RE: WG next steps

["Christian Huitema" <huitema@windows.microsoft.com>]

> > It is undesirable for a large enterprise to use a multi-address/host
> based
> > solution without a mechanism to interact with the network topology
to
> > perform path selection.  Network operators know more about the
viability
> > of paths (whether by technology or manual policy) than the hosts do.
> > Left-side longest-match is hardly a sufficient routing protocol
and/or
> > policy control for a multi-provider topology like the Internet.  The
> > network *MUST* be a part of the solution -- enterprises want policy
> > control, and policy control should be centralized.
> 
> Succinctly put.  However I don't think size matters as much as does
the
> existence of policy.  Only the hardware/software/wetware at the edge
of
> the
> enterprise is likely have the necessary information to enable rational
> path
> (and, therefore, source and destination address) selection.

I would not be so sure about "rational". There is some documented
evidence that the path selected by the routing fabric is often not the
shortest route between two points; this is easy to understand because
BGP is about policy, not metrics; the interference between PBG policies
and multi-homing can be interesting, to say the least. The routers have
some information, and will probably detect unreachability faster than
the host; on the other hand, the host is likely to get better
information about available capacity and round trip time. Regardless,
the real question there is how to affect source and destination address
selection in a rational way. 

There is an existing mechanism to convey information from router to
host, the router advertisements; it can be used, at a minimum, to prefer
the source prefixes that correspond to functional providers over those
for which connectivity has been lost; it is easy to use this mechanism
in a single link network; using it in a more complex network requires
exchange of information between the site routers. If that mechanism is
not sufficient, nothing prevents us from designing a better one,
possibly some variation of ICMP Redirect between site exit routers and
host on the site.

-- Christian Huitema