BCP: Slicing a /32 for an ISP

[Doug Barton]

Mark Tinka wrote:

...

> * From each /48 assigned to each PoP

...


Mark,

Thanks for this in-depth analysis. In my opinion this type of thought is 
what has long been missing from the "it's /48s all the way down" dogma 
that has been the default answer to IPv6 allocation discussion for too 
long.

Two small contributions I'd like to throw in to the thread. As much as I 
am a proponent of not using the "abundant" IPv6 address in a profligate 
manner, even I would suggest using public address for everything, 
including router loopbacks, etc. Someone else already mentioned the 
issues with traceroute, which IMO is sufficient justification on its own.

The other thought is about the name server issue. We are generally 
conditioned to think of "the name server" as the box that the software 
runs on, however I try to get my clients to think in terms of "name 
server instances" which you can define as different instances of named 
(or whatever you use) running on the same box/virtual server, or even 
the same instance of named listening on multiple IPs, whether v4, v6, or 
both. For most enterprises, especially for resolving name servers, it's 
unlikely that they will be working hard enough to justify one named 
instance per piece of hardware given how much horsepower even the 
low-end commodity equipment has nowadays. Thinking in terms of "one name 
server instance per IP" may create some logistical hurdles while getting 
everything set up, but it pays dividends down the road with greater 
agility. (It should go without saying that I'm not suggesting lumping 
all your name server eggs into one hardware basket, you still need 
physical redundancy for critical infrastructure.)

As always, your mileage may vary.

Doug