Learning-based power prediction for data centre operations via deep neural networks

Abstract

Modelling and analyzing power consumption for data centres can diagnose potential energy-hungry components and applications, and facilitate in-time control, benefiting the energy efficiency of data centers. However, solutions to this problem, including static power models and canonical prediction models, either aim to build a static relationship between power consumption and hardware/application configurations without considering the dynamic fluctuation of power; or simply treat it as time series, ignoring the inherit power data characteristics. To tackle these issues, in this paper, we present a systematic power prediction framework based on extensive power dynamic profiling and deep learning models. In particular, we first analyse different power series samples to illustrate their noise patterns; accordingly we propose a power data de-noising method, which lowers noise interference to the modelling. With the pretreated data, we propose two deep learning based prediction models, including a fine-grained model and a coarse-grained model, which are suitable for different time scales. In the fine-grained prediction model, a recursive autoencoder (AE) is employed for short-duration prediction; in the coarse-grained model, an AE is used to encode massive fine-grained historical data as a further data pretreatment for long-duration prediction. Experimental results show that our proposed models outperform canonical prediction methods with higher accuracy, up to 79% error reduction for certain cases.