【作者】 曾彬；

【导师】 张大方；

【作者基本信息】 湖南大学 ， 计算机应用技术， 2009， 博士

【摘要】 随着网络技术的不断发展,网络规模不断扩大,承载在IP网络上的业务日益趋于多样化和复杂化,IP网络已经被要求从提供简单连接与控制的网络逐渐向高品质业务网络的方向发展。这对网络性能提出了更高的要求。但由于Internet固有的高度异构性、复杂性和动态性等特点,人们对网络整体的性能特征及其行为的透彻理解也变得越来越困难。网络性能监测通过直接采集网络性能原始数据,并以此为基础对各项指标进行量化,对网络性能进行特征化描述,是充分理解并正确认识网络的基本手段,是进行网络控制、管理,以及业务计费的重要依据,也是解决网络所面临挑战的前提。本文系统地总结了网络性能监测领域迄今为止的主要研究工作、进展及特点,指出了采用主动测试方式进行网络性能监测所面临的一些难点问题,针对各协议层次重要指标测量方法,流量数据采样方法,监测系统设计实现等问题进行了深入研究并取得如下创新性成果:(1)提出了一种低开销、快速收敛的可用带宽测量改进方法。针对PRM(Probe Rate Model)可用带宽测量方法存在测量开销大、收敛慢的问题,提出了一种可用带宽测量的改进方法WPathload,该方法基于时延变化的统计规律,改进发送速率调整算法,并采用周期流组到达目的端的速率代替周期流的发送速率,更新可用带宽上界。实验结果表明,改进后的方法能加快收敛速度,快速反映可用带宽的变化,增强了跟踪带宽变化的能力。(2)深入研究了可用带宽测量中的包对行为特征。PGM(Probe Gap Model)可用带宽测量方法中探测包对间隔容易受到路径上背景流量、数据包异常等因素的干扰,从而造成对可用带宽的高估或低估。本文结合理论推导和实验分析,定量分析这些干扰因素对探测包对间隔行为特性的影响,观察这些因素与测量结果精度之间的关联规律,分析得到了一系列有价值的结论,为PGM测量工具的改进与完善提供了依据。(3)提出一个基于四状态Gilbert丢包机制的TCP吞吐量模型。现有TCP吞吐量模型采用的简单丢包、Bernoulli丢包等机制已经成为制约提高吞吐量模型精确度的重要原因。本文采用四状态Gilbert模型来描述测量过程中端到端路径上的丢包行为,对TCP的拥塞控制过程进行建模,重新推导出一个基于四状态Gilbert丢包机制的TCP吞吐量模型。实验证明该模型具有更高的精确度,且在此模型上实现的吞吐量间接测量工具,不需要产生多余的探测流量,可以很好地支持长期在线测量。(4)提出了一种基于主动业务仿真的流媒体业务性能测量方法。目前已有的流媒体业务性能测量工具主要用于对服务进行压力测试,不能反映终端用户的实际使用性能。本文从端用户角度提出了合理反映流媒体业务性能的指标,借鉴主动测试技术,提出并实现了一种基于主动业务仿真的流媒体业务性能测量方法RealMeasure。并应用该工具,通过仿真实验,深入分析带宽、帧速率、丢包率和延迟对流媒体业务性能的影响,并在分析的基础上提出了一种流媒体业务性能预警方法;最后,提出一种流媒体客户端缓冲区监测管理机制,能有效提升缓冲区包容包失序的能力,精确有效地监测包失序的严重程度及其对业务性能的影响。上述工作能及时向流媒体服务器反馈信息,实现动态的质量控制,并且为流媒体业务系统的优化提供依据。(5)提出了一种自适应的网络流量数据采样方法。针对目前常用的流量采样方法无法在网络流量突发时及时、准确反映真实流量特征的问题。基于集合的观点,提出一种自适应的网络流量的采样方法ADSA(Adaptive Sample),该方法通过对已采取样本的分析,估计网络流量当前的变化程度,自动调整样本间的采样间隔和采样概率,实现对流量指标较细致准确的反映,能以较小的采样代价获知实际的网络流量行为特征。既达到了采样速率自适应于流量变化的目的,又可以控制对测量设备资源的消耗,而且该方法易于实现。(6)提出了面向网络行为特征分析的网络监测系统框架,设计并实现了分布式网络监测系统NetMonitor。该系统通过实时监测网络中的流量行为、端到端行为、路由行为、业务行为,提取网络行为的基本特征,构造并验证网络行为的数学模型。该系统能将网络监测技术和对网络数据分析技术有机结合起来,为网络行为的各种理论分析提供统一的研究平台。更多还原

【Abstract】 With the advances in the broadband networking technology, the network scale becomes increasingly magnitude and IP network services trend to be diverse and complicated. Besides, more rigorous schemes to operation, administration and maintenance bring the requirement of building network infrastructure with high QoS service, compared with functions that offered simple connections and controls in the past. However, because of the inherited high isomerizationess, complexities and dynamic natures of the Internet, it is more and more difficult for people to comprehend the performance characteristics and behaviors of network.Monitoring network performance can quantify metrics and describe network capabilities directly via collecting raw data directly. It is not only the basic method for acknowledging network correctly, but also the important foundation for network control, management and business fee counting. Moreover, it is also the premise to address challenges facing network. In the thesis, we summarized systematically the state of the art in network performance measurement and the past progresses as well as important achievements in this area. In addition, we point some critical problems of measuring network performance based on active measurement. Besides, we delve into measuring important metrics from network layer to application layer, traffic data sampling methods and the design and implement of large scale network performance monitoring system. The main innovative achievements of this thesis are as following:(1) This thesis proposes an improved bandwidth measurement algorithm with low cost and rapid convergence. PRM (Probe Rate Model) is one of the important methods for measuring available bandwidth. Unfortunately it still exist some problems, e.g. the long convergence time and large probe traffic. In order to solve these problems, WPathload, a modified available bandwidth measurement algorithm based on delay jitter statistical law, was proposed. WPathload replaces the sending rate of periodical traffic by its arrival rate, and updates the upper limit of available bandwidth, resulting in shortening convergence speed and decreasing the measurement cost. The experimental result indicates that the ameliorated algorithm can sense changes of available bandwidth rapidly and strengthen the capacity of tracing bandwidth changes.(2) This thesis comprehensively studies behavior characteristic of packet pair during available bandwidth measurement. Because PGM (Probe Gap Model) technique, a current available bandwidth measurement method, can be significantly affected by some inference factors, including cross traffic, packet pathology and clock precision, the result of available bandwidth may be overrated or underrated. With theoretical analysis and experimental data, the impact of these factors is quantitatively analyzed in this paper, and the relationship between these factors and measurement result accuracy is also discussed. Finally, we draw many valuable conclusions, which can be used for ameliorating the PGM.(3) This thesis presents a new TCP throughput model based on Gilbert four-state packet loss pattern. The existing TCP throughput model, which adapts simple packet loss or Bernoulli packet loss mechanism, has become an important factor of restricting the accuracy of throughput model. In this thesis, Gilbert four-state model is used for describing the packet lost behavior of end-to-end internet path, and constructing a new model for TCP congestion control. The simulative result shows that the new model is capable of more precision in actual network. In addition, because the throughput measurement tool based on the proposed model will not produce large measurement traffic, it exerts no impact on the real network and can be used for long period on-line measurement.(4) A streaming media service performance monitoring method Based on active application simulation is also presented in this thesis. From the perspective of users, we propose several metrics reflecting the performance of streaming media application, and accomplish measurement tool—RealMeasure, which is based on active business stimulation. The simulative experiment is used for analyzing the influence on streaming media performance caused by bandwidth, frame rate, packet loss and delay. And according to the analysis, an approach to warning a streaming media performance is discussed. At last, to arrange the out of order packets and improve the client performance, we come up with a streaming receiver buffer monitoring and management algorithm to quantitatively measure the degree of reordering in sequences of packets. According to the experimental data, our work can accomplish dynamic control of quality and provide basis for optimizing the streaming media business system.(5) This thesis presents an internet traffic data sampling methodology based on aggregated time-serial evolution trend. Current popular sampling algorithms can’t provide accurate approximation of real characteristics under the circumstance of network traffic bursting. The paper develops an adaptive network traffic sampling methodology ADSA (Adaptive Sample) based on aggregated time-serial evolution trend. The ADSA algorithm adaptively adjusts the time interval between consecutive samples according to the evolution trend of network traffic which is estimated from the obtained samples. As a result, it is able to approximate the measured traffic more precisely. Compared with existing popular sampling algorithms, ADSA algorithm can acquire real traffic characteristics via few samples. Besides, it can not only adjust the sampling rate adaptively according to traffic changes, but also can reduce the device cost.(6) This thesis develops a Behavior Analysis-oriented IP network monitoring system. To reflect Internet running behavior characteristic roundly, we design and implement a network monitoring system for the analysis of the behavior characteristics. By monitoring and analyzing traffic behavior, end-to-end behavior, routing, and application behavior characteristic, the system can acquire the basic characteristics of network behavior, which are used for constructing and validating math model of network behavior. In addition, this system can combine network monitoring technology with network data analyzing technology, and bring about a research platform for various theoretic analysis on network behavior.更多还原