一种基于强化学习的云应用弹性伸缩算法  

AN ELASTIC SCALING ALGORITHM FOR CLOUD APPLICATIONS BASED ON REINFORCEMENT LEARNING

作　　者：帅斌 龙士工[1,2] Shuai Bin;Long Shigong(College of Computer Science and Technology,Guizhou University,Guiyang 550025,Guizhou,China;Guizhou Provincial Key Laboratory of Public Big Data,Guiyang 550025,Guizhou,China)

机构地区：[1]贵州大学计算机科学与技术学院,贵州贵阳550025 [2]贵州省公共大数据重点实验室,贵州贵阳550025

出　　处：《计算机应用与软件》2022年第9期285-290,共6页Computer Applications and Software

基　　金：贵州省科技计划项目(黔科合重大专项字[2018]3001)。

摘　　要：针对以基础设施服务(IaaS)为主的云服务提供商,对部署在其上的Web应用以虚拟机为弹性粒度进行了弹性策略研究,提出一种基于强化学习的弹性伸缩算法PDS-lambda。该算法综合考虑用户服务违约,应用的当前负载及虚拟机数量,利用决策后状态(PDS)把算法需要学习的信息划分为动态已知与动态未知,使算法只学习动态未知的那一部分并且采用多步更新来快速达到收敛。算法对Web应用进行弹性扩张与弹性收缩操作来调整其拥有的虚拟机数量,使其在满足用户服务质量同时尽可能降低运行成本,提高云平台可靠性。仿真实验结果表明,该算法同已有的强化学习相关算法相比能更快达到收敛,且平均成本更低,用户服务违约更少。Aiming at the cloud service provider based on the infrastructure as a service(IaaS), the Web application that deployed on it was researched on the elastic strategy of the virtual machine with elastic granularity. An auto-scaling algorithm based on reinforcement learning(PDS-lambda) was proposed. The algorithm considered user service defaults, the current load of the application, and the number of virtual machines. The post-decision state(PDS) was used to divide the information that the algorithm needs to learn into dynamically known and dynamically unknown, so that the algorithm learned only the part of the dynamic unknown and used multi-step updates to achieve rapid convergence. The algorithm performed elastic expansion and contraction operations on the Web application to adjust the number of virtual machines it owns, so as to meet user service quality while reducing operating costs as much as possible, and improving cloud platform reliability. Simulation results show that this algorithm can achieve convergence faster than existing related reinforcement learning algorithms, and has a lower average cost and fewer user service defaults.

关 键 词：弹性伸缩 强化学习 云计算 负载预测 WEB应用 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]