基于LSTM-Z的云平台主机负载预测方法  

Host load prediction method based on LSTM-Z in cloud

作　　者：谢同磊 邓莉[1] 曹振 梁晨君 李超[2] XIE Tong-lei;DENG Li;CAO Zhen;LIANG Chen-jun;LI Chao(College of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Department of Information Development and Management,Hubei University,Wuhan 430062,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]湖北大学信息建设与管理处,湖北武汉430062

出　　处：《计算机工程与设计》2023年第9期2561-2568,共8页Computer Engineering and Design

基　　金：新一代信息技术创新基金项目(2020ITA01005)。

摘　　要：由于主机负载具有短期突变性和非线性等特点,主机负载中短期突变的信息难以被捕获。为提高主机负载预测的准确性,设计并实现一种基于Zoneout的LSTM(long short term memory with zoneout,LSTM-Z)主机负载预测方法。该方法能适应具有波动性特点的主机负载预测模式,通过遗传算法在迭代进化过程中探索最优的历史窗口权重向量,充分利用历史数据依赖关系,提高预测的准确性。通过在谷歌和阿里云两个真实的云平台数据上进行单步和多步预测实验,验证了其有效性。Due to short-term mutation and non-linearity of cloud host load,it is difficult to capture information about short-term sudden changes in host load.To improve the accuracy of host load prediction,an LSTM host load prediction method based on Zoneout(long term memory with Zoneout,LSTM-Z)was designed and implemented.This method was appropriate to host load prediction with volatility characteristics,and the optimal historical window weight vector was explored using genetic algorithm.The historical data dependence was fully utilized and the accuracy of prediction was improved.Experiments were conducted on two real cloud traces of Google and Ali.Single-step and multi-step predictions were carried out to verify the effectiveness of this method.

关 键 词：云平台 主机负载 预测 进化算法 深度学习 长短期记忆网络 正则化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]