of application workloads, where a separate power model can be used for each distinct workload.||