ITR/SI(CISE):Optimal and Robust TCP Congestion Control

互联网正在经历一场规模、多样性和覆盖面都前所未有的大变革，通过整合通信、交通、娱乐、公用事业和金融等网络，对我们的科学、社会、经济和政治生活的各个方面产生了深远的影响。这一重要基础设施的稳定性和健壮性需要一个严格的理论来理解当前的协议，并发展它们以应对新出现的挑战。本文提出了一种TCP拥塞控制理论，并将其应用于TCP拥塞控制，以提高现有协议的稳定性、鲁棒性和最优性。不同的TCP和主动队列管理（AQM）方案对应于不同的效用函数和不同的算法来最大化它们。首先，我们将发展一个新的TCP拥塞控制的理论模型的基础上对偶优化和多变量鲁棒控制。该理论将阐明Tahoe、里诺和Vegas等源算法以及DropTail、RED和REM等主动队列管理算法在网络控制中的作用，并建立当前协议的性能极限;将解释时延、拓扑、容量和负载增加对稳定性的影响;将提供TCP/AQM算法的反馈稳定性不受这些影响的条件。事实上，这样的理论已经在我们最近的工作中出现，即使在目前的初步阶段，它已经提供了对当前协议的一些广泛观察到的性能和公平性行为的基本理解，并揭示了新的和令人惊讶的稳定性问题。例如，当网络容量增加时，现有的协议会变得不稳定，并产生扩展分支。此外，保持稳定的capacityscales任意施加了严重的约束源如何调整他们的速率（TCP）和什么拥塞信息反馈（AQM）。目前的协议不满足这种稳定不变性的条件，因此可能不适合未来的网络，在应用需求的拉动和技术进步的推动下，容量将是巨大的。我们的研究的第二个组成部分是在理论的基础上设计实用的TCP和AQM协议，并开发原型和实验来证明它们的有效性。我们将使用该理论来识别当前协议中延迟、网络规模、容量和流量负载增加时的不稳定性来源。我们将设计两个增强功能，逐步发展目前的协议，并彻底新的协议，具有强大的理论承诺的稳定性。作为我们的算法的一个具体应用，我们将应用它们来提高无线链路上的TCP性能，这既是因为它们无处不在，也是因为它们可能仍然是未来网络中最重要的瓶颈。

Internet is undergoing an overhaul unprecedented in size, diversity, and reach, with profound im-pact in all aspects of our scientific, social, economic and political life through the integration ofnetworks of communication, transportation, entertainment, utilities, and finance. The stabilityand robustness of this vital infrastructure demands a rigorous theory to understand the currentprotocols and evolve them to meet emerging challenges. We propose to develop such a theory forTCP congestion control, and use it to drastically improve the stability, robustness and optimalityof the current protocols.A key insight is to view congestion control as a distributed asynchronous computation to maxi-mize aggregate source utility over the Internet; different TCP and active queue management (AQM)schemes correspond to different utility functions and different algorithms to maximize them.Our research hastwo components. First, we will develop a new theoretical model of TCPcongestion control based on duality in optimization and multivariate robust control. The theory willclarify the role of source algorithms, such as Tahoe, Reno and Vegas, and active queue management,such as DropTail, RED and REM, in the control of networks and establish performance limits ofthe current protocols; it will explain the effect on stability when delay, topology, capacity, and loadscale up; and it will provide conditions under which the feedback stability ofTCP/AQM algorithmsare invariant to these effects. Indeed, such a theory is already emerging from our recent works.Even in its currently preliminary stage, it already provides a fundamental understanding on somewidely observed performance and fairness behavior of the current protocols, and uncovers new andsurprising stability problems. For example, it shows that the current protocols become unstable andexhibit bifurcation when network capacity increases. Moreover, maintaining stability as capacityscales up arbitrarily imposes severe constraints on how sources adjust their rates (TCP) and whatcongestion information is fed back (AQM). The current protocol does not satisfy the conditionfor such stability invariance, and hence may be ill suited for future networks where, pulled byapplication demand and pushed by technological advances, the capacity will be large.The second component of our research is the design of practical TCP and AQM protocolsbased on the theory, and the development of prototypes and experiments to demonstrate theireffectiveness. We will use the theory to identify the sources of instability in the current protocolswhen delay, network size, capacity, and traffic load scale up. We will design both enhancements thatincrementally evolve the current protocols, and drastically new protocols that have the stronglyrobust stability property promised by theory. As a concrete application of our algorithms, we willapply them to improve TCP performance over wireless links, both because they are ubiquitous andbecause they are likely to remain the most important bottlenecks in future networks.