Network traffic gets in trouble primarily when a high-speed segment meets a low-speed segment. IV was designed to run over links with low throughput rates--low by today's standards, that is. The variation between the slowest links and the fastest ones was perhaps the difference between no bits/sec and 2,400 bits/sec--a ratio of slightly more than 20 to 1.

Compare that ratio with one you might observe on a typical router configuration today: A router with a 56Kbit/sec frame relay interface and a 100Mbit/sec LAN interface has a ratio of approximately 2,000 to 1. The same router with a 1Gbit/sec LAN interface has a ratio of 20,000 to 1. If a thin pipe fills up and a fat pipe full of traffic is destined for the slow interface, the only thing the router can do is start throwing packets away or putting them into a buffer or queue. If the throughput mismatch is large, the buffers had better be large.

Large buffers, however, introduce other problems. Queuing can cause delays in a flow, which some types of traffic--live audio and video, for example-cannot tolerate. Furthermore, the jitter characteristics of a flow can deteriorate because of buffering. This trouble with queuing led some developers to consider more subtle ways of smoothing out the flow of network traffic. The resultant technique for this is TCP rate control.

Rate Control and Queuing There are two ways that bandwidth monitor can control network traffic: rate control and queuing. Rate control is designed to improve the default behavior of TCP connection endpoints. It identifies traffic flows according to their destinations, sources, TCP ports, or some other characteristic. Rate control products calculate what the values of TCP window sizes and other parameters ought to be to make a traffic flow perform within the desired constraints; they then plug those values into the packets as they go through the device. Traffic flows that would, by default, nonchalantly hog the available pipe can be throttled down so that they leave room for other flows.

Enforcing explicit limits on certain flows will often improve the jitter characteristics of the flows, as well as the delay and jitter of other flows using the same link. One of the principal advantages of the rate-control approach is that the bandwidth management device is made aware of the full end-to-end-and-back path of traffic (at least with TCP flows), and can to some extent anticipate and smooth out throughput changes.

Queuing, which creates multiple queues and allocates priorities among them according to some algorithm, is the other approach to bandwidth management. For example, high-priority traffic flows will go to a queue with the least backlog, while lower-priority flows will find themselves in a large, slow queue. The trick with this approach is to refrain from blocking flows that are worthy of best-efforts service, while providing the greatest possible throughput to high-priority flows. Some of the queuing techniques that solve this problem are weighted round robin, deficit weighted round robin, Weighted Fair Queuing (WFQ), and Class-Based Queuing (CBQ).