Using Fluid Models for AQM Evaluation

Abstract

In spite of the congestion management mechanisms included in the TCP/IP protocol, the Internet still suffers from a lack of optimization in the way network overload is managed. One of the reasons for this is that the original congestion avoidance techniques specified by TCP/IP are implemented at the edges of the network. This means that the critical mission of reacting to link saturation relies on hosts that don't have a global view of the network. Moreover they do not have any obligation to respect a set of common rules. One possible solution to this problem is to introduce congestion management inside the core routers of the network. One way of doing this, Active Queue Management, consists of designing intelligent algorithms that start dropping packets before the routers' queues get full, so that they are always ready to accommodate bursts of traffic.

The rapid growth of the Internet has presented many challenges for those wishing to design and validate new protocols. Almost all such advances are evaluated through simulation (e.g. using the ns packet-level simulator). Because the internet is growing very quickly, researchers had to invent new methods to simulate its activity. Indeed, it has been quite a few years since the processing capacity of regular packet-level network simulators was sufficient to simulate the behavior of a network like the Internet. One alternative is to simulate the network traffic at a flow-level. Such a higher-level simulator would be a useful tool for network protocol design, because this kind of simulator is able to scale a lot more easily than a packet-level one. This project looks at a fluid-based model for flow-level network simulation, which can be easily extended to support a variety of Active Queue Management schemes. The work consisted of validating, optimizing and extending an implementation of this model, while also making sure that it is suitable for evaluating Active Queue Management techniques. Ultimately, the simulator was successfully used to evaluate a number of these schemes.