基于网络层析的网络性能测量与推测方法研究

【作者】 陈希；

【导师】 隆克平；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2007， 硕士

【摘要】 当今Internet是一个庞大的分布式网络,随着网络的不断扩大,网络的业务种类和复杂性也迅速增长。如何了解当前网络状态和性能,以便更合理、有效的管理网络,成为网络管理者和网络服务供应商都非常关注的问题。由于网络不断朝异构化、不协作特征发展,使得网络性能测量工作的研究极具挑战性。现有的许多测量工具都需要网络内部节点之间的协作,这会影响真实的网络业务量,并且会引发安全问题。因此,近年来人们提出了一种新的网络性能测量与推测技术,称为“网络断层分析”。该技术是通过网络端到端测量,对内部链路性能进行推测估计,对网络性能测量工作来说有重要研究意义。本文首先介绍了近年来网络测量和推测技术的现状,然后详细阐述了网络层析技术中基于组播测量的链路丢包率推测算法――直接推导法(Direct-MLE)和期望最大化算法(EM-MLE)。由于并非所有的网络都支持组播,且路由器处理组播数据包与单播数据包的方式也明显不同。为了避免组播的局限性,普遍采用背靠背单播数据包对进行丢包性能的测量推测。本文在基于单播测量的基础上提出了一种新的基于TCP监测的被动测量机制,被动测量机制具有不向网络注入探测包,不会影响实际网络流量的特性,可用于高负载的大规模网络性能测量。本文通过OPNET和Matlab仿真建模,对比了主动测量中使用直接推导法和EM算法推测链路丢包率的准确性、收敛性等特征,仿真证明EM-MLE算法优于Direct-MLE算法。在被动测量仿真中,对比了固定搜索间隔和动态搜索间隔对估计值的影响,实验证明,采用动态的搜索时间间隔比固定搜索间隔得到的估计值更逼近真实链路丢包率。另外本文还研究了网络层析技术用于时延测量估计的方法,分别对基于累积生成函数和有限高斯混合密度函数的时延推测算法进行了分析。并通过Matlab仿真证明了有限高斯混合模型中EM算法的有效性。更多还原

【Abstract】 Today’s Internet is a massive, distributed network which continues to explode in size as all kinds of traffic expanded rapidly. How to get the network performance on the existing infrastructure and manage network more reasonable or more effective is a challenge task. Especially, it is essential for the network managers and ISP. The heterogeneous and no-corporative structure of the Internet renders the tasks such as network behavior measurement extremely challenging.Many measurement tools require the cooperation with network nodes; they may affect the network real traffics and cause the security problems. For these reason, a promising technique named“Network Tomography”is emerging in recent years. It infers the network internal performance by end-to-end measurements, and makes a good performance in network fault detection.In this paper, we first introduce the recent research in network tomography, and then review the measurement and inference algorithm we could use Direct-MLE and EM-MLE. Because multicast protocols are not supported by significant portion of Internet and the routers treat differently between multicast packets and unicast packets. It is popular to use unicast back-to-back pairs as the probes to avoid those problems in multicast network. In this paper, we propose a new measurement based on TCP monitoring to get the packet pairs. The passive measurement does not influence the real network flows, so it is useful in heavy load network measurement.In this paper, we use the OPNET and Matlab model simulation to analyze the inference algorithms. We compare the difference on accuracy and convergence characteristic of direct inference algorithm and EM algorithm in active measurements. Simulation results show that EM algorithm has better convergence and can infer the internal performance effectively. We also analyze the searching space between two packet pairs which influences inference algorithm’s accuracy. The conclusion shows that dynamic searching space is better than fixed searching space. At last, we introduce the delay inference in network tomography. We discuss the CGF model and MFMM methods in delay inference respectively. With Matlab simulation, it shows that EM algorithm in MFMM model is effectively.更多还原