Energy measurement and prediction for multi-threaded programs
Thomas Rauber, Gudula Rünger, Michael A Schwind · ERef Bayreuth (University of Bayreuth) · 2014
For many years, runtime performance has been the main concern in high-performance and parallel computing. However, power and energy consumption become more and more important, especially for large-scale supercomputer systems. Accordingly, energy-saving methods, such as dynamic voltage frequency scaling, have been integrated in recent processor architectures. In this article, we investigate the power and energy consumption for the execution of multi-threaded programs on recent Intel multicore processors, using the PARSEC and the SPLASH benchmarks as examples. In particular, we perform experiments for measuring the power and energy consumption with two different methods. Based on the measured energy values, we investigate power and energy models suggested in the literature as well as a new heuristic model and compare the models with respect to their prediction capabilities. We show that especially the heuristic model is able to quite accurately predict the power and energy consumption depending on the scaled frequencies.