[e2e] Net-unfriendly solutions (was Is a non-TCP solution dead?)

[Jim.Gettys at hp.com]

> Sender: end2end-interest-admin at postel.org
> From: "Armando L. Caro Jr." <acaro at mail.eecis.udel.edu>
> Date: Wed, 30 Apr 2003 15:57:20 -0400 (EDT)
> To: <end2end-interest at postel.org>
> Subject: [e2e] Net-unfriendly solutions (was Is a non-TCP solution dead?)
> -----
> Not to re-start an old (and long) thread, but this mail is a side-note to
> one of Alex's mails...
> 
> On Mon Mar 31 22:57:08 2003, Alex Cannara wrote:
> > I found it interesting some years ago, when consulting for a large storage
> > apps company, that their TCP traffic was violating most TCP rules -- no
> > slow start, window shutdowns ignored...  It was amazing to watch with a
> > Sniffer(r) the vast data rates they achieved among big Sun servers.  I was
> > ready to list the violations as problems with their stacks when one of
> > their development guys explained that they knew enough about how the OS
> > handled IP that they didn't need to use the kludges added to TCP to
> > 'protect' the net.  Many on this list have likely used this vendor's
> > systems in one way or another.
> 
> What worries me is how much praise such violations receive. Network World
> (issue 04/21/03) has an article that praises a startup
> called Netli, which has developed a producted called NetLightning.
> NetLightning is a CDN architecture which essentially establishes three
> transport layer connections: (1) client to local VDC (Virtual Data
> Center), (2) local VDC to remote AAP (Application Access Point), and (3)
> remote AAP to HTTP server. Connections (1) and (3) use the normal
> HTTP/TCP, but connection (3) uses Netli's prioprietary transport protocol.
> 
> According to Netli, their transport protocol "uses the reliable, secure,
> and network-friendly techniques of TCP, but is enhanced to provide optimal
> performance for communication between a VDC in one geographic region and
> an AAP in another." However, Tim Greene of Network World reports more
> details which reveal just how network-friendly Netli's protocol is. For
> example, Greene reports "that the [Netli] protocol can reduce from 31 to
> two the number of round-trip interactions needed to send a 70K byte Web
> page with 25 objects on it." Also, the Netli Protocol "eliminates the
> slow-start feature of TCP," because "these and other refinements of TCP
> reduce the time it takes to download pages and to interact with Web
> applications."
> 

HTTP/1.1, properly implemented, does this round trip reduction.  This
isn't saying much at all....  One round trip to get the base HTML page,
and then another for all the embedded images.  Best performance is
done pipelined, over a single TCP connection, beating out all multiple
connection hacks.

See: http://www.w3.org/Protocols/HTTP/Performance/

Unfortunately, most browsers have not done decent implementations (doing 
full pipelining of requests). IE does not do pipelining, the last I knew, 
and only recent Mozilla versions started trying.  I haven't looked at 
packet traces to see how good/bad Mozilla is now doing.  Opera has done 
it well for quite a while, and I believe this accounts for some of its 
performance advantage over other browsers.  Doing it well is unfortunately 
a pain in the *ass* due to things like broken add servers that serve up 
HTTP .9 protocol stuff to you.

I'll leave the comments about messing with slowstart to others to flame...
                         - Jim



--
Jim Gettys
Cambridge Research Laboratory
HP Labs, Hewlett-Packard Company
Jim.Gettys at hp.com