RE: Multihoming by IP Layer Address Rewriting (MILAR)

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

> From: Iljitsch van Beijnum [mailto:iljitsch@muada.com]
> 
> On Tue, 4 Sep 2001, Christian Huitema wrote:
> 
> > > Are you saying that if there is a good way to implement
> multihoming
> > > that requires a globally distributed hierarchical database system,
> we
> > > should create a new one?
> 
> > Well, you should first prove that we actually need a globally
> > distributed hierarchical database. I don't think so.
> 
> The problem is that there is no clear "best" way to do multihoming in
> IPv6. So router people want the hosts to solve it, layer 4 people want
> layer 3 to solve it, and so on. Who will be the tiebreaker? Someone
> people are bound to be unhappy, whatever we end up implementing.

Uh, I don't know of any such thing as layer 3 or layer 4 in the TCP-IP
model, I would rather speak of transport and Internet. Anyway, I am a
bit puzzled by your comment about layered implementations. There is a
rhetorical question about whether multi-homing could be entirely hidden
by the network so that hosts don't have anything to do, but the truth is
that we have multi-homed hosts to manage, which means the question is
very much settled. Then, there is the layer issue. In my own small
boutique, there is not all that much distance between the implementers
of TCP and those of IP, or IPv6; I think that the same is true for most
host implementations. It is then more a question of practicality, e.g.
using ICMP messages or IPv6 end to end options versus using TCP options
and doing it again for UDP. 
 
> 
> > We start with the
> > assumption that hosts have multiple addresses, but that the
> > corresponding host only knows one of them.
> 
> For multi-address multihoming to work, we need to know more than one
> address the time of session establishment. Preferably before setting
> up a session, but certainly well before the setup times out. However,
> traditional multiple A/AAAA records should suffice for this purpose.

Yes, but what the host will do here is select one of a set of DNS
provided addresses to send the initial SYN. It certainly does not follow
that the host can use all the addresses in the set to balance traffic or
reroute the connection: in many cases, the addresses belong to different
"boxes." Also, we could just as well use alternatives to DNS for getting
the initial address.

-- Christian Huitema