【作者】 郑宏云；

【导师】 陈常嘉；

【作者基本信息】 北京交通大学 ， 通信与信息系统， 2007， 博士

【摘要】 目前Internet进行资源分配和拥塞控制的时候，通常做如下假设：Internet是面向无连接的分组交换网络，提供尽力传递的网络服务，当分组丢失概率与流速率无关时TCP端到端的拥塞控制机制能够保证网络收敛到平稳状态，获得有效率而且公平的资源分配。但是，上述假设条件并非在所有的网络情形下都满足。在网络中，的确存在着一些情形，即当分组丢失概率与流速率相关时，TCP难以收敛到平稳状态；也存在着新的技术方式，诸如光突发交换不再使用传统的Internet分组交换工作方式来实施网络资源的分配。论文工作围绕上述两个方面的若干问题展开。一方面，研究了在分组丢失概率与流速率相关的两种网络情形下TCP的拥塞控制特性；另一方面，研究了在光突发交换网络中信道优化调度问题。这些问题不同于以往的资源分配和拥塞控制问题，对传统的以带宽为目标的资源分配和拥塞控制机制提出了挑战。本论文对上述问题进行了深入探讨，并给出确实可行的技术解决方案。论文的主要贡献包括以下三点。研究了TCP在交换结构中的拥塞控制性能。指出在交换结构的环境中，TCP拥塞控制存在着不同寻常的新问题：单靠传统的TCP／队列管理机制无法兼顾资源分配的有效性和公平性。分析了丢尾和丢头队列管理机制与TCP配合进行拥塞控制失效的原因。提出把交换结构的仲裁机制和对输入缓存的队列管理结合起来的基本思想，并给出一种简单的启发式公平仲裁和随机提前丢头算法，通过不均匀地仲裁分组转发和随机地提前丢弃队头分组来兼顾资源分配的有效性和公平性。仿真结果表明算法确实有效。研究了TCP在光突发交换网络中的拥塞控制性能。指出在光突发交换网络中，TCP吞吐量存在着极大的不公平性，原因在于光突发交换网络中突发丢失特性与TCP速率回退机制的相互作用，即不同速率的流经历不同概率的突发丢失，速率小的流总是比速率大的流经历更多的突发丢失，而突发丢失常常使TCP由于超时而进入慢启动，从而加重流之间的吞吐量差异。分析TCP进入超时慢启动的原因。给出突发丢失的数学模型解释TCP吞吐量的不公平性。分析偏置时间对流速率的控制作用，提出一种自适应偏置时间算法，根据流的当前速率动态地设置偏置时间，让速率大的流具有小的偏置时间，仿真实验结果表明算法达到预期目标。研究了光突发交换网络的信道优化调度。发现离线调度的OBS-GS算法的信道利用率甚至劣于LAUC算法。提出一种以信道利用率为优化目标的信道优化调度算法，将信道调度问题转换为图的着色问题来求解。将这个问题描述为整数线性规划形式，利用间隔图的优良特性可在多项式时间内求解。给出可适用于任意波长个数的信道优化调度问题的通用的等效图论描述形式。讨论了对短突发的性能歧视。数学分析给出提出的优化算法所带来的信道利用率增益的上下限。更多还原

【Abstract】 Generally we consider resource allocation and congestion control issues in the Internet assuming that it is a connectionless, packet-switched network with best effort service; and TCP end-to-end congestion control mechanism makes the Internet converge to steady state with effective and fair resource allocation assuming that packet loss probability is not a function of flow rates. However, in no all scenarios of network do assumptions above come into existence. On the one hand, there indeed exist some network situations where packet loss probability is a function of flow rates and it is difficult for TCP to converge to steady state. On the other hand, some new network technologies, for optical burst switching (OBS) as an example, use different ways to allocate network resources from those used in packet-switched network like Internet.In this dissertation several problems involved the two hands are studied. For the first hand TCP congestion control performance in two network situations where packet loss probability is a function of flow rates is studied. For the second hand the channel scheduling problem in OBS network is discussed. These problems are quite different from those existing, thus the traditional resource assignment and congestion control mechanism, in which link bandwidth and buffer space in routers are viewed as the target resources, will be challenged. In this dissertation the problems above are looked into, and relevant technology solutions are carried out. The contributions are three-fold as following.The first contribution is that TCP performance in the input-queued switch is studied for the first time. It is found that the TCP congestion control performance in the input-queued switch is different from traditional one, and that it is impossible to achieve efficiency and fairness among TCP flows at the same time only by queue management. The reasons why drop tail and drop front queue management mechanisms do not work are analyzed. The queue management and switching arbitrating must be in conjunction, and a scheme, which combines heuristic fair arbitrating (AFS) and queue management mechanism of early drop front randomly (rEDF), is proposed. In proposed scheme, switch arbitration strategy of hFS unevenly allows input ports to transfer packets to output ports while packets at head of any other input ports involved in conflicts have to be dropped by the policy of rEDF with a probability. Simulation results prove that the proposed scheme can achieve better tradeoff between throughput and fairness. The second contribution discusses TCP congestion control performance in OBS network. It is found that significant throughput unfairness exists among TCP flows that share the OBS network through edge nodes. The cause is that the interaction of burst dropping and TCP transmitting rates back-off mechanism. In other words, in OBS network flows with different rates undergo different burst dropping probability, i.e., a TCP flow with a lower rate will undergo more burst dropping than one with a higher rate, and burst dropping always makes TCP enter slow start due to retransmission timeout, which worsen throughput difference between flows with different rates. The reasons why TCP does retransmission timeouts are analyzed. Burst dropping models are proposed to explain TCP throughput unfairness. It is observed that offset time will be a good choice to control TCP fairness, and then an adaptive offset time scheduling (AOS) algorithm is proposed. AOS assigns bursts offset time value adaptive to flow rate, i.e., makes a burst with shorter offset time for the recent flow rate increasing. The simulation results show that the fairness can be significantly improved by proposed AOS scheme.For the third contribution, the optimization problem about channel scheduling in OBS network is studied. It is found that for the channel utilization the offline OBS-GS algorithm is even worse than online LAUC algorithm. An optimal scheduling scheme is proposed to maximum channel utilization. The problem of optimum channel scheduling is mapped to a problem of colouring graph, which is formulated as an integer linear programming problem solvable in polynomial time by making use of interval graph good characteristics. Discrimination to smaller burst and the alleviation with adjusted weight are discussed. Analysis results present the minimal low bound and maximal bound of channel utilization of proposed schemes.更多还原