This Time Self-Hosted
dark mode light mode Search

More on IPv6 feasibility

I didn’t think I would have to go back to the topic of IPv6, particularly after my last rant on the topic. But of course it’s the kind of topic that leads itself to harsh discussions over Twitter, so here I am back again (sigh).

As a possibly usual heads’ up, and to make sure people understand where I’m coming from, it is correct I do not have a network background, and I do not know all the details of IPv6 and the related protocol, but I do know quite a bit about it, have been using it for years, and so my opinion is not one of the lazy sysadmin that sees a task to be done and wants to say there’s no point. Among other things, because I do not like that class of sysadmins any more (I used to). I also seem to have given some people the impression that I am a hater of IPv6. That’s not me, that’s Todd. I have been using IPv6 for a long time, I have IPv6 at home, I set up IPv6 tunnels back in the days of having my own office and contracting out, and I have a number of IPv6-only services (including the Tinderbox).

So with all this on the table, why am I complaining about IPv6 so much? The main reason is that, like I said in the other post, geeks all over the place appear to think that IPv6 is great right now and you can throw away everything else and be done with it right this moment. And I disagree. I think there’s a long way to go, and I also think that this particular attitude will make the way even longer.

I have already covered in the linked post the particular issue of IPv6 having originally be designed for globally identifiable, static IPv6 addresses, and the fact that there have been at least two major RFCs to work around this particular problem. If you have missed that, please go back to the post and read it there because I won’t repeat myself here.

I want instead to focus on why I think IPv6 only is currently infeasible for your average end user, and why NAT (including carrier-grade NAT) is not going away any year now.

First of all, let’s define what an average end user is, because that is often lost to geeks. An average end user does not care what a router does, they barely care what a router is, and a good chunk of them probably still just call them modem, as their only interest is referring to “the device that the ISP gives you to connect to the Internet”. An average user does not care what an IP address is, nor cares how DNS resolution happens. And the reason for all of this is because the end user cares about what they are trying to do. And what they are trying to do is browse the Internet, whether it is the Web as a whole, Facebook, Twitter, YouTube or whatever else. They read and write their mail, they watch Netflix, NowTV and HBO GO. They play games they buy on Steam or EA Origin. They may or may not have a job, and if they do they may or may not care to use a VPN to connect to whatever internal resources they need.

I won’t make any stupid or sexist stereotype example for this, because I have combined my whole family and some of my friends in that definition, and they are all different. They all don’t care about IPv6, IPv4, DSL technologies and the like. They just want an Internet connection, and one that works and is fast. And with “that works” I mean “where they can reach the services they need to complete their task, whichever that task is”.

Right now that is not an IPv6 only network. It may be, in the future, but I won’t hold my breath for a number of reasons, that this is going to happen in the next 10 years, despite the increasing pressure and the increasing growth of IPv6 deployment to end users.

The reason why I say this is that right now, there are plenty of services that can only be reached over IPv4, some of which are “critical” (for some definition of critical of course) to end users, such as Steam. Since the Steam servers are not available over IPv6, the only way you can reach them is either through IPv4 (which will involve some NAT) or NAT64. While the speed of the latter, at least on closed-source proprietary hardware solutions, is getting good enough to be usable, I don’t expect it being widely deployed any time now, as it has the drawback of not working with IP literals. We all hate IP literals, but if you think no company ever issue their VPN instructions with an IP literal in them, you are probably going to be disappointed once you ask around.

There could be an interesting post about this level of “security by obscurity”, but I’ll leave that for later.

No ISP wants to receive calls from their customers that access to a given service is not working for them, particularly when you’re talking about end users that do not want to care about tcpdump and traceroute, and customer support that wouldn’t know how to start debugging that NowTV will send a request to an IPv4 address (literal) before opening the stream, and then continue the streaming over IPv4. Or that Netflix refuse to play any stream if the DNS resolution happens over IPv4 and the stream tries to connect over IPv6.

Which I thought Netflix finally fixed until…

Now, to be fair, it is true that if you’re using an IPv6 tunnel you are indeed proxying. Before I had DS-Lite at home I was using TunnelBroker and it appeared like I was connecting from Scotland rather than Ireland, and so for a while I unintentionally (but gladly) sidestepped country restrictions. But this also happened a few times on DS-Lite, simply because the GeoIP and WhoIs records didn’t match between the CGNAT and the IPv6 blocks. I can tell you it’s not fun to debug.

The end result is that most customer ISPs will choose to provide a service in such a way that their users feel the least possible inconvenience. Right now that means DS-Lite, which involves a carrier-grade NAT, which is not great, as it is not cheap to run, and it still can cause problems, particularly when users use Torrent or P2P heavily, in which case they can very quickly exhaust the 200-ports forwarding blocks that are allocated for CGNAT. Of course DS-Lite also takes away your public IPv4, which is why I heard a number of geeks complaining loudly about DS-Lite as a deployment option.

Now there is another particular end user, in addition to geeks, that may care about IP addresses: gamers. In particular online gamers (rather than, say, Fallout 4 or Skyrim fans like me). The reason for that is that most of the online games use some level of P2P traffic, and so require you to have a way to receive inbound packets. While it is technically possible to set up IGD-based redirection all the way from the CGNAT address to your PC or console, I do not know how many ISPs implement that correctly. Also, NAT in general introduces risks for latency, and requires more resources on the passing routers, and that is indeed a topic that is close to the heart of gamers. Of course, gamers are not your average end user.

An aside: back in the early days of ADSL in Italy, it was a gaming website, building its own ISP infrastructure, that first introduced Fastpath to the country. Other ISPs did indeed follow, but NGI (the ISP noted above) stayed for a long while a niche ISP focused on the need of gamers over other concerns, including price.

There is one caveat that I have not described yet, but I really should, because it’s one of the first objections I receive every time I speak about the infeasibility of IPv6 only end user connections: the mobile world. T-Mobile in the US, in particular, is known for having deployed IPv6 only 3G/4G mobile networks. There is a big asterisk to put here, though. In the US, and in Italy, and a number of other countries to my knowledge, mobile networks have classically been CGNAT before being v6-only, and with a large amount of filtering in what you can actually connect to, even without considering tethering – this is not always the case for specialised contracts that allow tethering or for “mobile DSL” as they marked it in Italy back in the days – and as such, most of the problems you could face with VPN, v4 literals and other similar limitations of v6-only with NAT64 (or proxies) already applied.

Up to now I have described a number of complexities related to how end users (generalising) don’t care about IPv6. But ISPs do, or they wouldn’t be deploying DS-Lite either. And so do a number of other “actors” in the world. As Thomas pointed out over Twitter, not having to bother with TCP keepalive for making sure a certain connection is being tracked by a router makes mobile devices faster and use less power, as they don’t have to wake up for no reason. Certain ISPs are also facing problems with the scarcity of IPv4 blocks, particularly as they grow. And of course everybody involved in the industry hates pushing around the technical debt of the legacy IPv4 year after year.

So why are we not there yet? In my opinion and experience, it is because the technical debt, albeit massive, is spread around too much: ISPs, application developers, server/service developers, hosting companies, network operators, etc. Very few of them feel enough pain from v4 being around that they want to push hard for IPv6.

A group of companies that did feel a significant amount of that pain organized the World IPv6 Day. In 2011. That’s six years ago. Why was this even needed? The answer is that there were too many unknowns. Because of the way IPv6 is deployed in dual-stack configurations, and the fact that a lot of systems have to deal with addresses, it seems obvious that there is a need to try things out. And while opt-ins are useful, they clearly didn’t stress test enough of the usage surface of end users. Indeed, I stumbled across one such problem back then: when my hosting provider (which was boasting IPv6 readiness) sent me to their bank infrastructure to make a payment, the IP addresses of the two requests didn’t match, and the payment session failed. Interactions are always hard.

A year after the test day, the “Launch” happened, normalizing the idea that services should be available over IPv6. Even though that the case, it took quite a longer while for many services to be normally available over IPv6, and I think, despite being one of the biggest proponents and pushers of IPv6, Microsoft update servers only started providing v6 support by default in the last year or so. Things improved significantly over the past five years, and thanks to the forced push of mobile providers such as T-Mobile, it’s a minority of the connections of my mobile phones that still connect to the v4 world, though there are still enough not to be able to be ignored.

What are the excuse for those? Once upon a time, the answer was “nobody is using IPv6, so we’re not bothering supporting it”. This is getting less and less valid. You can see the Google IPv6 statistics that show an exponential growth of connections coming from IPv6 addresses. My gut feeling is that the wider acceptance of DS-Lite as a bridge solution is driving that – full disclosure: I work for Google, but I have no visibility in that information, so I’m only guessing this out of personal experience and experience gathered before I joined the company – and it’s making that particular excuse pointless.

Unfortunately, there are still “good” excuses. Or at least reasons that is hard to argue with. Sometimes, you cannot enable IPv6 for your web service, even though you have done all your work, because of dependencies that you do not control, for instance external support services such as the payment system in the OVH example above. Sometimes, the problem is to be found in another piece of infrastructure that your service shares with others and that requires to be adapted, as it may have code expecting a valid IPv4 address at some particular place, and an IPv6 would make it choke, say in some log analysis pipeline. Or you may rely on hardware for the network layer that just still does not understand IPv6, and you don’t want to upgrade because you still have not found enough of an upside to you to make the switch.

Or you may be using an hosting provider that insists that giving you a single routable IPv6 is giving you a “/64” (it isn’t — they are giving you a single address in a /64 they control). Any reference to a certain German ISP I had to deal with in the past is not casual at all.

And here is why I think that the debt is currently too spread around. Yes, it is true that mobile phones batteries can be improved thanks to IPv6. But that’s not something your ISP or the random website owner care about – I mean, there are websites so bad that they take megabytes to load a page, that would be even better! – and of course a pure IPv6 without CGNAT is a dream of ISPs all over the world, but it is very unlikely that Steam would care about them.

If we all acted “for the greater good”, we’d all be investing more energy to make sure that v6-only becomes a feasible reality. But in truth, outside of controlled environments, I don’t see that happening any year now as I said. Controlled environments in this case can refer not only to your own personal network, but to situations like T-Mobile’s mobile data network, or an office’s network — after all, it’s unlikely that an office, outside of Sky’s own, would care whether they can connect to NowTV or Steam. Right now, I feel v6-only network (without NAT64 even) are the realm of backend networks. Because you do not need v4 for connecting between backends you control, such as your database or API provider, and if you push your software images over the same backend network, there is no reason why you would even have to hit the public Internet.

I’m not asking to give a pass to anyone who’s not implementing v6 access now, but as I said when commenting on the FOSDEM network, it is not by bothering the end users that you’ll get better v6 support, is by asking the services to be reachable.

To finish off, here’s a few personal musings on the topic, that did not quite fit into the discourse of the post:

  • Some ISPs appear to not have as much IPv4 pressure as others; Telecom Italia still appears to not have reassigned or rerouted the /29 network I used to have routed to my home in Mestre. Indeed, whois information for those IPs still has my old physical address as well as an old (now disconnected) phone number.
  • A number of protocols that provide out-of-band signalling, such as RTSP and RTMP, required changes to be used in IPv6 environments. This means that just rebuilding the applications using them against a C library capable of understanding IPv6 would not be enough.
  • I have read at least one Debian developer in the past giving up on running IPv6 on their web server, because their hosting provider was sometimes unreliable and they had no way to make sure the site was actually correctly reachable at all time; this may sound like a minimal problem, but there is a long tail of websites that are not actually hosted on big service providers.
  • Possibility is not feasibility. Things may be possible, but not really feasible. It’s a subtle distinction but an important one.

Comments 6
  1. I’m not sure what your suggested way forward is. I’m pretty sure it’s not “it’s hard, therefore we shouldn’t do it”.You specifically call out IPv6-only networks as being a bad idea, but I’m not sure why. Or rather, if the “end user” is software developers then we can’t assume that it’s the same individuals who complained two years ago that something doesn’t work with IPv6 as it is this year. Essentially it’s giving hands-on experience to people going to a developer conference that the industry needs to get its shit together and support IPv6. It’s things like this that make them put “IPv6 support” on their RFP when they go back home.I’ve been part of setting up 3G and 4G core networks, and when new equipment comes in I’d test their IPv6 functionality. And *before* we sign the acceptance test sheet from the vendor I’d make sure that IPv6 is fully tested. We need more people to try this now, instead of saying “maybe we can get IPv6 after we’ve refreshed our core network in 5 years”, because that’d be 5 years that the future is deferred.Of course I’m coming from a place of “it’s in everybody’s best interests that we *eventually* all run IPv6 everywhere”, and we can’t get from here to there by waiting. We need to invent the future in order to get there. If you disagree with that then that’s a different conversation.When I say it’s in everybody’s best interest I mean that every bps shifted to IPv6 is a ~linear cost saving on CGN hardware, which is a huge amount of money in CAPEX and OPEX. In Sweden traditionally most mobile telcos actually used public addresses (thus no NAT), but I think you’re overlooking the HUGE increase in mobile network speeds over the last few years. 10 years ago you could run an entire country’s mobile network NAT on a single desktop PC. Hell, 10 years ago you could run all of Europe’s mobile data core network from a single rack (they didn’t, but that’s because telco industry is weird compared to internet industry. My point is valid though). Traffic volumes, customers, and state just weren’t that high. Today the backhaul cables to individual cell towers are getting saturated. Doing 10x NAT is more than 10x harder and expensive (costs passed on to the consumer). We need to break the star shape of the networks, and NAT gets in the way of that. Or at least we need to *enable* breaking the star shape networks. (star can be continent-sized today, by the way)So for end users it’s cost, battery life, latency, and throughput. (the two latter combining to “page load speed”). For ISPs/Telcos it’s CAPEX (less powerful NAT, and eventually no NAT), flexibility, OPEX (e.g. your access network doesn’t need 100 networks for address capacity reasons). Latency matters for games, but also for VoIP. I know people who can’t use Hangouts or Skype, and have to call each other internationally because the latency is so high they keep interrupting each other. The latency is for them a showstopper.So to me clearly where we want to be eventually is IPv6-only. And we can’t get there by not doing anything.But can we throw away IPv4 today for non-technical end users (except in special cases like T-Mobile)? No, probably not. Hell, as I’ve learned by putting my Windows machine on a whitelist-only HTTP(S) proxy[1] and firewalling other traffic that you can’t even do that in the general case because there’s plenty of programs (including parts of Windows itself) that ignores the system wide proxy settings.Now if your question is “so when can we drop IPv4 from consumer connections?”, then I think we’re some years off. There may be a long tail of IPv4-only-and-broken-with-NAT64, but honestly when Steam and Netflix, and maybe at most 50 more sites get their shit together then it’s probably fine. Or maybe the future should have a PAC that at least would solve it for (most) web traffic. No matter if we do it in 1, 10, or 100 years, there will be stragglers. But in the end we’ll have to do it so we’re really just talking about at what point the long tail is sacrificable, not if.[1] https://github.com/google/s

  2. I totally agree with what you’re saying (and that’s why I said we’ve been violently agreeing). And I don’t really have a good answer, except for saying that any ISP still stuck on v4-only should get off their seat and go *at least* to DS-Lite. More DS-Lite users would compel the v4-only services into adding support to v6, and that should make the long tail more apparent quickly.The whole point of why I started the long discussion on Twitter was whether we’re ready to abandon v4 entirely. We’re not, though I wish, and the final users don’t care. Final users don’t care about NAT, whether it is CGNAT on v4 or NAT64, and that’s not going to change in by itself. The “enemy” of adoption is the “no NAT or no v6” stance that I have seen too many people take against DS-Lite as an option. DS-Lite gives us at least something to work with.

  3. By the way, I just realized that the thread with Hector on twitter actually has an answer of what is very likely needed now: AWS and GCP to support v6!Whether we like it or not (I shall not comment), lots of new startup services/apps rely on AWS, GCP and similar offers. The two major players not supporting v6 is sad 🙁

  4. I thought AWS had it now, but it seems that maybe not Internet-facing. https://aws.amazon.com/blog…But yeah, it’s sad indeed to see products launched even in 2017 without IPv6 support.But on Twitter I think I was not so much apparently disagreeing with you as disagreeing with this other person who appeared to have opinions while not know at all what they were talking about, and even admitted as much in at least one tweet.

  5. Uh, I think we then crossed-messages because I was mostly ranting at a third person who complained about NAT and the fact that customers care about it (they don’t, providers do, they suffer, as you pointed out), and thus DS-Lite is no solution.The person who admitted to not knowing was initially honestly asking questions 🙂

  6. Most AWS have IPv6 support, but the individual services using AWS have to opt in.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.