by Mihai Budiu
This interview was published in the March issue of the Romanian computer magazine Netreport.
Q: Can you tell us a little about yourself?
A: Sure. I am a computer scientist. I am a professor at Carnegie Mellon (CMU). I grew up in Illinois, in the center of the US, but in my childhood I spent many months living in Eastern Europe: Bulgaria, Yugoslavia, and I even spent the summer of 1969 in Romania. I went to university at MIT, in Boston; I received my undergraduate degree and my Ph.D. there; after that I served as a postdoc at MIT and then I went to NEC labs, a big research institute, established by the Japanese company with the same name. After spending three and a half years there I came to CMU. I have been here for seven years, although I took three semesters off to a company called Akamai that started. While at Akamai I was the vice-president for research and development and I was in charge of managing all engineering efforts. Apart from Computer Science I like to play ice hockey and I play softball, which is a version of American baseball. I also like to run and ride bicycles.
A: I worked in the general area of parallel computing. In particular, i studied the problem of ``what should the general structure of a parallel computer be?'' You have many independent processors connected with some sort of communication network: how should the network look like, and how should the processors make use of it? More recently, though, I've become interested in networking. After my stay at Akamai I've shifted my focus. I was mostly a theoretician before, but now I am working on problems that are somewhat more immediate. In the end I think that most of the work I will do will have some sort of theoretical justification.
Q: This is great, because I have many questions about the relationship between theory and practice, and how they blend. Let me ask exactly that: how do you view the relationship between systems research and theory? They seem to be rather disjoint; the two communities don't talk to each other that often. What do you think?
A: I think that there are a few misconceptions, and in fact there is a fair amount of interaction between the two. But finding common ground, where work done from one point of view is immediately beneficial to the other is a challenge. The theoretician can apply their most powerful tools when there's a nice, clean, abstract model and the system builders can take advantage of theoretical ideas when the implementation details are manageable and not horrendously complicated. It's true, though, that people tend to view themselves as being in one camp or the other. Therefore some real effort has to go into making connections with the other side. And there is only a small number of people who are well-known in both communities. Other common misconceptions are that theoreticians are impractical and systems builders are incapable of theoretically sound designs; I think that both of these are wrong. I think that theoreticians, when they've been properly ``educated'' with the real systems issues in a particular domain, can develop the right models and apply the right tools to improve on the best designs that could be constructed without having a theoretical basis. I also think that the people in systems can appreciate and even derive good theoretical designs. I have seen, at least at Akamai, that theoreticians were able to rapidly and eagerly become integrated into big systems building projects. Details of say, how Linux implements some networking protocol cannot be easily learned by a theoretician in a short time, but by being in the trenches and working with the system builders you get to know these details, and have a feel for why things are the way they are. It's possible for theoreticians to have a big impact on the design of really big systems. Especially when what is cast into software will have to scale to magnitudes of thousands of machines, billions of hits per day and terabytes of data collected.
Q: You're pushing me to ask you about Akamai... what can you tell us about Akamai?
A: Akamai is the leader in Internet content delivery. What that means is that Akamai serves, for example, the images that appear on the most popular web sites in the world. Akamai has maybe 1400 customers. The customers use Akamai for the actual delivery of the bits that make up the images or the video streams rendered on the end-user workstation. Akamai has a lot of high-profile customers, such as Yahoo, Microsoft, Apple, CNN, and these customers have outsourced the problem of quickly delivering the data to the end-users to Akamai. Akamai has established servers all over the world; over 8000 servers in more than 50 countries at around 500 locations. When an end-user goes to visit say Yahoo, [using a web browser], the pictures will be sent to the user not from the central location, where Yahoo has its master web servers, but from one of Akamai's locations, which is usually much closer to the end-user than the central location.
Q: How do you know which one is closer to the client? The Internet doesn't provide any infrastructure to determine where the clients are?
A: That's a good question. First of all, it is not true that the Internet doesn't provide any infrastructure. There's some information in just knowing the client IP address. The Internet is not a completely flat or disorganized network: it's an interconnection of independent networks, and from the client's IP address you know the network the client is on and that tells you something about where the client is located. Furthermore, Akamai now has a very widely distributed presence. With servers at over 500 locations we're able to survey the Internet and build in real time a very comprehensive model of what's going on: in particular which network is connected to which network, and where the performance bottlenecks are located. This is one of Akamai's core competencies. Knowing when an end-user makes a request not only where the end-user is located, but also where that end-user would be best served from.
Q: Is this whole process transparent for the client?
A: Yes, completely transparent. You've seen at some web sites where you want, say, to download some software, that you've been offered a choice to select from a site in one country or another. That model requires the client to participate in making the decision. The problem with that model is that broad geographical classifications don't necessarily lead to good network predictive power. It's not the case that the site that's closest to you in number of kilometers is the one that can offer the best performance. This [kind of choice] is a burden on the end-user. The Akamai system is completely transparent. In essence, when an end-user makes a Domain Name System (DNS) request about an Akamai name for an image that is going to appear in the browser, the Akamai DNS system is examining who is making the request and decides where that client should be served from, based on the client location.
Q: So the images in the web pages [at Yahoo] have URLs at Akamai?
A: That's right. The DNS request has to go to Akamai so that Akamai can direct the client to the optimal Akamai location. You may have noticed that when you visit certain web sites, at the bottom of your browser the word akamai flashes briefly; this is an indication that the web browser is contacting Akamai to resolve URLs that contain the name akamai and then fetching the corresponding images from Akamai. There are a number of ways that a content provider can use Akamai's service without actually labelling their web page with akamai [URLs], and more and more of Akamai's content providers are using DNS mechanisms that mask the fact that Akamai is involved. Not only is the browser automatically contacting Akamai, but there's no indication on the screen that the content is served by Akamai.
A: I think it is misleading to say that routing tables are publically available: there's no way that you could at any time retrieve the routing tables of all the routers of a particular Internet Service Provider (ISP), let's say UUNET. These are not generally available.
Q: I meant inter-domain routing tables.
A: Even the inter-domain routing tables are only available if some network has agreed to share them with you; more correctly, technically, peer with you as a router. The answer though is ``Yes'' to both questions. Akamai has access to a very large number of routing tables in real time, so it has a complete picture of the connectivity of the Internet at any given time. There's more to performance than connectivity. So Akamai performs a variety of different measurements. Some are based on the performance clients are seeing and others are more active.
Q: Do you require special support from the ISPs you work with or are you just using the bare network infrastructure?
A: I cannot comment on that.
Q: You mentioned that Akamai is also providing video. Are you moving into the streaming media business?
A: Streaming media is very important to Akamai. We support all the popular streaming media formats, including Windows Media, Real Audio and Video and streaming Quicktime. We see this as a more and more popular media form, and one that is going to consume more and more network bandwidth. This kind of data requires more bandwidth, and, more importantly, is very sensitive to losses and delays; the longer the path from the server to the end-user, the more likely that something will go wrong and that the problem will be visible to the client. If the server is as close as possible to the client, it can get a qualitatively different streaming experience. Over the next few years, the bottleneck of the ``last mile'', which is often a slow modem, will go away, as people switch to broadband connections to the home, and then the advantage of having a server very close to the user will be even larger than it is today.
Q: For media which is archived, I can see how the model you described for images will work. But if you stream a live event, will you do something different? Then there is only one source [which cannot be replicated ahead of time].
A: Right. With streaming you have to make this distinction between video-on-demand, which is an event that is not live, that has been stored, and a live event. For the latter, there's an additional problem to be solved: that is reliably distributing the signal from the point of creation to all the servers that deliver it to the end-user. For this Akamai has a separate infrastructure, that is not required for static contents. One of the unique things about the Akamai system is that distribution of live content uses a highly redundant mechanism, so that, even in the presence of multiple network failures, the probability that the quality of the stream (as viewed by any particular end-user) is impacted is very small.
Q: There are many companies that seem to be doing something similar to Akamai, like web caching. Akamai, though, was a big hit from the beginning, because it had some very important clients right from the start, like CNN and Yahoo. Was this because of technical superiority, personal connections, strong investor backup or other reasons?
A: It's a complex set of reasons. The two most important are that:
Q: Wasn't the original work done in an academic setting and later expanded into a company?
A: Yes, the original work was done at MIT; this was before I joined Akamai, in 1995. This was spun-off into a company in 1998.
Q: Akamai is a very interesting start-up in that it employs a lot of Ph.D.s, and many with a strong theoretical inclination. Did this bear fruit in the business of the company? Did they help solve important problems?
A: That's a good question. It's true: Akamai has a relatively high number of Ph.D.s: there are approximately 55 Ph.D.s and many of them hold or once held tenured academic positions in Computer Science or Mathematics Departments. It's not the case that they're all theoreticians; there's a number of very prominent systems people who are today at Akamai. However, it is true that a large number has very strong reputations in the theoretical Computer Science community. One of the misconceptions I addressed earlier was that ``theoreticians are not practical.'' I think that the Akamai experience shows that good systems design can be done with collaboration between large numbers of theoreticians and systems researchers. Akamai is solving problems on a scale so large that solutions that can't be analyzed or can't be proven to scale will fail to succeed. Today Akamai is such an important part of content delivery on the Internet that there's no room for error. The solutions have to scale immediately to very large size and they have to be able to address the huge load, that grows from month to month. There's a large number of components in the system whose design is directly motivated by a theoretical analysis and which sometimes do things that are counter-intuitive, because that's theory says that's the way to solve the problem.
Q: I heard once [1986 Turing Prize winner] Robert Tarjan say that if you want to have a direct impact on the world, especially as a theoretician, the best place to go is a start-up. Can you verify this assertion?
A: I am not sure whether this is specific to theoreticians, except that a theoretician thrown into the mix of a start-up will not be afforded the luxury of carefully modelling and developing a large body of theorems. The demands of a start-up are more immediate, and the theoreticians get sucked into doing whatever is necessary for the company, even in the short run. But for anybody that wants to have an immediate impact, the start-up is attractive. This was one of the most rewarding parts of working at Akamai: work done one month could lead to impact just a few months down the road. It gave the opportunity to a computer scientist to develop a component of a system and within three months see his mother using it, even if she didn't know it. That, I think, is very rewarding to anybody.
Q: Did you bring from your work at Akamai some interesting problems to work on during your ``new'' academic life?
A: I did. In a start-up, or in any rapidly growing company, there's often little time to ponder a problem in the thorough fashion that we're used to in academia. You really appreciate this luxury once you're back in academia. During the course of the design of the Akamai system many technical problems arose for which there was an immediate solution that could be put into place and we had great confidence in, and yet intriguing problems were left unsolved because they did not have immediate priority. I brought back with me a number of problems motivated by phenomena that we saw in the design of the Akamai system, but that didn't warrant immediate development at the time. Although I was unable to perform my usual research duties at Akamai, I wouldn't view it as an interruption in my research. Today I have much better context into which to think about networking problems, and a lot more experience to draw on to understand both what problems are important and what kinds of solutions would have some feasibility.
Q: Has the transition academia --- start-up and back been smooth? Were there substantial ``cultural'' differences?
A: I would say that there are very large differences; larger than I had expected. I think that the time spent at Akamai was the most enjoyable and rewarding year and a half of my entire career. It's also the least healthy: I saw more sunrises than I had seen in my entire life, I ate more ice-cream and pizza and did little exercise. What appeared to be big sacrifices were at that time fairly easy, because the work was so compelling and the adrenaline was so high that it was easy to loose track of everything else. One thing I've missed since returning is the adrenaline, the knowledge that a mistake, even a minor one, could have a very large real impact on a company, all the employees, the customers and the investors. The stakes, at least the potential for a real loss was much higher than in academia, where you have time to really check your work, and where there's a process that is designed to ensure that the mistakes are corrected as soon as possible; things here move much more slowly. The first thing I noticed when I came back is that I have so much more freedom to spend time on my research; the pace of life much easier. One other [different] thing is a much stronger sense of teamwork within the company than within the university. I think that the reason for that is that in the university faculty and even graduate students are evaluated largely as individuals. Even though they they ``pull'' together for the sake of the department or university, in the end it is very important that each individual has very defined contributions. Within Akamai those individual contributions were not as critical. It was much more important that as a team we create a good product and that we satisfy our customers. It was especially pleasant that I was able to do that with team members whom for many years I had known as individuals and I collaborated with, but only in the standard academic style, where we had each our own research program. That was no longer an issue. There was no jockeying that we were working on something that's a little bit different than our colleagues at another university. Instead, we threw our effort into the project and there was no concern on who would get credit for what.
Q: Is that a result of the contract regarding Intellectual Property? You cannot attribute an innovation to an individual anyway, since it belongs to the company.
A: I don't think it's a result of that. It's true that what's done belongs to the company, but I don't think that people think in this way. I think that there's a genuine understanding that the reason we are at Akamai and we've taken a break from our academic careers, or we've changed our career paths was because we're trying to build a company based on a service; to make the company succeed we had to provide the best service to satisfy the customers. Once everybody understands their role, the teamwork comes naturally. It's true that sharing of information is much more restricted at a company, and especially at a start-up, where it is critical to maintain a competitive advantage. The publication rate out of Akamai is extremely low: you may be able to say that you never had that many top professors publish so few papers. But I think this is motivated by the same reasons that made us work together as a team.
Q: A graduate student friend of mine would like to do an internship at Akamai. I am not trying here to pull strings, but I would like to ask about the Non-Disclosure Agreement (NDA): will my friend be able to bring interesting research problems from a stage at Akamai? Is the NDA very strict?
A: I think that Akamai has a fairly standard NDA; it is not primarily targeted at preventing researchers from conducting research. It is more like a standard industry NDA, so that when there are collaborations between members of different companies, or between Akamai and consultants, everybody understands what Intellectual Property (IP) must remain confidential. I think that in general Akamai is supportive of academic research: our roots come from the academia. We do our best, especially when students are involved, to limit the impact that we would have on their future careers. It would really depend on the specifics of the case: it wouldn't be possible for a student to write a thesis that described Akamai's entire working infrastructure and then explained new research ideas that could be used to improve specific pieces. There will obviously have to be some limits. By signing the NDA the student agrees not to disclose critical IP that belongs to Akamai.
Q: The late [Turing & Nobel Prize winner] Herbert Simon said somewhere that he longs for the times when groups of professors were focused on an important research goal, and they were led by a senior faculty member. He said that nowadays most faculty seem to pursue more individual research objectives. This seems to match what you said about teamwork. Do you think something could be borrowed from the company model to use in the academia? Maybe to, say, waste less time on ``marginal'' research problems?
A: I think, if I understood the quote, the biggest concern is not whether the faculty are working independently, but whether there's someone that's leading the faculty and sending them in a certain direction. I believe it is very healthy for individual faculty members to choose the problems that they think are important and to work on them, and not to be constrained by the opinions of a more senior faculty member who sets the research direction. At the same time, it is definitely the case that it's possible to conduct research much more efficiently as a team, and that the academia, at least in the US, really rewards individual accomplishments, especially early in the faculty member's career. To gain tenure at a research university a faculty member has to establish an independent reputation. If the faculty were to participate only in large collaborative efforts that would work against the current reward system and against the long-term faculty member's career aspirations. This aspect of our academic system discourages teamwork. I don't think there's an easy solution. When giving senior faculty the authority to lead other, you must know that the right individual has been empowered.
Q: Let me go to a few more technical questions about Akamai. The word is that you use Linux a lot. Is is true? Why Linux?
A: It is true that most of Akamai's servers are Linux servers. However we also run a large number of Windows 2000 servers, in particular the servers delivering Windows Media format. Akamai began with Linux servers, and there are a variety of reasons:
Q: How do you administer, maintain and debug so many machines in 50 countries?
A: That's an interesting question. First of all, we don't man any locations where our servers are located. All of the servers are administered over the network. This is a really challenging task: when you have 8000 servers on any given day you are bound to have hardware problems. Some will be located in places that lose connectivity to the rest of the Internet. The software that runs on them has to be designed with the understanding that there are periods of time when it is not possible to access the servers and there are periods when servers have difficulty communicating with other parts of the system. This is really an open-ended question. Basically most of Akamai's core IP addresses this issue: how to build and maintain a really large distributed system. There are some interesting rules of software development which arise when building such a system. For example, any new software must be backwards compatible with software out in the field. The reason is that, even though we do upgrade the software simultaneously on all the servers, the process may take a long time depending on the network connectivity. Parts will have been upgraded while parts haven't, and yet the system must be able to operate during this transition. The system also has to be completely de-centralized. There can be no one location on which the effectiveness of the entire system depends. Since it is so large, it is really important to minimize the number of manual operations that have to be performed in order to administer a system. If human intervention was needed to keep the system working, as more and more and more servers are added, it would require more and more human power. We have a really sophisticated monitoring system, which gives us a total view of our system. The design methodology was to build a system that was always self-organized and de-centralized.
Q: Let's say that tomorrow we can go to the Internet Engineering Task Force (IETF) and submit a proposal on what service the network should provide. What would you wish the network to provide to ease your job?
A: That's a good question. I think that the biggest thing would be that the inter-domain routing protocols to incorporate some notion of performance when making route selections. Today this is based primarily on the number of hops to get from one autonomous system to another. That is only very roughly correlated with performance. What we see throughout the Internet on any given day is that some connections may become congested, although this is not advertised in any routes propagated using the Border Gateway Protocol (BGP). Because of this, traffic may go through the congested links, and this causes big complications, because Akamai promises its content providers it will do a good job. Because the information is not incorporated into routing decisions we have to perform measurements to work around the default routes.
Q: I will wrap-up with some more personal questions. Say I am an undergraduate looking towards a research careers in grad school. What research area would you advise me to follow?
A: I don't want to pick a research area for anybody. I think you should work on something that you believe is important and interesting. Unless you're compelled to do the research because you find it interesting, your career can be very burdensome. You must have some strong motivation from the beginning. Choosing an university can be sometimes easy, sometimes hard. You may choose a university because there's a particular faculty member that you want to work with. That's risky, because often there's no guarantee that the faculty member will be able to take you on as a supervisor. As a general rule, it would be best to choose an university with a reputation for high-quality research results. This can be measured, well, by your opinion of different research papers. If you look at some papers, and you find some that you think are good, look where those authors are from. This may be difficult for a student who has yet to begin a research career: then the advice of faculty members at your undergraduate institution can help. I think it is generally a good idea, although I didn't follow this advice, to study for a graduate degree at a different school than your undergraduate institution, because you'll get a different point of view from the faculty, as you will meet many new potential collaborators. In choosing a graduate school the most important thing is that you understand what it is that you want to study.
Q: Is there a great CS book which is overlooked by too many?
A: I haven't written any book, so the answer doesn't come to mind immediately [laughs].
Q: Thank you very much!