Re: A connection-based Internet?

Noel Chiappa <jnc@ginger.lcs.mit.edu> Wed, 18 December 1996 17:48 UTC

Received: from cnri by ietf.org id aa29567; 18 Dec 96 12:48 EST
Received: from murtoa.cs.mu.OZ.AU by CNRI.Reston.VA.US id aa18998; 18 Dec 96 12:48 EST
Received: from mailing-list by murtoa.cs.mu.OZ.AU (8.6.9/1.0) id EAA02736; Thu, 19 Dec 1996 04:34:47 +1100
Received: from munnari.OZ.AU by murtoa.cs.mu.OZ.AU (8.6.9/1.0) with SMTP id EAA02707; Thu, 19 Dec 1996 04:23:27 +1100
Received: from ginger.lcs.mit.edu by munnari.OZ.AU with SMTP (5.83--+1.3.1+0.56) id RA08094; Thu, 19 Dec 1996 04:23:18 +1100 (from jnc@ginger.lcs.mit.edu)
Received: by ginger.lcs.mit.edu id AA15345; Wed, 18 Dec 96 12:12:03 -0500
Date: Wed, 18 Dec 96 12:12:03 -0500
From: Noel Chiappa <jnc@ginger.lcs.mit.edu>
Message-Id: <9612181712.AA15345@ginger.lcs.mit.edu>
To: huitema@bellcore.com, mo@uu.net
Subject: Re: A connection-based Internet?
Cc: big-internet@munnari.oz.au, flows@research.ftp.com, jkr@netstar.com, tagswitch@cisco.com
Precedence: bulk

    From: "Mike O'Dell" <mo@uu.net>

    the fundamental problem is that the model of "ip destination-only"
    forwarding is not powerful enough to build the networks required for
    the current Internet, much less the future.

Much as I am very interested in the basic question under debate here, I think
this discussion is fundamentally pointless, if not "out of order", *in this
forum*.

Experience in the IETF has shown, over and over again, that we'll argue about
this for months, and when the dust settles, very few minds will have been
changed, and the situation will not have been resolved. In the meantime, much
time/energy of the WG-to-be will have been wasted.

So, I will say again: this is not an effort to formally get the IETF as a
whole to agree to switch to flows as a fundamental paradigm. It is a forum for
people who like a certain set of ideas to come up with specifications which
implement those ideas.

If you think the ideas are crazy/impossible, that's fine, just please sit
quietly and watch as other people waste their time, and let the rest of the
group get on with doing their thing.


    nobody is arguing for "end to end VCs". that's just silly.

One of the things that continually annoys me no end is the apparent inability
of some people to grok that there is a lot (anything?) in the middle of the
spectrum between the extremes of pure-stateless-datagram (a la IPv4), and
old-style-virtual-circuits (a la X.25).

It seems like anytime someone stands up and says "maybe we need to do
something other than pure datagram", there are always people accuse you of
wanting to do VC's. The fact that the proposer is well aware of the problems
of pure VC's, and doesn't want to do a pure VC network, is usually completely
missed. They also don't usually seem to have bothered to take the time to
understand what it is you actually *are* proposing. Needless to say, after a
while it all, especially the latter bit, gets pretty unfuriating.

Someone at the just-passed IETF described this as a "four-legs-good,
two-legs-bad" model of reality, and that's right on target. I find, over and
over again in the IETF (and it was very obvious with the past debate on
variable length addresses), that many minds are already made up, and no real
objective, open-minded, thoroughgoing, from-scratch analysis of the
engineering good and bad points of new approaches are made. Instead, they are
dismissed with an immediate, simplistic, and unstudied "two-legs-bad" kind of
reaction that's all to apparent, after you've been on the receiving end enough
times.

The good thing about reactions like that is that you soon figure out that
since there is little deep analysis behind them, you can just blow them off.
If you think reasoned debate is going to change them, think again - been there,
tried that.

The packet world is a victim of its own sucess. The people who would, in the
60's, have been *outside* saying "it's new, and therefore won't work" are now
*inside* saying "it's new, and therefore won't work".


    what we are talking about is switching flows, where there is some
    efficiency to be gained in establishing soft state in the forwarding
    paths.

Be careful with the use of the term "soft state". As far as I can tell, it
means different things (in terms of operational considerations like who
establishes it, who maintains it, who removes it, what happens when it's
missing, etc) to just about everyone who uses the term.

In fact, it seems to have fallen victim to "four-legs-good" disease, in that
schemes that someone likes are inevitably described as being "soft state",
whereas schemes they don't like are always described as "hard state". (One's
own scheme is *always* "soft state", no matter how the definition of that has
to be twisted to fit.)

Given that the routing tables used in the current pure-stateless-datagram
model are the hardest of hard state, I wonder exactly how they fit into the
simplistic "hard-state-bad" view of reality that seems to be common among
those who adhere most tenaciously to the PSD model, but I digress.


    the ability to place bandwidth between the points where it needed, as best
    approximated by where you can actually get it, and then place the traffic
    of interest on that path without going crazy screwing with IP metrics

Some of us might point out that the current fundamental routing architecture
of the 'Net, one inherited basically unchanged from the Baran work in the
early 60's, and one intended for far smaller networks with different
requirements, is really not the paradigm we ought to be working within - but
that's a different WG! :-)

    in other networks ... the flows are much larger, aggregate objects which
    have little to do with any particular IP prefix, other than it and a bunch
    of others directly connect to the infrastructure off a particular superhub.
    (this is where the mapping between randomly-assigned IP address and physical
    proximity happens.)

Too bad addresses in packets don't exclusively reflect something that's
actually useful to the packet-forwarding fabric, like where you are, or
anything. Nah, that's too obvious.


    link-state technology tries to duck the question by letting everyone
    pretend they know everything. there's something deeply unscalable about
    this notion of "just tell everyone everything".

Depending on exactly what you mean by "link-state", sorry to disappoint you,
but this statement has been untrue for at least 10 years. ("Four legs
good...") See:

	Josh Seeger and Atul Khanna, "Reducing Routing Overhead
                in a Growing DDN", MILCOMM '86, IEEE, 1986.

for an early one, and of course the IETF's own OSPF also somehow seems to
work without global knowledge.

It's this kind of simplistic statement which leads less-knowledgeable people
down the path, and I hope people will cease and desist with them?


	Noel