Random Early Detection (RED) aims to provide feedback to TCP connections before congestion occurs. It does this elegantly by marking packets probabilistically when the queue size is between an upper and lower water mark, and completely beyond that. The scheme uses simple mechanisms to estimate the running average of the queue length. RED results in small queue sizes and hence limited packet processing latency, as well as fair allocation of router resources (heavier flows probabilistically have more marks).
RED can be used either to drop or mark packets. Marking packets assumes well-behaved citizens in the network who will use the markings to reduce their flows. Dropping packets will require no changes to end-hosts and piggybacks on TCP’s interpretation of a drop as a sign of congestion. Clearly, RED would be “unfair” if nodes do not respond to marked packets.
I think the idea in this paper is neat and elegant, and something that can be easily deployed in the real-world (Is it?). But I was baffled that such a simple, neat idea needed such a long paper! While going to depth is definitely valuable, it might be better to flag the “less important” sections.