Re: Transport level multihoming

[Greg Maxwell <gmaxwell@martin.fl.us>]

On Wed, 4 Apr 2001, Geoff Huston wrote:

> The bgp table has a number of roles:
> 
> - connectivity maintenance
> - policy negotiation
> - destination-influenced inter-domain traffic engineering
> 
> While the push to the end system for multi-homing control can allow some 
> degree of benefit in the number of prefixes used to maintain connectivity 
> of the inter-domain space, it complicates policy negotiation as we know it 
> today, and it certainly alters what we currently think of as inter-domain 
> traffic engineering. The issue is that inter-domain traffic engineering and 
> policy assertion will not go away - if it is no longer possible to use 
> announcements of AS + specific routes to achieve these two objectives, then 
> doubtless other mechanisms will be found. Now if you regard the total 
> inter-domain routing space as an information negotiation across all three 
> spaces, the real question in my mind is whether this approach reduces the 
> total amount of information sloshing around the inter-domain space or not? 
> Somehow I'm not convinced that this allows a real information reduction.

Perhaps, if you consider all sloshing as equal cost you statement holds.

However, it isn't. It's a lot more cost to slosh routes through every
default free router on the Internet then it is to have clients obtain a
routing policy along with their DHCP lease or router discovery.

I'm not saying that we should take all policy out of the routers, what I'm
saying is we should reduce their statefulness as much as possible. 

Consider some policies that would be trivial with transport level
multihoming but very hard with centerlized multihoming, for example,

Equal cost multipath: Hosts pick their endpoint addresses based on
RTT and TTL. This would accomplish well balenced at all time (if a link
becomes congested (signaled by recpt of ECN or long queueing) or fails,
hosts move off of it, if they are equally loaded the traffic could flow
over the path with the fewest hops. The path can easily be decided for
each direction of the traffic separately.

This would lead to a much smoother bidirectional utilization of multiple
equal links. The above simple formula could easily be weighed to favor
differnt links, and the cost decisions could easily be made seperatly for
differnt flows requiring differing quality of service.

Routers could enforce policy by refusing to carry traffic that violates,
responding with an ICMP host unreachable type message to inform the
misbehaving end-node.

Transport level multihoming also makes it much simpler to remove the
router as a single point of falure.

The potential is certantly there, and the strongest argument I've heard
against it so far is that it isn't what we do today, though it is
certantly possible that I havn't dug deep enough yet. References are
greatly appreciated.