IPv6 adoption starting to add up to real numbers: 0.6 percent

In a paper presented at the prestigious ACM SIGCOMM conference last week, researchers from the University of Michigan, the International Computer Science Institute, Arbor Networks, and Verisign Labs presented the paper "Measuring IPv6 Adoption." In it, the team does just that—in 12 different ways, no less. The results from these different measurements don't exactly agree, with the lowest and the highest being two orders of magnitude (close to a factor 100) apart. But the overall picture that emerges is one of a protocol that's quickly capturing its own place under the sun next to its big brother IPv4.

As a long-time Ars reader, you of course already know everything you need to know about IPv6. There's no Plan B, but you have survived World IPv6 Day and World IPv6 Launch. All of this drama occurs because existing IP(v4) addresses are too short and are thus running out, so we need to start using the new version of IP (IPv6) that has a much larger supply of much longer addresses.

The good news is that the engineers in charge knew we'd be running out of IPv4 addresses at some point two decades ago, so we've had a long time to standardize IPv6 and put the new protocol in routers, firewalls, operating systems, and applications. The not-so-good news is that IP is everywhere. The new protocol can only be used when the two computers (or other devices) communicating over the 'Net—as well as every router, firewall, and load balancer in between—have IPv6 enabled and configured. As such, getting IPv6 deployed has been an uphill struggle. But last week's paper shows us how far we've managed to struggle so far.

In an effort to be comprehensive, the paper (PDF) visits all corners of the Internet's foundation, from getting addresses to routing in the core of the network. The researchers also got their hands on as many as half of the packets flowing across the Internet at certain times, counting how many of those packets were IPv4 and how many were IPv6.

The authors focused on content providers, service providers, and content consumers. For each of these, the first step toward sending and receiving IPv6 packets is to get IPv6 addresses. Five Regional Internet Registries (RIRs) give out both IPv4 addresses and IPv6 addresses. Looking at 10 years of IP address distribution records, it turns out that prior to 2007, only 30 IPv6 address blocks or address prefixes were given out each month. That figure is now 300; the running total is 18,000. IPv4, on the other hand, reached a peak of more than 800 prefixes a month in 2011 and is now at about 500. Although IPv6 is close on a monthly basis, IPv4 had a big head start and is currently at 136,000 prefixes given out.

Once a network obtains address space, it's helpful if packets manage to find their way to the holder of the addresses. Prefixes must be "advertised" in the global routing system so routers around the world know where to send the packets for those addresses. The number of IPv6 prefixes in routing tables in the core of the Internet was 500 in 2004 and 19,000 now—a 37-fold increase. IPv4, on the other hand, went from 153,000 to 578,000 in the same time, an increase of a factor four. Network operators often break up their prefixes into smaller ones, so this number is higher than the number of prefixes given out by the RIRs. Note that this is purely the number of address blocks, regardless of their size.

With routing and addressing out of the way, we need to start thinking about the DNS. When a system wants to talk to a service known by a domain name, it can ask the DNS system for the IPv4 addresses that go with that domain and have an A (address) record query. If the system also wants to know the IPv6 addresses that go with the domain, it has to perform a separate AAAA record query. These queries can be performed over either IPv4 or IPv6.

The researchers also looked at the DNS queries received by the .com and .net top-level domain nameservers. These nameservers need to be consulted along the way for every address lookup for every .com and .net domain, so they get several billion queries per 24-hour period. It turns out that about a third of IPv4 nameservers perform AAAA queries and about three quarters of the IPv6 nameservers do so. However, these figures include a lot of "small" nameservers that aren't very active. Of the big, active IPv4 ones—such as those at an ISP serving many customers—more than 90 percent perform AAAA queries. Of the big, active IPv6 nameservers, it's 99 percent.

The researchers also looked in more detail at IPv4 and IPv6 routing and network topology, which showed that IPv6 is "largely deployed" in the core of the Internet (the big service providers). That's not so much the case at the edges, where the content and the consumers live. Out of the Alexa top 10,000 websites, currently about 3.5 percent are IPv6-enabled. This number was less than half a percent in early 2011 before World IPv6 Day. (The paper doesn't mention that this number is much higher, at 23.4 percent, for the Alexa top 500.) Google tracks the number of its users that have IPv6 enabled, a figure that has been going up by a factor 2.5 every year the past few years, and it's listed as 2.5 percent in the paper. The current numbers are actually just over four percent during the weekends and a little under during weekdays. The paper also looks at the performance difference between IPv4 and IPv6. The result: IPv6 is 95 percent as fast as IPv4.

With all of the above out of the way, let's have a look at some actual packets. The researchers scraped together traffic information from as many as 260 sources, estimating that they got to look at a third to half of the Internet's traffic at times. Rest assured that this is merely Netflow data that counts the various types of traffic flowing through a router.