Improving the Accuracy of Energy Predictive Models for Multicore CPUs Using Additivity of Performance Monitoring Counters

Abstract

Energy predictive modelling using performance monitoring counters (PMCs) has emerged as the leading mainstream approach for modelling the energy consumption of an application. Modern computing platforms such as multicore CPUs provide a large set of PMCs. The programmers, however, can obtain only a small number of PMCs (typically 3–4) during an application run due to the limited number of hardware registers dedicated to storing them. Therefore, selection of a reliable subset of PMCs as predictor variables is crucial to the prediction accuracy of online energy models. State-of-the-art methods for selecting the PMCs are largely based on their correlation with energy consumption.

Recently, Additivity is introduced as a property of PMCs that appears to have significant impact on the accuracy of energy predictive models. It is based on an experimental observation that energy consumption of serial execution of two applications is equal to the sum of the energy consumption of those applications when they are run separately. In this work, we demonstrate how the accuracy of energy predictive models based on three popular techniques (Linear regression, Random forests, and Neural networks) can be improved by selecting PMCs based on a property of additivity.