This section is meant as an introduction to the queuing at backbone networks, which often involves >100 megabit bandwidths, which requires a different approach than your ADSL modem at home.
The normal behaviour of router queues on the Internet is called tail-drop. Tail-drop works by queueing up to a certain amount, then dropping all traffic that 'spills over'. This is very unfair, and also leads to retransmit synchronization. When retransmit synchronization occurs, the sudden burst of drops from a router that has reached its fill will cause a delayed burst of retransmits, which will over fill the congested router again.
In order to cope with transient congestion on links, backbone routers will often implement large queues. Unfortunately, while these queues are good for throughput, they can substantially increase latency and cause TCP connections to behave very burstily during congestion.
These issues with tail-drop are becoming increasingly troublesome on the Internet because the use of network unfriendly applications is increasing. The Linux kernel offers us RED, short for Random Early Detect, also called Random Early Drop, as that is how it works.
RED isn't a cure-all for this, applications which inappropriately fail to implement exponential backoff still get an unfair share of the bandwidth, however, with RED they do not cause as much harm to the throughput and latency of other connections.
RED statistically drops packets from flows before it reaches its hard limit. This causes a congested backbone link to slow more gracefully, and prevents retransmit synchronization. This also helps TCP find its 'fair' speed faster by allowing some packets to get dropped sooner keeping queue sizes low and latency under control. The probability of a packet being dropped from a particular connection is proportional to its bandwidth usage rather than the number of packets it transmits.
RED is a good queue for backbones, where you can't afford the complexity of per-session state tracking needed by fairness queueing.
In order to use RED, you must decide on three parameters: Min, Max, and burst. Min sets the minimum queue size in bytes before dropping will begin, Max is a soft maximum that the algorithm will attempt to stay under, and burst sets the maximum number of packets that can 'burst through'.
You should set the min by calculating that highest acceptable base queueing latency you wish, and multiply it by your bandwidth. For instance, on my 64kbit/s ISDN link, I might want a base queueing latency of 200ms so I set min to 1600 bytes. Setting min too small will degrade throughput and too large will degrade latency. Setting a small min is not a replacement for reducing the MTU on a slow link to improve interactive response.
You should make max at least twice min to prevent synchronization. On slow links with small Min's it might be wise to make max perhaps four or more times large then min.
Burst controls how the RED algorithm responds to bursts. Burst must be set larger then min/avpkt. Experimentally, I've found (min+min+max)/(3*avpkt) to work ok.
Additionally, you need to set limit and avpkt. Limit is a safety value, after there are limit bytes in the queue, RED 'turns into' tail-drop. I typical set limit to eight times max. Avpkt should be your average packet size. 1000 works OK on high speed Internet links with a 1500byte MTU.
Read the paper on RED queueing by Sally Floyd and Van Jacobson for technical information.