Design, Theoretical Validation, and Empirical Evaluation of a Class of Multicast Congestion Control Schemes: Marriage of Feedback Control Theory and Multicast Congestion Control

研究人员观察到，在互联网上向多个接收者传输连续媒体应用程序的情况出现了爆炸性增长。例如，连续媒体服务器、数字图书馆、远程医疗诊断和远程学习。由于大多数基于Internet连续媒体的组播应用不支持端到端的拥塞控制，这些应用的广泛应用会产生严重的负面影响，从自主控制的TCP流的饥饿到拥塞崩溃的可能性。因此，该方案的主要目的是设计、实现、分析验证和经验性地评估一类具有以下设计目标的组播拥塞控制方案：可伸缩性，以分析可证明的方式调整源发送速率以实现TCP友好和(加权)公平性的能力，处理相同分组的独立丢失的能力，处理动态业务/成员变化的能力，以及最小的路由器支持。具体地说，研究人员将为基于速率的组播拥塞控制制定一个分析框架。特别是T1。研究人员提出了一种新的ACK聚合方法，以便为发送者提供关于组播树中拥塞状况的简单但全面的视图。T2研究人员将拥塞状态描述为三个阶段：无拥塞、拥塞警报和拥塞避免，并将设计一种简单而有效的机制，供发送方根据收到的聚合确认中的参数和本地保存的参数来诊断其连接处于哪个拥塞阶段，以及是否发生了相同分组的独立丢失。T3。研究人员将基于鲁棒反馈控制理论设计出一类速率调整方案，它(I)对TCP友好，(Ii)鲁棒稳定，(Iii)在竞争的组播会话之间实现(加权)公平，特征(Ii)对于具有反馈环路的系统尤其重要，但在大多数组播拥塞控制工作中可能除了[10]之外还没有得到广泛的解决。因此，所提出的控制理论工作将围绕基于速率的多播流量控制中出现的鲁棒稳定性和干扰抑制问题展开。T4.为了促进所提出的方案的广泛部署，研究人员将调查是否以及如何在终端主机上实现操作(可能除了确认聚合)。如果某些操作在没有路由器支持的情况下无法完成，研究人员将研究可能的轻量级实现方法。T5。为了经验性地评估所提出的组播拥塞控制方案的性能，研究人员将在ns-2[81]中进行广泛的仿真来测试它们的TCP友好性、公平性和可扩展性，并将它们与已有的工作进行比较。在OARnet(Internet2的俄亥俄州GigaPOP经理)人员的帮助下，研究人员还将在FreeBSD中实施所提出的方案，并使用以从MR和CT获取的大量患者图像为中心的数据库服务器作为代表性的组播应用，以经验性地测量Internet2上原型软件的关键功能、可伸缩性和自适应特性。

The researchers have observed an explosive growth in transporting continuous media applications to multiple recipients on the Internet. Examples include continuous media servers, digital libraries, remote medical diagnosis, and distance learning. As most Internet continuous media based multicast applications do not support end-to-end congestion control, wide deployment of these applications have severe negative impact, ranging from starvation of self-controlled TCP flows to the potential for congestion collapse. The main intent of this proposal is thus to design, implement, analytically validate, and empirically evaluate a class of congestion control schemes for multicasts with the following design objectives: scalability, capability to adjust source sending rates to achieve TCP-friendliness and (weighted) fairness in an analytically provable manner, capability to handle independent losses of the same packet, capability to deal with dynamic traffic/membership changes, and minimal router support. Specifically, the researchers will lay out an analytical framework for rate-based multicast congestion control. In particular, T1. The researchers propose a novel approach for ACK aggregation so as to provide the sender with a simple but comprehensive view of congestion conditions in the multicast tree. T2. The researchers characterize the congestion status with three phases: congestion free, congestion alert, and congestion avoidance, and will devise a simple, yet effective mechanism for a sender to diagnose, based on the parameters in the aggregated acknowledgment received and parameters locally kept, which congestion phase its connection is in and whether or not independent losses of the same packet has occurred. T3. The researchers will devise, based on robust feedback control theory, a class of rate adjustment schemes that are (i) TCP-friendly, (ii) robustly stable, and (iii) achieve (weighted) fairness among competing multicast sessions, feature (ii) is especially important for systems with feedback loops, but has not been extensively addressed in most of the multicast congestion control work perhaps except [10]. The proposed control theoretic work will hence center around robust stability and disturbance attenuation problems that appear in rate-based flow control for multicasts. T4. To promote the wide deployment of proposed schemes, the researchers will investigate whether or not, and how, the operations (perhaps except acknowledgment aggregation) can be realized at end hosts. If some of the operations cannot be accomplished without router support, the researchers will look into possible light-weight implementation methods. T5. To empirically evaluate the performance of proposed multicast congestion control schemes, the researchers will conduct extensive simulation in ns-2 [81] to test their TCP-friendliness, fairness, and scalability properties and compare them against existing work. With the help of OARnet (the Ohio GigaPOP manager of Internet2) personnel,the researchers will also implement the proposed schemes in FreeBSD, and use a database server centered on large amounts of patient images acquired from MR and CT as a representative multicast application to empirically measure the key functional, scalable, and adaptive characteristics of the prototype software over the Internet2.