Cyberspace weather. Internet delays, as measured round trip between Austin, Texas, and hosts around the United States.

Net Hogs Slow the Web

Any regular reader of ScienceNOW has experienced the spates of congestion that afflict the Internet. As computers send volumes of data from server to server, phone lines fill up, causing Internet traffic jams--and making Web browsers chug away fruitlessly. Then, moments later, the congestion abates. In today's issue of Science,* two physicists present mathematical and computer models that explain these Internet "storms." The models show how millions of users who have no incentive to economize clog the Internet with data and then get discouraged, relieving the congestion--all at roughly the same time.

The researchers, Bernardo Huberman and Rajan Lukose of the Xerox Palo Alto Research Center in California, started by recognizing that the Internet tends to be overused because most users pay a flat rate for unlimited access. As a result, people have no incentive to limit the size of their downloads, their Web meanderings, their e-mail, or their Internet chatting. This collective display of self-interest, says Lukose, leads to "overusing and degrading the value of resources. That's the tragedy of the commons." But unlike the gradual deterioration of other common resources--for example, the atmosphere or the oceans--the Internet's congestion is sporadic. "There are short spikes of congestion," says Huberman, "on the order of seconds or tens of seconds."

To explain this behavior, he and Lukose created a model of Internet use in which each user behaves rationally, overusing the Internet most of the time but logging off when congestion becomes too great. The model predicted the statistical properties of the network delays as many users logged on and off. Most of the time, the delays were small. But every so often the delays spiked in an Internet storm as a large number of users put a load on the system at the same time. Huberman and Lukose tested these predictions by timing how long it took to send packets of data from Stanford to England and back. Sure enough, they measured spikes of congestion distributed roughly the way the model predicted.

"You look at the Internet and say, 'My god, it's a mess; nobody's going to understand it,' but Huberman gets qualitative insights into very complicated problems," says Kenneth Steiglitz, a computer scientist at Princeton University. Explaining Internet storms may prove easier than controlling them, however. Huberman thinks the answer will lie in new pricing schemes, such as a pay-per-packet scheme or a priority-pricing method (the Internet equivalent of Federal Express). He hopes to put his model to work studying how various pricing schemes would affect congestion. But one thing is already clear, he says. "I don't think the idea of a [flat-rate] Internet will go on forever."

* For more details, Science Online subscribers can link to the full text of the Report.