RE: network controls are neccessary

["Michel Py" <michel@arneill-py.sacramento.ca.us>]

Erik,

> Erik Nordmark wrote:
> If the exit routing policy can be expressed with a few rules
> it would essentially be additional rules in that table (and a
> protocol by which the hosts can learn those rules).
> Hence my questions on the list (so far without answer) about
> reasonable sizes for the exit router selection policy/routes.

It's not that simple, because the policy can be the combination of
several fractional policies applied at different points. For example, a
network administrator could tag a route or a summary a certain way to
make traffic flow in a general direction, then the site-exit router that
happens to be in that direction would make another policy decision based
on the AS-PATH, that kind of thing.

The end-result of TE, which is the policy that determines which
interface on which router is the one being used can be the result of
decisions made in several routers. It can be a tree, not a flat policy.
 

>> Michel Py wrote:
>> a) The hosts to have the same routing capabilities as routers
>> currently have.

> Sorry I don't follow the logic. The hosts don't route. Hosts
> with multiple source locators that talk to a node with
> multiple destination locators just do a selection e.g. when
> creating a new connection.

My bad, I should have said: "The hosts to have the same *decision*
capabilities as routers currently have".

You are correct, instead of choosing the egress interface the host would
choose among several addresses, which in turn would influence the egress
interface.

TE is not about routing, it's about choosing the egress interface (and a
few other things).

The point is about the criteria the host would use to make this choice.
The way TE is today, the router has a wide variety of things to help
decide, including the AS-PATH, MEDs, route tags, port numbers, diffserv
codepoints, name it it's used somewhere in a route-map.

Since we don't want hosts to run BGP, you would have to transform all
these criteria into something simpler that the host would understand.
This not something I know how to do.

A simplistic approach says that since we can't have the hosts receive
the same quality of information than routers, the solution is a querying
mechanism where the host would query the routing system with a list of
possible source addresses, a list of possible destination addresses, a
protocol and port number, and have the routing system reply with a
primary and possibly a backup pair.

In short: I think I will concur with what Joel and Michael have
contributed, and I think host solutions will need a protocol that is an
interface between the routing system policies and the choice of the
right address or address pair.


> I think the edge thing is also important to explore in parallel
> with a host-based scheme.

I have heard of another ML that has made it part of its charter :-)

Michel.