Congestion in a network is loosely defined as a condition where demand for resources (usually bandwidth or CPU time) exceeds capacity. Congestion avoidance tries to prevent demand from exceeding capacity, while congestion recovery tries to restore an operative state. It is possible for a router to contribute to both of these mechanisms. A great deal of effort has been spent studying the problem. The reader is encouraged to read [FORWARD:2] for a survey of the work. Important papers on the subject include [FORWARD:3], [FORWARD:4], [FORWARD:5], [FORWARD:10], [FORWARD:11], [FORWARD:12], [FORWARD:13], [FORWARD:14], and [INTERNET:10], among others.
The amount of storage that router should have available to handle peak instantaneous demand when hosts use reasonable congestion policies, such as described in [FORWARD:5], is a function of the product of the bandwidth of the link times the path delay of the flows using the link, and therefore storage should increase as this Bandwidth*Delay product increases. The exact function relating storage capacity to probability of discard is not known.
When a router receives a packet beyond its storage capacity it must (by definition, not by decree) discard it or some other packet or packets. Which packet to discard is the subject of much study but, unfortunately, little agreement so far. The best wisdom to date suggests discarding a packet from the data stream most heavily using the link. However, a number of additional factors may be relevant, including the precedence of the traffic, active bandwidth reservation, and the complexity associated with selecting that packet.
A router MAY discard the packet it has just received; this is the simplest but not the best policy. Ideally, the router should select a packet from one of the sessions most heavily abusing the link, given that the applicable Quality of Service policy permits this. A recommended policy in datagram environments using FIFO queues is to discard a packet randomly selected from the queue (see [FORWARD:5]). An equivalent algorithm in routers using fair queues is to discard from the longest queue or that using the greatest virtual time (see [FORWARD:13]). A router MAY use these algorithms to determine which packet to discard.
If a router implements a discard policy (such as Random Drop) under which it chooses a packet to discard from a pool of eligible packets:
Advanced methods of congestion control include a notion of fairness, so that the 'user' that is penalized by losing a packet is the one that contributed the most to the congestion. No matter what mechanism is implemented to deal with bandwidth congestion control, it is important that the CPU effort expended be sufficiently small that the router is not driven into CPU congestion also.
As described in Section [4.3.3.3], this document recommends that a router SHOULD NOT send a Source Quench to the sender of the packet that it is discarding. ICMP Source Quench is a very weak mechanism, so it is not necessary for a router to send it, and host software should not use it exclusively as an indicator of congestion.