Congestion Avoidance and Control

This paper describes congestion control techniques for TCP built around reaching an equilibrium between packets entering and leaving the network and maintaining that equilibrium:

reaching the equilibirum

slow start — ramp up sending speed slowly so the correct sending speed won't be overshot. Prior systems would not recover from a mismatch in their local network rate and the least capable link to their end host because they would continuous overshot.

linear window size increase — be conservative about increasing the amount of data in flight to avoid massively overdriving the network, from which it is hard to recover.

don't retransmit until packets is out of the network

RTT variance estimate — because of queues, etc. round-trip time varies in the network a lot. This must be taken into account, which the authors do by creating an estimator of RTT variance.

backoff after retransmission enough to avoid more losses

exponential backoff — motivated by some theoretical results (about what sort of backoff is required for stability) and a desire to be as general as possible with what network topologies they support, the authors choose exponential backoff.

Despite the authors attempt at generality in these algorithms, since as they note there were TCP implementations for systems ranging from 56Kbit/s to hundreds of Mbit/s, there are some interesting assumptions about the environment. The most glaring one is that losses due to damage are are rare, which seems a strange assumption when one is considering microwave links, etc. It is hard to see, though, how this sort of congestion control (only at the ends and where loss types are not distinguishable) could function very well without assuming that at least most losses are due to congestion.