[e2e] TCP in outer space
faber at ISI.EDU
Fri Apr 13 10:57:24 PDT 2001
On Thu, Apr 12, 2001 at 05:10:53PM -0700, Alex Cannara wrote:
> My fault finding is more with attitude than
IETF attitude? Seriously, it's unclear to me whose attitude you're
finding fault with. I thought you meant the original designers of the
Internet, but you tell me that's not so.
> As far as what's ad hoc and what's not, we all know much of the Internet
> indeed has been addressed in ad hoc fashion over the years. Given the
> number of people involved worldwide, the great resources made available
> over the years, and the 'fixes' now being made, there's every reason to
> suggest a more broadly evaluative approach than parroting things like
> "Changes must be TCP-friendly", etc.
Well, as practical engineers trying to maintain the viability of a
global infrastructure, it's hard to fault the IETF for being
conservative. Creating a fair, efficient congestion control system
that works across the range of delays, bandwidths, and technologies in
the Internet is daunting even if you restrict yourself to long TCP
connections. TCP-friendliness is a way to keep the problem vaguely
tractable for the folks whose phones ring when congestion breaks
I think people would be open to systems that violated the letter of
the (largely unwritten) TCP-friendliness law, but maintained or
improved Internet stability, but maybe I'm naive. IMHO, the mantra is
"stability first," but apparently "TCP-Friendliness" is catchier.
> As an example, even in low RTT
> environments (vastly easier than spacecomm), TCP has large performance
> penalties for its users when loss is not purely congestive.
There are many environments where TCP gives suboptimal performance.
There are few environments where it gives unusable performance. The
fact is that more value is given to TCPs robustness than its
performance in the IETF, and, IMHO, reasonably so.
We agree that TCP has the properties you describe; we disagree about
how serious that is. The IETF seems to have made the value judgement
that robustness is more important than performance (one I agree with,
but won't defend). Either you have to change their institutional
mind, or you have to design your performance enhancements within their
constraints. Or do your work in the other Internet.
> And, we
> won't even bring up the IP addressing debacle, saved only by NAT boxes,
> which themselves sacrifice network security for addressibility.
I don't know enough about the details of IP addressing to make a
reasonable statement, but I'm no fan of NATs.
> As far as design goals, one would be to eliminate the discussion title:
> "TCP in Outer Space".
Pick one and put it in the subject line if you reply.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 230 bytes
Desc: not available
Url : http://www.postel.org/pipermail/end2end-interest/attachments/20010413/5f8ed268/attachment.bin
More information about the end2end-interest