Applicability of the empirical mode decomposition for power traces of large-scale applications

Abstract

Current trends in HPC show that exascale systems will be power capped, prompting their users to determine the best combination of resources to satisfy a power budget. Hence, performance and energy models must interplay and aid users in this resource selection based on the desired application parameters. While existing performance models may predict application execution at a scale, current power models are inadequate for this propose due, in part, to the variability of instantaneous dynamic power and the need to handle large amount of power measurements at the runtime to populate the models. In this paper, the latter challenge is tackled by selecting certain power measurements and applying to them the empirical mode decomposition (EMD) technique, which itself already deals with instantaneous variability of power during the runtime. Specifically, it is proposed here to apply EMD to segments of a power trace to rapidly generate a quadratic model that describes overall time, power, and thus energy simultaneously. The proposed models have been applied to several realistic applications. The error across the proposed models and the measured energy consumption is within 5% for the smaller segments consisting of 2,000 trace samples and is about 2% for the segments of 6,000 samples.