Re: An idea: GxSE

[smd@ebone.net (Sean Doran)]

Mohan Parthasarathy <Mohan.Parthasarathy@eng.sun.com>

| > Why not just use SCTP?  Why reinvent the wheel?
|
| In SCTP, i think you have to know all the addresses at the beginning
| of the connection. Here, the new address is not known until
| renumbering happens.

My interpretation of Ohta-san's draft proposal (see list archives)
understands that it calls for the addition and subtraction of
addresses *known inside the host* during the lifetime of a connection,
based on learned information about the behaviour of nonlocal topology.
In other words, Mohan is correct if "here" refers to Ohta-san's proposal.

Although his draft is the only one we've gotten so far, I do not
think we would want to exclude other GSE-like proposals, should
the arrive upon us in the form of a draft (hint, hint, hint to everyone).

By contrast, Paul Francis's idea (GxSE) proposes that things inside
multihomed sites do not need to know their entire set of addresses,
but that a nameserver will ("typically") know all of them.   To me,
this is Ohta-san's proposal with the topology knowledge pushed one
or more hops away from the host, and as a result of which introduces
an architecture which requires some small changes by host-developers
(e.g., don't make assumptions about the end-to-endianness of the 
GR (RG in Mike O'Dell's GSE) when doing things like checksums or
encodings of host addresses in upper-layer protocols).   PF notes this,
Ohta notes the opposite in his draft.

Without making any sort of judgement, I'd like to suggest that
although our initial volunteer team hasn't finished the draft of
a requirements doc yet, at least a couple of them read the list.
I suspect that (if it's not there already), the reqs will indicate
a desire to support a wide variety of "transport-layer" protocols.

Thus, while we're waiting on the reqs, I think it'd be particularly
nice if proposals (however informal) somehow could be "fitted-into" --
or at least take into consideration -- TCP, UDP, and so forth,
and ideally make implementing/managing IGPs (in the sites themselves
or in their respective inverse trees of providers) very much more
difficult than today (in v4 or v6).    Shim approaches (like SCTP)
are not the only -- or even necessarily the best (though maybe they are) --
way of taking these possible-requirements into consideration.

In other words, it's far too early to exclude prima facie any new
ideas, or any old ones, either.

Yours wishy-washily,

	Sean.