Content-Centric Networking (CCN) as an alternative to IP
I've been meaning to read the CCN article from Van Jacobson and friends for months now, but I finally got around to it on this lazy Sunday. It starts with mostly synthesis of other research in the area of content-named delivery, but has a few major innovations that make it very interesting. In particular, it is designed around the same core philosophy as IP: mainly statelessness, automatic congestion control, and routing tables based on prefixes (though in this case the prefixes are names rather than addresses). The routing tables go so far that they even suggest using (unmodified!) OSPF for discovering topology. (I'm not convinced that part would actually scale, but maybe.)
The basic architecture is that a client sends out an "Interest" packet to register its interest in a particular named bit of data. Your (content) router receives that and forwards it to one or more onward routers, based on its (content) routing table (aka "Forwarding Information Base"), and so on, recursively. Finally the Interest reaches the end provider of the data, who sends back a response. So far, so simple. But the key innovation is what happens when packets are dropped or duplicated. If you get the same Interest from more than one source, you only forward it the first time; if you get the Data response more than once, you only forward it once (well, to each existing Interest) and then throw away the Interest.
That sounds simple for the same reason that IP itself is deceptively simple. But as with IP, the end result of this simplicity seems like magic. First of all, it eliminates the problem of "routing loops"; if router X is configured to send Interests upstream to Y and Z, and it receives the same Interest from both A and B, then it will only upstream that Interest exactly once (to Y, or Z, or both, depending how its routing table is set up). So in that example, if A, B, Y, and Z are actually peers, you don't have to worry about Y and Z looping Interests back to A and B. That is, they will forward to A and B, but since A and B have already seen that Interest, the buck stops there.
Secondly, if nobody on the whole network provides a given bit of data, it's no big deal; nobody will answer it. Routers will keep the Interest in their tables for a while (so they know who to forward the Data response back to if it shows up), but they won't send a response, because there is no response. As in IP, you can't automatically know the difference between "nobody is there" and "packet got dropped."
This, in turn, is how flow control / congestion control works. Routers don't ever resend Interest packets on their own; that's the job of the client endpoint. (Retransmits include a "nonce" - just a random number - to ensure that they don't get eaten by the anti-routing-loop logic above.) So as with IP, an overloaded router can simply drop excess Interest packets to slow down network activity. And just like with TCP, a client endpoint can create a "sliding window" implemented by having multiple outstanding Interest packets at once. How many at a time? That's your window size, and it depends on observed retransmit characteristics.
Apparently it's actually better than TCP's retransmit characteristics in the sense that the flow control is on a link-by-link basis instead of end-to-end. I don't really know how or why that's better; there's a note in the doc that says "We will cover this topic in detail in a future paper," which I hope is not equivalent to a "remarkable proof of this theorem which this margin is too small to contain." I haven't looked - maybe they published it already. For now, I'll take their word for it.
The design also mostly-transparently deals with support for multiple network interfaces - like a wired, wifi, and 3G network all at once - by allowing you to just forward all your Interest packets on all the interfaces if you want. Then, using techniques similar to ethernet bridging, you adjust your routing table priorities based on latency/speed/loss of responses received. You occasionally do an experiment by requesting an answer from a supposedly non-optimal interface just in case things have changed; if you don't get an answer back on the supposedly optimal interface, maybe it's dead, so you can failover right away.
Now, we all know failover is nice for content-named networks (along with caching, that's basically the whole point). But this is where things get really fun. You can carry a content-named network over IP network sessions; in fact, that's how their prototype was built. But what if you could carry IP network sessions over a content-named network?
Well, according to a later paper by the same team about VoCCN (Voice Over CCN), you can! And the way it works falls out naturally from the design. You just send out a window of Interest packets for stuff that doesn't yet exist. Initially, you get no response. But the unanswered Interest packets are remembered by the router nodes for a short time, so as the live data is created by the server endpoint, it just responds to the outstanding Interests and the data gets distributed back to the original client(s).
The reverse path, data sent from the client to the server, is encoded by registering Interests in which the last segment of the path is the actual data you want to send. The server can then send a small Data (acknowledgement) packet to state that it was received and doesn't need to be retransmitted.
They point out that one neat side effect of all this is if your client is multi-homed and one of its network links drops out or moves - say if you move from one wifi network to another, or from wifi to wired - then you can still recover, because the intermediate content routers have actually cached the data you would have lost by hopping networks, and you don't have to "reconnect" because your connection was always about the content, not the endpoint.
Internet-wide multicast-like behaviour - with scalable retransmits!! - is an inherent part of this design. Want to send something to more than one client with minimal load? Don't do anything special. Just respond to Interests. If there's more than one Interest for a given bit of data, any content router along the way can receive just one copy and fan it out to all the other recipients.
Symmetrically (but separately), you can use multicast or broadcast on a LAN to send out Interests, so if someone nearby has already seen what you're asking about, they can send it to you. Your local content router(s) could also choose to multicast Data response packets on the LAN if more than one local machine has expressed Interest in it, using whatever heuristics or conventions you want to use. Clients and routers that receive unsolicited Data packets just ignore them.
Finally, note that, unlike many recent trends in content storage, content is not named directly after its hash. As they point out, the problem with that method is you need a separate name-to-hash conversion layer, and that a client can't request data which doesn't exist yet - which prevents web-style dynamic page generation (based on query parameters) as well as disallowing any kind of two-way communication channel like they created in VoCCP. Hash-based naming also ends up necessitating things like DHTs for routing, instead of allowing the prefix-based routing they recommend. I have to say, prefix-based routing does have a lot of appeal to me, after having considered hash tables and DHTs pretty extensively. The problem with hash tables is you lose all locality of reference, so you end up with related data (for example, consecutive blocks of a big file, or several files in one directory) scattered all over the place instead of (if your cache is written carefully) stored consecutively on disk.
On the other hand, naming your content after its hash makes verification super easy; if you request block 234283123, and the hash of the received block is not 234283123, then you reject it. CCN, by comparison, if I understand their paper correctly, has on the order of 325 bytes of security crap for every 1024 bytes of content. (32% overhead?!) Maybe I'm reading it wrong, but I suspect not; the security section of their paper seems to be a complicated multi-layer abomination, leading me to suspect it was either an afterthought, or designed by someone totally different from the people who designed the core of CCN.
So the security part is definitely going to need more work, but I think it's not unresolvable, for the same reason that security over IP (itself not secure) is resolvable (by adding SSL). For example, with something like VoCCP, you could just rely on the datastream to itself be encrypted, using (literally) SSL, and leave it out of the transport layer altogether. That leaves you with a problem where an attacker could insert fake data into your SSL stream, which would be rejected by SSL, probably aborting the connection and leading to a denial of service. That's already possible with TCP, and hasn't been a huge problem, although it would be even easier to create this kind of attack in the case of a network that has lots of untrusted caches. (That is, you wouldn't need to be a man in the middle. Or if you prefer, there are more men in the middle.)
Anyway, there are lots of really great ideas in these two papers. The best part is the core concepts - datagram-based Interests and Data packets - that can be applied separately from the other parts (routing protocols, security protocols, tunneling protocols). Again, just like with IP.
I think something like CCN really may be the future of networking. It will take a bit of work first though. :)
Addendum: Replace IP? Really?
Yes. This is not as crazy as it sounds, although of course it would take a long time to complete. You might recall that I think IPv6 replacing IP as the core of the Internet is rather unlikely, but that's because of chicken-and-egg problems. IPv6 won't provide any utility to the Internet until all the clients and all the servers have switched; until then, it's just twice as much work for everyone. Yes, you can tunnel IPv6 in IPv4, and vice versa, but if your personal workstation and the server aren't both speaking IPv6, you don't benefit from IPv6.
CCN (or variants of it) are different, because they're designed from the start to be carryable over IP (IPv4 or IPv6). That means they cooperate with the existing system. Moreover, you can tunnel IP (IPv4 or IPv6) over CCN, so if you did replace all your IP routers with CCN routers, your computer could still talk IP if it wanted, and still reach everyone on an IP network. (This is very different from other content-based networking proposals, which lack that property and thus could never totally replace IP.) You could even host IP-based services and advertise them using a subdomain prefix on the CCN network (using OSPF or whatever they settle on), so incoming connections would work.
Most importantly, though, CCN can be implemented entirely in userspace. Your browser could support it even if your OS kernel doesn't. That's unlike IPv6, where you could theoretically have an "IPv6 proxy server" accessible from a user space process over IPv4 (or vice versa), but you wouldn't do it that way because if you're using a proxy server anyhow, you'd probably use a session-level (ie. HTTP) proxy server, skipping the IPv6 layer altogether. CCN is different in this case because it provides actual value above and beyond what IP provides. If a CCN server exists for certain kinds of things - video files, for instance, like Youtube or iTunes - then your life, and the life of your ISP, and the lives of the people serving the files - will get incrementally better as people add more CCN servers, routers, and clients. That encourages incremental adoption because there's an actual reward for each increment. With incremental deployment of IPv6, your only reward is the knowledge that you made the world just a little bit more complicated today. Yay you.