On Apr 21, 2009, at 2:04 PM, Melinda Shore wrote:
You can call it "foo" for all I care, but much of what's
been discussed so far is policy. From the proposed
charter:
"A host connected to multiple networks has to make decisions about
default router selection, address selection, DNS server selection,
choice of interface for packet transmission, and the treatment of
configuration information received from the various networks. Some
configuration objects are global to the node, some are local to the
interface, and some are related to a particular prefix. Various issues
arise when multiple configuration objects that are global to the node
are received on different interfaces. At best, decisions about these
matters have an efficiency effect. At worst, they have more
significant
effects such as security impacts, or even lead to communication not
being possible at all."
That's all policy.
My shaman once said "For God, everything is just a question of
policy." But, we need to reduce this problem to a more mortal scope,
and I'm not quite certain that the proposed charter text accomplishes
this goal.
There's a tremendous amount of complexity inherent in the problem. I
suspect as worded that it is roughly isomorphic to the peering
problem, with all the business-driven policy consideration that entails.
Consider that peering policy is often driven by things that are well
beyond the scope of protocol. Its potential range of expression is
unlimited; in fact driven by a natural-language contract and heuristic
operations on underspecified constraints derived from that natural-
language contract.
The only alternative I can see here is to try and reduce the range of
expression. One such approach might be to develop a policy language
having the property of an algebra, so that policies can be written and
interpreted in a defined manner, and so that the combinations and
interactions of multiple policies can be evaluated and applied
predictably in automata and simulation.
This requires a substantial body of precursor work: What things
should the language be able to express? What combinatorial operations
are required? What are the relative priorities of linguistic
statements? What sorts of fundamental things should the policy
describe (we have things like route selection, prefix manipulation,
addressing, access rules, name services, name defaults, and so on to
consider as options).
Even with this sort of model, I am not confident at our ability to
achieve success, at least without even more substantive reductions in
the scope of the effort.
Can we narrow it down some more?
--
Dean Willis
_______________________________________________
Ietf mailing list
Ietf@ietf.org
https://www.ietf.org/mailman/listinfo/ietf
