Dynamic Power Consumption Prediction and Optimization of Data Center by Using Deep Learning and Computational Fluid Dynamics

Abstract

Simply by using computational fluid dynamics (CFD) and a power consumption model incorporating each piece of equipment including servers and air conditioners, we built a power consumption simulator to predict the total power consumption of a data center that can have any device configuration, without having to learn the entire data center in advance. The error of the power consumption model built by this deep learning method was at most 8%. An algorithm with a data center optimizer (DCO) that incorporates the power consumption simulator found the optimum operation parameters of the air conditioners and optimum workload allocation that minimizes the total power consumption. In an actual implementation, the total power consumption fell within 1 second by 8% from the initial state with a uniform workload allocation. The DCO constructed in this research exhibited potential as a practical dynamic optimal task allocation management system for data centers of any size and make up and is applicable not only to dynamic migration within the data center but also to migration between data centers located at different sites.