Measuring IPv6 Deployment Geoff Huston George Michaelson research@APNIC.net this is a story about trying to fit data to a question rather than a well planned exercise of finding a question and then generating a test rig and gathering data -0 i.e. it’s the “scientific method” applied backwards!

The story so far… In case you hadn’t heard by now, we appear to be running quite low on IPv4 addresses! I think Erik gave me special dispendation to talk about ipv4 exhaustion

and tyo talk about treain wrecks – this is the gare du nord I believe in Paris of course. I can just see the picture when the engine driver gets home “Had a good day on the trains today dear?” “Well, you wouldn’t believe what happened to me today…” 

IANA Pool Exhaustion Total address demand Prediction IANA Pool Looks like the train managed to do a fill 360 wheelie! Advertised IANA Pool Unadvertised RIR Pool

In this model, IANA allocates its last IPv4 /8 to an RIR on the 15th April 2011 This is the model’s predicted exhaustion date as of the 10th March 2009. The predictive model is updated daily at: http://ipv4.potaroo.net what is holding the train up? How the hell can you suspend a few tons of steel in mid air like that? Oh year, and we’ll run out of Ipv4 addresses soon,

Ten years ago we had a plan … IPv4 Pool Size IPv6 Deployment Size of the Internet Dublin – only the Irish can attempt to emulate the french and just not get it right! IPv6 Transition using Dual Stack 6 - 10 years 2000 2006-2010 Time

Oops! We were meant to have completed the transition to IPv6 BEFORE we completely exhausted the supply channels of IPv4 addresses! Its like elephants mating isn’t it?

What’s the revised plan? Today IPv4 Pool Size Size of the Internet Ahah – must be a record – this crash killed a plane, a car AND a horse. Now none of those are high speed vehicles so you really have to wonder how it happened! ? IPv6 Transition IPv6 Deployment Time

Its just not looking good is it? Buster Keaton – did all his own stunts.

IPv6 Deployment The new version of the plan is that we need to have much of the Internet also supporting IPv6 in the coming couple of years You just have to wonder how it corners! Or if!

How are we going today with this new plan?

How are we going today with this new plan? OR: How much IPv6 is being used today?

Can the data we already collect be interpreted in such a way to provide some answers to this question? ok – on to the subject at hand – can we use long line data series to say anything useful about Ipv6 uptake – i.e. not just take a snapshot of the net and say something but observe some aspect of the network for years and years and then attempt to say something useful about Ipv6?

How much IPv6 is being used today? At APNIC we have access to dual-stack data for: BGP Route table DNS server traffic WEB Server access and the data sets go back over the past 4 years What can these data sets tell us in terms of IPv6 adoption today? so here’s the data sets we can use @ apnic that go back for many years. DNS measurements back to 2002, Weblogs even further.

The BGP view of IPv6 1800 start with BGP table size – bgp Ipv6 table size, snapshot taken every hour since 2004. 2008 looks good doesn’t it. 400 2006 2008 2004

The BGP view of IPv4 300K 120K 2004 2006 2008 same data – but for IPv4 note the y axis scale 120K 2004 2006 2008

BGP: IPv6 and IPv4 300K putting the two together – the Ipv6 line looks like one of those hospital machines when the patient has expired – flat line. 2006 2008 2004

BGP IPv6 : IPv4 0.6% 0.3% 2004 2006 2008 so here’s the ratio So what this says is that IPv6 in 2008 is growing faster than IPv4 across most of 2008 2006 saw the demise of the 6bone and 2007 was, well, confused, for most of the year. interestingly the economic downturn is not readily visible in these numbers 0.3% 2004 2006 2008

What’s this saying? Since mid-2007 there appears to have been increasing interest in experience with routing IPv6 over the public Internet you can read it as well as I - yes?

What’s this saying? V6 is 0.6% of IPv4 in terms of routing table entries Growth is 0.22% p.a., linear IPv6 deployment will reach IPv4 levels in 452 years But the routing domain of IPv4 is heavily fragmented, while IPv6 is not Assuming IPv6 will exhibit 1/3 of the routing fragmentation of IPv4, then IPv6 deployment will fully span the Internet in about 149 years! assuming that the relativities of growth remain constant then the growth will be linear, and this is the outcome of the linear projection – 150 years at best.

This seems highly implausible! What’s this saying? V6 is 0.6% of IPv4 in terms of routing table entries Growth is 0.22% p.a., linear IPv6 deployment will reach IPv4 levels in 452 years But the routing domain of IPv4 is heavily fragmented, while IPv6 is not Assuming IPv6 will exhibit 1/3 of the routing fragmentation of IPv4, then IPv6 deployment will fully span the Internet in about 149 years! This seems highly implausible! 150 years, or even 149 years, is a ludicrous outcome.

What’s this saying? Routing is not traffic - the relative level of IPv6 use cannot be readily determined from this BGP announcement data maybe this is the wrong data set – bgp table entries have all kinds of things in them including traffic engineering, routing policies, etc.

Lets refine the question How much of the Internet today is capable of running IPv6? One way to answer this is to look at IPv6 routing on a per-AS basis so can we change the question to something more useful?

IPv6 AS Count 1400 AS’s announcing Ipv6 prefixes 300 2004 2006 2008

IPv4 AS Count 32K ditto for ipv4 16K 2004 2006 2008

AS Count IPv6 : IPv4 4.4% and the ratio – much the same story as BGP table entries isn’t it. 2.2% 2004 2006 2008

What’s this saying? The number of AS’s announcing IPv6 routes has risen from 2.5% to 4.2% from Jan 2004 to the present day 4.2% of the networks in the Internet are possibly active in some form of IPv6 activity as it says.

What’s this saying? At a relative rate of update of 0.8% per year, a comprehensive update to IPv6 is only 120 years away. but again the extrapolation does not give a brilliant result does it? 120 years!

This too seems highly implausible! What’s this saying? At a relative rate of update of 0.8% per year, a comprehensive update to IPv6 is only 120 years away. This too seems highly implausible! I don’t buy it!

That 4.2% is not uniform In IPv4 4,002 AS’s are transit networks and 26,874 are origin-only Of the 4,002 IPv4 transit AS’s 687 also have IPv6 routes 440 of these IPv4 transits are IPv6 stub ASs 17.1% of V4 Transit AS’s also route IPv6 Of the 26,874 V4 stub AS’s 630 also route IPv6 49 of these IPv4 stubs are IPv6 transit ASs 2.3% of V4 Origin AS’s also route IPv6 If you split the AS’s into stub and transit the numbers look a little better

What’s this saying? The proportion of IPv4 transit ASNs announcing IPv6 prefixes has risen by 3.3% in 12 months At this rate comprehensive Ipv6 deployment in the “core” will take only 25 more years. and the extrapolation looks better too!

Even this seems highly implausible! What’s this saying? The proportion of IPv4 transit ASNs announcing IPv6 prefixes has risen by 3.3% in 12 months At this rate comprehensive Ipv6 deployment in the “core” will take only 25 more years. Even this seems highly implausible! but 25 years is still too long isn’t it!

Capability vs Actual Use As 17% of the number of transit AS’s are announcing IPv6 address prefixes, does this mean that 17% of the Internet’s “core” is running IPv6 right now? and can we “believe” these numbers anyway – does announce = traffic?

Capability vs Actual Use As 17% of the number of transit AS’s are announcing IPv6 address prefixes, does this mean that 17% of the Internet’s “core” is running IPv6 right now? This seems highly implausible! I don’t think so.

Interpretation of BGP data Transit-AS uptake appears to be the most relevant indicator of IPv6 uptake in the ISP sector An investigation of topology comparison of IPv4 and IPv6 inter-AS network should be undertaken Is the IPv6 BGP network isomorphic to the IPv4 BGP network? Or is this network being constructed as an overlay using a different set of connectivity enablers and constraints?

DNS Server Stats APNIC runs two sets of DNS servers for the reverse zones for IPv4 and IPv6 One set of servers are used to serve reverse zones for address ranges that are deployed in the Asia Pacific Area The second set of servers are used as secondaries for zones served by RIPE NCC, LACNIC and AFRINIC DNS servers which are “high” in the apex of the in-addr delegation chain, and take of the order 5000 q/sec per instance, including a high proportion of PTR requests: either nodes with no reverse DNS, or nodes where the resolver doesn’t have up to date cache and is going to follow the additional section to the delegation masters

DNS Reverse Query Load Examine the average query load for reverse PTR queries for IPv6 and IPv4 zones for each of these server sets This is the longest baseline measure in DNS we hold at APNIC, based on a sampling methodology, 1min samples every 15min across all our servers.

DNS Reverse Query Load 100K 100 0.001 2004 2009 PTR queries per second IPv4 100 Caution: Log Scale! IPv6 Its immediately apparent from this slide that both V4 and V6 have been increasing (mostly) consistently since 2004, and that the initial apparent rate of increase in V6 is encouraging. But, the log scale makes this more attractive than it really is. Its better to see the ratio of V6 to V4.. 0.001 2004 2009

Relative DNS Query Load 2% 1% Once you go to to a relative V6/V4 view, you can see that the percentages under 2%, and there is a marked increase in 2008/2009, Linear Scale 2004 2009

What’s this saying? Reverse DNS queries for IPv6 addresses are around 0.2% of the IPv4 query load AsiaPac IPv6 query load was higher than for other regions to 2008, now lags Query load has increased since 2007 The interactions of forwarders and caches with applications that perform reverse lookups imply a very indirect relationship between actual use of IPv6 and DNS reverse query data

DITL 2008 to Present AP 2009 1% 2008 0% V6 transit improvements For a slightly different take, this is a graph of the relative v6/v4 counts of querying source Ips, comparing the DITL whole-day capture from 2008, with the current figures using our ongoing packet capture. The thicker lines for the 2009 data are due to the 3 day sample and a higher variance, but the increase is visible: from at best <1% to 1.2% year on year. Which appears to be less than the increase which I believe stems from radically improved IPv6 transit to the Brisbane node (where ns1 is) which happened during the same time. 0%

DITL 2008 to Present RoW 2009 1% 2008 The relative improvement in the same period for the Rest of the World hosts is far less. This looks to be a factor of the ‘in region’ quality of DNS reverse query: there is a strong in-region preference, and nodes which use the asia-pacific located NS for queries about Asia/Africa/Middle-East IP addresses appear to be less impacted by improved IPv6. –But, there is also a smaller relative improvement in Ipv6 uptake 2008 to 2009. 0%

What’s this saying? Best-case improvement in V6/V4 ratios from 2008 is 2x increase in V6 in a year Arguably more improvement if V6 transit improved than from ‘growth’ in V6 AP saw bigger increases than RoW Local RTT preference? If we intuit this as a normal ‘it doubles every year’ Internet model, then use of IPv6 in infrastructure DNS is increasing, but the most optimistic rate for it to get to equivalence with IPv4 is of the order 6 to 7 rounds of doubling. –Which also seems very implausible, but we continue to measure this, and will review as time goes on.

Web Server Stats Take a couple of dual-homed web servers: http://www.apnic.net http://www.ripe.net Count the number of distinct IPv4 and IPv6 query addresses per day Not the number of ‘hits’, just distinct source addresses that access these sites, to reduce the relative impact of robots and crawlers on the data and normalize the data against different profiles of use Look at the V6 / V4 access ratio What proportion of end host systems will prefer end-to-end IPv6, when there is a choice? As it says!

APNIC Web Server Stats 8% 0% 2004 2006 2008 And the data…. That is a HUGE spike 0% 2004 2006 2008

What happened on the 12th September 2008? So lets ‘zoom in’ on that spike event. –A very specific date, and then a long tailoff in load. What caused this?

here’s the culprit!

How do you say “we’ve been slash dotted in Chinese?”

RIPE NCC Web Server Stats 1.2% Here’s RIPE‘s data 0.0% 2004 2006 2008

Combined Stats 1.4% The two data sets show some good correlation over time 0.0% 2006 2008 2004

Combined Stats 1.4% 0.0% 2008 2004 2006 APNIC Meetings RIPE Meetings But they are small data sets – even the RIR meetings distort the numbers 0.0% 2004 2006 2008

What’s this saying? Relative use of IPv6 when the choice is available is 0.2% in the period 2004 – 2006 Relative use of IPv6 increased from 2007 to around 1% today Is interest in IPv6 slowing picking up again? Increased use of auto-tunneling of IPv6 on end host stacks?

Use of V6 Transition Tools APNIC Web Server Stats 100% 50% hmm – the chinese slashdot effect has drowned out the data 0% 2006 2008 2004

Use of V6 Transition Tools RIPE NCC Web Server Stats 100% 50% But RIP’s is interesting 6to4 is NOT going done – strange Teredo continues on – equally strange (hacker client) 0% 2004 2006 2008

Use of V6 Transition Tools Combined WebStats 0% 50% 100% 2004 2006 2008

Transition Tools in DNS Combined Stats 50% 25% This is totally bizarre: there is a measureable (if small) volume of Infrastructure DNS happening from Teredo: again, the ‘hacker’ community, that can only use this if they deliberately co-erce it into their hosts, because Teredo in Vista is an even lower preference than IPv4. –Yet, reverse-DNS queries are consistently seen from Teredo 2001:0: sources, and there is arguably even growth tracking the overall Ipv6 growth rate. -6To4 is obviously far more prevelant, and of the same order as for the web, at 25% of the seen Ipv6. Which suggests that tunneled IPv6 usage is consistent in both domains of interest. 0% 2008 2009

What’s this saying? Up to 25% of IPv6 clients in the Euro/ Mid East Region appear to use access tunneling techniques across an edge Ipv4 infrastructure The use of IPv6 clients using access tunneling is lower in the Asia Pac region Infrastructure DNS is using tunnels Even Teredo (lower pref than v4 in Vista)

Where are we with IPv6? The “size” of the IPv6 deployment in terms of end host IPv6 capability is around 10 per thousand Internet end hosts at present so the web server stats say 1% of hosts will use Ipv6 when there is a choice But this is on the high end of the guess range for the reasons explained here.

Where are we with IPv6? At most! This may be too optimistic: Widespread NAT use in IPv4 undercounts IPv4 host counts These web sites are tech weenie web sites. More general sites may have less IPv6 clients So perhaps the current IPv6 deployment level for end users may be closer to 5 – 7 per thousand so the web server stats say 1% of hosts will use Ipv6 when there is a choice But this is on the high end of the guess range for the reasons explained here.

What’s the revised plan? IPv4 Pool Size 100% Size of the Internet ? Today IPv6 Transition IPv6 Deployment Time

What’s the revised plan? IPv4 Pool Size 100% Size of the Internet So I’d guess that at present it would be say to claim that around one half of one percent of the Internet is ipv6 capable. ? Today IPv6 Transition IPv6 Deployment 0.5% Time

Thank You! research@apnic.net