基于图神经网络的网络性能智能预测  

Intelligent prediction method of network performance based on graph neural network

作　　者：李奕江 叶会标 谢仁华 楼佳丽 庄丹娜 李传煌[1] LI Yijiang;YE Huibiao;XIE Renhua;LOU Jiali;ZHUANG Danna;LI Chuanhuang(School of Information and Electronic Engineering(Sussex Artificial Intelligence Institute),Zhejiang Gongshang University,Hangzhou 310018,China;Zhejiang Branch of China Telecom Co.,Ltd.,Hangzhou 310020,China)

机构地区：[1]浙江工商大学信息与电子工程学院(萨塞克斯人工智能学院),浙江杭州310018 [2]中国电信股份有限公司浙江分公司云网监控维护中心核心网室,浙江杭州310020

出　　处：《电信科学》2022年第3期143-157,共15页Telecommunications Science

基　　金：国家自然科学基金资助项目(No.61871468);国家自然科学基金国际合作与交流项目(No.62111540270);浙江省新型网络标准与应用技术重点实验室资助项目(No.2013E10012);浙江省重点研发计划基金资助项目(No.2020C01079)。

摘　　要：传统网络性能预测技术存在网络状态获取不够全面及网络性能评估准确性欠佳等问题,利用图神经网络学习推理网络关系数据的特点,结合捕获的网络全局信息,提出了一种基于图神经网络的网络性能智能预测方法。通过网络系统抽象及网络性能建模,将复杂的网络信息转化为形式化的图数据进行描述,利用图空域卷积处理图网络节点的消息传递过程,实现网络信息之间的关系推理,研究了实现网络性能预测的图神经网络模型,提出了一种能处理流量矩阵、网络拓扑、路由策略、节点配置的图神经网络体系结构,最后通过实验论证了模型能更好地实现包括时延、抖动和丢包率的网络性能的准确预测。There are some problems in the traditional network performance prediction technology,such as incomplete network state acquisition and poor accuracy of network performance evaluation.Combined with the characteristics of graph neural network learning and reasoning network relational data and the captured global information of the network,on the basis of the current network performance prediction methods,an intelligent prediction method of network perfor-mance based on graph neural network was proposed.Aiming at the complex network information,through the research of network system abstraction and network performance modeling,the network information can be transformed into the graph space convolution was used to process the message passing process of graph network nodes to realize the relationship reasoning between network information.The graph neural network model for network performance prediction was studied,and a graph neural network architecture which could deal with traffic matrix,network topology,routing strategy and node configuration was proposed.Finally,the experiments show that the model can better achieve accurate prediction of the network performance including delay,jitter and packet loss rate.

关 键 词：图神经网络 网络性能预测 网络建模 网络分析 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]