【作者】 王晖；

【导师】 季振洲；

【作者基本信息】 哈尔滨工业大学 ， 计算机系统结构， 2013， 博士

【摘要】 TCP（Transmission Control Protocol）的拥塞控制是Internet稳定运行的基础，围绕着TCP协议的拥塞控制一直是Internet研究的一个热点，吸引着诸多的学者。随着网络通信流量的急剧增加和各类实时业务流量对QoS要求的大幅提高，仅仅依靠端到端拥塞控制己经难以满足网络需求。事实上，在Internet这样复杂的异构网络环境中，希望所有用户都兼容端系统拥塞控制也是不现实的。必须发展路由器等中间网络设备的控制，以增强拥塞控制的效果。路由器位于拥塞的发生点，所以在路由器上进行拥塞控制是非常有意义的。队列管理机制就应运而生了。主动队列管理（Active Queue Management，AQM）机制通过对拥塞的预判和主动丢包，实现对拥塞的控制，成功避免了死锁、全局同步等现象。IP拥塞控制机制的研究是当前拥塞研究的热点。本文将路由器参与的显式拥塞控制也归为IP拥塞控制机制。网络业务流自相似性的发现和研究推翻了早先网络流量短相关的基础假设，由于网络流量突发性更为突出，直接导致拥塞发生更为频繁和加剧，这使得网络流量的统计特征提取、排队性能分析和缓冲空间设置以及拥塞策略的设计均有所变化。自相似模型的引入给原本复杂的拥塞控制带来新的问题，但它同时也会带来新的解决方法。往返时延RTT是网络拥塞控制机制有效运行赖以维系的节奏，网络时延不易准确估算也是造成网络拥塞机制偶尔失效的主要原因。故而需要分别对往返时延RTT建模为常数、常函数、随机过程等不同的数学形式，利用数学方法分析RTT对网络拥塞控制机制的影响。进一步提出基于网络排队延时的显式拥塞控制算法QDCN，该算法通过路由器监测队列长度，进而得出排队延迟，实时更新RTT，以显式方式通知源端改变拥塞窗口，从而实现拥塞避免。进一步看到在恒速网络业务流下，本来性能比较优越的SFPID-RED和QDCN算法性能有所下降，并且时延并不是造成算法性能下降的诱因。实际上网络流量的自相似性（突发性）才是这些算法失效的根本原因，时延的抖动只是它的一种表现形式。于是提出一种基于自相似流量的随机早检测算法——STRED。该算法采用时间槽作为操作单位，以减少计算量，降低网络参数更新速率；根据时间槽记录观测参数进而预估自相似系数（Hurst系数），并依据相关函数调整RED算法丢弃概率，增强RED算法对自相似网络流量的适应能力，从而实现对自相似流量的拥塞控制。然而学界关于网络流量是具有泊松特性的短相关还是具有分形自相似的长相关的争论一直不断。虽然有大量的网络测量实验分析得出网络自相似的结论，但也有证据表明泊松特性依然存在。网络流量模型经历了短相关，长相关，多重分形的发展演变，现在有回归短相关的一种可能。实际上网络流量的长相关和短相关两种特性同时存在。有鉴于此本文基于网络自相似性和TCP/AQM流体流模型，提出一种新型Lévy随机过程，并建立一种TCP/AQM二象性流体流模型，可以同时描述网络的自相似性和包驱动特性，试图从理论上对网络流量的两种特性进行统一，并进一步研究了二象性模型的有关性质。为以后基于自相似网络的拥塞控制研究奠定了基础。总之本课题主要针对网络拥塞问题，剖析了一种现有应用较为广泛的网络流量模型——TCP/AQM流体流（Fluid Flow）模型，并以此为基础结合网络自相似提出一系列的研究方案。更多还原

【Abstract】 Congestion control in TCP (Transmission Control Protocol) is the basis for theInternet steady operation. The TCP protocol congestion control has been a focusedresearch area of the Internet, which is attracting many scholars. Due to dramaticallyincreasing network traffic and quality of service (QoS) requirements for a variety ofreal-time traffic, it is difficult to meet with the network requirements, dependingsolely on end-to-end congestion control. In fact, in a complex and heterogeneousnetwork environment, it is unpractical to keep compatible congestion controlscheme in end-to-end systems for all users. Control of routers and other intermediatenetwork devices must be involved in order to enhance the effect of congestioncontrol. It is at the router where congestions occur actually; therefore it ismeaningful to conduct queue management mechanism at the router. With earlycongestion predicting and active packet dropping, active queue management (AQM)mechanisms could enforce the congestion control and avoid the deadlock and theglobal synchronization. This is the so-called IP congestion control. IP congestioncontrol mechanism is a current research focal point. In this research, the explicitcongestion control mechanism with routers involved is also classified as an IPcongestion control mechanism.The study of the self-similarity of network traffic overthrows the basicassumption that the network traffic is with short-range dependence. As the result ofmore remarkable burstiness of traffic, more congestion occurred frequently andaggravatingly. It is destined to make the analysis of statistical characteristics of thenetwork traffic and the performance of the router queuing, and the set of the buffersize at the router, different from the original assumptions. The introduction ofself-similar model is not only bound to bring new challenges to congestion controlwhich is inherently complex, but also new solutions.The round-trip time (RTT) is the prerequisite to effectively operate congestioncontrol mechanism followed the rhythm step by step. The imprecision of theestimation for the network delay also plays a primary role while network congestionmechanism occasionally fails. Therefore it is necessary to model the RTTrespectively to be a constant, a common function, a stochastic process and suchdifferent mathematical formats. By mathematical analysis, in this research theimpact of the RTT is analyzed against congestion control mechanisms in networks.Further, an explicit congestion control algorithm named as QDCN is proposed basedon network queuing delay. In this algorithm, the router queue length is monitored,and then the queuing delay is obtained to update RTT in real-time. By explicit notice sent by the router，the source end decided to change the congestion window in orderto avoid congestion.Further, the SFPID-RED and QDCN perform gradationally, which do perfectlyat a constant rate network service before. However, the delay is not the primarycauses of the performance degradation. In fact, the self-similarity of network traffic(burstiness) leads to the failure of these algorithms while the delay jitter is just amanifestation. Therefore, a self-Similar-Traffic-based RED algorithm named asSTRED is proposed. In this algorithm, time slots are adopted as the operating unit toreduce the workload of calculation and the updating rate of the network parameters.The self-similarity coefficient (Hurst coefficient) is estimated based on time slots toobserve and record values, and the packet drop probability in the RED algorithm isadjusted based on several relative functions in order to enhance the adaptability toself-similar network traffic and to achieve the congestion control under this traffic.However, the continuous academic debates are carried on about whether thenetwork traffic performes Poisson characteristics (short-range dependence, SRD) orself-similar fractal (long-range dependence, LRD). Although there is a large numberof measuring experiments in kinds of networks so that the conclusion is obtainedwhich networks are self-similar. However, there are also evidences that Poissoncharacteristics still exist. Network traffic models have experienced a SRD, LRD,multi-fractal evolved evolution, and now show a tendency back to SRD. In fact,both SRD and LRD characteristics of networks traffic are coexisted. Hence, in thisresearch a new type of Lévy stochastic process is proposed based on networks withself-similarity and TCP/AQM fluid flow model. Meanwhile, a TCP/AQM fluid flowmodel with wave particle duality is also established, which can describe theself-similarity in networks and the characteristic of a packet driver (SRD) at thesame time. This model tried to unify the two theories on the characteristics of thenetwork traffic, and further to understand the nature of the model with duality,which builds up the foundation for the algorithm design of congestion control infuture based on the self-similar network.In summary, this research mainly focused on the network congestion problem,analyzed a widely applied network traffic model named as TCP/AQM fluid flowmodel. Based on this analysis combined with characteristics of self-similar networks,a series of research approach has been proposed.更多还原