Application-level Prediction of Program Power Dissipation

UCSB Technical Report #2002-10

Rich Wolski, Chandra Krintz, and Ye Wen

Abstract

In this paper, we investigate the degree to which power dissipation induced by program execution can be measured and predicted by application-level software tools. Application control of the power it uses while executing on a processor is critical to both the next generation of super-dense machine architectures (e.g. IBM Blue Gene) and battery-powered mobile devices that are an integral to any realization of ubiquitous computing.

Our work investigates the use of instruction-level power dissipation measurements to make whole-program power-consumption estimates and statistical techniques that predict battery death strictly from observed dissipation history. We demonstrate the prediction accuracy associated with each approach. As such, this work establishes an initial set of bounds on the accuracy with which software power management tools can expect to predict application power dissipation.