A predictive system shutdown method for energy saving of event-driven computation

Abstract

We present a system-level power management technique for power saving of event-driven applications. We present a new predictive system shutdown method to exploit sleep mode operations for power saving. We use an exponential-average approach to predict the upcoming idle period. We introduce two mechanisms, prediction-miss correction and pre-wakeup, to improve the hit ratio and to reduce the delay overhead. Experiments on four different event-driven applications show that our proposed method achieves high hit ratios in a wide range of delay overheads, which results in a high degree of power saving with low delay penalties.