Just-in-time component-wise power and thermal modeling

Abstract

As computer systems increasingly focus on balancing the performance and power efficiency of software applications together with temperature variations of the machine, they need to understand how software applications utilize the various architecture components differently. This paper develops a power and temperature modeling framework to provide such timely feedback, which can then be used to support a dynamic optimization system to attain better energy efficiency for applications. In particular, we present a framework that combines McPAT [17], a cycle accurate architecture simulation model, with runtime hardware performance counter statistics, to attain component-wise power consumption breakdown of applications while running at GHz speed. Our framework is able to consistently achieve 98% accuracy when compared to the actual system-level power consumption measured using a real-time power meter [1]. Finally, we present a preliminary study to demonstrate the potential of using our framework to support the optimizations of applications for better energy efficiency.