Rough transcript of 18-minute presentation by Mike O'Dell, Chief Scientist & Vice President, UUNET, at the May 16, 2000 symposium at Stanford, "Optical Internet: The next generation." Title of O'Dell presentation in program: The future of the Internet Title of 10-slide O'Dell presentation deck: Racing with an exponential or the dangers of linear thinking in an exponential world MODERATOR: The, the last panel is convened to answer the question: What are we going to do with all the bandwidth that all these other guys have created? A, a seemingly simple question, ah, so we've, we've brought together some of the, sort of, most interesting and knowledgeable and opinionated folks from both the service provider and application developer, ah, communities that, ah, we were able to find, and the first, ah, is, Mike O'Dell, the chief scientist at UUNet, and I was very amused to learn, ah, that Mike actually worked at Bellcore, for three years, prior to, prior to joining UUNet. So he's, he's done, he's gone both ways. Anyway, Mike, take it away! O'DELL: Thank you. Yes, I spent three years learning that you don't want to work for a phone company. So, at least not seven of them, anyway. Ah, it was very interesting listening to some of the talks earlier today, because, ahm, it is very interesting to hear pieces from three-year old talks coming out of other people's mouths. It's really entertaining to, to see that happening. So maybe it means that some of the baying at the Moon actually stuck and I am not completely crazy, so ... But most people would argue for that ... So anyway, today I will be talking of racing with an exponential or the dangers of linear thinking in an exponential world. And, this is about scaling: first, last, and always. Ah, I've got, got a few gee-whiz slides here because I always get asked these questions. [points at slide of network in mid-1997] That's the network, as it looked in mid-1997, and it was all DS3-based, and those are lots and lots of wires, the white lines are DS3s, but there's often a lot more than one of them. The way we measure the capacity of the network is in OC12c-equivalent route miles, so a DS3 is a quarter, er, an OC3 is a quarter of an OC12, one mile of an OC12 is an OC12 route mile, and a 48 would be four of them. So, you can do the math, it's pretty simple. But that's the way we measure network capacity, ahh, because it turns out that you want the speed-distance product, because a, a 10 to the 20th bit per second link that was a millimeter long would, would have fairly limited applications, right. It's not completely uninteresting, but it's, it's sort of limited in what you can do with it. So, so you need a distance, distance, distance-speed product, and that's the number, when you, when you hear us talking of the capacity of the network sort of doubles every four months or so, that's the unit that we're talking about, is that number. So if we go to the next one, [move to mid-1998 slide] a year later it looks like this, there are three hundred, let's see, thirty eight thousand route miles, it's gotten a lot bigger, ah, the fiber's wet, lot of stuff, we put four hundred million dollars under the Atlantic Ocean to do that. It turns out if you take twenty dollar bills and tape them end to end, and then roll them across the Atlantic Ocean twice, that's just about four hundred million dollars, that's pretty interesting [laughter from audience]. Doesn't work as good as the fiber, but it's about the same [laughter]. [move to mid-1999 slide] Mid-1999, ah, we had a little over a quarter million route miles, ah, but, the way we put it, is, if we were to suspend service on Earth, we could actually serve the Moon, ah. But the good news is, is that four months later we would serve the Earth again, so, as well as the Moon, so, but that's, again, that, that was running OC48, all the green stuff is all OC48. We had 48s across the, across the Atlantic. Across the Pacific is much harder. The people that made the original investments in the capacity out there did not unroll nearly enough string, and they sure didn't put enough wax on it, so there is a lot of money going into putting capacity into the Pacific. Ah, this is the 192c network rolled out at the end of this year, that's, that's in the process of being rolled out, and. The conversation on lasers this morning was very near and dear to my heart, both because somehow I was the guy who did not run fast enough and ended up on the board of the OIF where the people are arguing about various short reach optics stuff, but the real issue is that to do this 192 deployment as it exists today, requires four times the world's total manufacturing capacity of 192 lasers, -- for a year. So, there's, there's, there's a timing problem there, we want to do this in a year, it takes four times the world's production of 192 lasers, and the reason is exactly, Graeme [Fraser, of Cisco, speaker during the preceding session] showed the picture, big expensive hocking laser to drive 300 feet through the fiber, that's a real waste. So, getting, getting the very short reach optics is actually really, really, really critical going forward for us, to do something that looks like that. So, ah, that's that's public record stuff that we showed analysts and things, ah, but again, that's ah, that's what that is, you can see some metro stuff in there. Well, will the Internet stop growing? That's the key question I keep getting asked by analysts, reporters, and people like that, and I like to, like to, like to explain why it won't happen. OK, the Internet, ah, is the death, is, is the death of one of the last vestiges of centralized planning. I mean, what two interesting events happened in 1917? Well, the Russian Revolution, the October Revolution, was one of them, and Theodore Veil got, got the monopoly from AT&T, was the other one, right? And, they both died in the 1980s, right, which is like, this is like, there is no relationship, but it is the most amazing irony I know of today, ah. [laughter] But, but, this death of centralized planning is really important because, because what's happening today is exactly paralleling what happened to the PC and the mainframe, OK. PC won over the mainframe, not because mainframes were intrinsically evil, right? Back when computers were rare, the, the number of people devoted to thinking of new things to do with the computer was, you know, there were a lot more people than computers, so that ratio was sort of right. But with the advent of the computer, the computer, the PC, computers got in the hands of lots and lots of people, and suddenly the ratio of, of people to computers in terms of thinking of new things, the department of new ideas was tragically undermanned, right. And when the PC got out there, it unleashed what I have referred to as the code crazies, people who did not know any better go off and try to solve all kinds of problems using this new computer, right. Well, the, the thing to observe is that nature operates by doing millions of experiments, all in parallel, most of them die, right. But a few of them get, win, you know, and get to play again, right. So, the reason the mainframe lost to the PC, the mainframe lost to the PC, was because of the transition of innovation from a centralized planning model, where there was a designated group of people to do new things, right, to a biological model, where nature is running the show and new things just happen, right, and we are seeing exactly the same thing in communications, right, the Internet did for communications what the PC did for the mainframe. In that back when I worked at Bellcore, I was actually part of the people in charge of all the new ideas, right: Yes, citizen, stand in that line right over there and you too can get ISDN and have some of our fine cardboard shoes while you are at it, right, [laughter] you know, if you live long enough. Ah, well, now, with the Internet, the barrier to experimentation is essentially zero. You don't need permission to try something new. The only difference between, you know, the only difference between, you know, an idiot and a genius is six months and an IPO, OK. [laughter] Right, so, but the barrier, people, economists talk of the barrier to entry, but I think actually the term barrier to experimentation is actually more important, because that is how you find out whether you're, whether it is even worth entering the market, whether you have something. So, so, in, in the Internet-enabled world, basically millions and millions of people are trying all the new ideas, you know, there are millions and millions of people in charge of thinking up new things. So, we have transitioned from the centralized planning model to a biological model, right, so that everybody is in charge of things growing. So, what we've done, again, is, this is creativity writ on a very very large scale, and everybody's helping push, right, so that's why, that's that's why it's very different. Now, as Graeme and some other people have known me for years and have sold me hardware, ah. I'd like to talk about the difference between an eccentric and a madman. Because some of the things I'm going to talk about in a couple of minutes, ah, you will be certainly convinced I'm mad, and that's why, ah, again, the comment about hearing old versions of my talk coming out of other people's mouths, ah, was entertaining, because, like, I was considered mad Vin. But again, the difference between a madman and an eccentric? [pause] An eccentric has a checkbook, right? [laughter] You'd be amazed, what POs with seven zeros on them, eight zeros, you'd be amazed at how quickly they decide maybe you're crazy but they want your money, right. But, so, but this transition that's happened, ah, in terms of this transition to a biological growth model, ah, has, has changed the role of large network, communication network operators. In a way I think of the job we do is, that we basically run a Petri dish, right. A Petri dish is the little glass thing that you put Agar in and you grow bugs in it in a laboratory. What we do, right, is we have this gigantic Petri dish, we keep it warm, we pour nutrients, and people pay us for the right to sit in the Petri dish, to see who multiplies, to see who fights with the next colonies, and to see who wins, right. And as it fills up we make more, we, you know, we put in more nutrient, we keep it warm, but basically, we are not in the business of I mean I, UUNet we are not in the business of picking winners and losers, right. You know, you want to sit in the Petri dish, you know, pay the toll, sit in the dish, and whatever happens, happens, I no longer, you know I used to get real concerned about, whether or not, oh, the right thing is it going to happen over here. Well, that's not my job, nature sorts all that out. Ah, so, but, but this notion, this notion of the Internet as a, a biological growth medium is I think an interesting point of view. Hm, so, ah, petabit networks, the new frontier. [this was the title of slide #7 in the O'Dell presentation deck] Am, basically my job is, is to try to make sure that the future turns out to be one we can live through. Ah, because we've been growing a lot and we've been growing ever since, basically from 93 to 99, ah, we scaled the network 10^6 in six years. So, right, I say to myself, what do you think is going to happen in the next five to six years? Right. The answer is, well, what I am not willing to bet, is that we don't need to do it again, OK? So, the assumption is that we need to grow the network by 10^6, 10^7 in the next five years. Now, if you remember the hockey stick graphs, right, ahh, if you don't draw them on log paper, right, then, then, the thing you figure out is that you really have got to have the velocity right, and you have to have it pointed the right way, because, when the curve gets steep, it gets real hard to steer, so, what that means is we actually have to be deploying the technology platforms which will scale these very frightening numbers within the next three years, and this slide is six months old, so it's two and a half. So, if you go back, go back to the slide that I've had, it seems I can actually back up the slide, ah, oops, going the wrong way, here we are. This, the network is roughly speaking broken up into roughly nine areas of North America, right. So those are what we call those are what we call regions, North American regions, and, at, at the point where, at the five year point, if you look at a cross-section of the trunk, that connects the regions together, a trunk is connected with three to four other regions, ah, a region is connected with three or four others with trunks, and the capacity within each region is roughly equal to the trunking capacity at the edge of the region. A cross-section of one of those trunks is roughly 10^15, 10^16 petabits, ah, bits, which is one to 10 petabits, OK. That's a big number, right, a petabit is a thousand terabits, right. And, so if you think about, that's what we have got to get to, right, and we don't have a lot of time to do it, you sort of have to start making a list of problems that you need to solve, because there is actually a pretty good sized list of them and the clock is running pretty fast, right. Now, one of the things when, when you are going to do something which is patently absurd, or patently impossible, ah, one of the things you'd like to do is pick one thing which you believe is not completely nuts, right. And I am glad, glad that the Nortel chap showed the slide on what they're doing with optics, because I didn't know whether I could quote his numbers, because of NDAs and things, so I am glad you spilled the beans. Ah, but, but the assumption that we're making in the design going forward, is we're going to have a hundred gigabit coherent on a lambda, a hundred lambdas on a fiber, and that way if a trunk is only a hundred fiber pair going to a thousand fiber pair. And while a thousand is certainly inconvenient it is not impossible, right. However, there is one big gotcha, OK. DWDM is a very poweful genie, but it is also a quite malicious genie. Because once you get it stuffed into the fiber, you cannot ever let it escape, right, because the last thing in the world you could do, is let one of them WDM fibers explode into a hundred patch cords, right. You can't do this external WDM thing where you have transponders, so that a hundred-head Medusa, you know, your one little fiber becomes a hundred Medusa (to moderator: I've got five, OK), ah, because it doesn't work. So the first rule is that anything that arrives on a single fiber has to terminate on that fiber, and has to terminate in one slot. So Dr. Miller's [Professor David Miller of Stanford, presenter in the first session in the morning] pictures of like magically doing WDM inside, that actually was quite nice. So, again, you've probably heard this quote from, around, the only real problem is scaling, all others inherit from that one, ah, in the object-oriented sense, so, this is all about scale. [Slide #8: caption: O'Dell's universal problem statement. body of slide: The only real problem is scaling. All others inherit from that one.] [Slide #9: caption: Petabit network challenges. body of slide: list of: optics, routing, trunking, reliability, software architecture, manageability, can't hire 10^6 more people] Network challenges, I'll touch on the list here, optics we've been talking a lot about. Ah, trunking, reliability, and software architecture, those are two sides of the same coin. Ah, a lot of the incumbents in this space, ah, in the telephony side actually have some experience building reliable software, not everybody else in this space has that history. Ah, it'll be interesting to see who learns from whom. Also, we can't hire 10^6 more people to build the network this big, because then the entire population of China would work for UUNet, and the entire revenue stream would be employee discounts, and how does that work [laughter]. Ah, so, but, but routing is the one I want to touch on really quick, ah, in that I assert, ah, that the most complicated computation ever attempted by mankind is the global distributed routing algorithm that runs the Internet. In fact, if anybody thought about it very hard, before we started, they would've been too scared to try. Ah, because it runs in near real-time, it's an online algorithm, it's, it runs, it runs on a multimillion node multicomputer, of an arbitrary topology, built by lots of people who have never met each other, right. And, ah, it's a very very complex computation because it's piecewise constructive, there is a lot of local consistency constraints, there is a bunch of global correctness criteria that are occasionally satisfied, and yet the thing mostly works. Which is astounding, when you actually look at what's going on. And, to give good credit, to give lots of credit for Cisco and all the guys who over the years tried very hard. Compared to what we need to build a network like this, the tools we have to specify this computation are basically, only a few millimeters above toggling an octal on console switches, right. And that's because I honestly believe that this problem is so hard that it has taken us twenty years to figure out the problem statement, much less to figure out how to do it and again, lots of credit to everybody over the years that has, that has worked very very hard to move this rock. But, ah, an example, ah, physics place, right, a good physics example. There is a number of problems in physics that picking the right notation makes all the difference in the world of whether or not you can understand the problem or solve it, right. Ah, Manhattan Project, solving the, the critical mass equation. Ah, quantum chromodynamics and Feynman diagrams, right. You know, when Feynman did it, three people in the world understood it, and now graduate students learn learn about it, right. Because you have a notation that lets you do that, right. We do not have a moral equivalnet of Feynman diagrams for routing, so that you can write down the behavior of a network from a routing perspective and know what it is going to do, right. Because engineering is the art of knowing how much smoke comes out of each hole before you turn it on, right. [laughter] Not, any, any fool can observe the obvious, right, but, but the trick is, the trick is knowing what is going to happen, and talk to anybody who does routing engineering in large networks, and it's really, intellectually the most difficult job I know of. So, again there is a whole lot of things here, I can talk for hours, and he says I can't. [laughter] So, ah, but again I can talk at great length about all these problems, but these, these are the things we need because, like I said, we spent twenty years understanding the problem definition and, ah, the great, the notion the notion that the existing networks is one percent of what the network needs to be, he's right, but he's off by a zero, right. In all, I love, I love a lot of the projects talked about this morning, but all the dates are like half a decade too late, right. This is real, right, this is coming. Again, my job is not to make sure that we need this capability, my job is to make sure we have it when we need it, right, it's the job of the sales people to make sure we need it. But that's, that's a separate problem, right, you beat those guys. But, but, throughout throughout the history of the Internet it's been: they're still coming, so we have to make more, right. And there's absolutely no evidence that that's going to slow down, much less stop. [slide #10: UUNet engineering credo. body of slide: If you aren't scared, you don't understand.] So, I will close with the, the original version of the quote that got used earlier, and I want my royalty payment, ah, because's that's my line: The UUNet Internet credo, if you aren't scared, by what I've just told you, you have't thought about it hard enough, you don't understand. So, here's to petabit networks, and I cede the field. [applause]