Traffic Management in IP Networks

互联网中存在对服务质量(Qos)的巨大需求。服务质量的一个关键要素是流量管理。由于网络流量是突发性的，如果不对流量进行适当的控制，就很难做出任何服务质量保证。目前，互联网协议(IP)仅具有最低限度的流量管理能力。当队列超过缓冲区容量时，分组被丢弃。传输控制协议(TCP)使用分组丢弃作为拥塞的信号，并减少其负载。虽然这一战略在过去取得了令人满意的效果，但出于两个原因，现在需要更好的战略。首先，很大一部分流量，特别是语音和视频流量不使用TCP。连续媒体流量使用用户数据报协议(UDP)。UDP流量所占比例的增长速度快于TCP流量。UDP流量是拥塞不敏感的，因为UDP源不会响应拥塞而减少其负载。其次，网络的带宽和距离都在增加。对于距离带宽非常高的乘积网络，丢包不是最佳的拥塞指示。在检测和响应分组丢失所需的时间内，可能会丢失几兆字节的数据。因此，需要一种更好的IP网络流量管理策略。互联网工程任务组(IETF)认识到需要直接反馈拥塞信息，提出了一种针对IP路由器的显式拥塞通知方法。当路由器拥塞时，IP报头中的位被设置。ECN比现有路由器使用的简单分组丢弃指示功能强大得多，适用于高距离带宽网络。不幸的是，为了充分发挥ECN的潜力，需要回答几个问题。在这个研究方案中，PI提出了一个全面的IP网络流量管理研究计划。他们建议开发一套新的基于显式拥塞通知机制的IP网络流量管理算法。总共将分析18个不同的问题。私人投资促进机构已为上述每个问题确定了可能的解决办法和办法。具体地说，他们建议为路由器开发一种新的拥塞检测和缓冲区管理方案，这是一种让TCP对来自网络的ECN消息做出反应的机制。本研究的一个重要目标是使基于往返时间和拥塞网关数目的TCP流量管理算法不存在任何偏差。PI的方法将是对简单场景的正式分析，对更复杂场景的启发式分析，以及使用模拟进行验证。重点将是开发简单的解决方案。然而，将从理论上分析以简单性为代价而损失的性能。流量管理是提供服务质量的关键。目前，NSF、DARPA和其他研究资金以及网络方面的精力都花在了服务质量问题上。当部署基于服务质量的解决方案(集成服务、差异化服务或多协议标签交换)时，流量管理的需求将变得明显，PI期望看到对流量管理适当方法的迫切需求。因此，这项建议是及时和重要的。

There is an immense demand for quality of service (QoS) in the Internet. One key element of quality of service is traffic management. Since the network traffic is bursty, it is difficult to make any QoS guarantees without proper control of traffic. Currently, Internet Protocol (IP) has only minimal traffic management capabilities. The packets are dropped when the queue exceeds the buffer capacity. The transmission control protocol (TCP) uses the packet drop as a signal of congestion and reduces its load. While in the past, this strategy has worked satisfactorily, there is need for better strategies for two reasons. First, a large part of the traffic, particularly, voice and video traffic does not use TCP. Continuous media traffic uses User Datagram Protocol (UDP). The proportion of UDP traffic is increasing at a faster pace than TCP traffic. The UDP traffic is congestion insensitive in the sense that UDP sources do not reduce their load in response to congestion. Second, the bandwidth of the networks as well as the distances are increasing. For very high distance-bandwidth product networks, packet drop is not the optimal congestion indication. Several megabytes of data may be lost in the time required to detect and respond to packet losses. Therefore, a better strategy for traffic management in IP networks is required.Recognizing the need for direct feedback of congestion information, the Internet Engineering Task Force (IETF) has come up with an Explicit Congestion Notification (ECN) method for IP routers. A bit in the IP header is set when the routers are congested. ECN is much more powerful than the simple packet drop indication used by existing routers and is suitable for high distance-bandwidth networks. Unfortunately, to realize the full potential of ECN, several questions need to be answered.In this research proposal, the PIs propose a comprehensive program of research on traffic management in IP networks. They propose to develop a new set of traffic management algorithms for IP networks based on Explicit Congestion Notification mechanism. A total of 18 different issues will be analyzed. The PIs have identified potential solutions and approaches for each of these issues. Specifically, they propose to work on a new congestion detection and buffer management scheme for routers, a mechanism for TCP to react to ECN messages from the network. One of the important goals of this research is to make TCP traffic management algorithm free of any bias based on round trip time and number of congested gateways traveled.The proposed research will be based on theoretical analysis and simulations. The PI's approach will be a formal analysis of simple scenarios, heuristic analysis of more complex scenarios and validation using simulations. The emphasis will be to develop simple solutions. However, the performance lost in exchange of simplicity will be theoretically analyzed.Traffic management is the key in providing QoS. Currently, a significant amount of NSF, DARPA, and other research funding as well as energy in networking is being spent on QoS issues. When QoS based solutions (integrated services, differentiated services, or multiprotocol label switching) are deployed, the need for traffic management will become apparent and the PIs expect to see an immediate need for proper methods for traffic management. This proposal is, therefore, timely and important.