Re: routing protocols will provide path-MTU

jnc@PTT.LCS.MIT.EDU (Noel Chiappa) Tue, 27 February 1990 16:46 UTC

Received: from decwrl.dec.com by acetes.pa.dec.com (5.54.5/4.7.34) id AA08127; Tue, 27 Feb 90 08:46:53 PST
Received: by decwrl.dec.com; id AA23174; Tue, 27 Feb 90 08:46:41 -0800
Received: by PTT.LCS.MIT.EDU id AA00908; Tue, 27 Feb 90 11:45:56 EST
Date: Tue, 27 Feb 90 11:45:56 EST
From: jnc@PTT.LCS.MIT.EDU (Noel Chiappa)
Message-Id: <9002271645.AA00908@PTT.LCS.MIT.EDU>
To: deering@pescadero.stanford.edu
Subject: Re: routing protocols will provide path-MTU
Cc: mtudwg

	Steve, here in the future we won't have routing 'tables' (or anything
like what people now think of as 'updates') any more. What will be distributed
are maps of the network, which allows entities to compute routes using
whatever special criteria and algorithms they feel like using. Boxes which do
routing will compute routes on demand, during a flow setup phase (which will
happen in parallel with sending the first packet), and will keep a cache of
recently used source-destination-TOS-etc tuples.
	Moreover, it won't be the case that a host will ask 'what is the MTU
to this destination', it will ask 'can you get to this destination with an
MTU of X'? A route with that MTU (if one exists) will be constructed on demand
(much as trucks with large objects on them may take special routes now).
Alternatively, you could ask 'what is the MTU of the path with the lowest
$ cost', or something similar.
	So, given this view, it is reasonable to suspect that this mechanism
will not have an infinite life, but probably (!!) wiil be superseded after
some years. It is thus reasonable to take the position that it is not
necessary to be absolutely on target (I know this is asking for something to
happen to make a liar out of me :-). If it's not criticial, I'd like to see a
final decision taken about what mechanism to use and get on with writing!
We do seem to be going around in circles a bit...

	Noel

PS: With regard to using a bit (as opposed to an option); there was some talk
at one point of a general 'path attribute discovery' mechanism. I suggested
then that the option have a format which would allow other things to be added
later, but that we stick to MTU for now. Using the bit of course makes this
impossible. (I'm still not convinced the overhead of the option is all that
bad if you only send it every 100 packets or so. :-)