* Readers offer more opinions on the current state of TCP
Editor’s Note: Starting next week, the Network Optimization newsletter will have a new author: Ann Bednarz, Network World’s associate news editor. As Network World’s applications expert, Ann will bring to the newsletter a wealth of knowledge about applications and their potential impact on network performance. You will still be able to keep up with Denise Dubie’s insights into the industry as next week, she takes over as author of the Network/Systems Management newsletter. To subscribe to that newsletter, please go to the Subscription Services link below. Thanks for reading!
A recent newsletter stirred up a lot of opinions from industry watchers and network managers on the current state of TCP and what needs to be done to improve the protocol today. Research firm Gartner called for improvements to TCP, but the responses vary as to what might work to enhance performance of the protocol.
Network Optimization readers say the premise of an updated TCP appeals to them, but it’s unclear how well a new protocol will fair considering how embedded TCP is in today’s networks. The topic started around a company called Data Expedition, Inc., or DEI, and its proposed replacement for TCP in its products, but has broadened to a discussion around TCP. It’s fair to say none of the respondents here are overly familiar with the start-up or well-versed in its technology. The responses seem more about particular readers’ long-term experience in this industry and the frequency of stagnating standards.
One reader writes, “I think this has a lot of potential, but has it been submitted to IETF for comments? Unless it emerges as an RFC it will be useless.”
Another gets more specific, pointing out just how much of today’s networks depend on TCP and how difficult it would be for another protocol to become quite as engrained in all elements of the IT environment.
“I think the problem you would have with this protocol is endpoint OS acceptance. TCP is an endpoint protocol, not something in routers. So, you’d need all the PCs, servers, blades, storage gear and everything else that talks TCP and have this new transport protocol enabled on it. That could take a long time to work through all the bugs, interoperability, support issues, etc.
“Now, as you mentioned, you could enable [DEI’s Multipurpose Transaction Protocol for the Internet or MTP] in the ‘network’ inside routers or blades, but that’s just a different technique for application acceleration, not a new model. And putting it on standard routers (i.e. no special blades) would require significantly more processing power. Furthermore, how do the end hosts then know to talk TCP to the router instead of setting up a session with the remote server? Even if you wrap the traffic in MTP, the hosts, which make the traffic (routers don’t make traffic), are still using TCP.”
Industry watchers say that the transition from today’s TCP to an updated version or a new protocol altogether will take a lot of time, standards approval and industry acceptance.
“Seems like there’s no pervasive memory in our industry. … It’s going to take nearly 20 years to migrate from one fully accepted and standardized version of IP to the next one (i.e. the migration to IPv6),” says Mark Fabbi, a vice president with Gartner. “As a bit of history – the first presentation I did as a Gartner analyst in February 1996 talked about IPv4 vs. IPv6.”
Fabbi’s colleague at Gartner, Joe Skorupa, a research director with the firm, goes on to say that a nonstandard protocol won’t solve the problem. While there is still work to be done, there are some measures, Skorupa says, enterprise IT managers can take today to improve TCP on their networks.
“If you want to improve TCP, just implement the fixes detailed in a series of IETF standards (including RFCs 3390, 2582, 2883 3517 and 4138). They include fast ramp, selective acknowledgement, large transmit and receive windows and explicit congestion notification. These are now standards, are well understood and have been tested by a large community and have been shown to work well,” Skorupa explains. “Additionally, many [WAN optimization controller] vendors incorporate most of these features and Microsoft’s new TCP stack for Vista/Longhorn has excellent support for this approach.”
