Toward the Computer Utility:

A Career in Computer System Architecture

Jack B. Dennis

Professor of Computer Science and Engineering, MIT, Emeritus

 

When I entered MIT in September 1949, the von Neumann concept of the stored program computer was only a few years old: it was the subject of a summer school program at the University of Pennsylvania Moore School of Electrical Engineering in 1946. It was around 1949 that realizations of the stored program concept went into operation at Cambridge and Manchester, both in England, followed closely by machines at the Moore School, at the Princeton Institute for Advanced Study, and the Whirlwind computer at MIT.

 

My introduction to computers came through two undergraduate friends: Bill Eccles, a fellow EE, who had some experience with the IBM "card programmed calculator" in Frank Verzuh's shop, and Ken Ralston, who entertained us with stories of his experiences with the Whirlwind computer, then under construction. I became a Whirlwind programmer, first to do optimization calculations for Prof. Saunders who was visiting from Berkeley, and then to implement an algorithm for the "transportation problem" under Prof. Bill Linvill. I prepared programs written in assembly language on punched paper tape using Frieden "Flexowriters", and stood aside watching the myriad lights blink and flash while operator Mike Solamita fed the tapes. When my program crashed, I could watch while the octal contents of memory were dumped onto a CRT screen and snapped by a camera for my later perusal using a microfilm reader.

 

That was 1954. Fifty years later much has changed: A room full of vacuum tubes has become a tiny chip with millions of transistors. A phenomenon once limited to research laboratories has become an industry producing commodity products that anyone can own and use beneficially.

 

But much of the fundamentals have not changed. A computer program is still a sequence of instructions obeyed by a processor as if their effects were accomplished one at a time. The only exception to this is the "interruption" of the program sequence to allow the processor to pay attention to another program, or to some input-output device calling for service. Even our favorite programming languages deal only with the sequential aspect of computation, and provide no means for expressing actions that require use of program interruptions; that is left to operating system calls, or, at best, library packages that encapsulate usage patterns of operating system facilities. This situation leaves programming languages unable to express large applications in a modular style such as that supported by the familiar procedure concept in sequential programming languages.

 

I have seen this history from inside a major research university, as a teacher of computer science, and as a researcher in what I like to call "computer system architecture".

 

In 1960 Professor John McCarthy, now at Stanford University and known for his contributions to artificial intelligence, led the "Long Range Computer Study Group" (LRCSG) which proposed objectives for MIT's future computer systems. I had the privilege of participating in the work of the LRCSG, which led to Project MAC and the Multics computer and operating system, under the organizational leadership of Prof. Robert Fano and the technical guidance of Prof. Fernando Corbató.

 

At this time Prof. Fano had a vision of the Computer Utility – the concept of the computer system as a repository for the knowledge of a community – data and procedures in a form that could be readily shared – a repository that could be built upon to create ever more powerful procedures, services, and active knowledge from those already in place. Prof. Corbató's goal was to provide the kind of central computer installation and operating system that could make this vision a reality. With funding from DARPA, the Defense Advanced Research Projects Agency, the result was Multics.

 

I am proud of the role I played in the activities that led to the construction of Multics. From the work of the LRCSG we envisioned that the hardware for Multics would be a symmetric multiprocessor (several processors having equal access to several banks of main memory). To support the sharing of processor and memory resources by independent programs running for many users at the same time, I advocated the combination of segmentation (inspired by the Burroughs B5000 system) and paging (inspired by the Manchester Atlas computer). This would allow segments containing program modules and units of structured data to be shared by many users without the need for making copies. This concept was adopted for Multics.

 

A team, composed of Prof. Corbató, Ted Glaser, Robert Graham, and me, spent much of the 1963-64 academic year visiting most of the major US computer manufacturers to see which of them might be able to meet the requirements we had formulated. IBM put in a valiant effort toward convincing us that System 360 was the right choice, and we had to explain to their top management how the 360 architecture failed to address the problem of achieving the rapid reallocation of resources demanded by the time-sharing environment. Nevertheless, we found a willing collaborator in John Couleur of the General Electric Company, and developed the Multics hardware, the first computer system with hardware support for large numbers of paged segments of virtual memory, as the GE 645 system, a major variant of the GE 635 product.

 

Rather than work on the software operating system for Multics, I chose to do independent research following the intellectual ideas and issues that arose from my experience with the Multics effort, and the time-sharing system some students and I had built using a DEC PDP-1 computer. During the 1960s Project MAC had very generous support from DARPA, and the MIT Computer Science faculty and graduate students could choose any topic to study, so long as it had some relation to computing. I formed the Computation Structures Group and focused on architectural concepts that could narrow the acknowledged gap between programming concepts and the organization of computer hardware. I found myself dismayed that people would consider themselves to be either hardware or software experts, but paid little heed to how joint advances in programming and architecture could lead to a synergistic outcome that might revolutionize computing practice.

 

Nevertheless, in the 1970s I found it easy to get government funding. The agencies were willing to fund pretty wild ideas, and I was supported to do research on "data flow" architecture, first by NSF and later by the DOE. This work inspired related projects at several companies and research institutions around the world, and earned me the Echert-Mauchly Award in 1984.

 

During the 1980s things changed. Computer Science Departments had proliferated throughout the universities to meet the demand, primarily for programmers and software engineers, and the faculty assembled to teach the subjects was expected to do meaningful research. To manage the burgeoning flood of conference papers, program committees adopted a new strategy for papers in computer architecture: No more wild ideas; papers had to present quantitative results. The effect was to create a style of graduate research in computer architecture that remains the "conventional wisdom" of the community to the present day: Make a small, innovative, change to a commercially accepted design and evaluate it using standard benchmark programs. This style has stifled the exploration and publication of interesting architectural ideas that require more than a modicum of change from current practice. The practice of basing evaluations on standard benchmark codes neglects the potential benefits of architectural concepts that need a change in programming methodology to demonstrate their full benefit.

 

Today, much of the excitement in computer science has shifted to various important application domains: medicine, speech processing, support for advanced human interfaces, communications, graphics, etc., and less funding is available for academic work in the core areas: programming languages, operating systems, and computer architecture. In fact, there are people who consider that these core areas have reached the limit of their potential for innovation. However, this is belied by the popularity of Java, a new language that incorporates important features long advocated by computer science academics.

 

The real gains in programmer productivity yielded by modern computer hardware are due mainly to the increasing size of physical memories and the universal adoption of single-user virtual memory support. Of course, major advances in programming have resulted from ideas such as structured programming and object-oriented design and the influence of these ideas on programming languages. However, these advances have failed to address the issues of high-level programming for computations using multiple processors.

 

Now, the design of computer processor chips has reached an impasse. It is no longer practical to make a faster processor by adding transistors to the logic, and it is very expensive to further increase the clock rate of the processor. To any hardware engineer the obvious answer is to put multiple processors on a chip, and this is now being done. However, the way provided for processors to cooperate with each other is ad hoc with little attention paid to how a beneficial methodology of parallel programming could be supported.

 

In September 1988 the MIT Laboratory for Computer Science celebrated the thirty-fifth anniversary of Project MAC. At that event I pointed out the limitations of conventional multiprocessor architecture:

 

"Yet present multiprocessors are very limited in their effective application. Their programming tools are absurdly limited and primitive in contrast to those of Multics. There is no automatic management of memory by the system on behalf of its users. Moreover, within the massive research effort now devoted to parallel architecture, hardly any effort is devoted to the problem of improving programmability in any fundamental sense."

 

Some wild ideas may be the key to a breakthrough: Functional programming; the idea of a memory that directly supports creation and access to data objects, but does not permit updates; hardware-supported allocation and garbage collection of memory. They need to be seriously explored.

 

At the Project MAC anniversary I explained:

 

"... the key idea ... is functional programming: getting away from the burden of sequential programming concepts embedded in our popular programming languages and computer architectures. The ideas to accomplish this advance exist; in fact they have been known for some time. Yet they are not in favor. Why? They do not fall into the current main stream of computer science. They do not solve the multiprocessor cache problem--instead they make it irrelevant. They do not solve the problem of shared objects in object-oriented programming--they eliminate the problem. They do not minimize the overhead of processor synchronization--they make it disappear altogether."

 

Fifteen years later these remarks remain valid. I concluded by saying:

 

"The computer systems of today do not realize our original vision from the inception of Project MAC. Yet the opportunity to make our dreams come true is still there. The vision is not obsolete. It is one that will be achieved. I believe the day will come when the ideas are widely accepted and we can move forward to build the Computer Utility. I hope to contribute to its realization and I look forward to enjoying its fruits."

 

There are wonderful opportunities for a new generation of talent to work at the core of a most relevant and rewarding area of intellectual effort.

 

Written for the 50th reunion book of the MIT class of 1953

Belmont, MA, August 2003