Active Queue Management is the process of anticipating congestion at the router level in an attempt to reduce the queue size and increase throughput. This can be done only if the routers can communicate network congestion states to the end users by either dropping or marking packets. This paper designs and analyzes an algorithm called the Adaptive Virtual Queue (AVQ), which addresses the problem of how to mark or drop packets in an intelligent manner so the that the current state of the network can be conveyed to the end users.
AVQ is a virtual queue that immitates a real queue. Packets can be enqueued and dequeued, and they are dropped when the capacity of the virtual queue is exceeded. When a packet in the virtual queue is dropped, the corresponding packet in the real queue is either dropped or marked. The capacity of the virtual queue is less than that of the real queue, and that virtual capacity can be updated based on a desired utilitization ratio (< 1), a value that represents the total flow in to the link, and a smoothing factor.
On a mathematical level, the paper shows that parameters can be selected to keep the system stable. Several theorems are stated to show how to select parameters. At a stable level, simulations show that AVQ outperforms many other AQM schemes. This means fewer packet drops and better utilization. It is important to note that AVQ works well in situations where there are a lot of small network flows. As the current Internet is predominately HTTP flows, how an AQM scheme deals with them is especially relevant.
I give the paper a rating of 4 for significant contribution because it gives a stable queue management scheme that performs better than other AQM schemes. It also addresses the situation of small network flows, which previous AQM scheme did not do. I think the simulations are good, though they are limited in scope, addressing simultaneously at most only two types of flow traffic. The use of TCP-Reno as the dominant TCP implementation may not reflect the current state of the network, especially because TCP-NewReno and TCP-SACK are more dominant now. What I don't understand, also, is why the authors used TCP-Reno on all experiments except in one, where he used TCP-NewReno. This is not consist. In general, I think this is a good paper that presents a good AQM scheme that works well with small network flows.