Archive for March, 2013

Predicting IPv6 transition

March 29, 2013

Recently someone asked if I had any predictions about when/how the world will transition from IPv4 to IPv6.  After writing up my response, I thought it might be worth posting.  If nothing else, this will make it easy to check back in a few years and see how far off I was.  😉

So, the only predictions I can make with much confidence are qualitative, not quantitative.  As we’ve expected for awhile now, the the first step in IPv6 transition was always going to be the dual-stacking of major backbones and big content.  That is now essentially complete for the biggest content providers, and others can easily come on board with very little effort (if nothing else, they can put an IPv6-capable CDN like Limelight in front of their website).

But where it really gets interesting is to see where people actually start transitioning *off* of IPv4.  To date the only big deployments I know of there are people like Comcast renumbering their internal addressing (of set-top-boxes and the like) from private IPv4 to IPv6.  But moving actual customer traffic off of IPv4 onto IPv6 is much harder: Comcast has publicly admitted that due to lack of home router IPv6 support in anything but high-end routers like the Airport Extreme, they will only be providing IPv6 service to customers who plug directly into their modems and don’t use a router for the most part.
Mobile is another story, though.  As it happens, that is where almost all of the growth is (in number of subscribers, if not bandwidth), and as Cameron Byrne has presented at several conferences recently, there is a working IPv6-only solution for handsets called NAT464XLAT that allows them to use IPv6 on the internal network, and only use IPv4 on the external NAT pools for connections to the IPv4 Internet.  Since network and phone support are already there, I suspect they will begin gradually rolling it out mostly to test customers, with a plan to start cutting production customers over as soon as they make their last IPv4 request from ARIN.
Similar technologies (like DS-LITE) also exist for residential deployments, but will not be practical until IPv6-capable home routers begin shipping en masse.
It’s not clear to me how networks in Asia are dealing with their new inability to get new IPv4 addresses.  I suspect that since they were built from the ground up around heavy NAT, they are simply putting more subscribers behind the same number of addresses.  I’m not sure if they’re also deploying IPv6 much or not.
The other aspect I haven’t discussed is enterprise networks.  They tend to be very comfortable with NAT as well, so likely aren’t feeling a lot of pressure.
Many content providers will initially be fine with their existing space for organic growth, as new servers tend to be higher-capacity and require fewer IPs for the same traffic.  But large fast-growing content providers will probably place a high value on addresses (when you’re spending $10k/server, $10/IP is a rounding error) and will be willing to pay to acquire whatever addresses they need.
For some time, I think the supply of addresses will come mostly from organizations that are defunct or no longer need the addresses at all.  But if IPv6 transition drags on for awhile (as it probably will), we will get to a point where the marginal supply will be provided by organizations using the money to finance renumbering out of existing space onto some sort of NAT64 technology (like those highlighted above) with a much smaller pool of IPv4 kept for NAT and static-IP services.  As Lee Howard highlighted at a recent NANOG, that demand should be fairly elastic, as there are lots of people who can renumber out of IPv4 at the right price.
But getting back to more quantitative predictions, I think IPv6 deployment will follow a classic S-shaped curve, starting exponential, growing slowly at first, moving through the middle fairly rapidly, and then asymptotically approaching 100%.  As we’re still in the first exponential phase, it’s difficult to fit a good curve to the data and predict when it will take off.  Just guesstimating I’d say 5 years, or a range of 2-10 years (CI 90%), to reach the point where half of the devices are running IPv6 (either native or dual-stack).