Learning-Directed Dynamic Voltage and Frequency Scaling for Computation Time Prediction

Dynamic voltage and frequency scaling (DVFS) is an effective technique for reducing power consumption. A number of DVFS researches apply learning methods in an attempt to approach the DVFS prediction model instead of using complicated mathematical models. In this paper, we propose a lightweight learning-directed DVFS technique using Counter Propagation Networks (CPN) to identify the task behavior and predict the corresponding voltage/frequency setting precisely. An adjustable performance mechanism is also provided to users that have diverse performance requirement. The algorithm has been implemented on the Linux operating system and used a PXA270 development board. The results show that the learning-directed DVFS method could accurately predict the suitable frequency, given runtime statistics information of a running program. In this way, the user can easily control the energy consumption by specifying allowable performance loss factor.