Runtime energy management of multi-core processors
Charles Leech · ePrints Soton (University of Southampton) · 2018
Performance requirements of emerging applications and tighter power consumption constraints of mobile and embedded platforms mean that runtime management software is required to control these systems efficiently. In order for embedded systems to maintain their optimality, especially in dynamic environments, runtime software must be capable of learning and adaptability. This thesis investigates and develops runtime modelling methods, including their experimental validation, to reduce energy consumption in homogenous and heterogeneous multi-core processors. A multiple linear regression model is established to predict power and performance and drive runtime adaptations of an application and platform to maximise energy efficiency whilst meeting performance targets. The proposed method is further validated with a parallel stereo matching application, which is developed to investigate the use of core scaling and analyse trade-offs between power and performance through runtime adaptation for energy saving. Experimental results obtained from a 61-core Intel Xeon-Phi platform show that the same performance can be achieved with an average reduction in power consumption of 27.8% and increased energy efficiency by 30.0%, in comparison to baseline dynamic power management techniques. To make energy management independent of the application and platform, this thesis presents a holistic approach to runtime management in the form of a runtime framework that is both application- and platform-agnostic. The framework unites the hardware and software layers by embedding the runtime management layer at the centre and enables cross-layer interactions through an API and dynamic knobs and monitors. The framework is demonstrated experimentally across multiple applications and two heterogeneous platforms, the Odroid-XU3 and Cyclone V. Two state-of-the-art runtime management approaches are validated that reduce the energy consumption of application execution by 18.2% and 17.2%. Trade-offs between power, performance and accuracy are presented in three application-platform scenarios.