Re: I-D ACTION:draft-palet-v6ops-tun-auto-disc-00.txt

[Brian E Carpenter <brc@zurich.ibm.com>]

Pekka Savola wrote:
> 
> On Wed, 21 Apr 2004, Brian E Carpenter wrote:
> > Section 4 puzzles me. My understanding is that anycast is best
> > if used in single-shot mode: send an anycast UDP packet, pick one of
> > the replies, and then use unicast to establish a session with the host
> > that sent that reply. You won't use the anycast address again after
> > that initial discovery packet. It's the same as the initial broadcast
> > in DHCP or the initial multicast in SLP.
> >
> > Admittedly, that's not the model in RFC 3068 (Anycast for 6to4). But 6to4
> > is a special case, since there is no tunnel negotiation.
> 
> Shared-unicast DNS root servers and other load-balanced services also
> use the similar technique as 6to4.
> 
> I.e., when you just use one address, it's closest to you whereever you
> roam in the network where there are multiple anycasting servers.
> 
> Obviously, this only works if you don't need to store state in the
> anycast server, or if it's OK to lose the state if the routing changes
> or you move.
> 
> There doesn't have to be stored state if the IPv4 address is embedded
> completely or encoded in a reversible fashion in the IPv6 address.

Which isn't the case in configured tunnels, hence my puzzlement at why
we would do it this way.

   Brian