Analyzing the Energy Efficiency of the Fast Multipole Method Using a DVFS-Aware Energy Model

We present a semi-empirical model for analyzing where a program spends its energy. This model combines measured features of an application, such as its instruction mix, with measured costs of the platform, such as the energy per operation specialized to different operations. The model permits costs to vary with voltage and frequency, as they would on platforms that support dynamic voltage and frequency scaling (DVFS). Our experiments indicate that the model accurately predicts the energy consumption of both microbenchmarks and a proxy application, which implements the fast multipole method for n-body problems. It is accurate enough to identify near-optimal voltage and frequency settings that maximize energy efficiency and permits an analysis of where potential hardware and software energy bottlenecks might be in our proxy application. We believe our overall methodology could be applied more broadly by other performance analysts.