Tuning dynamic power management for mobile devices

Abstract

Mobile devices have rapidly reached almost ubiquitous adoption amongst the global population. Smartphones have been the catalyst for introduction of high-performance System-on-Chips to mobile devices bringing with them the capability to execute ever more demanding applications but also widespread power management challenges. Traditionally, the foremost power management challenge was extension of battery lifetime. The emergence of sustained performance applications including mobile gaming, Virtual and Augmented Reality has presented a new challenge in constraining performance to within a sustainable thermal envelope. Cooling techniques, limited to passive technologies in mobile devices, have proved insufficient to maintain device skin temperatures below thresholds the human skin can tolerate. Dynamic Power Management policies have been developed to reduce mobile device power consumption to meet both energy and thermal constraints. This thesis proposes and then explores a new area of research in systematic tuning of Dynamic Power Management policies for mobile devices. Static and dynamic configuration of Dynamic Power Management policy parameters are compared to quantify the potential energy and performance improvements. Experimental results from a modern mobile device across four applications suggest up to 10% reduction in dropped frames and a 25% reduction in CPU energy consumption. Interactive performance degradation from Frequency Capping - the Dynamic Power Management lever used in device skin temperature throttling - was shown to induce up to a 43% increase in dropped frames. A new Dynamic Power Management lever - Task Utilisation Scaling - is proposed and validated to mitigate the performance degradation by reducing the number of dropped frames by up to 8.5% compared to Frequency Capping. Evaluation of interactive performance metrics such as dropped frames was shown to be up to 750x slower than CPU energy consumption. An investigation into the root cause of this problem lead to a new metric for estimating dropped frames - PPPx - which can accelerate evaluation by up to 43.3x. PPPx was validated in a tuning experiment that lasted 8 days which would otherwise have required 8 months.

More information

Identifiers

ORCID for James Robert Benjamin Bantock: orcid.org/0000-0002-0766-7532

ORCID for Geoffrey Merrett: orcid.org/0000-0003-4980-3894

Catalogue record

Export record