大尺度IP骨干网络流量矩阵估计方法研究  

Approach of Traffic Matrix Estimation in Large-scale IP Backbone Networks

作　　者：蒋定德[1,2] 王兴伟[1] 郭磊[1] 许争争[1] 陈振华[1] 

机构地区：[1]东北大学信息科学与工程学院,辽宁沈阳110004 [2]电子科技大学宽带光纤传输与通信网技术重点实验室,四川成都610054

出　　处：《电子学报》2011年第4期763-771,共9页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61071124;No.61070162;No.71071028;No.60802023;No.70931001);高等学校博士学科点专项科研基金(No.20100042120035;No.20100042110025;No.20070145017);新世纪优秀人才计划(NCET-08-0095);中央高校基本科研业务专项资助(No.N090404014;N090504003;N090504006)

摘　　要：流量矩阵估计是当前的热点研究问题,它被网络操作员用来进行负载均衡、路由最优化、流量侦测、网络规划等等.然而,流量矩阵估计本身固有的高度病态特性,使得精确地估计流量矩阵成为具有挑战性的研究课题.本文研究大尺度IP骨干网络的流量矩阵估计;基于RBF(Radial Basis Function)神经网络,提出一种新的估计方法TMRI(Traffic Matrix Recurrence Inference).TMRI利用RBF神经网络强大的建模功能来建模流量矩阵估计问题,将这一问题的病态特性克服于RBF神经网络的训练过程中,从而避免复杂的数学建模过程.并在所建立的估计模型基础上,将流量矩阵估计描述为约束条件下的最优化过程,通过迭代寻优,TMRI能进一步克服这一问题的病态特性.仿真结果表明TMRI能精确地估计流量矩阵和追踪它的动态变化,与以前的方法相比,具有更强的抗噪声性能和显著的性能改善.Traffic matrix estimation is an interesting research problem at present.Network operators use it to conduct load balancing,route optimization,traffic detecting,network dimensioning and so on.However,the highly ill-pose nature of traffic matrix estimation itself makes it being a challenging research subject to estimate accurately traffic matrix.This paper studies traffic matrix estimation in large-scale IP backbone networks.Based on RBF（radial basis function） neural network,a novel estimation method,namely TMRI（traffic matrix recurrence inference）,is proposed.TMRI exploits the powerful modeling ability of RBF neural network to model traffic matrix estimation.The ill-pose nature of this problem will be overcome in the process of training the RBF neural network.Accordingly,the complex process of mathematic modeling can be avoided.Built on this estimation model,traffic matrix estimation is described into the optimal process under the constraints.By seeking the recurrent optimal solution,TMRI can further get rid of the illpose nature of this problem.Simulation results show that TMRI can accurately estimate traffic matrix and track its dynamics,and in contrast to previous methods,it holds the stronger robustness to noise and more evident performance improvement.

关 键 词：流量矩阵估计 非平稳流量 RBF神经网络 病态特性 最优化 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]