Re: ipv6 anycast

[Mark Smith <ipng@69706e6720323030352d30312d31340a.nosense.org>]

On Wed, 01 Sep 2010 21:58:40 +0000 (GMT)
Michael Smith <mksmith@me.com> wrote:

> 
> 
> On Sep 01, 2010, at 02:53 PM, Mark Smith <ipng@69706e6720323030352d30312d31340a.nosense.org> wrote:
> 
> On Wed, 01 Sep 2010 22:48:36 +0900
> Randy Bush <randy@psg.com> wrote:
> 
> > >>> I use them for my network. RFC5969.
> > >> 
> > >> aha! do you mean
> > >> 
> > >> 5. Troubleshooting and Traceability
> > >> ...
> > >> The 6rd CE and BR SHOULD support the IPv6 Subnet-Router anycast
> > >> address [RFC4291] for its own 6rd delegated prefix. This allows, for
> > >> example, IPv6 ICMP echo messages to be sent to the 6rd virtual
> > >> interface itself for additional troubleshooting of the internal
> > >> operation of 6rd at a given CE or BR.
> > > 
> > > indeed. and note that this makes no requirement that _any_ type of
> > > interface has to support this.
> > 
> > btw, that is an rfc, not a network :)
> > 
> > do you know if free.fr and/or others actually deployed and tested this?
> > among other things, i worry about code paths that are rarely, if ever,
> > used.
> > 
> 
> Here is a post of somebody who used it to provide "poor mans" HSRP.
> 
> http://lists.cluenet.de/pipermail/ipv6-ops/2010-June/003639.html
> 
> Â
> That's how I use it on GSR's that will never have HSRPv6 support in 12.0S. ÂIt works pretty well, although the failover time is +/- 30 seconds.
> 

Which isn't any worse than traditional spanning tree (not that that is
great these days, but it's better than nothing, and seems to still be
used a fair bit).

Another possibility would be to provide last hop redundancy for an
application on a single LAN segment, where the host people aren't
trusted to inject routes into the network by the routing people e.g.
colo scenarios, or where the network and the hosts are operated by
separate organisations with separate SLAs. Recovery time isn't as good
as HSRP/VRRP etc. obviously, but its better than nothing, and should be
commonly available as baseline IPv6 functionality, unlike IPv4 which
doesn't have a "within prefix anycast" mechanism at all.

It could possibly be used for multiple NTP servers on a LAN
segment instead of multicast. 30 second recovery from failure wouldn't
be an issue (IIRC, NTP only issues time queries once every 60 seconds
or so once the local clock has converged, and of course the local clock
keeps ticking even if the NTP server goes away), and it would avoid
periodic multicast NTP queries from the clients being flooded through
the segment.

Regards,
Mark.