Sep 23 447 lns
The Age of Computing: A Personal Memoir
By Nick Metropolis
-----------------------------------------------------------------------------
In the history of modern technology, computer science must figure as an
extraordinary chapter, and not only because of the remarkable speed of its
development. It is unfortunate, however, that the word "science" has been
widely used to designate enterprises that more properly belong to the domain
of engineering.
"Computer science" is a glaring misnomer, as are "information science,"
"communication science," and other questionable "sciences." The awe and
respect which science enjoys and which engineering is denied is inexplicable,
at least to one who sees the situation from the other side.
The popular image of science has changed little since it was invented by
Jules Verne and H.G. Wells. "Science" represents the search for knowledge,
the conquest over nature, the discovery of some very few fundamental laws
that will free mankind from worry and toil; this is as true today as it was
at the turn of the century.
The word "engineering," however, carries less exciting connotations. I
recall a pleasant evening at the house of the American Academy of Arts and
Sciences. In the great hall, conversation and gossip flowed freely in
anticipation of a brilliant lecture.
A distinguished lady, a pillar of Cambridge society, was expressing her
admiration for Professor S. She extolled his discoveries and his brilliant
insights. "And what department at MIT does he belong to?" she finally asked,
by way of indicating that our brief exchange was coming to an end.
"Mechanical engineering," I answered.
A look of horror crossed the lady's face. "Why, I thought he was a
scientist!" she blurted out before she could cover up her gaffe.
I saw in her eyes the image of a man in a dirty gray frock, a pair of
pliers in his greasy hands, bent over some Chaplinesque contraption of gears
and pulleys.
APPLICATIONS OVER THEORY
But contrary to the lady's prejudices about the engineering profession, the
fact is that quite some time ago, the tables were turned between theory and
applications in the physical sciences. Since World War II, the discoveries
that have changed the world were not made so much in lofty halls of
theoretical physics as in the less-noticed labs of engineering and
experimental physics.
The roles of pure and applied science have been reversed; they are no
longer what they were in the golden age of physics, in the age of Einstein,
Schrodinger, Fermi and Dirac. Readers of Scientific American, nourished on
the Wellsian image of science, will recoil from even entertaining the idea
that the age of physical "principles" may be over.
The laws of Newtonian mechanics, quantum mechanics and quantum
electrodynamics were the last in a long and noble line that appears to have
somewhat dried up in the last 50 years. As experimental devices (especially
measuring devices) are becoming infinitely more precise and reliable, the
wealth and sheer mass of new and baffling raw data collected by experiment
greatly exceeds the power of human reason to explain them.
Physical theory has failed in recent decades to provide a theoretical
underpinning for a world which increasingly appears as the work of some
seemingly mischievous demiurge. The failure of reason to explain fact is also
apparent in the life sciences, where "theories" (of the kind that physics has
led us to expect) do not exist; many are doubtful that this kind of
scientific explanation will ever be successful in explaining the secrets of
life.
SUBJECTIVE HISTORY
Historians of science have always had a soft spot for the history of
theoretical physics. The great theoretical advances of this century --
relativity and quantum mechanics -- have been documented in fascinating
historical accounts that have captivated the mind of the cultivated public.
There are no comparable studies of the relations between science and
engineering. Breaking with the tradition of the Fachidiot, theoretical
physicists have bestowed their romantic autobiographies on the world,
portraying themselves as the high priests of the reigning cult.
By their less than wholly objective accounts of the development of physics,
historians have conspired to propagate the myth of science as being
essentially theoretical physics. Though the myth no longer described
scientific reality 50 years ago, historians pretended that all was well, that
nothing had changed since the old heroic days of Einstein and his generation.
There were a few dissenters, however, such as the late Stanislaw Ulam, who
used to make himself obnoxious by proclaiming that Enrico Fermi was "the last
physicist." He and others who proclaimed such a possibility were prudently
ignored.
Physicists did what they could to keep the myth alive. With impeccable
chutzpah, they went on promulgating new "laws of nature" and carefully
imitated their masters of another age. With dismaying inevitability, many of
these latter-day "laws" have been exposed as quasi-mathematical
embellishments, devoid of great physical or scientific significance.
Historians of science have seen fit to ignore the history of the great
discoveries in applied physics, engineering and computer science, where real
scientific progress is nowadays to be found. Computer science in particular
has changed and continues to change the face of the world more thoroughly and
more drastically than did any of the great discoveries in theoretical
physics.
The prejudices of the academic world have stood in the way of the
historian. One wonders whether a historian of contemporary engineering could
get a teaching job at a respectable university.
For some reason, histories of long-obsolete discoveries, such as the steam
engine, are acceptable in academia: Dozens of such histories have been
written and, undoubtedly, dozens more will be written now that the field has
become an established one. However, a history of the transistor is still
beyond bounds (no such history has even been attempted, to the best of my
knowledge).
Thanks to the joint public relations efforts of historians and physicists,
the white mane of Albert Einstein remains the unquestioned symbol of genius.
It is scandalous, however, that virtually no cultivated person has ever heard
of John Bardeen, whose discoveries may have revolutionized the world at least
as much as Einstein's. Bardeen's midwestern background and his having taught
in Urbana, Ill., were fatal flaws that prevented his ever being recognized.
It would be tempting to conclude, after an inspection of empty library
shelves, that the absence of engineering histories, recounting major
discoveries, is due in part to the difficulty of gaining access to essential
facts. Practical discoveries are not as easily traceable to research papers
as are theoretical discoveries. Such a conclusion, however, would not be
warranted. The development of any discovery of even the slightest practical
value is generally thoroughly documented in reports, replete with names,
careful attribution given to who did what, when, with the funding sources and
dollar amounts given.
Unfortunately, access to such documents, at present, is severely restricted
by bureaucratic barriers deliberately placed in the way of those who have no
"need to know." Only the top managers of major business corporations, certain
officials of the federal government, and, in times past, selected members of
the KGB in the late Soviet Union were privileged to pursue such documents.
In our rapidly changing political and international climate, it is possible
that such restrictions will soon be lifted. When that happens, it will be
inexcusable for a historian of science to neglect the history of the great
technological discoveries of our time, including, obviously, the history of
computer science.
In offering some random remarks on the possibilities of such a history in
the future, I would like to suggest that the history of computer science --
if and when it comes to be written -- will establish a new and different
paradigm for history writing. It may indeed rid us of certain stereotypes
common to the history of science, with its overemphasis on the history of
theoretical physics.
SIMPLE IDEAS
In contrast to physics, the fundamental ideas that underlie the development
and implementation of large-scale computers are almost commonplace. The
principles of computer science are now so well known that they are thought to
be few and simple. They are unlikely to fire the imagination of a reading
public spoiled by science fiction; nor are they revolutionary ideas on which
movie scripts can (or will) be written. In fact, they sound pedestrian,
predictable and instrumental, reminding us of the old adage about
mathematics, that the ugliest theorems find the best applications, and vice
versa. In computer science, simple ideas requiring little or no intellectual
or scientific background have often worked out better than the more complex,
subtle and scientifically inspired proposals.
In universities today, students of computer science are the least
historically minded group in a student population not known for its
historical concerns. They seem to believe that the current concepts in the
field have existed from time immemorial, like a patrimony that all have the
right to access. Priorities in discovery have been unjustly attributed;
individuals who had no part whatever in the development of the field, such as
Alan Turing, are now given the status of heroes, while the names of those who
did the hard work, like John von Neumann, are scarcely remembered.
The phenomenon of obsolescence is particularly acute in computer science;
it works against the historian's task. In the age of the microchip, the
history of the vacuum tube has only limited appeal.
The discovery of a new computer model surrounds memories of all preceding
models with a thick web of irrelevance. In examining a computer of 10 or 20
years ago, our first reaction is not one of curiosity mixed with wonder and
admiration, as it should be, but of embarrassment, revulsion, almost
irritation. The inspection of the creations of our masters elicit smiles, or,
more often, giggles. The work of our predecessors has little to teach us, not
even in the lessons derived from what we perceive to be their clumsiness. In
computer science, obsolescence means a total break with the past, which
uniquely distinguishes this field from all others.
RELATIONSHIPS
The relationship between computer science and mathematics scarcely
resembles that which exists between physics and mathematics. The latter may
be best described as an unsuccessful marriage, with no possibility of
divorce.
Physicists internalize whatever mathematics they require, and eventually
claim priority for whatever mathematical theory they become acquainted with.
Mathematicians see to it that every physical theory, sooner or later, is
freed from all shackles of reality and liberated to fly in the thin air or
pure reason.
Computer science, in a very different mode, turns to mathematics in much
the same way that engineering always has. It freely borrows from already-
existing mathematics, developed for altogether different purposes or, more
likely, for no purpose at all. Computer scientists raid the coffers of
mathematical logic, probability, statistics, the theory of algorithms, and
even geometry.
Far from resenting the raid, each of these disciplines is buoyed by the
incursion. Statistics will never be the same given what the processing of
large samples by supercomputers has made possible. The Monte Carlo method,
without which computer simulations of neutron diffusion would have been
impossible, was developed by Ulam and myself without any knowledge of
statistics; to this day, the theoretical statistician is unable to give a
proper foundation to the method.
In a similar way, the theory of algorithms would amount to very little
without the needs of computer software. The rebirth of Euclidian geometry in
the most classical vein can be traced to the requirements of computer
graphics.
Like any other engineer, the computer scientist does not stop to work on
whatever mathematics he or she may need. Rather, a segment of the
mathematical population, relabeling itself "theoretical computer scientists,"
meets the mathematical needs of the other computer scientists. This shift, if
nothing else, has been financially beneficial.
SERIES AND PARALLEL CONNECTIONS
Two branches of mathematics have been wholly revamped, indeed given a new
lease on life by being required to meet the needs of computer science.
Mathematical logic is one. The other is the once-obscure chapter of
probability theory, now called "reliability theory." The beginning of this
transconfiguration may be traced to a master's thesis written by Claude
Shannon at MIT in 1939. A brief summary of his principal idea will illustrate
my point.
Computers are made up of circuits consisting of large numbers of replicas
of identically behaving units. Once upon a time, the units were vacuum tubes;
later, they were transistors; today, they are chips.
Every chip processes electric signals which enter at one point and exit at
another. Signals going through various chips can be connected in essentially
two ways: in series or in parallel.
Two chips A and B are said to be connected in series when the exit point of
A is soldered to the entrance point of B, so that a signal entering through
the entrance point of A will automatically be routed through B, and finally
exit through the exit point of B. On the other hand, chips A and B are said
to be connected in parallel when the entrance points of A and B are soldered
together, as well as the exit points of A and B. In this way, a signal
entering at the joint entering point of two chips connected in parallel has a
choice of whether to go through A or through B before exiting at the common
exit point.
Shannon's fundamental insight was that series and parallel connection of
chips are analogous to the connectives "and" and "or" of mathematical logic.
Indeed, when A and B are connected in series, the resulting circuit will send
a signal through if and only if both A and B are processing the signal. When
A and B are connected in parallel, the resulting circuit will send a signal
through if and only if either A or B is processing the signal, not
necessarily both.
By this analogy, any logical expression involving "and" and "or" (as well
as the third essential logic connective, "not," covered by a rather ingenious
trick) can be replicated by circuits. Simple as Shannon's observation was, it
ushered in the age of computing. The design of expert systems in our day
further exploits the basic idea that circuits can be made to perform logical
operation, for example, by developing circuit-theoretic devices that render
the Fregean quantifiers "for all" and "there exists."
RELIABILITY THEORY
Shannon's idea of relating series and parallel connection with the two
basic connectives of logic was to bear fruit in a direction that has proved
central to computer engineering. In the logical interpretation of electric
circuits, truth and falsehood correspond to whether or not a chip processes a
signal. A more realistic assumption, however, is that the chip will work or
not with a certain probability, depending on several factors, including the
age of the chip.
A realistic model for this situation is to assign to each chip in a circuit
an exponentially distributed random variable. Random variables corresponding
to distinct chips can be assumed to be independent. Thus motivated,
probabilists were led to develop a remarkable calculus, which is now known as
"reliability theory."
The principles of reliability theory are simple. If chip A has probability
p of failure and chip B has probability q of failure (we disregard the
possibility of these probabilities varying with time), then the probability
that the series connection of A and B will fail is 1 - (1 - p)(1 - q), and
the probability that their parallel connection will fail is pq.
When p and q are restricted to the extreme values 0 or 1, one finds, as a
limiting case, Shannon's interpretation of the logical connectives. Any
series-parallel circuit has a certain probability of working, which can be
computed by iterating the above two rules. Such a probability is called the
reliability of the circuit.
Reliability theory is concerned with the design of circuits of high
reliability at a minimum cost. No computer circuit can be designed without
allowing for the possibility that one or more components may fail (what von
Neumann was the first to call the "synthesis of reliable circuits from
unreliable components").
Soldering two or more chips in parallel will increase the reliability,
since a signal will still go through even if one or the other fails. If chips
cost nothing, we could achieve perfect reliability by soldering together in
multiple parallel connections. In practice, however, the costs of such a
design would be prohibitive.
Soldering chips in series decreases the cost of the circuit, but it also
decreases the reliability. In computer design, the engineer is forced to fall
back on his or her own wits (or on those of mathematicians) to design (or
"synthesize") circuits of high reliability at a minimum cost.
The design of complex systems of high reliability -- whether airplane
wings, telephone networks or computers -- is a daunting task. It is
unquestionably the central issue of today's computer science. Some of the
most ingenious mathematics of our day is being developed in response to the
needs of reliability theory.
COMPUTATIONAL IMPLEMENTATION
Although the basic rules for the computation of reliability were long
known, it took several years during and immediately after World War II for
the importance of the concept of reliability to be explicitly recognized and
dealt with. Only then did reliability computation become an essential feature
in computer design.
The late Richard Feynman was one of the first to realize the centrality of
reliability considerations in all applied scientific work. In the early days
of the Manhattan Project in Los Alamos (in 1943 and early 1944), he tested
the reliability of his first program in a dramatic fashion, setting up a day-
long contest between human operators working with hand-operated calculators
and the first electromechanical IBM machines.
At first, human operators showed an advantage over the electromechanical
computers; as time wore on, however, the women who worked with the
calculators became visibly tired and began to make small errors. Feynman's
program on the electromechanical machine kept working. The electromechanical
computers won out by virtue of their reliability.
Feynman soon came to realize that reliable machines in perfect working
order were far more useful than much of what passed for theoretical work in
physics, and he loudly stated that conviction. His supervisor, Hans Bethe --
the head of T-Division (T for theory) at the time and a physicist steeped in
theory -- at first paid no attention to him.
At the beginning of the Manhattan project, only about a dozen or so hand-
operated machines were available in Los Alamos; they regularly broke down,
thereby slowing scientific work. In order to convince Bethe of the importance
of reliable computation, Feynman recruited me to help him improve the
performance of the hand-operated desk calculators, avoiding the week-long
delays in shipping them to San Diego for repairs. We spent hours fixing the
small wheels until they were in perfect order. Bethe, visibly concerned when
he learned that we had taken time off from our physics research to do these
repairs, finally saw that having the desk calculators in good working order
was as essential to the Manhattan Project as the fundamental physics.
Throughout his career, Feynman kept returning to the problem of the
synthesis of reliable computers. Toward the end of his life, he gave a
remarkable address at the 40th anniversary of the Los Alamos Laboratory where
he sketched a reliability theory based on thermodynamical analogies.
In contrast to Bethe, John von Neumann very quickly realized the importance
of reliability in the design of computers. It is no exaggeration to say that
von Neumann had some familiarity (in the 1950s) with all the major ideas that
have since proved crucial in the development of supercomputers.
Von Neumann realized very early the advantage of parallel computation over
series computation. He knew that the day would come when series computations
would reach their physical limit, namely, the velocity of light, and that
only a computer based on the principles of parallel computation could exceed
that limit.
Curiously, however, his choice of series computation in preference to
parallel computation (now referred to as the "von Neumann computer") was the
result of his negative experiences with the first experiments he devised to
test the effectiveness of parallel computation. Repeatedly frustrated by his
inability to achieve the required synchronicity in a simple parallel
computation experiment that he set up (an impossible task in his time), the
failure kept him at a distance from all ideas of parallelism for the rest of
his life.
THE ENIAC
The first large-scale electronic computer to be built, the one that may be
said to inaugurate the computer age, was the ENIAC. It was built at the Moore
School of the University of Pennsylvania by an engineer and a physicist --
Presper Eckert and John Mauchly. Their idea, trivial by the standards of our
day, was a revolutionary development when completed in 1945.
At the time, all electromechanical calculators were built exclusively to
perform ordinary arithmetic operations. Any computational scheme involving
several operations in a series or in parallel had to be planned separately by
the user. Mauchly realized that if a computer could count, then it could do
finite-difference schemes for the approximate solution of differential
equations. It occurred to him that such schemes might be implemented directly
on an electronic computer, an unheard of idea at the time.
They managed to sell their idea to the U.S. Army, which authorized funding
of the project, on the condition that the machine be used at the Aberdeen
Proving Grounds for ballistic computations. A Capt. H. Goldstine was chosen
by the Army to supervise the project and was to benefit greatly from the
interaction with Eckert and Mauchly.
Alone among the large computers of the time, the ENIAC was designed with
paramount concern for reliability. It consisted of 18,000 vacuum tubes wired
together, with full allowance made for redundancies that would increase
reliability. Most of the maintenance work involved the replacement of vacuum
tubes that went out of order.
To many observers unfamiliar with reliability computations, it seemed a
miracle that the ENIAC worked at all. Enrico Fermi, who later was to become
one of the first physicists to perform large computer experiments, made only
one incorrect prediction so far as I know: He mistakenly computed the
reliability of the ENIAC on the basis of the mean free time between vacuum
tube failures; he announced that the machine could never work, scarcely
realizing that the ENIAC was far more reliable than the counting apparatus in
his lab.
In spite of all predictions to the contrary, the computer worked for
periods of several hours without error. The designers of the computer
resorted to all manner of precautions to keep the vacuum tubes from failing,
including keeping "heaters" on at all times.
I remember distinctly the time when the ENIAC was dismantled and packed for
transportation to the Aberdeen Proving Grounds. Each of the wires was
carefully marked and then clipped; I never believed that Mauchly and Eckert
would be able to put it back together again. They did, and the ENIAC proved
to be a great success.
At the time the ENIAC was installed, von Neumann was a consultant at the
Aberdeen Proving Grounds. Realizing that the ENIAC was being underused, he
proposed that it be put to work on a computation that would simulate a one-
dimensional thermonuclear explosion, following on the notions of Edward
Teller's group at Los Alamos. The computation was finally made, and the ENIAC
came through with flying colors. The experiment came to be known as the
"shakedown cruise" of the ENIAC.
THE MANIAC AND UNIVAC
At the end of the war, von Neumann and I began to plan the building of a
more powerful computer in Los Alamos, which would benefit from the experience
of the ENIAC and the reliability lessons that it had taught us. I spent a
year the Institute of Advanced Study in Princeton to discuss detailed plans
with von Neumann. Edward Teller, who was then beginning to do his
calculations on thermonuclear reactions, enthusiastically encouraged us to go
ahead with the project.
The MANIAC took several years to build. It was finally operational in 1952,
and a more realistic computation of a thermonuclear reaction was finally
tried on it, with great success.
Of all the oddly named computers, the MANIAC's name turned out to be the
most unfortunate: George Gamow was instrumental in rendering this and other
computer names ridiculous when he dubbed the MANIAC "Metropolis And von
Neumann Install Awful Computer."
Fermi and Teller were the first hackers. Teller would spend his weekends at
the laboratory playing with the machine. Fermi insisted on doing all the
menial work himself, down to the least details, to the awed amazement of the
professional programmers. He instinctively knew the right physical problems
that the MANIAC could successfully handle.
His greatest success was the discovery of the strange behavior of nonlinear
systems arising from coupled nonlinear oscillators. The MANIAC was a large
enough machine to allow the programming of potentials with cubic and even
quartic terms. Together with John Pasta and Stanislaw Ulam, he programmed the
evolution of a mechanical system consisting of a large number of such coupled
oscillators. His idea was to investigate the time required for the system to
reach a steady state of equidistribution of energy. By accident one day, they
let the program run long after the steady state had been reached. When they
realized their oversight and came back to the computer room, they noticed
that the system, after remaining in the steady state for a while, had then
departed from it, and reverted to the initial distribution of energy (to
within 2 percent).
The results were published in what was to be the last paper Fermi published
before he died. Fermi believed this computer-simulated discovery to be his
greatest contribution to science. It is certainly the first major scientific
discovery made by computer, and it is not fully understood to this day
(though it has spawned some beautiful ideas).
In the same year that the MANIAC was inaugurated, 1952, the first public
demonstration of computer reliability was instrumental in convincing the
public of the importance of computers. Howard K. Smith employed the UNIVAC on
television to predict the outcome of the presidential election. Shortly after
the polls closed (within half an hour, actually), the UNIVAC predicted an
Eisenhower landslide. The programmers' disbelief that immediately followed
this prediction and their subsequent retraction made the computer's
prediction all the more astounding. The rise of computer science can be
traced to that day.
The history of computer science since 1952 is far more complex. The
underlying mathematical and engineering ideas were already known at that time
and have since varied only in detail. The gap between these ideas and their
implementation, however, was to grow wider as the demand for speed and
reliability increased. In fact, the discontinuous leaps forward in computer
design went hand in hand with advances in chemistry and material science. The
discovery of the transistor, and later the introduction of the miraculous
chip, are the two main stages that mark turning points in computer science.
It is my hope that a historian of computing will some day tell the
fascinating stories of these inventions.
-----------------------------------------------------------------------------
Nick Metropolis is Senior Fellow Emeritus at Los Alamos National
Laboratory.
-----------------------------------------------------------------------------
Reprinted by permission of Daedalus, Journal of the American Academy of
Arts and Sciences, from the issue entitled, "A New Era in Computation,"
Winter 1992, Vol. 121, No. 1. Copies may be obtained for $6.95 plus $2.00
shipping and handling ($3.00 for surface mail) from the Daedalus Business
Office, 136 Irving St., Suite 100, Cambridge, MA 02138. As of October 1,
1992, the price for a single copy increases to $7.95.
-----------------------------------------------------------------------------
Your comments on this and other Daily News items are welcome in Your
Feedback on the News. For information on how to add your voice to the
discussion, see About News Feedback under Your Feedback on the News.
Copyright 1992 by HPCwire. All rights reserved.