August 23, 1999

Multimedia Transmissions Drive Net Toward Gridlock

By SARA ROBINSON

ike a nation's highway system, the Internet has traffic rules that prevent chaos and gridlock. But an increasingly popular type of multimedia traffic is turning out to be a cyber road hog that flouts the rules and creates traffic jams.

This scofflaw traffic, known as streaming media, is widely hailed as the future of home entertainment, capable of eventually delivering high-quality audio and video programming over the Internet.

Unfortunately, network experts warn, it may drive everyone else off the road.

Streaming media is analogous to broadcast media in that the audio or video material is produced as soon as a computer receives the data over the Internet. The problem is that the streaming data does not respond to network congestion in the way other Internet traffic does.

When a computer sending conventional data encounters congestion, it significantly slows its own transmission rate, but a computer sending streaming data will reduce the flow only slightly. So if streaming traffic competes with conventional traffic for the same congested strip of roadway, the streaming traffic, like some VIP motorcade, assumes the right of way and lets all other data traffic pile up.

The difference, said Van Jacobson, chief scientist for Cisco Systems Inc., a leading maker of Internet traffic routers, is that conventional traffic "is polite; this stuff is impolite."

"People writing software for that traffic -- they don't care," Jacobson said of streaming media. "In the long term, it's a problem."

By its very nature, streaming media has to flow continuously to the user's computer, so it cannot follow the same traffic rules as conventional data. But even so, it is possible for packets of streaming data to interact civilly with other traffic on the Internet. The reason they do not, Jacobson said, is that streaming media providers have no incentive to comply with traffic rules.

Today, he said, "if Real Networks is polite and Microsoft isn't, then Real looks crummy."

Even elbowing all other data aside, today's streaming media produces a very low quality of entertainment most of the time.

Much of that lack of quality today is a result of slow modems at the user's end. In three to five years, when cable modems and souped-up digital telephone lines are expected to be common, most Internet users may be listening to live Webcasts or playing high-quality radio on their computers. In 10 years, movies and commercial television might very well be carried over Internet channels.

This increasing demand will add vast amounts of streaming traffic to the Internet and could lead to what Jacobson calls "congestion collapse" -- the Internet equivalent of gridlock.

Internet service providers are also concerned.

Tony Blake, vice president for marketing and business development for AT&T Labs Research, calls the trend disturbing, adding, "What people are saying is, 'If my stuff gets through, that's all I care about."' He said AT&T and other Internet service providers were pressuring the producers of streaming media to head off the problem before it grows out of control.

Streaming media traffic is also of growing concern to the Internet Engineering Task Force, an industry group that is the closest thing the Internet has to a governing technical body. Jacobson, an active task force participant, said the group has not yet formally addressed the issue, because "it's not big enough to kill the Net."

Not yet anyway. But a number of Internet service providers are getting nervous.

Jacobson said he would like to see streaming media use congestion controls that are compatible with those used for conventional data-packet traffic.

Today, according to measurements by the National Laboratory for Applied Network Research at the University of California at San Diego, none of the existing streaming media providers are using such controls.

Real Networks, the largest streaming media provider, declined to respond to the growing criticism, although a spokesman, Jay Wampold, said the company was trying to address congestion problems. Over the last year, he said, Real Networks had installed a network of computers for delivering the company's broadcasts that bypasses the Internet's main traffic routes.

Kevin Unangst, lead product manager for Windows Media, the streaming media division at Microsoft Corp., said the company was working with network companies to route its streaming data around the Internet's main arteries.

"We know bandwidth and optimizing bandwidth is a huge, huge issue," Unangst said, using an industry term for transmission capacity, "and we're covering all the bases."

But Jacobson said that while actions like those taken by Real Networks and Microsoft might help, they would not resolve the problem. The bottlenecks are not on the main thoroughfares, he said, but on the side streets and access roads.

"As you get out to the edge of network, closer to users, that's where the bandwidth gets limited," he said.

While cable modems and digital subscriber lines will increase bandwidth to the home, a truly significant increase in Internet bandwidth would require ripping out 300 million copper wires and replacing them with fiber.

"That's not going to happen fast," Jacobson said.

The remedy, experts say, lies in providing incentives for streaming media providers to comply with traffic rules.

With the proper incentive structure in place, streaming material from both Microsoft and Real Networks would look slightly worse than it does now, Jacobson said, but at least the competition for bandwidth would be fairer.

As a standards-setting organization, the Internet Engineering Task Force does not have legal authority to force software developers to adopt congestion-control measures.

But standards, which enable all the computers linked to the Internet to talk to one another, can sometimes be used to discourage disruptive behavior.

"It's the tragedy of the commons," said Sally Floyd, a researcher at the AT&T Center for Internet Research at the International Computer Science Institute in Berkeley, Calif. "Only with the right incentive structure do the right things happen."

The "tragedy of the commons" is an example from economic game theory. If a group of farmers must graze their cattle in the same small patch of grass, or commons, it is in the farmers' collective interest not to overgraze, lest all the grass die out and all the cows starve. The problem is that it is in each individual farmer's interest to act counter to the common good.

Ordinary Internet traffic -- e-mail and data transmissions to and from Web sites, for example -- follows a basic set of traffic laws known as the transmission control protocol, or TCP.

Since the late 1980s, when rising Net traffic led to severe congestion problems, TCP has included a congestion-control mechanism that instructs a transmitting computer to slow the data flow by half as soon as the data it is sending encounters congestion. Only when the congestion clears does the data speed up again.

"What's interesting is that congestion itself is effectively the error signal that gives the sender the feedback to speed up or slow down," said Robert J. Berger, president of Internet Bandwidth Development, a consulting firm in Saratoga, Calif. "It means that there will almost always be at least a bit of congestion, but that's OK."

But it is not OK when traffic that conforms to TCP runs into traffic that does not follow those same rules. When the point of congestion is reached, TCP traffic cuts its sending rate in half. But streaming traffic just barrels through the stoplights. This in turn creates more congestion, which further slows the law-abiding traffic.

Streaming media traffic "needs to back off," said Scott Shenker, head of the AT&T Center for Internet Research in Berkeley. "I think most researchers feel it's a flaw in the Internet architecture that there's no way to deal with this."

Both Shenker and Jacobson believe the solution to the streaming media problem is to build incentives into the architecture to encourage good behavior. But they have very different ideas of what form those incentives should take.

Shenker prefers to plant a mechanism within Internet routers that he calls "fair queuing," which would force packets sent from different sources to wait their turn to be sent.

Jacobson's proposal, developed with Ms. Floyd in 1989 and only now being tested in Cisco routers, uses a kind of virtual penalty box. When the router experiences congestion, it takes a random sample of its traffic. If a certain host computer is overrepresented in that sample, its packets are placed at the end of the line.

This creates the right incentive structure, he said, because the Internet service providers do not have to persuade Real Networks or any other company to obey the rules. Rather, he says, "The customers do instead, because the quality" of their audio and video "gets really crummy."

Fair queuing and penalty boxes rely on software fixes to change the incentives. Another approach, favored by many service providers, is to create economic incentives for fair sharing of the roadways. The basic idea is that heavy users should be paying for the extra bandwidth that they consume.

This model, known as differential service, is already being deployed on the Internet. But there are many questions about how it should be adapted to streaming media -- especially about who should pay.

"The real puzzle about streaming media is who is going to make the money -- if we knew this, then we could answer a lot of the technical questions," said David Clark, a senior research scientist at the Massachusetts Institute of Technology's Laboratory for Computer Science.

Whatever solution is eventually adopted, the streaming media problem is certain to be resolved, these experts say.

"An Internet that's growing and can handle streaming media is in everyone's interests," said Blake of AT&T Labs. "We poke each other with sharp sticks all the time, but slowly we move forward."

---

Related Sites
These sites are not part of The New York Times on the Web, and The Times has no control over their content or availability.

• Cisco Systems Inc.

• Real Networks

• Microsoft

• AT&T Labs Research

• Internet Engineering Task Force

• National Laboratory for Applied Network Research

• AT&T Center for Internet Research

• Internet Bandwidth Development

• Massachusetts Institute of Technology's Laboratory for Computer Science