Report ID
2002-10
Report Authors
Rich Wolski, Chandra Krintz, and Ye Wen
Report Date
Abstract
In this paper, we investigate the degree to which power dissipationinduced by program execution can be measured and predicted byapplication-level software tools. Application control of the powerit uses while executing on a processor is critical to both the nextgeneration of super-dense machine architectures (e.g. IBM Blue Gene)and battery-powered mobile devices that are an integral to anyrealization of ubiquitous computing.Our work investigates the use of instruction-level power dissipationmeasurements to make whole-program power-consumption estimates andstatistical techniques that predict battery death strictly fromobserved dissipation history. We demonstrate the prediction accuracyassociated with each approach. As such, this work establishes aninitial set of bounds on the accuracy with which software power managementtools can expect to predict application power dissipation.