ITR: Active Queue Management for Scalable Network Services: Theory and Internet Practice

当今Internet面临的两个最大问题是：(1)为需要某种形式的带宽可用性和/或端到端延迟的“保证”的应用程序提供服务质量问题，以及(2)避免传统的最佳努力流之间的拥塞问题。分组调度一直是传统上用于服务质量保证的机制，而传输层的终端系统适应性一直是拥塞控制的主要形式。通常情况下，拥塞控制和服务质量的问题在很大程度上是作为独立的问题来解决的，并且为每个问题开发了单独的机制。这个项目的本质是研究拥塞控制和服务质量之间的相互作用，以及为实现这两者所采用的机制。本文的出发点是研究在网络路由器中使用主动队列管理（AQM）技术来为IP流提供拥塞控制和服务质量。主动队列管理是指在路由器的出站网络接口上操纵数据包队列以使网络流的性能产生偏差的实践。例如，从队列中丢弃数据包是一种主动队列管理机制，用于RED（随机早期检测）拥塞避免机制。对于单个AQM机制如何影响网络流的性能，目前还没有任何基本的理论认识。从Internet服务提供商的角度来看，这个问题由于对如何调优AQM机制以实现特定性能目标的理解很少（如果有的话）而更加复杂。AQM机制对更广泛的网络性能度量的影响是未知的，因为分析AQM性能的框架尚不存在。这与其他服务质量的方法（如数据包调度）形成对比。数据包调度器已经被仔细地分析过，并且存在用于计算性能指标（如延迟、延迟抖动、数据包丢失等）边界的框架。该项目的主要目标是发展理论基础和工程原则，以指导互联网AQM的构建和部署。这需要对AQM机制进行分析和实证研究，以便了解使用AQM进行拥塞控制和服务质量的成本、收益和可伸缩性限制。该项目还有助于对路由器资源管理原理的全面理解。此外，还开发了新的AQM方案，为服务质量和拥塞控制提供了新的可扩展解决方案。具体目标包括：开发一个分析框架来理解AQM机制的性能。一个新颖的方面是考虑以网络为中心的性能指标（如链路利用率）和以终端系统或用户为中心的指标（如响应时间）。2. 开发和分析新的AQM和混合AQM包调度方案，实现一系列服务质量和拥塞控制服务。这里的一个关键贡献是演示了在可伸缩的服务实现中使用分组调度与AQM的成本和收益。3. 通过管理研究网络和支持先进、实时、分布式虚拟环境应用的案例研究，展示AQM方案的有效性。该项目的结果可能对互联网的发展产生重大影响，以支持新的服务质量水平和控制拥塞，同时确保大量用户和设备的可扩展性。

Two of the largest issues facing the Internet today are the problems of (1) providing quality-of-service to applications that require some form of "guarantee" of bandwidth availability and/or end-to-end delay, and (2) the problem of avoiding congestion between traditional best-effort flows. Packet scheduling has been the mechanism traditionally used for quality-of-service guarantees while end-system adaptations in the transport layer have been the dominant form of congestion control. Typically, the problems of congestion control and quality-of-service have been addressed as largely independent concerns and separate mechanisms have been developed for each . The essence of this project is to study the interplay between congestion control and quality-of-service and the mechanisms that have been employed for realizing each. The starting point is an investigation of using active queue management(AQM) techniques in network routers to provide both congestion control and quality-of-service for IP flows. Active queue management refers to the practice of manipulating the queue of packets at an outbound network interface on a router to bias the performance of network flows. For example, discarding packets from the queue is an active queue management mechanism that is used in the RED (random early detection) congestion avoidance mechanism. There does not yet exist any fundamental theoretical understanding of how individual AQM mechanisms effect the performance of network flows. From the perspective of Internet service providers this problem is compounded by the fact that there is little, if any, understanding of how AQM mechanisms can be tuned to realize specific performance goals. The impact of AQM mechanisms on broader measures of network performance is unknown because a framework for analyzing AQM performance does not yet exist. This is in contrast to other approaches for quality-of-service such as packet scheduling. Packet schedulers have been carefully analyzed and frameworks exist for computing bounds on performance metrics such as delay, delay-jitter, packet loss, etc. The main goal of this project is to develop the theoretical underpinnings and engineering principles to guide the construction and deployment of AQM for the Internet. This requires an analytic and empirical study of AQM mechanisms for the purpose of understanding the costs, benefits, and scalability limitations of using AQM for congestion control and quality-of-service. The project also contributes to the overall understanding of principles of resource management in routers. In addition, novel AQM schemes that provide new, scalable solutions for quality of service and congestion control are developed.Specific objectives include the following: 1. Develop an analysis framework for understanding the performance of AQM mechanisms. A novel aspect is consideration of both network-centric performance metrics such as link utilization, and end-systemor user-centric measures such as response time. 2. Develop and analyze novel AQM and hybrid AQM-packet scheduling schemes that realize a spectrum of quality-of-service and congestion control services. A key contribution here is a demonstration of the costsand benefits of using packet scheduling v. AQM for scalable implementations of services. 3. Demonstrate the effectiveness of the AQM schemes through a case study in managing a research network and supporting an advanced, real-time, distributed-virtual-environment application. The results of this project can have a significant impact on the evolution of the Internet to support new levels of service quality and control congestion while ensuring scalability for very large numbers of users and devices.