from SIAM News, Volume 32, Number 10
`
`Akamai Technologies:
`A Mathematical Success Story
`
`By, Alan Edelman‘
`
`What a success story for mathematics! No, it’s not a proof of the Riemann hypothesis, but a multibillion-dollar mathematical
`event. It’s the story of a research problem that blossomed into one of the hottest recent IPOs on Wall Street. And it’s all due to the
`tremendous impact of the Internet.
`The mathematicians are Tom Leighton (ofthe applied mathematics group and the Laboratory for Computer Science at MIT) and
`Danny Lewin (often referred to as “Tom’s graduate student” but a visionary in his own right). Together, they created Akamai
`Technologies. Akamai is now a large company, with hundreds ofemployees (many ofwhom are SIAM members; Leighton himself
`is a SICOMP editor).
`Like any good h1ghtech company, Akamai1s run by someone with business experience George Conrades, the CEO, is a former
`senior vice president ofIBM and more recently a venture partner at Polaris Venture Partners; hejOIned Akamai last spring Jonathan
`Seelig, another Akamai founder, recognized the opportunity while at MIT’s business school; it is probably fair to say that he became
`a business school sensation
`.7
`
`'
`
`AkamaiIS very much the talk ofMIT these days, and the news has spread rapidly among mathematicians everywhere Some are
`thrilled with the great news and others are inspired, while some at MIT are concerned about the implications for academics. At least
`five faculty membersin MIT’s appliedmathematics group are working at Akamai, on a part-time basis or more.
`
`A Limosine Service for the Web
`
`Akamai speeds up the delivery of Web pages with its “FreeFlow” service. On entering Akamai’s Web site, once you’ve passed
`the flashy “Shockwave” graphic, you will see the slogan: “Internet Content Delivery: Guaranteed.” Mathematically, if T is the
`download time for a Web page, they strive to make T as small as possible, and they guarantee that T < oo. Backing up this claim
`are mathematical cleverness and programming talent. The policy is that if Akamai fails to perform, the customer does not pay. No
`excuses.
`
`I recently began working at Akamai part time, while on sabbatical from the mathematics department at MIT. I had never expected
`to find the atmosphere of MIT reproduced in industry—~people working around the clock to exceed ordinary standards of quality.
`(Akamai has talented people from many places, not only MIT, but I think it is fair to say that the MIT culture and the Akamai‘culture
`are indistinguishable.) Readers will see that my enthusiasm level is high as I write this; it has been great fun going to Akamai.
`Understanding Akamai requires some understanding ofthe nature of the Internet. The Internet is much like the streets ofBoston,
`or perhaps any old, unplanned city. Connections were based at first on local needs, and thoroughfares were superimposed later.
`Routes are constantly clogged or unavailable due to repairs. Similarly, no central authority built the Internet; that’s why it is so
`widely available today, and it’s also why it is so messy and chaotic. Akamai1s something like a limousine service that guarantees
`a short ride no matter what the traffic conditions.
`
`It18 easy to tell if a Web page is “Akamaized.” My favorite example is The New York Times. Go to www.nytimes.com and view
`the source. (As many of you know, the official term for a Web address is a URL, or Universal Resource Locator, sometimes called
`a Universal Record Locator or Uniform Resource Locator. Most URLs begin with the prefix http:// and most browsers assume the
`http:// prefix by default.) If you search the source for Akamai, as I have just done, you will find a long URL: httpzl/
`a836.g.akamaitech.net/7/836/967/3 cee073de96781/graphics.nytimes.com/images/nytlogo12.gif. A few clicks away is the famil-
`iar neWspaper masthead logo:
`
`.
`
`The URL seems terribly long, but of course nobody ever types it. (Internally, an Akamaized URL is known as an ARL.) Just for
`fun, you can grab the last three fields from the above URL: graphics.nytimes.com/images/nytlogo12.gif. Again, with just a few
`clicks, you will see the same image. So why did my image come from the long Akamaized URL rather than the shorter one?
`The shorter URL produces the image “live” from New York, which might not be the most convenient location for every user.
`The New York server could also slow down or die in the event of an overload ofrequests or some other failure. But Akamai servers,
`located around the world, grab the image once from New York and serve it to readers worldwide from convenient locations near
`them. Akamai customers can choose to Akamaize some or all of their content.
`'
`, Even more fun is the Star Wars trailer: http://www.apple.corn/quicktime/trailers/fox/episode-i/480.html. The long version ofthe
`URL for the quicktime movie that is being downloaded is http://a1912.a.kamai.com/7/1912/52/3847dda752238/
`downloadakamai.com/apple/menac480mov.
`If you try to download the movie, you will probably notice the akamai on the bottom of your browser. A large number of well-
`MIT 2009
`MIT 2009
`Limelight v. MIT
`Limelight v. MIT
`IPR2017-00249
`|PR2017-00249
`
`

`

`known content providers, including CNN, Yahoo, and ESPN, are now
`Akamaized.
`
`
`
`~
`
`BesServed
`
`3
`Tues Wed Thurs
`
`q
`
`q
`Sat
`
`Fri
`
`Mon
`
`Dynamically Changing Traffic Patterns
`Many readers may be familiar with the simple idea of “mirroring.” For
`two servers with exactly the same content—one in the U.S. and the other
`in Europe, for example—mirroring reduces the need for ongoing trans- ,
`Atlantic communication. More sophisticated is the notion of caching—
`remote servers fetch a page from the original server only the first time it
`is requested; thereafter, the “caches” serve the requests until the page
`expires or the least popular material is removed from the cache to make
`space. In summary, mirroring is concerned with geographic locality
`(really network locality), while caching makes it possible to handle large
`demand. (Mirroring would be enough in a world of low demand for Web
`pages. Cachmg would be enough in a tiny world With high demand.)
`M1rror1ng and cachlng are stlll not enough in a world of dynamically
`changing traffic patterns. Returning to our U.S./Europe mirror example:
`During the morning hours in Europe, when people in the U.S. are still
`asleep, a European mirror could well be overloaded. In this situation, new
`requests might best overflow to an American server. Thisis but one example of the problems that must be solved because of the
`dynamic nature of the Internet The real world ofthe Internet1s very messy: Messages get lost for inexplicable reasons, servers go
`up and down, content becomes more popular and then less popular, traffic can surge, lull, and then surge again.
`Akamai has a network of 1475 servers (and the number is growing) spread around the entire planet. These servers and the entire
`Akamai system are dedicated to optimizing the combined issues of network locality, user demand, Internet traffic, and hardware
`reliability. On arriving at Akamai’s headquarters, in Cambridge, Massachusetts, a Visitor sees a command center that makes this
`author think of something like NASA’s Mission Control. There, the Internet is monitored and warning lights activate when trouble
`ar1ses.
`
`Sun‘
`
`F599“ 1 - quber Ofbytes served to an Akamai customer
`dunng a typ’ca' week'
`'
`
`.
`
`
`
`,
`
`_.
`
`_
`
`The system details are proprietary, but roughly speaking, a request for a Web page is mapped to a server. This server is likely
`to be geographically nearby, but Internet traffic and general bandwidthissues are really the deciding factor. If the Web page is in
`the cache, it is served, otherwise, the first step
`is to fetch the page. Dynamic routing occurs based on
`immediate knowledge of real-time Internet
`monitoring and the requirement to serve fresh content.
`Measuring the Internet can also be a chal—
`lenging problem. Isuspect that it is easier than obtain-
`ing data about automobile traffic patterns,
`because a lot can be done with software. We can
`already make predictions: For most Ameri—.
`can content providers, demand is lowest at 5 in the
`morning, Eastern time. By then, most Ameri-
`cans are probably asleep and Europeans have not yet
`reached their lunch breaks. The cycle of lunch breaks tends to start the surge. The issue of measurement fascinates me. Probably
`SIAMNews readers would not be surprised to learn that I consider Matlab the right data playground for Internet data exploration.
`Just for fun, a graph of a typical week for one of Akamai’s customers is shown in Figure 1. Every fiveminutes, we take a count
`ofthe number ofbytes served. Each ofthe vertical (shaded and white) regions represents six hours; a day begins at midnight, Eastern
`Standard Time. The daily human cycles are easily observed. We can match one weekday against another, or one week against the
`next, to predict the daily trends and the weekly growth.
`.
`“Consistent hashing” is the mathematical idea underlying the caching ofWeb pages. “Hashing” itselfrefers to the fast generation
`of a relatively short identifier for a larger piece of text. The hash value serves as a pointer to the item itself. If every cache had a
`copy of every item, there would be no room left on disk. If the items are distributed, then a quick addressing scheme is needed to
`locate the right cache. Any scheme used should require a small number of changes when new objects need to be cached and old
`objects are deleted. This is the mathematical problem that began the whole story. The problem was originally posed by Tim Berners-
`Lee, founder of the World Wide Web, whose office in MIT’s Laboratory for Computer Science is on the same floor as Tom
`Leighton’s.
`
`Will Academia Suffer?
`
`I would like to close with some observations about industry and academia. A large number of technical employees at Akamai
`are MIT students and faculty. Many work part time, although Akamai is so invigorating that it is easy to spend more time there.
`Concerns have been voiced that students, hooked by Akamai, are neglecting their studies or social life. Similar concerns are
`expressed about faculty. Many ofus are aware that industry in general is attracting many of the best minds, and academia does not
`always seem to be able to compete.
`.
`Some gloom-and-doom predictors believe that academia will suffer because of the allure of Internet start-ups. Indeed, many
`talented people are leaving academia, and classes are harder to staff. The five mathematicians and two computer scientists who are
`consulting or on temporary leave at Akamai retain their commitments and relationship with MIT. I believe that we will all return
`with the benefit ofa real-world applied mathematics and computer systems experience that will enhance our research and teaching.
`Applied mathematics will only have gained.
`
`

`

`Further Reading
`D.R. Karger, E. Lehman, T. Leighton, M. Levine, D. Lewin, and R. Panigrahy, Consistent hashing and random trees: Distributed caching
`protocolsforrelieving hotsp‘ots on the World Wide Web, Proceedings ofthe 29th ACM Symposium on Theory ofComputing, ACM, ACM Press,
`May 1997, 654—663.
`D. Karger, A. Sherman, A. Berkheimer, B. Bogstad, R. Dhanidina, K. Iwamoto, B. Kim, L. Matkins, and Y. Yemshalmi, Web caching with
`consistent hashing, Eighth World Wide Web Conference, Toronto, Canada, May 1999, http://www8.org.
`
`Alan Edelman is an associate professor ofapplied mathematics and a member ofthe Laboratoryfor Computer Science at MIT.
`
`