I think my distaste for NAT is pretty well known by now. I’ve talked time and again about how I believe that NAT is a bad idea, especially where IPv6 is concerned. I’d said my peace and had time for good conversations with my friends Ivan Pepelnjak (@ioshints) and Ed Horley (@ehorley) about the subject. I was content to just wear my “I HATE NAT” t-shirts to conferences and let bygones be bygones. Then, suddenly…
Some of you have seen my responses before. Maybe you’ve even been amused by them. Coupled with the fact that I tend to lean toward the snarky side of things, I can see where I might come off as a bit of a smart ass. But “belitted?” “chastened?” Wow. Maybe I’ve let my passion blind me to the plight of the Small-to-Medium Enterprise/Business (SME/B) network/server folks. Maybe we really should stop trying to undo years of duct tape patches to networking and embrace the fact that NAT is a great thing because it allows the little guys to spend less time changing ISPs and deciding to renumber their internal networks on a whim. In fact, I’m even considering calling all my buddies at the IETF and rescinding the whole idea of IPv6. I mean, after all what good is renumbering the Internet if it breaks such a fundamentally important protocol like NAT?
Oh, sorry. I just couldn’t keep a straight face anymore…
In all seriousness, Trevor Pott (@cakeis_not_alie) brings up some very interesting points in his discourse on the impact of IPv6 for the Small-to-Medium Enterprise/Business (SME/B). I’m even willing to admit that we might have glazed over some of the lower-end aspects of what a change like this will mean to people on the razor’s edge of IT. But in the article, the painting of networking professionals as uncaring dictators of fiat laws is almost as silly as characterizing me as a belittling jackass. Yes, I write some pretty pointed posts about NAT and IPv6. Sure, I have a lot of fun playing the heel. But I would hope that my points are made from somewhat sound networking reasoning and not simply blind hatred of NAT and those that use it. Yes, especially those on the edge of networking.
When I was an intern at IBM Global Services in 2001, I had my first real exposure to the way networking operates. I spent a lot of time configuring static IP addresses on devices and using DHCP on others. I got a real eye opening experience. It even colored my perception of networking for a few years to come, although I didn’t know it at the time. You see, as one of the “old guard” networking companies, IBM has their own registered /8 (Class A) network prefix. Everything inside IBM runs on 188.8.131.52/8. Apple similarly has 184.108.40.206/8. These folks have the luxury of a globally routable IP space large enough that they never (realistically) have to worry about running out. For many years afterwards, I couldn’t understand why I was unable to reach my 192.168.1.0/24 network at home from my university network. It wasn’t until I really started learning more about networking that I realized that RFC1918 existed and NAT was in place to allow ever-growing LANs to have Internet access in absence of registered IP space like I had enjoyed at IBM. As time moved on and I started becoming involved with more and more network services that are affected by NAT, I began to see what IBM’s /8 could offer an enterprise. The flexibility of being static. By having their own IP space, IBM didn’t have to change their address structure to suit the needs of users. They never had to worry about changing ISPs and renumbering their network. Everything just stayed the same and we went on with our lives. But, as Trevor Pott pointed out in the article, IBM and Cisco and Juniper and Apple are enterprises. They aren’t…small.
On the polar opposite end of the scale, we have the little guys. The folks that keep law offices running on a SOHO router/firewall/DHCP server. The accounting offices that can only get a /28 or a /29 IPv4 block from their ISP. Folks that look at duct tape as a solution and not a patch. The “cost conscious” customer as one might say. I can identify with many of these customers because their are my customers in my day job. I’ve had to renumber a publicly addressed network on the fly on a Saturday morning. I’ve had to reconfigure firewalls because the ISP decided to reclaim IP space from a customer. It’s a giant pain in the exhaust port. It’s not glamorous or fun. It’s a necessary evil. But is it a reason to rail against IPv6?
In my previous posts, I talked about the issues with IPv6 on the small end of the scale. Sure, we’ve got a lot of addresses to hand out. We’ve also got a lot of configuration to do. We have to reexamine our stand on firewalls and routes and DNS and a lot of other things. Yes, I will freely admit that this isn’t going to be cheap. I’ve already started building the costs of these analyses into the contracts I sign with my customers for the coming year because I know it will need to be done and I don’t want them to be surprised when they get the message from their provider that the time has come to renumber. But I’ve also got another solution in mind for some of my most “cost conscious” customers and readers. I don’t really want to spill the secret sauce, but here goes:
If it’s going to bother you that much, don’t use IPv6.
Plain and simple in black and white (and red). Unless your ISP is going to start charging you an inordinately high monthly fee to keep an IP block you’ve had for years, don’t change. Stay on IPv4. There’s a whole world out there that is never going to move from IPv4 and be perfectly happy. People who run equipment that will never have enough memory or CPU power to process a naked IPv6 packet (let alone a NATed or NPTed packet). People who are mandated to use translated addresses because of some kind of regulatory oversight like the Payment Card Industry (PCI). I really don’t mean to sound like I’m snorting derisively with this advice. If the additional cost of maintaining an IPv6 network with things like link local addresses and proper DNS resolution and multiple firewall translations isn’t worth it to you and your customer base, then stay where you are. No one will come to your office and put a gun to your head to make you move. The issues we have with address space exhaustion inside enterprises are a wholly different animal than keeping the small office going. Honestly, you folks will stay in business for years to come serving a subset of the Internet populace. There may come a time when there is an opportunity cost of being able to reach new customers that are IPv6-only, but that cost will either be balanced by the need to trade out your “low cost” equipment for something that will run newer IPv6 features or it will be balanced out by your need to get any business to offset falling revenues due to IPv6-only customers no longer being able to reach your site on IPv4.
If you’re an SME/B network admin that’s still reading this, I’d highly recommend that you take a moment to think about something though. Is IPv6’s insistence on one-to-one communications and move away from NAT really disrupting the way the Internet works? Or is it moving back to the way things were before? Setting right what once went wrong? One of the funny things about information technology that I’ve noticed can best be summed up by a quote from the new Battlestar Galactica, “All of this has happened before. All of this will happen again.” Think about mainframes. We used to do all of our work from a dumb terminal that gave us a window to a large processing system that housed everything. Then we decided we didn’t like that and wanted all the processing power to live locally in minicomputers and client/server architecture. Now, with the advent of things like virtualization and virtual desktop infrastructure (VDI), we’ve once again come back to using dumb terminals to access resources on large central computers. All of this has happened before. And when we get constrained on our big hypervsior/VDI servers, we’ll go right back to decentralized processing in a minicomputer or client/server model once more. All of this will happen again.
In networking, we moved from globally routable address space on all of our nodes to running them all behind a translated boundary. At first we did it to prevent exhaustion of the address pool before a suitable replacement was ready. But as often is the case in networking (and information technology for that matter), we patched something and forgot to really fix the problem. NAT became a convenient crutch that allowed network admins to not have to worry about address renumbering and creating complex (even if appropriate) firewall rules. I’m just as guilty of this as anyone. It was only when I realized that many of the things that I want to do with networking, such as video conferencing, require more effort to work with NAT than they would otherwise. We spent so much time trying to patch things to work with the patch that we forgot what things looked like before the patch. I’d argue that getting back to end-to-end communications to “fix” protocols like SIP and FTP is just as important as anything. Relying on Skype to do VoIP/video communications just because it doesn’t care about NAT and firewalls isn’t good design. It’s just an inexpensive way to avoid the problem for a little longer. The funny thing about IPv6 is that while there is a huge amount of configuration up front and a lot of design work, when things are configured correctly, stuff just works. Absent of a firewall in the middle, I can easily configure an end-to-end connection directly to a system. Before you say that something like that is only important to an enterprise, think about something like remotely supporting a small office. I no longer have to poke holes in firewalls and create one-to-one NAT translations to remotely connect to servers. I can just fire up my RDP client (or your screen sharing tool of choice) and connect easily. No fuss, no muss, and no NAT needed.
I’ve also said before that I now see that there is a use case for Network Prefix Translation (NPT). Ivan has talked about it before and showed me the light from the networking side. Ed Horley has also given me a perspective from the Microsoft side of things that has changed my mind. But exhorting NPT as a solution to all of our NAT problems in IPv6 is like using a butter knife as a screwdriver. NPT was designed to solve one really huge issue – IPv6 multihoming. Announcing address space to two different upstream providers, which is easier to do with NAT in IPv4 than it currently is in IPv6 absent of the solution provided in RFC6296. NPT for multihoming is a good idea in my mind because of the inherent issues with advertising multiple address spaces to different providers and configuring those addresses on all the internal links in an organization. I also believe that NPT is a transition mechanism and will allow us to start “doing it right” down the road when we’ve overcome some of the thinking that we’ve used with IPv4 for so long. One-to-one NAT makes no sense to me in IPv6. Why are you hiding your address? The idea is that the device is reachable, whether it be a web server or a video conferencing unit. Why force a translation at the edge for no apparent reason? Is it because you don’t want to have to re-address your internal network devices?
Absent the aforementioned multihoming issues, let’s talk about renumbering for a second. How often to you really renumber your internal network? At the company that I work for, we’ve done it once in ten years. That’s not because we were forced to. It was because we ran out of space and needed to move from a /24 to a /23 (and now maybe to a /22). We didn’t even renumber half the devices in the network. We just changed a couple of subnet masks and started adding things in new subnets that were created. Now, granted, that was with an RFC1918 private address space internally. However, with SLAAC/DHCPv6 and IPv6, renumbering isn’t that big of a pain. You just change the network ID that is being handed to your end nodes. Thanks to EUI-64 addressing, the host portion of the address won’t change one bit. And Trevor Pott points out in the article that enterprises assume that DNS resolution will take care of the changeover just before he snorts derisively about how no one has managed to make it work yet. I’d argue that he’s more right than he knows. I have the IP addresses of hundreds of customers memorized. Most of them are RFC1918. Some are not. All of them are dotted decimal octets. I know that when I move these customers to IPv6, I will be relying on DNS resolution to reach these end nodes. My days of memorizing IP addresses are most definitely coming to a middle. And for those that might scoff at the ability of a DHCP server to register and maintain a database of DNS-to-host address mappings, you might take a look at what Active Directory has been doing for the last twelve years. I say that because in my experience, many SME/B networks run some form of Microsoft operating systems, even if it is just for directory services.
I’d like to take a moment to talk about “small” enterprises versus “large” enterprises. For most people, the breaking point is usually measure in costs or in devices. As an example, if you have more than 1000 devices, you’re large. If you have less than 50, you’re small. Otherwise, you’re in the middle (medium). Me? I don’t like those definitions. 10,000 devices in a flat Layer 2 network is (relatively) simple. A 10-person shop doing BGP multihoming and DMVPN is more of an enterprise than the previous example. For those networking admins that are running tens or even hundreds of servers, ask yourself what you really consider yourself to be. Are you a small enterprise because you have a Linksys/D-Link/Netgear Swiss Army Box at your edge? Or are you really a medium-to-large enterprise because of what you’re doing with all that horsepower? Now ask yourself if you want your network to be easy to configure because that’s the way networks should be, or is it really because you’re understaffed and running far more infrastructure that you should be? I’m not going to sit here and say that you just throw more people at the problem. That’s never the right answer. In fact, it’s usually the one that gets you laughed at (or worse). Instead, you should examine what you’re doing to see why wholesale renumbering or network changes are even necessary in the first place. One of the main points of the article is that IPv6 will allow network admins to finally be able to create hundreds of VMs on a single physical server and make them reachable from the global Internet. I would counter with the idea that if the only thing truly holding you back from doing that has been address space, the SME/B that you work for has really been a large enterprise wolf in small enterprise sheep’s clothing all along.
Now, if you’re still with me this far you should congratulate yourself. I’ve expounded a lot of thoughts about the technical reasons behind the way IPv6 behaves and why there are difficulties in applying it to the SME/B. I also wrote all that in isolation on an airplane. As soon as I stepped off and got my Internet lifeline back, I checked up on the original article and noticed that Trevor Pott had clarified his original intent at the bottom of the post with a long comment. Being no stranger to this myself, I read on with measured intent. What I came away with galvanized my original thoughts even further. Allow me to restate my original point a little more pointedly:
If “cheap” and “simple” are your two primary design goals, IPv6 probably isn’t for you.
We’ve gone through this whole problem before in the infancy on the Internet. Last year, Vint Cerf gave a talk at Interop about the problem of protocol adoption. One of the stories I love from this talk involved Mr. Cerf’s attempt to spur the adoption of TCP/IP over the then-dominant NCP protocol. He needed to drive people away from NCP, which wouldn’t scale into the future, and force them to adopt TCP/IP. But adoption rates plateaued quite often as network operators just became comfortable that NCP would always be there to do all the work. Mr. Cerf eventually solved his adoption issues. How did he do it? He turned off NCP for a couple of hours. Then for a day. Then for a week. He drove adoption of a better protocol through sheer force of will and an on/off switch. Now, we all know that we can’t do that today. The Internet is too vital to our global economy to just start shutting things off willy-nilly. Despite that, “cheap” and “simple” aren’t design goals for the Internet core or even the ISP distribution layer. We have to have a protocol that will scale out to support the explosion of connected devices both now and in the coming years. Enterprise providers like Cisco and Juniper and Brocade are leading the charge to provide equipment and services to support this in-state transition. There will be no shutdown of IPv4. This is a steady-state parallel migration to IPv6. These kinds of things don’t come without a cost of some kind. It may not be in the form of a purchase order for a new network core. It may not even be in the form of a service contract to a consultant to help engineer a renumbering and migration plan. The cost may be extra hours reconfiguring servers. It may be taking more time to read RFCs and understand the challenges inherent in reconfiguring the largest single creation in the history of mankind at a fundamental level.
Economies of scale are a good thing. They bring us amazing products every day. They also enable us to spend less time configuring or working and spend more time on creating solutions. The first time you tried to ride a bicycle was probably difficult. As you practiced and progressed it became easier. Soon, you could ride a bike without thinking about it. You might even be able to ride a bike with holding the handlebars or ride it standing on the seat (I never could). That kind of practice and refinement is what is needed in IPv6. We have to make it work on a large scale first to get the kinks worked out. Every network vendor does this. Yes, even the ones that only sell their wares at the local big box retailer. Once you can make something work on a big scale, you can start winnowing down the pieces that are necessary to make it work on the small scale. That’s where “cheap” and “simple” come from. No magic wand. No easy button. Just hard work and investment of time and money.
Spurring us “priestly” networking people to change the way things work is a very valid goal and should be lauded. Doing it by accusing us of being obstinate and condescending is the wrong way to do things. I don’t consider myself to be a member of the Cabal of IETF High Priests. I’m not even a member of the IETF. Or the IEEE. I’m a solutions guy. I take what the academics come up with and I make it work in real life. Yes, much like Trevor Potts, I’m a blogger. I like to take positions on things and write interesting articles. Yes, I lampoon those that would seek to hobble a protocol I have high hopes for with thinking from fifteen years ago for the sake of making things “simple”. I’d rather be spending my time working on ways to reduce the time and effort needed to roll out IPv6 everywhere. I’d rather focus on ways to make it easier to renumber the “hundreds” of VMs I typically see at my local small business. In the end, I want what everyone else wants. I want an Internet that works. I know that it may take the rest of my career to get there. But at the end of the day, if I’m forced to choose between making the best Internet I can for the sake of everyone or making it “cheap” or “simple”, then I’ll sacrifice and pay a little more in time and costs. It’s the least I can do.