基于深度强化学习的服务功能链映射算法  

SFC mapping algorithm based on deep reinforcement learning

作　　者：金明 李琳琳 张文瑾 刘文 Jin Ming;Li Linlin;Zhang Wenjin;Liu Wen(School of Operational Support,Rocket Force University of Engineer,Xi’an 710000,China;96941 Troops of PLA,Beijing 100089,China;96872 Troops of PLA,Baoji Shaanxi 721000,China)

机构地区：[1]火箭军工程大学作战保障学院,西安710000 [2]解放军96941部队,北京100089 [3]解放军96872部队,陕西宝鸡721000

出　　处：《计算机应用研究》2020年第11期3456-3460,3466,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目。

摘　　要：针对服务功能链映射对网络时延和部署失败率的影响,提出了一种基于深度强化学习的服务功能链映射算法DQN-SFC。首先构建了一个多层次NFV管理编排架构,以满足算法对资源感知和设备配置的需求;然后基于马尔可夫决策过程建模,对SFC映射问题进行形式化描述;最后构建了一个深度强化学习网络,将网络平均时延和部署失败产生的运维开销作为奖惩反馈,经过训练后可根据网络状态决定虚拟网络功能的部署位置。通过仿真实验,对该算法的正确性和性能优势进行了验证。实验表明:与传统算法相比,该算法能有效降低网络平均时延和部署失败率,同时算法运行时间具有一定优势。This paper proposed an algorithm for SFC mapping based on deep reinforcement learning which was called DQN-SFC,aiming at reducing the influence of SFC mapping on the average time delay and deployment failure ratio in the network.Firstly,it constructed a multi-layer NFV management and scheduling architecture to meet the requirements of resource awareness and equipment configuration of the algorithm.Secondly,based on Markov decision process,it formally described the SFC mapping problem.Finally,it constructed a deep reinforcement learning network,which used the average network delay and the operation expense of the deployment as reward and punishment feedback.After training,the target position of the virtual network function where to be deployed can be determined according to the network status.The simulation experiment verifies correctness and performance of this algorithm.Experiment shows that this algorithm can effectively reduce the average network delay and deployment failure ratio,and has certain advantages in algorithm running time.

关 键 词：网络功能虚拟化 服务功能链 深度强化学习 网络时延 网络运维开销 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]