基于在线学习的数据中心节能方案设计  

Design of Data Center Energy-saving Scheme Based on Online Learning

作　　者：张乾 徐欢乐 刘敬民 ZHANG Qian;XU Huanle;LIU Jingmin(School of Computer Science and Technology,Dongguan University of Technology,Dongguan 523808,China;School of Computer Science,Chinese University of Hong Kong,HongKong 999077,China;Vertix Technology(Xi'an)Co.,Ltd.,Xi'an 710018,China)

机构地区：[1]东莞理工学院计算机科学与技术学院,广东东莞523808 [2]香港中文大学计算机学院,中国香港999077 [3]维谛技术(西安)有限公司,陕西西安710018

出　　处：《东莞理工学院学报》2022年第5期86-94,共9页Journal of Dongguan University of Technology

基　　金：国家自然科学基金资助项目(61802060)。

摘　　要：近年来,随着计算机互联网等技术的迅速发展,与之对应的数据中心等基础设施的功耗也在持续增长。在数据中心耗能中,冷却系统的耗能占比甚至达到了30%。为了对数据中心冷却控制系统进行优化,提出一种基于在线学习的优化方案,采用基于高斯过程的多摇臂赌博机模型在线决策数据中心的空调开启台数以及采用XGBoost回归模型实时决策空调的设置温度。合适的空调设置可提升冷却系统效率,从而降低冷却系统的能源消耗。这一方案使用两个机器学习模型将空调设置的决策空间解耦,将决策空间缩小近十倍,加快收敛速度,并能根据环境的变化自动调整决策。In recent years,with the rapid development of technologies such as computer Internet,the corresponding power consumption of the infrastructure such as data centers also continues to grow.In the data center energy consumption,cooling systems even account for 30%.In order to optimize the data center cooling control system,the paper proposes an online learning-based optimization scheme,which adopts the Gaussian process-based multi armed bandit algorithm to determine the number of air conditioning opening in the data center and the XGBoost regression model to determine the setting temperature of the air conditioning in real time.Suitable air-conditioning settings can improve the efficiency of the cooling system,thus reducing the energy consumption of the cooling system.This scheme uses two machine learning models to decouple the decision space of air conditioning settings,reduce the decision space by nearly ten times,accelerate convergence,and automatically adjust decisions according to changes in the environment.

关 键 词：数据中心 多摇臂赌博机 高斯过程 置信上界(UCB) 电能利用效率(PUE) 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP308[自动化与计算机技术—控制科学与工程]