[e2e] on the need for more research in core networking

George Michaelson ggm at apnic.net
Wed Jan 15 02:28:41 PST 2003

Loved the slides.

But I think we need to encourage as much attention on process/management stuff
as on technology.

You touch on overlay networks, but both 6bone and multicast represent problems
as well as solutions in some senses: transition 'away' from overlay to native
is deeply painful sometimes, and there is a 'third party' issue in the tunnel
which gets into the whole economic/greatest-good question about the converged
(or lack of converged) interests in a loose confederation of competing
parties. You can't make overlay go away, but you also have to cope with the
goals of the overlay being orthogonal to interests underneath. 

Against which overlay/tunnel/encapsulation is probably as much a good thing
for the libertarians as it is for technologists. We all hate NAT, but we've
learned to live with it, and SSH and other ad-hoc userspace tunnelling is
pretty well a given now. (which is interesting, since it becomes a route-free
mobile-IP mechanism which doesn't seem to suffer any inherent scaling
problems, as long as you accept the downsides)

I think that we should be looking to the US and Australian de-regulated energy
utility crisis and a 25+ year political/social agenda about what *is* the
right mix of technology, societal coordination and cost:benefit. Nobody ever
joins in my threnody about this, but I really don't see why the discussion of
a replacement for BGP or mechanisms for enhancement divorce themselves from
reconsideration of how we *organize* the network. We don't dig streets, or
waterpipes, the way we deploy 802.11, but we don't deploy 802.11 the way we
interconnect fibre either. If real world bricks and morter demand compliance
with social processes, why should this be seen as untenable in the digital

Some of the stuff floating around about signed routing updates is getting into
this space, since group key management is 'hard' one can expect that there are
going to be top-down authority models coming back here, and with the
inevitable re-convergeance of transit into one or two cable-owning agencies
per nation, noises about how useful the ability to randomly plug any two nets
together are going to need to balance off against the social agenda anyway.

Geoff Huston makes very strong noises about the research/measurement of what
is and is not going on in the network being a few questions short of the depth
we need to be asking. 

	we don't really understand the dynamics of route propagation. -Look at
	the mistakes that have eventuated from a single timeout value in BGP
	flap damping carrying forward for more than 5 years.

	we don't really know the rate of change of routing, aggregation,
	stability of announcement, fragmentation, effects of multi-homing.
	(curve shape fitting in arrears doesn't have any predictive value
	 outside the short term)

I still believe that the majority of network connectivity is almost completely
static, and that we've deployed a network based around 'dynamic re-healing'
which is way way too complex.

I also still believe that confederations, and local topology is the place to
concentrate the effort on rich routing, and that re-structuring the
interconnect into a layer-8 or layer-9 process with intensely long-lived
assertions of routing would be both simpler to manage, and grow.



More information about the end2end-interest mailing list