Network Computingwww.networkcomputing.com

Although TCP and satellites aren't exactly strangers, the protocol's space flights primarily have been long scientific transmissions--not the short transactions of today's Web. Short transmissions are governed by the "slow start" TCP implementation revision that became necessary in the late 1980s, when the Internet was brought to its knees by congestion. Slow start improved network performance by replacing a cacophony of competing bandwidth requests with a process in which each communication slowly tests the network waters for congestion.

But now that more than half the Internet's traffic consists of short transactions, slow start leaves many satellite users cooling their heels--and unable to tap otherwise available bandwidth. When congestion or errors are encountered midtransmission, delay becomes even more pronounced. Highly interactive applications probably suffer most. Even a half-second delay can result in death by avatar, no ma tter who pulled the trigger first.

Satellites: An Alternate Internet? Satellite promoters like Victor Barajas, a systems scientist with Hughes Spaceway, say because of legitimate concerns about congestion , it would be foolhardy simply to remove slow start. He's interested in proposals that would mitigate slow start's effects--and in adequately testing any changes over both terrestrial and satellite networks.

The vision put forward by Barajas is that satellites may one day relieve Internet congestion by letting business users use them to interconnect directly with content providers. Applications such as audio, video and multicast could be handled at much greater speeds. In December, when satellite advocates broached key Internet Engineering Task Force (IETF) officials with their TCP agenda, they got a thumbs-up. Fred Baker, chairman of the IETF, says tackling this issue is wise, especially given the satellite-dependence of third-world nations. Baker, however, believes the issues can be resol ved without taking the advice of some satellite advocates to move into the heart of the protocol specification, as some satellite advocates would prefer; instead, he suggests changing some of the Request for Comments (RFCs) and unofficial rules that govern TCP stack implementation. In January, Baker was preparing two draft RFCs--one his own design, the other on behalf of the Internet Society's End-to-End Research Group--to move the task along.

Opening the Satellite Window The first RFC suggests tweaking slow-start implementations to increase the size of the initial TCP window so that when the transfer of actual data begins, four segments could automatically move into the pipe with a burst of 4,096 bytes in the first 560 milliseconds. (Today, some initial TCP windows are mere peepholes, allowing one data segment at as little as 200 bytes.) Because slow start requires that this data undergo a round trip with acknowledgments before the window size can be doubled on the second trip, redoubled on th e third and so forth, these first two initial exploratory round trips can add up to a second in delay over satellite links. By increasing the initial window size, Baker believes, these two exploratory round trips can be lop ped off, especially since the typical Internet transmission is 4,096 bytes or fewer.

"By changing the initial window," Baker says, "we can help a lot of FTP, e-mail and Web transmissions by sending the entire file in the first burst."

Baker isn't the only one willing to reexamine slow start. Jon Postel, an Internet Society trustee and editor of the original TCP specification, agrees that slow-start changes are "reasonable," especially if they can be accomplished in new software, rather than requiring wide-ranging stack changeouts. In December, however, Postel was less than enthusiastic about upping the allowable number of parallel transmissions. Although users may benefit for a while, he says, there are trade-offs in server and Internet resources. The use of parallel transmissions, in fact, has been widely debated at the IETF following Netscape Communications Corp.'s decision to use such an approach as its browser default. Baker says the fact that the approach hasn't broken the Internet tends to be an argument for concluding it is safe.