Re: SUMMARY: Van Jacobsen's ``Fat Pipe'' modifications...

From: Andy Nicholson (droid@poplar.cray.com)
Date: Sat Oct 05 1991 - 10:59:43 CDT


>Folks,
>
> Sorry it took so long for me to summarise, I got busy for a while, and
>then forgot that I had promised to summarise.
>
> Anyway, here is the info I have gotten:
>
> 1. This code is not yet intended for general consumption -- it is
> currently intended to be used to test TCP-IP throughput, and used
> by people like Dave Borman at Cray and Van Jacobsen himself.
> Maybe some day it will be intended for general consuption, but
> that day doesn't look like its anytime soon.
>
> My observation is that if you are a true TCP-IP Guru (with a
> capital ``G''), or TCP-IP Speed Demon, then you might be able to
> request the stuff from Van or Dave and have reasonable
> expectations of getting their code. On the other hand, if you fit
> this mold, then you probably should have been conferring with Van
> and Dave for a long time before this and helped them write this
> stuff in the first place.
>

Hopefully Dave or Van will respond to your posting with more information
regarding the ultimate distribution of the TCP expanded windows code, but
since they have not posted anything, I'd like to clarify a few points now.

The expanded windows code, in Dave's and Van's separate implementations, are
part of the process of getting the RFC's defining how to do it on a
standards track. This process has been stalled somewhat by some problems
which have been discovered in the RFC's (1185 and 1072). This will be
rectified at a working group meeting at the next IETF (November 18-22 in
Santa Fe). This working group will produce an RFC for TCP expanded windows
which will then (hopefully) go on the standards track.

The idea behind all this is to make it fairly easy for anyone with reasonable
understanding of TCP to be able to use the RFC to make their own
implementation. However, I am sure that the prototype implementations will
be available in one way or another. Some Cray Research Inc. customers
have been running the code for some time now (as reported in "High Speed
Networking at Cray Research", Nicholson, et. al., _Computer Communications
Review_, January 1991). It is part of UNICOS 6.1, the latest release of
our operating system, which many of our customers are running. It will
continue to be available in future releases.

I would hate to think that you have to qualify as a TCP/IP "Guru" to take
part in the learning process involved in using TCP expanded windows.
Cray Research as a company and Dave Borman personally have made quite an
investment in time and effort to make TCP expanded windows something that
everyone can use and take advantage of. That investment is wasted if no
one takes advantage of the effort.

> 2. It's not really any good for speeding up Local Area Network TCP-IP
> throughput -- it's meant to squeeze the last ounce out of Wide
> Area Network throughput (like two or more Ethernet LANs connected
> by a 1.5Mbps T1 link or two or more FDDI LANs connected by a
> 45Mbps T3 link). If you are looking to boost LAN speed, then you
> should probably put the BSD 4.3 Reno networking code (which can be
> anonymously ftp'ed from uunet, but where it's hidden in their
> hierarchy I haven't the foggiest), or you should go to your Vendor
> and scream in their ear. Either way, the Van Jacobsen code won't
> help you too much.

I must disagree with the statement that it isn't really any good for
speeding up a LAN. TCP expanded windows will improve throughput any time
the round trip time equals or exceeds the time it takes for a host to fill
a peer's window. Certainly WAN's with long RTT's are the obvious case.
But the world is making a transition to faster backbones (like FDDI) such
that the signalling rate makes it possible to fill the window in one RTT or
less. At the 100 Mbps signalling rate, it takes a little over 5ms to fill
a 64k window. And I know that RTT's in excess of 5ms are not uncommon on
FDDI lans, especially on campus-wide nets with a router or two in the path.
LANs where TCP expanded windows are useful will become fairly common in the
next few years. TCP expanded windows anticipates and satisfies this need.

If I seem to be overreacting, I'm sorry. But there are still a lot of people
out there who think TCP is not good enough to solve real-world long-term
internetworking problems. In the Byte Magazine special issue that just
came out, Bob Metcalfe, (inventor of Ethernet, etc.) is quoted as saying
(not exact, it's at home) "TCP has reached it technological limits", and that
TCP must give way to "real" standards such as OSI (Really, that what it says!)

-- 
Andy Nicholson				droid@cray.com
Cray Research, Inc.			(612)683-5473
655F Lone Oak Drive, Eagan, MN  55121



This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:06:18 CDT