In October 1986 the early Internet suffered something that sounds like a traffic jam on a highway. Data throughput between two sites on the same campus, the Lawrence Berkeley Laboratory and UC Berkeley, only 400 yards apart, fell from 32 kilobits per second to 40 bits per second. The authors describe a drop by a factor of a thousand. This was the first of a series of congestion collapses.
Van Jacobson and Michael Karels investigated why. The problem was not broken hardware. It was that the protocol everyone used, TCP, reacted to trouble by sending more. Their fixes, published in 1988, are in almost every TCP stack today and are a model of how to think about shared resources.
When the network is overloaded, a sender should send less, not more. Start slowly and grow (slow start), probe gently for more capacity (add one packet per round), and cut the sending rate sharply when a packet is lost. Packets leaving the network should pace the packets entering it.
Why networks collapse
Picture a road where drivers who hit a jam respond by driving faster to get through. The jam gets worse. In the old TCP, a sender whose retransmit timer was set too short resent packets that were merely delayed. Under load, those duplicates filled the already full links, making real loss more likely.
The authors build their cure on one principle they call conservation of packets. A connection in equilibrium, with a full window of data in transit, should put a new packet into the network only when an old one leaves. Acknowledgements coming back are the clock that tells the sender a packet has left.
They find three ways this principle can fail: the connection never reaches equilibrium, a sender puts a new packet in before an old one has left, or the equilibrium cannot be maintained because of resource limits along the path. Each failure gets a fix.
