GAT-IL:一种基于图注意力网络与模仿学习的服务功能链部署方法  

GAT-IL:A Service Function Chain Deployment Method Based on Graph Attention Network and Imitation Learning

作　　者：范琪琳 牛岳 尹浩[2] 王天富 李秀华 郝金隆[1] FAN Qi-lin;NIU Yue;YIN Hao;WANG Tian-fu;LI Xiu-hua;HAO Jin-long(School of Big Data and Software Engineering,Chongqing University,Chongqing 400044,China;Beijing National Research Center for Information Science and Technology,Tsinghua University,Beijing 100084,China;School of Computer Science and Technology,University of Science and Technology of China,Hefei,Anhui 230027,China)

机构地区：[1]重庆大学大数据与软件学院,重庆400044 [2]清华大学北京信息科学与技术国家研究中心,北京100084 [3]中国科学技术大学计算机科学与技术学院,安徽合肥230027

出　　处：《电子学报》2024年第8期2811-2823,共13页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62102053,No.61972222,No.92067206,No.62072060);国家重点研发计划(No.2019YFB1706101)。

摘　　要：网络功能虚拟化通过将网络功能从专用硬件设备迁移到商用服务器上运行的软件中间盒中,简化了网络服务的配置和管理.在网络功能虚拟化的环境下,由一系列有序的虚拟网络功能组成的服务功能链正在成为承载网络服务的主流形式.将底层物理网络资源分配给服务功能链的需求称为服务功能链部署问题.对于基础设施提供商来说,在有限的资源条件下获得长期高回报是一个重要的挑战.本文形式化定义了服务功能链部署问题,提出了一种基于图注意力网络与模仿学习的服务功能链部署方法(Graph Attention Network and Imitation Learning,GAT-IL).该方法使用图注意力网络评估每个物理服务器的放置潜力,通过蒙特卡洛树搜索方法给出专家示范,并采用模仿学习方法进行智能体的训练,融入集束搜索策略优化解空间.大量的实验结果表明,本文提出的GAT-IL方法在平均收益代价比和接受率的性能指标上均优于现有代表性算法.Network function virtualization simplifies the configuration and management of network services by mi⁃grating network functions from dedicated hardware devices to software middleboxes running on commercial servers.Under the environment of network function virtualization,the service function chain(SFC)composed of a series of ordered virtual network functions is becoming a mainstream alternative to host network services.The SFC deployment problem is to allo⁃cate the underlying physical network resources to the requirements of service function chains.It is challenging for infrastruc⁃ture providers to obtain long-term high returns under limited resources.In this paper,we formally define the problem of SFC deployment and propose a novel method named graph attention network and imitation learning(GAT-IL)based on graph attention(GAT)network and imitation learning for SFC deployment.This method utilizes GAT to evaluate the poten⁃tials of each physical server,provides expert demonstrations through the Monte Carlo tree search algorithm,applies imita⁃tion learning to train the agent,and integrates the beam search strategy to optimize the solution space.Extensive experimen⁃tal results show that the GAT-IL method proposed in this paper outperforms existing representative algorithms on perfor⁃mance metrics of average revenue-to-cost ratio and acceptance rate.

关 键 词：网络功能虚拟化 服务功能链 图注意力网络 模仿学习 蒙特卡洛树搜索 集束搜索 

分 类 号：TP393.0[自动化与计算机技术—计算机应用技术]