Time Series Modeling Based Power and Performance Scaling Framework

This paper enhances a software based framework to dynamically scale power and performance with high accuracy in a resource limited embedded system, like cellular phones and PDAs. Key challenges for such a framework are accurate forecasting of dynamic resource demands inherent in the workloads. In this paper we describe three innovative methods: (1) smart forecast method based on linear and non-linear filtering models; (2) policy decision based on high fidelity memory and computation characterization; and (3) adaptive sampling period to adapt to dynamic changes in the workloads. Power and performance framework driven by the proposed algorithms reduces power consumption thus improving battery lifetime for the end-user.