The big bottleneck is this: the vast majority of top-tier cloud providers (AWS, Azure, etc.) do not support IPv6 and seem completely uninterested in doing so. Until this changes I don't foresee a tipping point.
I really don't understand why this is. None of these are old systems. IPv6 existed when AWS was built. Were they really designed from the ground up without IPv6 support?
Of course I kind of know the answer. While most of hackerdom is pretty open to trying new things, most networking engineers are (in my experience) ultra ultra conservative and terrified of change. The majority of the network engineering community has dug in its heels like a stubborn mule at IPv6 and must be dragged grumbling into it.
>Were they really designed from the ground up without IPv6 support?
Yup. It is simply amazing how many things I see devs push out that hard code badness (such as php ip2long, as an example). Intrusion detection that only will correctly resolve and send IPv4 to automatic firewall rules. Interfaces that truncate IPv6 addresses so information is lost. Who knows how many security problems are potentially added by this code that expects one thing and can get something else.
You've actually hit upon (part of) the answer right there. While IPv4 was designed around the (then) new 32 bit processors of the day, the committee that came up with v6 decided to ignore the underlying hardware reality. There's no corresponding type in C (or the underlying hardware, without going into SSE) that can store a v6 address natively. It just acts as a handbrake and makes it that much harder to implement. The question should be turned around, why didn't they design a protocol that could fit into a built-in type in C? After all, we've only just "run out" of v4 addresses about 20 years after it was first predicted to happen. Going to 64 bit would have kicked the can another few decades down the road, and could have led to a much faster implementation. You can't blame the "designers" of C (since there aren't any really, it's just the CPU manufacturers), it's the IETF that ignored the reality of hardware and programming languages.
You've actually hit upon (part of) the answer right there. While IPv4 was designed around the (then) new 32 bit processors of the day, the committee that came up with v6 decided to ignore the underlying hardware reality. There's no corresponding type in C (or the underlying hardware, without going into SSE) that can store a v6 address natively
Huh? That's what abstractions like sockaddr_storage are for. Who cares how the address maps to the underlying CPU word size?
It's the same reason you shouldn't be storing pointers in ints. Portability requires effort. That's just life.
IP64 would have worked, and would have kicked the can much further than another few decades. If I were asked to invent IP 2.0 I'd have suggested just extending the V4 address space by another 32 bits and calling it a day.
But I do see the advantage of big addresses. They give you room for meaningful cryptographic semantic information and stateless auto configuration, both of which are very valuable in big distributed systems and for security.
So yeah, could have been simpler but not a big deal. Things like struct sockaddr_storage make programming it pretty easy too, but most older devs aren't familiar with them.
IP64 would have worked, and would have kicked the can much further than another few decades. If I were asked to invent IP 2.0 I'd have suggested just extending the V4 address space by another 32 bits and calling it a day.
And you'd still have to rewrite all software to take advantage of it.
Everyone who thinks they could've come up with a better solution seems to conveniently forget that no solution would've both extended the address space while providing free backward compatibility with v4. Those are, quite literally, incompatible goals.
Any hacky solution you do come up with (like, say, NAT64) applies to IPv6, as well.
Meanwhile, IPv6 goes on to address a whole host of other issues, like the mess that is CIDR, simplified autoconfiguration, built in address randomization/privacy extensions, etc.
I don't think that's true - there were proposals that extended the address space in a backwards-compatible way but they were rejected in favour of a clean-slate approach that required rewriting everything and breaking existing networks.
How do you extend the address space such that old devices in legacy addresses can communicate with new devices in an extended address space without NAT64-like hacks.
This is probably the best historical document looking at the original discussions around IPng (which eventually became v6). Interestingly these debates were happening right around the time of the invention of the www so there isn't a lot online today.
Section 6 (Transition Plans) talks about plans to use new headers or IP Options to carry the extra information for the new protocol in an IPv4 packet. In the end they went for the alternative which was not to have a transition plan, they went for the dual stack approach instead.
The SIP proposal is interesting, it's basically the same as v4 but with 64 bit addresses. They were worried that it might not be enough, of course their predictions about exhausting the 32 bit address space were off by decades.
That section makes my point for me. From the text:
(a) That IPng hosts can also use IPv4 or
(b) There is translation by an intermediate system
And that "The transition plans espoused by the various proposals are simply different combinations of the above."
Those options are precisely equivalent to dual stack and NAT64 respectively. No method of extending the address space van avoid doing one of the two.
Amusingly, it even predicts the situation today, noting that "Experience would tend to show that all these things will in fact happen, regardless of which protocol is chosen."
The mistake the committee made was dismissing NAT immediately instead of standardizing and approach alongside dual stack, to provide operators with options.
An interesting thought experiment is "What would the Internet be like if everyone used 6to4, without the anycast gateways?" I think it would have been deployed more rapidly, at the cost of being architecturally hideous.
Personally I think there is still room for a new version of IP. v6 is still only in single digits. If something better came along (that was backwards compatible!) you could potentially outstrip it. Just because some committee made a bad decision 20+ years ago doesn't mean that the discussion is over. A lot of the people implementing these solutions were children when v6 was designed and they might have new ideas. A 64 bit, backwards compatible version of IP could be a winner. Part of the problem of course is that most people still don't see v4 as needing a replacement because it's still working.
> After all, we've only just "run out" of v4 addresses about 20 years after it was first predicted to happen.
Look, if by the 50's somebody make some calculations and said "if we don't change our ways, we'll be out of X by 2030", and X runs out in 2030, this person will be perfectly on time, not late by 80 years.
You have to rebuild the whole SDN layer to support it and then test it sufficiently to go to production. It's a lot of work, not just "bigger addresses".
I really don't understand why this is. None of these are old systems. IPv6 existed when AWS was built. Were they really designed from the ground up without IPv6 support?
Of course I kind of know the answer. While most of hackerdom is pretty open to trying new things, most networking engineers are (in my experience) ultra ultra conservative and terrified of change. The majority of the network engineering community has dug in its heels like a stubborn mule at IPv6 and must be dragged grumbling into it.