viewpoints

VDOI:10.1145/2822519 Len Shustek Interview An Interview with Fred Brooks ACM Fellow and A.M. Turing Award recipient Fred Brooks reflects on his career.

CM FELLOW FREDERICK (“Fred”) Brooks, recipient of the 1999 A.M. Turing Award, has made landmark contributions to computer Aarchitecture, operating systems, and software engineering. After earning a Ph.D. in Applied Mathematics and Computer Science from Harvard under the legendary Howard Aiken, he worked for IBM on several landmark computer systems, most notably the System/360 series that came to dominate main- frame computing for decades. He left IBM in 1964 to found the Computer Sci- ence Department at the University of North Carolina, from which he retired at the end of the Spring 2015 semester. Noted software architect Grady Booch conducted an oral interview of Brooks in Cambridge, U.K., in Septem- ber 2007. The complete transcripta of this interview is available in the Com- puter History Museum’s oral history archive; presented here is a condensed and highly edited version designed to whet your appetite. Fred Brooks with a statue created in his honor at the University of North Carolina —Len Shustek Department of Computer Science.

Falling in Love with Technology go to medical school, so we settled in I paid $35 for. I made my own McBee I was born in Duke University Hospital Greensboro, N.C. It was a great place card keysort systemb for keeping track because my father was teaching chem- to grow up because of the superb of my map collection. So I have al- istry at the University of North Caroli- school system. ways been fascinated with the kind of na, and that was the nearest hospital at My favorite subject was physics, but machines that process information. the time. When I was about six months I always had a fascination with busi- When I was 13, I read in a Time maga- old he decided to change careers and ness equipment. When the local corset factory went out of business, I bought b An edge-notched card filing system invented a http://www.computerhistory.org/collections/ filing cabinets for $4 apiece and a in 1896; see http://en.wikipedia.org/wiki/

catalog/102658255 Burroughs comptometer that I think Edge-notched_card CHAPEL HILL SEARS, THE UNIVERSITY CAROLINA AT DAN OF NORTH BY PHOTO

36 COMMUNICATIONS OF THE ACM | NOVEMBER 2015 | VOL. 58 | NO. 11 viewpoints viewpoints

zine in the little town library about the an improvement, and that was essen- Harvard Mark I computer and I knew tially the machine I designed for my from then on that was what I wanted When I was 13, I read dissertation, but there were no further to do. about the Harvard economies to be had by making it spe- cialized for payroll. Studying Computers at Harvard Mark I computer Aiken was a very impressive person, I did my undergraduate major at Duke and I knew from and did something I cannot do with my University in physics, but I also stud- then on that was students: he came to my office every V ied math, economics, and accounting day during the year I was working on as well as the humanities because that what I wanted to do. my dissertation and wanted to see the was also my interest. The head of the prose that had appeared since the day physics department was determined before. And guess what? Some had ap- that his best students should go to peared since the day before. Harvard to study physics, and two of us did. When I told my advisor there Helpful Summer Jobs that I really want to study computers, 10,000 words of program storage, and I was offered a summer job at Mara- he said, “Fred, you’re too late to get in 4,000 words of backup drum storage. thon Oil Company in Findlay, OH, us- on the ground floor, but you can catch There wasn’t an assembler, but Aiken ing an IBM 650 to mechanize payroll. the first landing.” That meant I had a had designed a relay box that enabled That was priceless experience. The chance to meet and really get to know you to encode it algebraically. Our first complexities in business data pro- J. Presper Eckert, John Mauchly, and program for determining the Markov cessing are that you are trying to lump Konrad Zuse. Howard Aiken was my frequencies ran about 2,500 instruc- a wide variety of cases under a single thesis advisor. So I got to really know tions. We were allowed two one-hour process, so typically on the 650 when the pioneers even though I was not in slots on the machine during the semes- it was doing business applications, that generation, and that has been a ter, so we did extensive desk checking. half of all instructions would be con- great joy. It is the only big program I ever wrote ditional branches. That is not true in I went into the computation lab, a that ran right the first time. scientific computing. very small group that was part of a di- It helped me as a computer archi- vision of engineering in applied phys- Howard Aiken as an Advisor tect that I was also doing scientific ics. There was Aiken, who was the At the end of my first year Aiken told computing. I spent a summer at North boss, and two young instructors, Ken Ken Iverson “I would like you to pre- American Aviation doing missile track- Iverson and Bob Minnick. Every day pare a course on the application of ing and database building. A summer the boss was in town the whole crowd computers to business.” Nobody had at Bell Labs trying to identify which gathered for coffee at 5 p.m. in the ma- ever taught any such course anywhere party on a four-party line was dialing chine room, which had the Harvard in the world. I had had trouble with a this call. A summer at IBM in Endi- Mark I on one side and the Harvard course in boundary value problems cott, NY, where I learned punched card Mark IV on the other side, each about and was not continuing my NSF fel- machines and was in the physics de- 60 feet long. We would talk for a half- lowship, so I went to Ken and said, partment doing acoustics. Those four hour or 45 minutes, and then he would “Can I be your teaching assistant, be- summers were a priceless part of my go home to supper and we would go cause that’s right down my alley?” Out education, and they exposed me to rad- back to work. of that came our book Automatic Data ically different corporate cultures and Processing, which went through an edi- radically different geographies. Programming the Harvard Mark IV tion based on the IBM 650 and then six We programmed the Mark IV start- years later an edition based on the 360. Working on IBM’s Supercomputer ing the first semester. My present col- Ken was fully as important in my edu- Steve Dunwell, who was project man- league Bill Wright and I undertook, cation as Aiken was. ager for Stretch, came through Harvard for our first year project, to write a For my dissertation, Aiken said “I in 1956 looking for people. His offer program that would analyze melodies want you to design a machine special- was very attractive: a chance to work on and create synthetic melodies using ized for payrolls.” That sounded like a the world’s fastest computer. I leapt at an eightfold Markovian process. We good topic, even though he was more the opportunity. My wife Nancy, whom chose common meter hymns because interested in the machine and I was I had met the first Sunday night at Har- we could find a lot of them that had more interested in the design method- vard, worked on the transistor circuits, the same metrical structure. We trans- ology—how you get from the require- and I worked with the architects under posed them all to the same key, then ments to the machine. Werner Buchholtz. we analyzed the transition probabili- His hypothesis was that by special- IBM had been thinking about it for ties of the melodies. It took us three izing a machine for payroll you could a year, and had just signed the contract years to get done, but we got music you make significant improvements in with the National Security Agency to could pass off on any choir. cost performance. That turned out build Harvest, a variant of Stretch for The Mark IV was programmed in not to be the case. By specializing for breaking wire-wheel cryptographic decimal absolute. It had 230 registers, serial file processing you could get codes. I was the only one of the archi-

NOVEMBER 2015 | VOL. 58 | NO. 11 | COMMUNICATIONS OF THE ACM 37 viewpoints

tects proper who was native-born, so I chines communications oriented. make sure my boys landed on their feet could get the security clearance to work In January we did an all-day presen- in the reassignment. I was on the way on it. Stretch, as the host, was about tation of the 8000 series with the brass out back to research when, to my utter 15-feet long and 5-feet high and 5-feet up from Armonk and White Plains. The amazement, Bob asked me to take the deep. Harvest was a “plug-in card” whole program was very well received, new product line. The crown jewels! that added another 20 feet of special- except for one fellow sitting in the back I was dubious. We had been fighting ized electronics. It processed about 4 who just got glummer as the day went pretty hard. But Jerry Haddad, who was million bytes a second, and had about on. That was Vin Learson, and that is Bob’s boss at the time, said, “I never 250 bytes of instruction to set it up that not who you want to get glummer, be- knew anybody that regretted working would take half a day to write. cause he was executive vice president for Bob Evans.” So I gave it a try, and The purpose of Stretch was to make of the company. Well, that night he it went really well. We clicked, and we the fastest machine we could, cost fired my boss. He brought in Bob Evans fought shoulder to shoulder, side by no object. That is both liberating and and told him “If it’s right, make it hap- side, the rest of the way. Bob is one of tempting, but we did not succeed. pen; if it’s wrong, turn it around.” four great bosses I had in my life. Dunwell had set out to make a ma- Bob spent three weeks looking chine 100 times faster than the IBM into it, then took me out to dinner at Designing the System/360 704. But the memories available were a restaurant in Poughkeepsie and told Architecturally we started out in the only six times faster, and the circuits me he had decided it was wrong. We data systems division pursuing a stack were about 10 times faster. The idea were losing market share to the Seven architecture, which is what Burroughs was that by using more of everything Dwarfs.c Everybody was out after us. We had done. It turns out that works great we could get there, and you can’t. We were obsolete. We were fundamentally if you’ve got a transistor register for sev- only got to 50 times faster. address-size limited. We could not at- eral levels, and it does not work great if It was declared a disaster and with- tach more memory to the 7090 or the you’ve got it all in memory, and you are drawn from the market. But later 7080, and the applications were hungry having to pull and push all the time. IBM recognized Stretch technology for more memory. The problem is addressing. If you enabled the 7090s, 7080s, and 7074s, He proposed to wait for a new semi- put addresses big enough on the littlest and that the concepts became cru- integrated circuit technology that was machine, your littlest machine is se- cial for the System/360. Tom Watson, going to be three years down the road. rial by byte and that means you spend to his credit, went back and got Steve The problem is, how do you hold the a lot of cycles fetching address bytes Dunwell out of disgrace and made a market in the meantime? My plan that you are never going to use, and special award to him recognizing the would get out there now, although it that compromises your performance. important influence the Stretch had had some fundamental difficulties. The stack architecture was a way of ad- on the company’s welfare. He was right and I was wrong. He had dressing that by not fetching as many two parts of his vision. One was that we addresses, but when we went through Figuring Out IBM’s Future ought not to do a new product line for the performance and cost estimation The Data Systems Division, which was the Data Systems Division; we ought to in March 1962, the stack machine was the middle of the market in computers, do a new product line for the IBM com- working fine from the middle size up recognized they had a product problem pany, big and little. The other was that but was not competitive below because with their many existing lines. The vice we ought to make the new product line of all this pushing data in and out of president for engineering appointed a coincide with the new technology. memory. We had accepted the assign- committee to study what a successor We fought back and forth. We went to ment of making the whole product line product line architecture might look the Corporate Management Committee upward and downward compatible like. I had moved to the research divi- in March, and I won. That did not slow with one architecture. sion, but I chaired that committee. Bob down a bit. Bob is unstoppable. So we scrapped it and had a design At the end of the summer I was He went to the Corporate Management competition, which was Gene Am- asked to come back to the Data Sys- Committee in May, and he won. It was dahl’s idea. I think there were 12 or tems Division as manager of architec- over. This meant stopping all the 8000 13 internal teams, mostly of three or ture. We undertook to design a new series projects in the Poughkeepsie lab four people. I said the decision of the product line called the 8000 series, and reassigning all the people. His plan judge will be final. Gene’s team and built around Stretch concepts. There was to do temporizing machines—the Jerry’s team came in with essentially was a small scientific computer, small 7094, the 7080 model 3, and so forth— the same concept, which was to use and mid-sized business machines, to hold the market as best we could un- the base registers that Philco had in- a grown-up version for high perfor- til we could get there with the new prod- troduced in a machine a year or so mance use by insurance companies uct line using the new technology. earlier. Base registers meant we could and utilities, and a grown-up version I had gone to a retreat up at Sara- get by with short addresses. for scientific computing. We worked toga Springs to spend a week ironing There was one very big difference. very hard, and in January of 1961 we out who is going to do what, and to Gene’s machine was based on the had cost estimates and market fore- existing 6-bit byte and multiples of casts, plus a plan that involved creat- c Burroughs, Sperry Rand, Control Data Corpora- that: 24-bit instructions and a 48-bit ing new markets by making these ma- tion, Honeywell, General Electric, RCA, and NCR. instruction or floating point. Jerry’s

38 COMMUNICATIONS OF THE ACM | NOVEMBER 2015 | VOL. 58 | NO. 11 viewpoints machine was based on an 8-bit byte competitors—because the project had 1964. We came back with a totally and 32-bit instructions, so 64-bit and been kept pretty well secret. And the different product plan that involved 32-bit floating point. This is not a orders just started pouring in. these memory levels, compatibil- real happy choice. There are strong ity among them, a variety of language arguments each way, but you want Operating Systems processors—two FORTRANs, a Report your architecture to be consistent. In 1962 we formulated a plan to have Program Generator, two levels of CO- You are not going to have an 8-bit four software levels, known as Ro- BOL—and a modular operating sys- byte and 48-bit instruction floating- mans I, II, III, IV. Because we had tem that would start at 16K. You have point word. hardware compatibility across the got to leave some space for the appli- It was our biggest internal fight. line, the software levels had to be dis- cation programs, so we said the oper- Gene and I each quit the company tinguished only by the memory size. ating system has to be resident in 4KB. once that week, but Mannie Piore, W.B. McWhirter, who was the Data It turned out we could not do that, so the senior scientist in the company Systems Division president, took the we ended up making 32KB the mini- and a person of great wisdom, got us software away from us, but they were mum size machine for OS/360. We back together. I had made the deci- busy, and they did not give the 360 had a tape-based operating system, a sion for the 8-bit byte. Gene appealed software much attention. card-based operating system, and a to Bob; but Bob affirmed it. Of all my Before the announcement I said to little disk operating system called Ba- technical accomplishments, making Bob, “Look, the machines are being sic Operating System. Those got deliv- the 8-bit byte decision is far and away released to the factory. The hardware ered on schedule in February 1965. the most important. The reason was part is done. The machines are on the But the big operating system was in that it opened up the lowercase al- track. Everything’s rolling. But the trouble. We patched and bailed, but by phabet. I saw language processing as software is in an utter mess. Between summer it was clear that we were still being another new market area that now and September, when I will leave in trouble. Tom Watson invited me we were not in, and could not get into to help found the Computer Science down to Armonk for one of these one- very well as long we were doing 6-bit Department at the University of North on-one luncheons in the executive din- character sets. Carolina at Chapel Hill, let me go over ing room and said, “Why don’t you stay When we announced the 360 in April there and bail, and just see what can be here?” I said that I really wanted to get 1964, I said to my team is “Tonight the done with the boat.” back closer to the technical level. “Well, lights will be on in the other guys’ of- I changed teams, and we had a will you stay another year and keep on fices”—those of our Seven Dwarfs retreat off in the woods in February bailing out the operating system? If you

NEW BOOKS FROM THE MIT PRESS

Handbook of ColleCtive Regulating tHe Cloud tHe little PRoveR intelligenCe Policy for Computing Infrastructure Daniel P. Friedman and Carl Eastlund edited by Thomas W. Malone and edited by Christopher S. Yoo and foreword by J Strother Moore Michael S. Bernstein Jean-François Blanchette afterword by Matthias Felleisen drawings by Duane Bibby "... ambitious and invigorating. Whether looking to start your research The emergence of the cloud as infrastructure: "... Friedman and Eastlund, two jolly career or simply curious to understand experts from a range of disciplines consider characters, eschew talk of bugs and what collective intelligence is about, policy issues including reliability, privacy, bombs, and strip [computational theorem you will find no better starting point consumer protection, national security, and proving] down to its essence as only a than this book." copyright. Little book can. Want proof? Look inside!” —Duncan J. Watts, Principal Researcher, Information Policy series | Paperback | $30 | £20.95 —Shriram Krishnamurthi, Professor of Microsoft Computer Science, Brown University Hardcover | $30 | £20.95 Paperback | $34 | £23.95 mitpress.mit.edu

NOVEMBER 2015 | VOL. 58 | NO. 11 | COMMUNICATIONS OF THE ACM 39 viewpoints

will, we’ll send somebody to Chapel dollars, which would be $4 billion to- Hill to teach your courses. You go one day roughly. week a month to get your department There is still an awful We recognized that building big organized. And when you need a new lot we do not know systems is different from building computer in Chapel Hill, we will help.” programs. There is still an awful lot That was a good offer, so I said yes. about how to build we do not know about how to build The workforce peaked at about complicated systems complicated systems and keep them 1,000. We did deliver the first version in under control. Tom Watson asked April or May of 1965, but it was slow. It and keep them me a question at lunch: “You’ve man- was November, really, before a respect- under control. aged the hardware program. You’ve able version was delivered. managed the software program. The worst decision, which is docu- What’s the difference?” I said, “Well mented in The Mythical Man Month,d I can’t answer that one but I’ll think was to take the software away from about it.” That’s where The Mythical the architecture group and give it to Man Month came from. As I say in the the operating system manager. The calls and parameter passing, which book, managing software is more like architecture manager had said, “If we had to master and adopt. An im- managing hardware than most soft- you leave it with me, it’ll be the same portant step forward in PL/I was the ware people believe and it is less like amount late, it will cost the same, but provision of compile-time mode, just managing hardware than most hard- it’ll be right.” He was right and I was like macro assembler. ware managers believe. wrong, and that was a multimillion- If we had been smart, we would dollar mistake. also have done a schedule-time mode On to Chapel Hill of PL/I instead of doing JCL, the Job I love to teach, and the atmosphere and Languages Control Language. But we weren’t the support at Chapel Hill in terms of my Assembly language, surprisingly smart. The worst mistake we made colleagues could not have been better. enough, posed a lot of technical was JCL. Its existence was a mistake. One of the areas we picked for em- problems. There were two entirely Building it on a card format was a phasis when we started the depart- different schools of thought. On the mistake. Building it on assembly lan- ment was interactive three-dimen- commercial side, the assembly lan- guage was a mistake. Thinking of it as sional computer graphics. I worked guage served as a platform on which only six little control cards instead of for 30 years with protein chemists, high priests in corporations wrote a language was a mistake, because it building tools to enable them to macros and the troops programmed had no proper subroutine facilities, construct protein structures from in the macro language. In the scien- no proper branching facilities. That crystallographic data. Then the vir- tific community, the scientists would we did not see it as a language was the tual reality concepts came along as write their own macros and they pro- fundamental problem; we saw it as a a natural extension of 3D interactive grammed in macro-enhanced assem- set of control cards. computer graphics. We have been bly language. The question of how to Except in the U.K., IBM waffled doing that for years, looking mostly resolve all this led me into the thickest about its support for PL/I. They did at serious applications as opposed to technical thickets I got into. the same thing with APL—on again, entertainment applications. The other set of technical decisions off again—when a wholehearted sup- I would advise someone consid- had to do with PL/I. One serious ques- port would have made it happen. Now, ering a career in the computer field tion is: Do we do independent evalua- could you have displaced FORTRAN? to look at the intersection between tion of expressions so that you can fac- No, I now think that would have been the computer field and biology. The tor out common sub-expressions? The impossible. I think you could have ablest young people today are often answer was yes, but that leads you to pe- displaced COBOL, whose community opting entirely to go totally into bi- culiarities such as one divided by three is more coherent. It is more top-down ology instead of computers, and bi- equals zero because of data typing. oriented, and PL/I is a better substitute ology offers the same promise today The basic concept was to make a for COBOL than it is for FORTRAN. that the computer field did to me in universal programming language that 1953. On the other hand, the key to would meld and displace FORTRAN How Big the Project Was much biology is information: where and COBOL. ALGOL was not really a My hardware development budget was it is hidden and how is it processed. factor, although part of our product about $100 million. That doesn’t count For the computer scientist the inter- plan included delivering an ALGOL all the I/O devices, the disks, the tapes, action with biology and biologists is compiler for government reasons. AL- the new keypunches to accommodate the golden opportunity. That is where GOL was popular in Europe and had 8-bit bytes and the larger character the fun is going to be. a lot of important concepts in it, par- set, and the capital for the factories. ticularly having to do with subroutine The whole program cost billions. The Len Shustek ([email protected]) is the chairman of the Computer History Museum. software budget for the OS, not count- d Frederick P. Brooks, Jr., The Mythical Man- ing DOS, I think was somewhere in the Month. Addison-Wesley, 1975, 1982, 47. neighborhood of $400 million in 1964 Copyright held by author.

40 COMMUNICATIONS OF THE ACM | NOVEMBER 2015 | VOL. 58 | NO. 11