RE: (multi6) requirements draft comments

["Tony Hain" <alh-ietf@tndh.net>]

Michel Py wrote:
> I think that the original intend of 3.1.2 for inbound was to prevent
> flaky destination address selection schemes such as
> round-robin. The way
> I read it is that, if a multihomed site as multiple PA
> addresses but no
> PI address, the address selection algorithm to pick which of the PA
> address is to be used must pick the best one for a specific source,
> which probably means that the address selection must know about the
> network topology, which means either a hook into BGP or some other
> sophisticated mechanism yet to be defined.
Having a host know the topology is the flaky scheme. DNS round-robin
works fine to isolate the address selection from the topology knowledge.
That section appears to be trying to describe the case that SiteA has
the ability to inject a long prefix into all ISPs. My concern is that it
states a site MUST be able to control traffic flow, when it is the
receiver. This is technically not possible since the holder of the
packet can choose any path it wants by ignoring the detailed
announcement from the receiver.

> There might be other ways (TCP hacks/stack hacks), but the PI address
> being the one that is configured on the hosts is the only proven one.
> Note that the PI address could (although I do not recommend it) be a
> site-local address with NAT.
The point was that the correspondent must perceive that the address
didn't change, else the interruption will be more than a simple packet
loss. Use of a NAT is orthogonal because the public side of the NAT has
the same issue.

Tony