ill

mm

ISSS
 Loughborough
College of Technology
Library
 The Computer Museum
History Center
library

<•
« OF TECHNOLOGY LIBRARY

000 2666 02
 CAMBRIDGE ENGINEERING SERIES
GENERAL EDITOR:
J. F. BAKER, F.R.S.

AN INTRODUCTION TO AUTOMATIC
DIGITAL COMPUTERS
 Digitized by the Internet Archive
in 2012 with funding from
Gordon Bell

http://archive.org/details/introductiontoau20live
 AN INTRODUCTION TO
AUTOMATIC DIGITAL
COMPUTERS
BY

R. K. LIVESLEY
Lecturer in the Department of Engineering,
University of Cambridge

V
<gll ill
WrSS
w yyv

KhfcS y flS'
rt
y

CAMBRIDGE
AT THE UNIVERSITY PRESS
1957
 PUBLISHED BY
THE SYNDICS OF THE UNIVERSITY PRESS
Bentley House, 200 Euston Road, London, N.W. 1
American Branch: 32 East 57th Street, New York 22, N.Y.

*•*« ioV\2z J** 1 - 4 * **

*QZ*W/oz

Printed in Great Britain at the University Press, Cambridge

(Brooke Crutchley, University Printer)
 1

CONTENTS
List of Plates page vi

Preface vii

Chapter 1. The Elements of Programming

1.1 Introduction 1

1.2 A simple calculation and its mechanization 2

1.3 A simple automatic computer 5

1.4 The control unit: jump instructions 9
1 .5 Modification of instructions 1

1.6 Instruction codes in real machines 13

Chapter 2. Input, Storage and Output of Numbers

2.1 Decimal and binary notation 15
2.2 The sign and size of numbers 17
2.3 Binary storage systems 18
2.4 Input-output equipment 23

Chapter 3. The Organization of Programmes

3.1 Dividing a programme into routines 27
3.2 Library routines 29
3.3 The organization of sub -routines 32
3.4 The development of a programme 34
3.5 Testing programmes 35
3.6 Automatic programming 37

Chapter 4. The Solution of Engineering Problems

4.1 Single problems 39
4.2 Repetitive problems 40
4.3 Factors in the design of standard programmes 40
4.4 Problems which have been solved on computers 43
4.5 Digital computers and mathematical methods 49
4.6 The human factor 50
4.7 Future prospects 51

References for further reading 53

v
 LIST OF PLATES
{Between pages 24 and 25)

I (a) Binary digits displayed on a cathode ray tube

(b) Part of a magnetic core storage system

II The first Manchester computer, built in 1948

III An English Electric DEUCE computer, built

in 1955

IV {a) The magnetic drum of a DEUCE computer

(b) Punched cards and punched paper tape

ACKNOWLEDGEMENTS
Plates and II are reproduced by courtesy of Ferranti Ltd.,
1(a)
Manchester; Plate 1(6) by courtesy of the Director of the Mathematical
Laboratory, Cambridge University; Plates III and IV (a) by courtesy
of the English Electric Co. Ltd., Stafford; and Plate IV(6) by courtesy
of the British Tabulating Machine Co. Ltd.
 PREFACE
This brief account of digital computers has been written mainly for
engineers and others who are faced with tedious calculations, so that
they can judge the possibility of using these machines in their own
numerical work. It describes in general terms what digital com-
puters can do and how they are made to do it, and also gives an
account of some of the calculations for which they have so far been
used. The emphasis is on the application of machines to routine
computing work rather than to research.
Since the book is merely an introduction and is not intended for
the specialist, I havemade no attempt to discuss the design of com-
puting circuits, nor have I tried to make the reader expert in the
operation of any particular machine. The reader who wishes to take
the subject further will find a more detailed treatment, together with
a comprehensive bibliography, in Dr M. V. Wilkes's book Automatic
Digital Computers (Methuen, 1956), while most computer laboratories
issue training manuals for those who wish to become skilled users of

their machines. When referring to existing computers I have used

the names or initials by which they are generally known. In many
cases the full title from which the initials were originally taken has
almost been forgotten.
I have avoided as far as possible any detailed discussion of mathe-
matical techniques. It is often assumed that a digital computer can
only be used effectively by those with very advanced mathematical
training, but this is largely because many of the problems so far
solved on computers have been mathematically complex. The pre-
paration of a normal engineering calculation for a digital computer
certainly requires care and precise thinking, but it does not usually
demand great mathematical skill.

Although I have tried to avoid giving undue prominence to the

idiosyncrasies of particular machines, the discussion of engineering
problems in Chapter 4 is largely based on my knowledge of work
carried out by members of the group using the Manchester University
vii
computer during the period 1951-6. I feel, however, that the con-
clusions I have drawn are valid for other machines as well.

The book is based on a course of lectures first given at Manchester

in the autumn of 1954 to post-graduate apprentices of the Metro-
politan-Vickers Electrical Co. Ltd. I am particularly indebted to
Professor D. R. Hartree for his comments and suggestions during the
preparation of the lectures for publication.

R. K. LIVESLEY
Cambridge, 1956
 Chapter 1

THE ELEMENTS OF PROGRAMMING

1.1. Introduction
The complex calculating machines which have become increasingly
well known during the last ten years may be classified as either
'analogue computers' or 'digital computers'.
In an analogue computer, numbers are represented by continuously
variable quantities such as lengths, angles or voltages. The electrical
or mechanical connexions of the computer are set up in such a way
that these quantities satisfy the same mathematical relationships as
govern the variables in the problem to be solved. The behaviour of
the computer is thus made directly comparable to the behaviour of

the physical or mathematical system under consideration.

In a digital computer, on the other hand, numbers are represented
by sequences of digits. Although many modern machines do not in
fact work in the decimal scale, it is often convenient to think of the
numbers which they handle as being composed of decimal digits,
stored in registers basically similar to the counting wheels of a desk
calculating machine. Since a digital computer can only store rational
numbers, the mathematical operations which it can perform are
merely those of ordinary arithmetic.
This book is concerned only with digital computers, and for the
sake of brevity they will generally be referred to merely as 'com-
puters'. Although these machines are sometimes called 'electronic
brains', their electronic nature is in one sense purely incidental.
Their importance lies far more in the fact that they are automatic,
being able to carry out lengthy sequences of numerical operations
without human intervention. Since even the most complex calcula-
tions can always be broken down into a series of simple arithmetical
steps, digital computers can be applied to many problems unsuitable
for analogue solution.
The technique of making a computer carry out a particular calcula-
tion is known as 'programming'. This involves first breaking the cal-
culation down into a sequence of arithmetical operations, and then
preparing a series of instructions which cause the computer to carry
' '

out the required operations in the correct order. Each instruction

specifies a particular step such as multiplication or addition, and
it is the programmer's job to see that the instructions are couched in
a language or code which the machine can interpret.
'
' A whole set
of coded instructions, arranged in the correct order, is known as a
'programme'. As there are almost as many instruction codes as
there are machines, a programme made up for one computer may
not be intelligible to another. It may, however, be used repeatedly
on the machine for which it was designed, carrying out any number
of similar calculations without the need for further programming.
A simple example of a computer with a permanently built-in
programme is a desk calculating machine equipped with automatic
division. Once the appropriate button has been pressed a series of
subtractions, sign tests and shifts occur without further human inter-
vention. Such an instrument, however, is like a musical box which
will only play a single tune. The type of computer described in this
book is more like a pianola, which can play any piece of music
provided that it is given a set of coded instructions in the form of
'
'

a suitably punched roll of paper.

It would be out of place in a book of this kind to give a detailed
account of the organization and instruction code of a real computer.
It is difficult, however, to give an idea of what programming involves
without taking particular examples. The remainder of this chapter
will be devoted, therefore, to the programming of some numerical
examples for an idealized computer with a very simple instruction
code.

1.2. A simple calculation and its mechanization

While there are several different ways of solving sets of simul-
taneous linear algebraic equations, one of the simplest methods is

that of successive approximation. Although this technique is only

suitable for certain classes of equations, it is very easy to programme
for a computer, and is chosen here for that reason.
The equations
I0x + 6y-z=19,\
3x + 9y + 2z = 27,l (1.1)

x-fy + 8z=llj
for instance, may be written in the form

ar=(19-6y + 2)/10, (1.2a)

y=(27-3a;-2z)/9, (1.26)

z = (ll-x + 4y)/$. (1.2c)

If an initial ' trial solution ' is chosen, such as x — y — z = \, equation
(1.2a) gives a 'corrected' value for x,

z = (19-6 + l)/10=l-4,

which may be used in equation (1.26) to give a new value of y,

*/=(27-3xl-4-2)/9 = 2-311.

The 'corrected' values of x and y may then be substituted in the

third equation (1.2 c) to give

z = (17- 1-4 + 4 x2-311)/8 = 3-106,

and the whole process repeated with the new values of x, y and z.

It is obvious that if the successive values of the variables converge

at all, they will converge to the correct solution of the original
equations. In fact, successive iterations give

Initial 1st
values iteration 2nd 3rd 4th
x 1 1-400 0-824 0-983 1-002
y 1 2-311 2035 1-997 1-999
z 1 3-106 3040 3-002 2-999

The results in the last column are in very close agreement with the
exact solution, which is x—\ 3 y = 2, z—Z.
In the case of a general set of three equations

a 1 x + b 1 y + c 1 z = d 1 ,\
a 2 x + b 2 y + c 2 z = d2 \ (1.3)

a 3 x + b 3 y + c 3 z = d 3 ,J

the equations for the nth iteration are

x n+l = (d 1 -b 1 y n -C1 Z n )ia 1 ,

yn+i = (d 2 -a 2 x n+1 -c 2 z n )/b 2 , I (1.4)

Zn+i = (d 3 -a.3 x n+1 -b 3 y n+1 )lc 3 .)

Alternatively, the process can be thought of as the calculation of

a series of successive corrections to the original values of x, y and z.

Thus equations (1.4) may be written

Xn+i-x n = (d 1 -a 1 x n -b 1 y n -c 1 z n )la 1 ,

Vn + x-yn = {d 2 -a 2 x n ^-b 2 y n -c 2 z n )\b 2 ,

-
(1.5)

z n+l~ z n = \d^~ a Z X n+l~ b z y n+1 — C z Z n )

jC Zi )
 :

and it will be noted that the numerators on the right-hand sides of

equations (1.5) are merely the differences between the right- and
left-hand sides of the original equations (1.3).

In order to carry out this calculation for a given set of coefficients,

using equations (1.5), a human being would need a pencil and paper
for writing down numbers, some kind of mental or mechanical
apparatus for doing arithmetic, and a written or memorized scheme
of analysis to follow. This scheme might be
{a) Form (d 1 -a 1 x n -b 1 y n -c 1 z n )/a 1 .

(b) Add this to x n to obtain x n+v

(c) Form {d 2 -a 2 x n+1 -b 2 y n -c 2 z n )jb 2 .

(d) Add this to y n to obtain yn+1 .

(e) Form {d z -a z x n+1 -b z y n+1 -c z z n )jc z .

(/) Add this to z n to obtain z n+1 .

(g) Go back to (a), testing for convergence on each occasion after

the first.

It is apparent that a human being given such a 'programme' of

instructions and a set of initial values x , y z could carry out the
,

calculation purely mechanically, even if he had never heard of

simultaneous equations.
In any pencil-and-paper calculation of this sort, the human com-
puter is involved in three kinds of activity, namely,
(1) Transferring numbers between his paper and his arithmetic
machine,
(2) Actually doing arithmetic,
(3) Looking at his list of instructions to see what to do next.
Of these three, it is probable that in many practical cases (2) con-
sumes even when carried out with the simplest type of
least time,
desk calculating machine. Therefore, even if a high-speed machine

which could add or multiply in a millisecond were available its use

would not greatly decrease the total time taken by a calculation.
A significant saving requires a reduction in the time spent on
activities (1) and (3) as well, and this can only be achieved if the
whole process is mechanized.
Now consider the features of a machine which could replace the
human computer entirely. It would need the following elements
(i) A 'store', corresponding to the human computer's sheet of

paper, in which numbers could be held.

(ii) An 'arithmetic unit' capable of addition, multiplication, etc.,
together with facilities for transferring numbers between it and the
store.
 Means for feeding numbers into the store and displaying results.
(iii)

A unit which would control the machine according to a list of

(iv)
'
instructions supplied by a human operator.
'

Such a machine would have little advantage over a human being

if only a single calculation had to be carried out, as it would probably

take as long to prepare the instructions as to do the calculation by

hand. However, if the instructions were made independent of the
actual numbers appearing in the machine, they could be used
repeatedly for carrying out similar calculations with different sets
of data, the speed depending only on the rate at which the computer
took in and obeyed the instructions.
These ideas form the basis of all automatic digital computers, and
were first put forward by Charles Babbage, long before the days of

electronics. Babbage envisaged an 'analytical engine' which was to

be controlled by holes punched in cards, just as Jacquard had used
punched cards to control the weaving of cloth. The specification of
this purely mechanical computer contained most of the logical
features found in present-day electronic machines, but although
parts of it were built it was never completed.

1.3. A simple automatic computer

Although the programme of instructions for solving a set of three
'
'

simultaneous equations which was developed in the previous section

may appear quite simple to a human being, it could only be fed into
a machine which understood both algebra and the English language.
To make an automatic computer a practical possibility it is necessary
to devise a much simpler form for the instructions. To fix ideas,
a simple computer using a numerical instruction code will now be
described. Although this computer is an imaginary one, it could in
fact be built without great difficulty.
The store of the machine consists of a hundred registers which for
convenience are numbered 1-100. The number by which a particular
register is known is number
called its 'address', while the decimal
stored in it is Thus the content of the
referred to as its 'content'.
register whose address is 59 might be the decimal number 3-14159.
The computer is equipped with an arithmetic unit similar to a desk
calculating machine, which can carry out addition, subtraction,
multiplication and two special registers
division. This unit contains
known and the accumulator. The computer
as the multiplier register
is able to read decimal numbers into its store from punched paper

tape, and to print numbers on a typewriter. It has two tape-readers,

5
one for numerical data and one for instructions. The arrangement is

shown diagrammatically in fig. 1.

When operating, the machine reads an instruction from the appro-

priate tape, obeys it, and then proceeds to the next instruction on the
tape. The instructions consist of pairs of numbers, each pair con-
sisting of a 'function number' / and an address s. The function

n Data Storage Multiplier

register Arithmetic
reader
unit
registers <*- Accumulator

1-100

Number
tape
Printer

100-

-]
n
Instruction

reader

¥
Instruction

tape

Paths taken by numbers » - Instructions — — *• — —

Fig. 1 . Arrangement of a simple automatic computer.

number specifies the type of operation to be carried out, while the

address specifies the register whose content is to be operated on.
The functions described in the following table will be sufficient for
a considerable proportion of numerical work. The symbol S indicates
the content of the register s referred to by the address section of the
instruction, M and A stand for the numbers stored in the multiplier
register and the accumulator respectively, while a dash indicates the
state of a register after an instruction has been obeyed. The contents
of registers not specifically mentioned remain unchanged, while the
number previously stored in a register is obliterated when a new
number is transferred to it.
Function
number Effect
1 Read one number into storage register s from the
number tape. {S' = tape entry number tape moves on.)
;

2 Print the content of register s.

3 Stop. (The address part of this instruction is irrelevant.)

4 Copy the content of register s into the accumulator.
(A' = S.)
5 Add to the content of the accumulator. (A' = A + S.)
6 Subtract from the content of the accumulator.
{A'=A-S.)
7 Copy the content of register s into the multiplier
register. (M' = S.)
8 Multiply the content of the multiplier register by the
content of the register s, and add the result to the
content of the accumulator. (A' = A + MS.)
9 As for function 8, but subtract the product from the
number in the accumulator. {A' — A — MS.)
10 Copy the number in the accumulator into register s
and clear the accumulator. (S = A, A' = Q.) /

11 Clear the accumulator. (A' = 0; the address part of

this instruction is irrelevant.)
12 Divide the content of register s by the number in the
multiplier register and place the result in the accumu-
lator. (A' = S/M the divisor is placed in the multiplier
;

register merely for convenience.)

With this code, a sequence for dividing the number in register 72 by the
content of register 49 and planting the result in register 86 would be
Function/ Address s Effect
7 49 The content of register 49 is copied into
the multiplier register.
12 72 The content of register 72 is divided by
the number planted in the multiplier
register by the previous instruction. The
result appears in the accumulator, the
previous content of which is obliterated.
10 86 The resultcopied into register 86 and
is

the accumulator cleared ready for the

next step.
It should be noted that these instructions do not alter the numbers
stored in registers 49 and 72.

7
 :

The instructions for carrying out the process described by equations

(1.5) can now be written entirely as a sequence of numbers. If the
twelve coefficients a v a 2 a 3 b v b 2 b 3 c ly c 2 c 3 d 1 ,d 2 d 3 are punched
, ; , ; , ; ,

in that order on the data tape, and are followed by the trial solution
x , y , z , the programme for a single iteration might be as follows

1.1 1
1.2
Beading instructions: a v ..a 3 placed in registers 1...3,
b 1 ...b 3 in 4. ..6, c x ...c z in 7. ..9, d x ...d 3 in 10. ..12, and x
y , z in 13, 14, 15.

1.15
4.10 A' = dt .

7.13 M' = x .

9.1
7.14 M' = y Q .

9.4 A = d -a 1 xQ -b
,
1 x y .

M
(i

7.15 = zQ f
.

9.7 A = d 1 -a
/
1
x -b 1 y -c 1 z .

10.16 Result copied into spare register and accumulator cleared.

7.1 M' = a x (ready for division).
12.16 A' — (d 1 — a 1 x — b l y — c 1 z )/a 1 = x 1 — x from equation (1.5).
5.13 A =(x 1 — Xq)-\-Xq-=x 1 .

10.13 New value of x copied into register 1 3 and accumulator

cleared.
2.13 New value of x printed.
4.11^1
7.13
9.2
7.14
9.5
Similar sequences for calculating yv

10.14
2.14
4.12
7.13
9.3
7.14
9.6
Similar sequences for calculating z v

10.15
2.15^
 : '

The reader will be well advised to work through this sequence in detail.
Alternatives are possible, some of which might require fewer instruc-
tions. The use of storage registers 1-16 is a purely arbitrary choice.

1.4. The control unit: jump instructions

Machines controlled directly by punched cards or paper tape are
quite feasible, and in fact have been produced commercially. They
suffer from two drawbacks, however, which have led many computer
designers to adopt a different system of control.
In the first place the speed of a machine which reads its instructions
direct from punched tape or cards is limited by the mechanical nature
of the input device. This will rarely work at more than 50 instructions
per second, while electronic circuits can easily be made to operate
at more than 1000 instructions per second. In the second place, such
a computer cannot easily go back to repeat a series of instructions,
nor can it use some numerical criterion to select one of two alternative
courses of action unless it has several instruction readers. It is found
in practice that facilities for repeating and selecting certain blocks of
instructions are extremely useful in designing programmes.
Since instructions can be given the form of numbers, there is no
reason why they should not all be placed in the main store of a com-
puter, together with the numerical data, before a calculation is actually
commenced. If this is done, it becomes necessary to add to the
machine described in § 1.3 a new unit which will be called the control '

unit, the function of which is to ensure that the instructions are

selected from the store and obeyed by the computer in the correct
order. This unit contains a register, known as the 'control register',
which stores a single number C specifying the address of the instruc-
tion being obeyed. It is arranged that when the operation defined by
an instruction is completed, the control number C is increased by
one, so that the machine passes on to the next instruction. In
symbolic terms, each instruction produces the effect C' = C+1.
This method of controlling the computer does not, of itself, increase
the overall speed of a calculation, since the instructions presumably
have to be fed into the machine from cards or tape in the first place.
It is now possible, however, to introduce instructions which alter
the content of the control register, thus causing the machine to jump
from one point in the programme to another. A simple type of jump
instruction is defined as follows
Function number Effect
13 Set the content of the control register equal to s. (C' = s.)
9
 :

It will be remembered that after a 'normal' instruction has been

number C is increased by one, and the next
carried out, the control
instruction taken from register 0+1. The effect of the instruction
13.5 is to cause the next instruction to be taken instead from

register s. As an example, the following sequence calculates

from the expansion

—
l
— = \+ x + x + x +
—x
2 z
...,

where the number 1 is assumed to be available in register 18. The

number x (< 1) is initially stored in register 19, and this register is
subsequently used to accumulate the terms of the series.

Address of
instruction Instruction Effect
50 7.19 M' = x.
51 4.18 A' = l.
52 5.19 A' - l + x.
53 ^10.19 On the first cycle, (1 + x) is placed in
register 19.
54 4.18 A'=\.
55 8.19 On the first cycle, A'=\ + x(\ + x).
56 —13.53 Return to the instruction in register
53 (C" = 53).

The reader who follows through this sequence will soon see the power
of a few instructions when they form part of a repeated cycle. The
second time the instruction in register 53 is reached, its effect is to
plant 1 The next cycle calculates 1 + x + x 2 + x?
+ x + x2 in register 19. y

and so on. A simpler method would be to calculate the individual

terms x, x 2 x z etc., and sum them, but this would require a few more
, ,

instructions.
The instruction introduced above can be used to jump forwards
as well as back, and its effect can be compared with the jumps which
occur in the game of 'Snakes and Ladders'. There is still no choice
of next instruction, however, and it is clear that once a loop of
instructions has been entered it can never be quitted. A more
powerful type of instruction is the 'conditional' jump, whose opera-
tion depends on the sign of a certain number, usually the content of
the accumulator. A typical example is defined as follows

Function
number Effect
14 Set the content of the control register to s only if A
is positive or zero. (C = s if ^ 0, = C + 1 if A C A < 0.)
10
In other words, the instruction is similar to the previous one, but
the jump only occurs if the number currently standing in the accumu-
lator is positive or zero ; otherwise the control number merely increases
by one and the next instruction is selected in the usual way. A simple
sequence for replacing the number x in register 19 by its modulus
would be

Address of
instruction Instruction Effect
50 4.19 A'— x.
'

51 —14.54 Jump to instruction 54 if x^

52 If 6.19\ Subtract 2x from A, i.e.

53 x^O. 6.19J A' — —x.

54 L>10.19 Place \x\ in register 19.
1

The conditional jump instruction provides a simple means of

repeating a loop of instructions a certain number of times, since
a count number can be set before the loop is entered, and reduced by
one each time it is traversed. If the number is set equal to n— 1,
after n cycles it have been reduced to —1, the conditional
will
instruction will cease to cause a jump, and the machine will com-
mence the next part of the calculation.
The reader should now be able to complete the programme for
solving a set of three simultaneous equations commenced in §1.3.
A suitable test for convergence might be to form the sum of the
squares of the corrections to x, y and z, and to test whether or not
this exceeded some fixed number e. A jump instruction conditional
on the sign of
{x n+1 - x n f + {y n+1 - y n f + (z n+1 - z n f - e
could be followed by a 'stop' instruction, so that the machine would
halt automatically when a sufficiently accurate solution had been
found.

1.5. Modification of instructions

In many calculations it is necessary to carry out the same numerical
operations on a series of different numbers. In a computer programme
this implies several groups of instructions which, while similar in
general pattern, refer to different addresses in the store. In the
programme constructed in §1.3, for instance, the instructions for
calculating yt and z 1 are identical with those for calculating x x except ,

that the addresses of the coefficients alt b lt %, d x are replaced by those

11
of a 2 b 2 c 2 d 2 etc. In the
, , , , same way, in the calculation of x x , the
sequence for forming d 1 — a^ — b x y — cx 2; contains a pair of instruc-
tions repeated three times, but referring to different addresses each
time.
It is clear that if the instructions for calculating x 1 were followed
by a sequence which added appropriate numbers to the addresses in
certain instructions, they could then be used again to calculate y v
On the second cycle the modifying sequence would alter the addresses
again and the instructions would cause the computer to calculate z v
In this way considerable savings can be made in the number of
instructions required for a given calculation. To sum the contents of
fifty consecutive storage locations, for instance, it is only necessary to
write down one addition instruction, and arrange for the address in
it to be increased by one each time it is used. A conditional jump
instruction can be used to count the number of additions in the
manner described in §1.4.
In the simple computer which has been described, the task of
modifying an instruction and replacing it in its original register
requires three instructions, so that modifying sequences will tend to
be lengthy if many instructions have to be dealt with. There is also
the difficulty of resetting altered instructions if it is necessary to
repeat a complete calculation involving several smaller loops of
instructions. A system in use on several machines, including the
Ferranti computers and the EDSAC, makes such modifications much
simpler. In this system, the content of one of a number of subsidiary
registers, known as ^-registers, is added to each instruction before
it is obeyed. This is done automatically as the instruction is selected
by the control unit, a special part of the instruction specifying which
B- register is to be chosen. do not require
(Instructions which
modifying can be arranged to select a whose content is
I?-register
zero.) By this means instructions can be made to operate on a series
of different addresses merely by altering the content of the appro-
priate ^-register. In some computers it is also possible to modify
the function number in an instruction, but this is a less useful
facility.

B- registers are also useful in counting repetitions, as conditional

jump instructions can be introduced which only cause a jump when
the content of the appropriate ^-register is non-negative. Thus
a cycle for summing the numbers in the first fifty registers of the
simple computer described in this chapter might use the basic
instruction 5.1 (add the content of register 1 into the accumulator)

12
 :

and modify it by a 2?-register which is initially set to contain 49.

The cycle would be
(1) Clear the accumulator.
(2) Set the ^-register to 49.
(3) Obey instruction 5.1 (modified by the jB-register).

(4) Subtract one from the ^-register.

(5) Test the JB-register, and return to (3) if it is non-negative.
(6) Stop.
The first time the control unit selects the instruction 5.1, the
actual instruction obeyed is 5.50. The second time it is 5.49, and

so on. Afterit has been obeyed as 5.1, the content of the ^-register

becomes negative and the cycle comes to an end.

Modified cycles of this type are extremely valuable to the pro-
grammer. They are most effective in calculations which are arranged
in a systematic form, and for this reason the programmer naturally
chooses numerical methods which involve as much repetitive com-
puting as possible. The effect which the properties of a computer
have on the choice of suitable numerical methods is discussed further
in subsequent chapters.

1.6. Instruction codes in real machines

The simple computer described in this chapter possesses most of
the important features found in real machines. The instruction code
which has been used is an example of a 'one-address' code, since
each single instruction refers to at most one register in the store.
Some computers use a two-address code which allows addition
operations of the type A = S 1 + 8 2 to be carried out by a single
f

instruction, while a three-address code is also possible, allowing

operations such as S3 = S 1
/
-\- S 2 In general, each instruction in
. a two-
or three-address code doesmore work, but is more difficult to write
down. Some machines, such as the ACE and DEUCE, do not have
a control unit of the type described in this chapter, but use an
instruction code in which each instruction has a section which
specifies the address of its successor.
The design of a digital computer involves a balance between the
which the user would like and those which the engineer can
facilities

reasonably provide. In some cases an instruction operates in what

may seem to be an inconvenient way merely to simplify the circuitry.
Most of the early computers, for instance, had no division instruction,
using instead a sequence of instructions which carried out the process
either by repeated subtraction or by a series expansion. The situation

13
is very similar to that which occurs in desk calculating machines.
Facilities for automatic multiplication, division, square-root extrac-
tion, etc., are very pleasant for the user, but if they are absent the
operations can easily be built up from the simple processes of addition
and subtraction.
Most computers have a larger instruction code than the one
described in this chapter. It is often necessary to scale a number up
or down in order to make it fit into a register, so that an instruction
which produces a left or right shift of one or more places is often
useful. Many machines also have an 'alarm' instruction, which can
be used in place of a stop instruction to
'
' summon the operator when
a check fails or a calculation comes to an end. The reader will find
detailed accounts of the instruction codes used in actual machines in
the 'programmer's handbooks' issued by most computing machine
laboratories.

14
 Chapter 2

INPUT, STORAGE AND OUTPUT OF NUMBERS

In the previous chapter it was shown how numbers could be used as
'
instructions ' to control the behaviour of a simple computer. Little
was said, however, of the forms taken by instructions and numerical
data in the various electronic computers now in use. This chapter
discusses the problem of number storage in more detail, and describes
some of the methods used in real machines.

2.1. Decimal and binary notation

Most readers will be familiar with the decimal storage registers of
a desk calculating machine, or at least with the mileage indicator on
a car dashboard. A register of this type consists of a number of
counting wheels, each labelled with the digits 0-9 and coupled to its

neighbours by a suitable carrying mechanism. There is no theoretical

'
'

reason why such registers should not be used to store numbers and
instructions in an automatic calculating machine, and in fact Bab-
bage's engine was designed to use them, but it is clear that the speed
of any computer with mechanical elements is restricted by the inertia
of its moving parts. The Harvard Mark I, which was the
for instance,
first automatic computer actually to be built, had 72 mechanical

registers each with 23 decimal counting wheels. It had an addition

time of 0-3 sec, and a multiplication time of about 6 sec. Thus,
although it was a landmark in the development of automatic com-
puters, it was no faster in carrying out individual operations than an
much higher speeds achieved
ordinary electric desk calculator. The
by machines are largely due to the development of non-
later
mechanical forms of number storage.
A decimal digit can have ten possible values, so that it requires
a storage element which has ten alternative states. Although such
elements are used in mechanical calculators, and faster ones using
valves called dekatrons can be built for electronic operation, they
are inherently slower and more complicated than stores with only
two alternative states. A two -state device like a relay, for instance,
is much simpler than a multi-state uniselector. Similarly, in trans-
mitting numbers from one part of an electronic machine to another,

15
 :

it is much easier to convey information by the mere presence or

absence of pulses than to use the size of each pulse to distinguish one
often alternative digits, even though the former system requires more
pulses for a given amount of information. The simplest method of
feeding numbers into an automatic machine is by using holes punched

in paper or card, and here again the presence or absence of a hole

defines the two possible conditions of a two-state storage element.
Magnetic material, too, can be used most easily as a two-state store,
since it is much simpler to measure sense of magnetization than
intensity.
For these reasons almost all high-speed computers are based on
two-state storage systems. Some, such as the ENIAC, the UNIVAC
and the Bell relay machines, nevertheless work on the decimal scale,
using groups of two-state elements to represent decimal digits.
A group of four such elements, for instance, can take up any one of
16 possible conditions, and ten of these can be associated with the
digits 0-9. It might be assumed that a decimal register of this type
is wasteful of equipment, since the six remaining storage conditions
associated with each group of elements cannot be used directly to
signify numbers, but in fact they are often useful for checking pur-
poses. Some machines actually use more than four two-state elements
for each decimal digit in order to facilitate checking.
Although it is obviously convenient for the programmer if his
computer works in the decimal notation to which he is accustomed,
the decimal scale is not the only way of representing numbers.
A simpler method of using two-state elements is to work in the scale
of two, or the binary' scale, as
'
it is often called. A number expressed
in binary form consists of a sequence of digits which are either or 1,
the convention being that the digit in the nth position is multiplied
by 2 n_1 This is exactly analogous to the decimal scale, in which the
.

digit in the nth position is treated as being multiplied by 10 n_1 .

The following examples will illustrate the principle

Decimal form of number Binary form

= 6xl0°)
6( 110(=lx2 2 +lx2 1 + 0x2°)
18(=lx 10! + 8x10°) 10010 (=lx2 4 +lx2 1
)

273 = 2 x 10 2 + 7 x 10 1 + 3 x 10°)
( 100010001 = 1 x 2 8 + 1 x 2 4 + 1 x 2°)
(

3-25 = 3 x 10° + 2 x 10-! + 5x 10~ 2

( )
11-01 (= 1 x 2 1 + 1 x 2°+ 1 x 2~ 2 )

Clearly the two possible conditions of a two-state element can be

used to represent the digits and 1, and n such elements coupled
together will form a register capable of holding any integer between

16
 and 2 n — 1. Since 2 n «10°" 3w ten binary digits are approximately
,

equal to three decimal ones. Most computers which work in the

binary scale have storage registers of between 30 and 40 binary digits,
which corresponds to about 12 decimal places.
The rules of binary arithmetic are very simple, the addition table,
for instance, being

+ = 0,0 + 1 = 1 + 0=1,
1 + 1 = 10 (i.e. a 'carry' of 1 to the next place).

It is not necessary, however, for the programmer of a binary machine

to be familiar with these rules, since he is rarely concerned with the
processes by which the computer carries out his instructions. He
need not even know how to translate numbers from decimal into
binary form, since the computer can do this for him.

2.2. The sign and size of numbers

Before a row of digits in a binary or decimal storage register can
be said to represent a definite number, the sign of the number and
the position of the binary or decimal point must be specified or
implied.
One method of dealing with signs is to associate with each register
an additional two-state element, which records the sign of the number
in the register. This system has certain disadvantages, however, and
in many computers a different method is used in which negative
numbers are represented by their complements. Anyone who has
used a desk calculating machine will know that if 1 is subtracted
from an empty register the result is a row of 9's. If 2 is subtracted,
the result is .99998, and so on. Thus it is possible to represent — 1
. .

by its complement 99999, etc., provided that a convention is intro-

. . .

duced to decide whether a particular sequence of digits is to be

interpreted in this way or as an ordinary positive number. Similarly,
subtracting 1 from an empty binary register produces a row of 1 's, and
this may be taken to represent — 1. The usual convention is that if
a binary register has in its most significant digit position its content
is treated as positive, while if the digit is a 1 it represents a negative
number in the sense outlined above. It is easy to arrange for negative
numbers to be converted into normal sign-and-modulus form on
output from the computer.
As in a desk calculating machine, the position of the binary or
decimal point in a computer storage register is not usually part of the
information stored, but is implied by the arithmetic operations carried

2 17 LDC
 '

out. Instructions merely operate on signed rows of digits, and it is

the programmer's job to see that they have numerical meaning. It

may be convenient, for example, to scale numbers so that they are
either integers or fractions less than unity, thus placing the binary
or decimal point at one end or the other of each register. The pro-
grammer must continually guard against the overflow of registers
and the loss of significant digits.
The problem of scaling can cause considerable trouble in calcula-
tions where it is difficult to assess the magnitudes of numbers. A type
of computer which is useful in such cases is one using floating point
'

registers. In floating binary notation, for example, a number x is

'
'

written in the form q2 p where q is a signed number lying between

,

certain limits and p is a signed integer. A floating binary machine

recently built at Manchester University stores numbers in registers
of 40 binary digits, each register being divided into two sections.
One number q as a signed fraction
section of 30 digits stores the
whose modulus between J and J, while the remaining 10 digits
lies

are used for the exponent p. Since p can range from +511 to —512,
the programmer is virtually free of all scale-factor worries. In the
same way, a number can be represented in floating decimal form as
'
'

q\0 p The Harvard Mark II computer, for instance, is a relay machine

.

using floating decimal registers in which the exponent p can vary

between +15 and —15. Since a floating point machine stores all
numbers to the full degree of precision, it may often be more accurate
than one using normal registers.

2.3. Binary storage systems

The size of store required for numbers and instructions in a com-
puter depends on the kind of calculation for which it is to be used.
Function tabulation, for instance, requires little more storage space
than that occupied by the programme, while the solution of 50 linear
simultaneous algebraic equations will use 2500 registers for coeffi-

cients alone. In general, it seems that a machine which can store

about 10,000 numbers is adequate for most reasonable needs, although
much useful work has been done by machines with far smaller stores
than this.
High-speed storage is expensive, however, and a machine with
10,000 high-speed registers would be very costly both to build and
to maintain. Even in the largest problems only a few numbers are
manipulated at any one stage, so that most of the registers in such
a computer would only be in use for brief periods during a calculation.

18
 :

—
Most large computers, therefore, have two types of store a relatively
small high-speed one where the actual computing is done, and a larger
auxiliary store which merely acts as a filing system, holding numbers
and instructions while they are not in use.
In most automatic computers, the high-speed store has between
50 and 500 registers capable of storing instructions or numerical data.
The time taken by the control unit to pick out the content of
a particular register is known as the 'access time', and in most
modern machines this is of the order of a millisecond.
The auxiliary store of a computer may have many times the
capacity of its high-speed store. Numbers contained in the auxiliary
store are not directly accessible to the control unit, but blocks of
information can be transferred to and from the high-speed store by
'copying' instructions inserted in the programme.

2.31. Types of high-speed store

One of the simplest two-state storage elements is the ordinary
electromagnetic relay. These were used in the Harvard Mark II
computer, the Bell Telephone Laboratory machines, and in a small
computer constructed some years ago at Imperial College. Although
reliable, a machine using relays is inherently slow and bulky, and
few modern machines use them for storage purposes. They do, how-
ever, find application in connexion with auxiliary equipment such as
printers, card and tape readers, etc., where speed and compactness
are not so necessary.
Another device which can be used to store a binary digit is a valve
circuit known as a 'flip-flop'. This has two stable states, and can be
triggered from one to the other by applying a pulse of short duration.
It is effectively an electronic switch, and can operate at much higher
speeds than a mechanical relay. Flip-flops were used for the main
store of the ENIAC (1946), the first all-electronic computer. Since
each binary digit requires at least two valves, however, a main store
built of flip-flops is extremely bulky and expensive; the ENIAC, for
instance, contains over 18,000 valves and consumes 150 kW. of
power. Later machines use flip-flops extensively in their computing
circuits, but other forms of storage are now available which are much

more suitable for a large number of digits. The commonest are

A. Delay lines
A form of storage which has been used extensively both in Britain
and the United States is the mercury delay fine. This consists of
19 2-2
a tube of mercury with a quartz crystal mounted at each end.
Electrical pulses are applied to one crystal and generate stress waves
in the mercury, which travel down the tube and are picked up at the
other end. These pulses are then amplified and fed back to the
driving crystal, so that they circulate continuously as long as the
power is switched on.
The speed of sound in mercury is about 5000 ft. /sec, so that a pulse
will take approximately 1 msec, to traverse a delay line 5 ft. long.
A delay line of this length can therefore be used to store 1000 pulses,
spaced at 1 /isec. intervals, and since each pulse can be made to
represent a binary digit this is equivalent to a set of 25 registers of
40 digits each.
The digits stored in a particular register appear as a chain of
electrical pulses at the pick-up crystal every time they circulate round
the system, and the content of a register may be altered by breaking
the connexion between the amplifier and the driving crystal at the
appropriate time. Since the contents of registers are only accessible
as they pass through the amplifier, the access time to a given register
depends on where the digits happen to be in the delay line when they
are required. If an instruction affects a number which has just
entered a line, it cannot be obeyed until the digits reappear at the
other end. In preparing programmes for a delay fine machine it is
therefore desirable to arrange the timing of numbers and instructions
so that the 'waiting time' before each instruction isobeyed shall
be as small as possible. This process is known as 'optimum pro-
gramming'.
Machines which use delay lines for their high-speed store are
usually equipped with several lines of different lengths. The DEUCE
(pi. Ill), for instance, has a main store comprising twelve delay fines,
each delay fine forming 32 registers of 32 binary digits. In addition,
there are several shorter lines providing 1, 2 or 4 registers each, which
allow rapid access to frequently required numbers and are also used
for carrying out arithmetic operations. The basic pulse frequency is

1 Mc./sec, so that the content of a register in one of the longer delay

lines appears at 1024/^sec. intervals. Numbers stored in the shorter
fines, however, are available at intervals of 32, 64 or 128/^sec.
Mercury delay lines are sensitive to temperature variations and
must be thermostatically controlled. In recent years delay lines
made of other materials such as nickel have been developed, some
of which are less affected by temperature. In a nickel delay line the

metal is formed into a thin tube or wire, and stress waves are generated

20
by passing a current through a coilwound on one end. The magnetic
field causes a distortion of the material known as magneto-striction,
and this causes a wave to be propagated down the delay line. A pick-
up coil at the other end operates in a similar manner. Nickel delay
lineshave been used in several small computers, as they are light and
can easily be made into plug-in units.

B. Cathode-ray tubes
A cathode -ray tube forms a convenient means of displaying the
contents of a binary register, and it can also be used to store binary
digits. The and l's can be represented either by dots and dashes,
O's
or by faint and bright dots on the tube's fluorescent screen (pi. la).
The electrons which cause the emission of light also produce a distri-
bution of surface charge on the screen, and this distribution can be
maintained by using the phenomenon of secondary emission associated
with fluorescent material. When acting as storage device, the cathode-
ray tube has a pick-up plate mounted in front of its screen, which is

scanned repeatedly by an electron beam in a manner similar to that

which occurs in a television tube. As the beam traverses the screen,
the digits stored in the form of surface charges produce a series of
pulses on the pick-up plate, and these pulses can be used to regenerate
the charge pattern which would otherwise gradually leak away.
A single cathode-ray tube can be used to store a large number of
digits. In the Ferranti Mark I computer, for example, the high-speed
store consists of eight cathode-ray tubes, each storing 1280 binary
digits arranged as 32 40-digit registers. The cathode-ray tube has an
advantage over the delay line in that all registers are immediately
accessible.To scan a particular register it is only necessary to switch
the electron beam to the appropriate point on the tube's screen.

C. Magnetic cores
Both the delay line and the cathode-ray tube are regenerative
systems, requiring a power supply merely to keep the stored informa-
tion circulating. A type of binary element which can store a single
digit without the need for external power is the magnetic core. This
consists of a small toroid of moulded ferrite, a material which has an
almost rectangular hysteresis loop. The material may be put into one
of the two possible states of remanent magnetism by applying a
current pulse of the appropriate sense to a coil Unking the core.
These two states may be used to represent the digits and 1, and
the number storedmay be 'read' by examining the voltage pulse
21
 '

produced in a second winding when a current pulse of standard

direction is applied to the first.

The cores used in practice are normally about 2 mm. in diameter, so

that a large number of digits can be stored in a very small space. The
'
writing and reading coils are often merely single wires which thread
' ' '

the cores and also support them in space (pi. 16), and the change-
over time from state ' '
to state 1 is of the order of a microsecond.
'
'

Magnetic cores were first used in a store of 1024 16 -digit registers

built for the 'Whirlwind' computer at M.I.T. in 1953, and more
recently they have been used in EDSAC II and the Ferranti Mark II
computer. It seems likely that they will very soon replace all other
types of store, at least for really high-speed machines.

2.32. Auxiliary storage

Magnetic material provides by far the commonest type of auxiliary
storage. It may be used either in the form of a 'magnetic drum', or
as tape or wire.
A '
magnetic drum ' (pi. IV a) consists of ametal cylinderwith a mag-
netizable surface, rotating in synchronization with the rest of the
computer. The surface of the drum is divided into 'tracks', spaced
along the axis of rotation, each track being equipped with a writing '

and a 'reading' head. The operation is very similar to that of a tape

recorder, the contents of a complete track being transferred to or
from the high-speed store in one rotation of the drum. As a typical
example, the Ferranti Mark I computer has a drum 10 in. in diameter
and 12 in. long rotating at 1955 r.p.m. There are 256 tracks along the
length of the drum, each track storing 2560 binary digits, corre-
sponding to the contents of 64 40-digit registers. The contents of a
track can be read from the drum to the high-speed store in 36 msec,
while the corresponding time for writing on the drum is 63 msec. No
facility exists for the transfer of the contents of single registers, but
this is not a noticeable drawback in practice.
Transfer appear in programmes along with the
instructions
ordinary arithmetic instructions. A transfer instruction specifies a
track number, a section of the high-speed store, and whether the
transfer is to 'read from' or 'write onto' the drum. In some com-
puters, such as the DEUCE, the calculation can be continued without
interruption, provided that it does not use the relatively small part of
the store concerned in the transfer during the transfer period.
While magnetic drums have been extensively used for auxiliary
storage in large computers, they often form the entire store of smaller

22
machines. Drum computers are relatively cheap to build, and while
they may not be as fast as machines with other types of storage, they
are usually very reliable. When used as a main store, the magnetic
drum suffers from the same disadvantage as the delay line, since the
content of a particular register is only accessible once every revolu-
tion, but this can be mitigated to some extent by optimum pro-
gramming techniques. A typical drum computer is the HEC 4E,
which has a drum store of 1024 registers, each holding 40 binary
digits. The drum revolves at 3000 r.p.m., so that the access time to
a particular register is 20 msec, or less.

Magnetic tape has been used as an auxiliary store, notably on the

UNI VAC and on the EDS AC. Its main disadvantage would seem
to be that not all blocks of information are equally accessible it —
might on certain occasions be necessary to run through a whole reel
of tape to reach a particular block of numbers. It is possible, how-
ever, to attach several tape units to one computer', and the total
storage available can be made extremely large.
The input -output medium of a computer can also be used as a
large -capacity auxiliary store. If a complete problem is too big for
a machine, for instance, intermediate results can be punched out on
cards or on paper tape, and fed back through the input device at the
appropriate time.

2.4. Input-output equipment

For the sake of reliability and economy, many computers use
standard or modified commercial data-processing equipment for the
input and output of information. The two commonest input media
are punched cards and teleprinter tape (pi. IV b) but magnetic tape
and wire have also been used. Small items of information can also be
fed into a computer by manually operated switches.

2.41. Punched cards

Hollerith cards store information in the form of 12 rows of 80
binary digits each, a hole in a digit position signifying a '1', while
a space signifies a When
used in conjunction with an automatic
'0'.

computer, each row of the card is usually made to correspond to

a single binary register; this means that part of the card may be
wasted (depending on the number of digits in the computer's
registers), but this is of little consequence.
A fast card reader can operate at a speed of about three cards per
second, corresponding to over 2000 binary numbers per minute, while

23
a standard output punch works at a slightly lower speed. Results
may thus be obtained rapidly in card form and printed out elsewhere
without interfering with the computer. A normal computer installa-
tion will include standard Hollerith equipment for preparing, sorting,
copying and tabulating cards.
would be a very tedious manual job to convert all data into
It
binary form and all results back into decimals. It is possible, however,

to punch numbers in a decimal code, using the various rows of a card

to represent different decimal digits. This pattern of holes is read by
the computer in the usual way, and converted into a binary number
by a short sequence of instructions supplied by the operator. In the
same way a binary-decimal output sequence can be made to produce
results in decimal form. In practice, all the necessary calculations
can be done in the intervals between the reading or punching opera-
tions, so that no machine time is lost. Finally, the decimal numbers
on the cards can be printed on a standard Hollerith tabulator. This
facility is only valuable if the results are required for human inspec-
tion. If they are merely to be fed back into the computer the
numbers are best left in binary form.

2.42. Teleprinter tape

Another common input medium is standard teleprinter tape. This
is paper tape with which holes can be punched
five or six positions in
across its width.Each row of a five-hole tape can store a 5-digit
binary number, and when used for telegraphy each of the 32 possible
digit combinations is made to represent a letter of the alphabet or
other symbol. This means that if standard equipment is used to

produce computer input tapes, instructions and data must be written

in a form corresponding to the characters on a teleprinter keyboard.
A feature of standard teleprinter tape which is sometimes a dis-
advantage is the fact that a single row of holes is not long enough to
specify an instruction. In the Ferranti Mark I computer, for instance,
the input medium is 5 -hole paper tape, while the instructions are
20-digit numbers. This means that each instruction appears as four
rows of holes on the input tape, and a special 'input programme' is
required to assemble the instructions and plant them in the com-
puter's store. Teleprinter tape, however, is quite suitable for presenting
decimal numbers to a computer, since each decimal digit can be
conveniently represented by a single row of holes.
Since paper tape can store far fewer digits per row than punched
cards, satisfactory input speeds can only be achieved by running the
24
(</) Binary digits displayed on a cathode ray tube

(b) Part of a magnetic core storage system

 PLATE II

The first Manchester computer, built in 1948

 PLATE III

An English Electric DEUCE computer, l.uilt in 1955

 PLATE IV

fm m

{a) The magnetic drum of a DE U CE computer

(b) Punched cards and punched paper tape

tape a great deal faster than is usual in normal telegraphy work.
Mechanical readers are in general too slow, and several photo-
electric readers have been developed especially for attachment to
computers. These are capable of reading speeds of up to about 200
rows of holes per second, which represents a digit rate similar to
punched card equipment.
Machines using paper tape for input are normally provided with
a tape perforator for output purposes. Standard telegraph punches
operate at up to 15 rows of holes per second, while special high-speed
perforators have been produced with speeds of up to 60 rows per
second. Many machines also have a direct- coupled typewriter for
printing results in a form which is immediately intelligible, and
several high-speed typewriters have been developed for this purpose.
The I.B.M. 701 computer, for instance, is equipped with a printer
which can produce 150 lines of printing per minute, each line com-
prising up to 72 decimal digits.
As in the case of punched card machines, a computer installation
based on paper tape will normally be provided with auxiliary equip-
ment for preparing and copying programmes, and for tabulating
numerical data.

2.43. Magnetic tape and wire

Magnetic tape has already been mentioned under the heading
of auxiliary storage, but it also provides an extremely rapid means of
feeding information in and out of a computer. In fact, in machines
which use magnetic tape the distinction between auxiliary storage
and input-output arrangements tends to become lost. Magnetic wire
has also been used successfully, notably on the SEAC its disadvan- ;

tage is that only one channel is available for recording information,

while on magnetic tape five or six can be used.
One drawback of magnetic material as an input-output medium
is that information is not immediately visible, and the equipment

needed to convert it into ordinary printing is more complicated

than is the case with cards or paper tape. It is, however, extremely
suitable for large-scale clerical work, since it provides a compact semi-
permanent storage for material such as accounts and insurance
policies.

2.44. Manual devices

In theory a computer needs hardly any manual controls except
those necessary to switch it on and off. In practice, however, most

25
 :

computers tend to acquire a complicated control panel. The controls

normally fitted can be classified as follows
(1) Means for switching on the computer, the auxiliary store, and
the input -output units.
(2) Keys for clearing the various storage registers (i.e. setting
their contents to zero). These are similar in effect to those on a desk
calculating machine.
(3) Keys which cause the computer to start or stop obeying
instructions. Such keys may allow the programme to proceed 'full-
speed ahead', or may stop the calculation at certain pre-assigned
instructions. The programmes.
latter facility is useful in testing

(4) Means amounts of extra data. Most com-

for feeding in small
puters have one or more registers which are directly controlled by
the operator, and which can be used to hold permanent parameters
or other items of numerical information. They can also be used to
make small corrections to a programme or to interpolate extra in-
structions. This allows the machine operator to correct programme
errors without leaving the machine in order to re-perforate a piece
of tape or a card.
(5) Means for displaying the contents of registers. A cathode-ray
tube is usually provided for this purpose, so that the operator can
inspect storage registers while the calculation is proceeding. Thus he
can see whether the programme is running according to plan.
The manual controls of a computer are used mainly in testing the
machine or its programmes. If both are functioning correctly, the
operator may not need to touch the machine except to supply fresh
data or to remove printed or punched results.

26
 .

Chapter 3

THE ORGANIZATION OF PROGRAMMES

In Chapter 1 it was shown how a simple piece of numerical work could
be carried out by a computer under the control of a sequence of coded
instructions. Short sequences of this nature are often merely parts
of larger and more complicated programmes, and when this is the
case they are usually called 'routines'. This chapter describes the
way inwhich large programmes are usually constructed, and con-
siders some of the devices which make programming easier.

3.1. Dividing a programme into routines

The aim of programming is always to break down a calculation
into a series of numerical steps which can be represented by machine
instructions. In simple problems, such as the one discussed in
Chapter 1, it may be possible to proceed direct from the mathe-
matical formulation to the coded instructions, but it is often more
convenient to carry out the work in two stages. The first stage con-
sists of writing down the general plan of the calculation in ordinary
human language, without considering in detail the way in which the
various operations are to be carried out by the machine. Once the
overall strategy has been decided, each section of the general plan is

considered in detail, and the individual operations translated into

machine code. The sequences for the various sections are then com-
bined to form the complete programme.
The process is best illustrated by an example. Consider, for
instance, the design of a programme to tabulate the function cos *Jx
over the range x — to x = 1, with an interval of 0-001. The first step
is to write down a statement of the various operations which the
programme must perform. A suitable scheme might be:
(a) The register which is to store the variable x is cleared. (This
starts the calculation with x = 0.)
(b) cos yjx is evaluated for the stored value of x.

(c) x and cos *Jx are printed.

(d) x is increased by 0-001.
(e) x is tested. The programme returns to operation (b) if #^1,
but comes to a halt if x > 1

27
 :

Some of these steps correspond to single machine instructions,

while others need sequences of instructions to carry them out. It is
now possible to consider the five steps as separate coding problems,
attacking each difficulty in turn.
If the programme is to be designed for the idealized machine
described in Chapter 1, the only step which might cause the pro-
grammer some trouble is operation (b). The human computer would
probably refer to a set of tables when calculating cos Jx, but, although
it would be possible to store appropriate tables within the machine,
it would be easier to use some kind of numerical approximation to
calculate each entry in turn. One might, for instance, use the series
expansion x x2 x3

taking sufficient terms to obtain accurate tabulation over the range

4
of x considered. Terminating the series after the term in a: gives an
approximation which may conveniently be written
cos V*«l-s(£-a(A:-a(7io -aferko)))),
in which form it is very simple to programme. A sequence for
evaluating this expression can easily be constructed using the in-
struction code described in Chapter 1.

computer to be used stores numbers in binary form, however,

If the
the operations of printing x and cos ^x will also require sequences of
instructions to convert the binary numbers into decimal ones. All
that is really necessary is one sequence which prints a single number
(say the content of the accumulator), since the programme can be
arranged so that this sequence is used twice.
Having prepared the two sequences described above, the pro-
grammer can now proceed as though the calculation of cos y/x and
binary- decimal conversion were functions available in the machine's
instruction code, like multiplication and addition. All he needs to
do is to write down a controlling sequence which will carry out the
other steps required, using suitable jump instructions to enter the
sequences for calculating cos y/x and for printing. Each of these can
be made to end with a jump instruction which causes a return to the
appropriate point in the controlling sequence. A suitable arrange-
ment for the controlling sequence would be
(a) Clear the register which is to
store x.
(b) Jump to the sequence for calcu-
lating cos y/x.

28
Return from this sequence to->(c) Copy x into the accumulator.
(d) Jump to the sequence for printing
the content of the accumulator.
Return from this sequence to->(e) Copy cos Jx into the accumulator.
(/) Jump to the sequence for printing
the content of the accumulator.
Return from this sequence to-^(gr) Add 0-001 to the value of x.
(h) Form and test (x—l). Return to
(6) if (a-lKO.

(i) Stop.

In this arrangement of the programme, the controlling sequence

described above is called a 'master routine', while the sequences for
calculating cos *Jx and for printing are known as sub -routines It is
'
'.

a characteristic feature of a sub -routine that when it has completed

its task the master routine carries on from the point at which it was
interrupted.
The concept of a programme as an arrangement of routines is an
extremely useful one in programming. programme a
It gives the
logical structure, and helps the programmer to isolate and overcome
singly any difficulties in coding which he may encounter. Even more
important, it allows many programmes to be constructed almost
entirely from pre -fabricated sub-routines. It is apparent that the
short master routine described above could be used to tabulate
another function merely by changing the function-generating routine,
while in the same way the two sub -routines could be used in any
other programme which required them. If a sub-routine is available
for a particular operation, it is effectively as though the computer had
been provided with an instruction for doing the job, and the pro-
grammer need not concern himself with the way in which the
operation is carried out.

3.2. Library routines

When a new computer is built, the first task of its users should be
to build up a library of routines for carrying out standard mathe-
matical operations. Once constructed, these routines are always
available to any programmer who may wish to use them. The
efficiency of acomputer depends to a large extent on the quality and
comprehensiveness of its library, so that it is worth spending con-
siderable time and effort on constructing and testing a well-integrated

29
set of routines. The majority of library routines fall into one of the
following categories:

A. Routines concerned with the input of information

These are of two kinds. The first is designed to organize the transfer
of instructions from the input medium to their correct places in the
machine's store. If parts of the programme are to be filed away in an
auxiliary store, it is this routine's job to carry out the transfers, and
to start the calculation when all the programme has been stored.
The second type of input routine is used for reading decimal
numbers, and in a binary machine its central feature is a decimal-
binary conversion sequence. It often appears as a sub-routine called
in at the beginning of a programme, and may be designed either to
read a single number and place it in a certain register, or to read
a series of numbers and place them in registers specified by the rest
of the programme. A computer will normally have input routines for
dealing with both integers and fractions.

B. Output routines
Output routines are mainly concerned with numerical information,
since it is not often that one requires a printed record of machine
instructions. Here the conversion required is from binary to decimal,
and most computers are equipped with several routines. These will
normally cover the output of single numbers and groups of numbers,
as integers, fractions or with pre-assigned decimal point. If a machine
has a direct- coupled output printer, its routines will usually include
provision for tabulation in columns and pages, so that tables can be
prepared with the minimum of human editing.

C. Function routines
These are sub-routines which generate standard mathematical
functions, such as Jx, sin x, sin -1 :r, log x, e x etc. The value of x is
,

provided by the sequence which calls in the sub-routine, and it is

convenient to standardize these routines so that they all use the same
storage registers for x and F(x). One commonly used arrangement
is for the sub -routines to be entered with x available in the accumu-

lator, and to be quitted with F(x) in the same place. In the notation
of §1.3 the effect of such a routine can be written simply as

A'=F(A).
Functions are normally calculated from a polynomial approximation
or by an iterative process.

30
 :

D. Operator routines
These carry out mathematical operations on functions, and
normally require a sub -routine for generating the function. A typical
example is a routine for the numerical integration of an arbitrary
function f(x). The routine is supplied with the limits of integration,
and works out suitable values x i at which to calculate the
first

integrand. These are then given to a sub-routine, which works out

the various values /(#J. Finally, the integration routine works out
an approximation to the integral from these function values, using
one of the well-known formulae for numerical integration.
Another example is a routine for integrating sets of ordinary
differential equations of the type

dx {
~dt~~

where a sub -routine is used to generate the functions appearing on

the right-hand sides of the equations. An integrating routine of this
type normally advances the integration by one step each time it is

entered, and requires a master routine to control it.

E. Routines for linear algebra

These routines carry out the processes required in matrix work.
The types of operation required are
(a) The solution of sets of linear simultaneous algebraic equations.

(b) The multiplication of matrices.

(c) The inversion of matrices.

(d) The extraction of the latent roots and vectors of matrices.

With these routines, operations on matrices can be carried out

almost as easily, from the programmer's point of view, as operations
on single numbers. All the programmer has to do is to organize the
input or calculation of the coefficients, and specify the appropriate
routines.

F. Test routines
These routines are used to test the correct functioning of a com-
puter. They are run before a machine is handed over by the main-
tenance engineers, and are available to an operator if he suspects that
a machine fault has developed. A good set of test routines includes
checks on each separate unit of a computer, and in some cases it is

possible to locate a faulty valve merely by watching the behaviour

31
of the routines. The design of test routines requires considerably more
knowledge of the circuitry of a machine than the average programmer

3.3. The organization of sub-routines

As mentioned in §3.1, a sub -routine always ends with a jump
instruction,which causes a return to the instructions which originally
called for the sub-routine. A possible arrangement is shown dia-
grammatically in fig. 2. It is apparent that while this allows the main
body of a sub-routine to be made completely self-contained, the final

Instructions of
main routine

*
1j Instructions of
sub-routine

Jump instruction '

Jump instruction

Fig. 2. Simple method of calling in a sub-routine.

jump instruction will depend on the point to which a return is

required. This means that if a sub-routine is required to operate at
several points in a programme, the final jump instruction must be
different in each case. In the simple programme described in § 3.1, for
instance, the printing sub-routine is followed by a return to operation
(e) on the first occasion, and to operation (g) on the second occasion.

One method of achieving this is as follows.

When it is necessary to call in a sub -routine, the jump instruction
in the main routine is preceded by an instruction which plants what
is known as a 'link' in a convenient register. The link is merely the

address of the instruction to which a return is required, and the sub-

routine is arranged to terminate with the instruction 'Jump to the
instruction whose address is contained in the link register'. This

32
 l

arrangement is shown in fig. 3. In this way a sub -routine can be made

completely independent of the instructions which call it in. It may
happen that a sub -routine has also sub -routines of its own. In such
a case it is necessary to keep a list of links in order to return to the
main routine, but the principle is the same.
If a computer is equipped with a magnetic drum, it is often con-
venient to store most of the programme there. Individual routines
can be stored on separate tracks, and it is possible to keep some of
the more commonly used library routines permanently stored on

Instructions of link register

main routine

instruction
t
sub-routine
i

S-2 Plant S in link register

S-1
S
Jump instruction

. _
Instruction causing

jump to address in

Fig. 3. Alternative method of

-. —Link register

calling in a sub-routine.

special isolated tracks whose contents cannot be over- written. In

this way only a small part of the high-speed store need be reserved
Each section of the programme can be transferred
for instructions.
from the magnetic drum as required, leaving the remainder of the
high-speed store free for numerical material.
With an auxiliary store of this type, a sequence calling in a sub-
must include appropriate instructions for transferring it to the
routine
high-speed store. Sub-routines can, however, be designed always to
occupy certain when being obeyed. When no auxiliary store
registers
have to be packed into the same store, so that
exists, all the routines
when a library routine is designed it is not known in advance exactly
which registers it will occupy. This difficulty is overcome on some
computers by using a system of 'floating addresses', in which those

3 33 LDC
instructions whose address sections depend on the position of a
routine in the store are automatically adjusted by a special input
sequence as the routine is fed into the machine.

3.4. The development of a programme

It is now possible to describe what happens during the design and
development of a complete programme.
The programmer's first task is to decide on a mathematical formu-
lation for his problem. In some cases this is already fixed, but in
others original mathematical research may be required. Even at this
stage, however, the programmer will try, if he can, to choose the
approach which leads to the simplest programming.
The second stage involves a choice of numerical method, and here
the programmer is very strongly influenced by the library routines
which he has available. In the example discussed in § 3- 1 for instance,
,

there would almost certainly be library routines available for calcu-

lating ^x and cos x. It would be simpler to make use of these than
to produce a new routine for calculating cos Jx, even though the
latter sequence might be slightly faster. The programmer is very
often faced with a choice between designing new routines and using
library routines which are not quite so well-suited to his programme.
It must be remembered that even if a new routine is faster than one
taken from the library, extra machine time will be necessary to test
it. In the tabulation programme under discussion almost all the
machine time would probably be spent on printing, so that there
would be little virtue in developing a new routine. Once the mathe-
matical problem has been reduced to a scheme of numerical analysis,
the calculation can conveniently be laid out in the form of a flow-
diagram, as shown in fig. 4. This helps to demonstrate the logical
design of the programme and shows how the work can best be split
up into routines of convenient size. At this stage the programmer
decides provisionally which library routines he will use.
A start can now be made on the task of machine organization. The
programmer allocates storage registers for instructions, numerical
data and working space, and decides how data shall be fed in and
results printed. He considers the approximate magnitudes of the
numbers which will appear in the calculation, and decides whether
any scaling is necessary to prevent registers overflowing or losing too
many significant figures. With these details fixed, he can prepare
detailed specifications of any new routines which may be required.
All that now remains is the detailed coding of the new routines,

34
 :

and the final assembly of the programme in card or tape form. As

each routine isprepared it is punched out and tested in the computer
—usually by feeding in a few test numbers and checking the results
by hand. When all the logical errors and coding slips have been
eliminated, the final programme is made up by combining the separate
parts on one pack of cards or a single length of tape. Master copies
of library routines are normally kept available in punched form, and
these can easily be copied on auxiliary editing equipment. Final
testing is usually simple if all the individual routines have been
checked separately.

Set i Form Form Print

Start
x= | ^/x cos \Zx"
X
1

i \
v

x-1^0 r n
Form Print
Stop 0-001
P "* x-1 cos Jx
x-1: >o to X
I

Fig. 4. Flow diagram for tabulation of cos ^x. Library sub-routines

shown with full outlines.

3.5. Testing programmes

Very few programmes work correctly the first time they are tried
out. The most common faults in programming may be classified as
follows

A. Slips
These are fundamentally due to human carelessness, and careful
checking of programmes will eliminate most of them. A common
example is the case where a programmer writes down some instruc-
tionsand then decides to change the location of a certain number.
He may easily forget to alter the address parts of the relevant
instructions, with the result that the computer carries out the right
operations on the wrong numbers. In the same way, the rearrange-
ment of a routine may omit an adjustment to a jump instruction, so
that the instructions are obeyed in the wrong order.
Slips can also occur in transferring instructions from paper to
punched form, and are often difficult to find since they do not appear
on the programmer's manuscript. Here again, careful checking before
the programme is fed into the machine is the only way to avoid
wastage of valuable machine time.

35 v2
B. Faults of organization
These are logical errors on the programmer's part, and usually-
mean that he has not foreseen all the consequences of an instruction.
They may result in the programme getting stuck in a closed cycle of
which were intended to repre-
instructions, or in sequences of digits
sent numbers being obeyed One fault which seems
as instructions.
to be fairly common among users of some machines is the use of a
conditional jump instruction 'the wrong way round', i.e. making
a jump occur only when a number is positive, say, when it should
have occurred only when the number was negative. Experience in
programming reduces the number of these faults, but very few
programmers avoid them entirely.

C. Numerical faults
These are usually the result of incorrect scaling. They result in
numbers either exceeding the capacities of their registers or losing
significant digits. The programme generally goes through the instruc-
tions in the correct order, but produces the wrong answer. Such
faults can usually be located by making the programme print out
intermediate answers at several points.

D. Special cases
It is usually impracticable to test a routine in all possible circum-
stances, and there are occasionally cases even in tested library
routines where the appearance of one particular number leads to a
wrong result. This may be due either to a logical fault in the routine's
—
design or to mathematical reasons division by zero is an obvious
example. This is a particularly difficult type of fault to trace, since the
programme may work correctly on one problem and fail on another.

Most computers have facilities for the inspection of their storage

registers,and the simplest way of testing a programme is to make
the machine go through it instruction by instruction, checking
visually on the effect of each operation. With this form of testing the
machine obeys a single instruction each time the operator presses
a switch, so that it is extremely wasteful of machine time. One way
of improving the process is to insert '
stop ' instructions at suitable
points in the programme. The operator can then run through the
tested sections, such as library routines, at full speed, changing to
single-instruction operation at the start of a suspected sequence. This

36
technique is not too time-wasting when carried out by an experienced
operator, but it should never be considered as a substitute for the
utmost care in programme preparation.
On some machines it is possible to print out a characteristic of
each instruction, such as the function number, as the instruction is
obeyed. In this way the programmer can quickly test whether the
order of the instructions is correct. This is especially useful in checking
the correctness of jump instructions. In the same way, arrangements
for printing out the contents of the accumulator at specified points
are extremely useful in tracing numerical faults.
It is evident that programmes constructed mainly from library
routines will, on the whole, require far less time for fault tracing, and
this is a strong argument for solving problems by standard mathe-
matical processes. Machine time spent on developing a programme
is just as expensive as time spent in obtaining results, and an extra
complication designed to save a few seconds in a calculation may
waste far more development time than it saves.

3.6. Automatic programming

The translation of instructions from normal English and mathe-
matical symbols into machine code is a somewhat tedious and
mechanical task, and it is sometimes suggested that a computer
might be able to write its own programmes. In effect, this implies
converting a code used by human beings into one used by a machine,
and the problems which arise are similar to those occurring in
problems of mechanical language translation. Programme -writing
routines fall into two groups.

A. Conversion routines
In a conversion routine, the instructions finally placed in the
machine's store are not the same as the 'instructions' punched on
the input medium. These 'instructions' are written in a code which,
while not as flexible as ordinary human language, is more familiar
to the programmer and has wider scope than ordinary machine code.
As each 'instruction' is read, the conversion routine works out the
corresponding machine instruction or group of instructions, and
places it in the store. When the programme of machine instructions
is complete, it is obeyed by the computer in the usual way.

A very simple example of a conversion routine is the input routine

mentioned in §3.3, in connexion with 'floating address' pro-
gramming. Conversion routines have also been devised for doing
37
operations on complex numbers, and for using a computer with
normal registers as though it were a floating-point machine.

B. Interpretive routines
With an interpretive routine, the 'instructions', which as before
are written in a special code, are actually planted in the computer's
store. Instead of being selected by the control unit in the normal
way, however, a special sequence selects each 'instruction' in turn,
decodes it, and calls in what is effectively a sub-routine for carrying
out the operation specified.

Experience has so far shown that machine-produced programmes

take longer to run and use more storage space than the equivalent
human product. Their main advantage lies in the reduction in pro-
gramming time, and the fact that virtually no testing is required. The
techniqueis very suitable, therefore, for small problems in which the

totalmachine time would be only a few minutes in any case. It does

not seem likely, however, that machines will ever be able to carry
out the broader aspects of programme design.

38
 Chapter 4

THE SOLUTION OF ENGINEERING PROBLEMS

This chapter discusses some of the points which arise in applying
computers to engineering calculations, and describes some of the
problems which have so far been solved.
Although the natural way of classifying calculations is according
to the basic mathematical processes involved, programme design is

also considerably influenced by the number of times a calculation is

likely to be repeated. In practice problems can be divided into two

distinct types, which may be called 'single' and 'repetitive', and
the differences between these will be considered first.

4.1. Single problems

In a single problem, the production of numerical results implies
that the problem has been completely solved, and that no similar
If a programme is prepared for
calculations will ever be required.
such a calculation, become redundant when the solution has
it will
been obtained, and must therefore pay for its creation during a
relatively short working life.

Many research problems are of this type. The end-product of the

calculation is usually a table of numbers, which may represent the
solution of a differential equation, the values of a mathematical
function, or standard solutions covering all practical parameter values
of a particular engineering product. A certain amount of repetition
may be involved, but the essential feature of a single problem is that
the number of repetitions required is known beforehand.
A direct comparison of time or cost will decide whether it is better
to programme such a calculation or to do it by hand on a desk machine.
In assessing the effort required to solve a problem on a computer,
time spent on programming and programme testing must be con-
sidered as well as actual machine running time. In many problems
the task of programming takes by far the greatest part of the time,
and may therefore be the determining economic factor. As men-
tioned in Chapter3, both programming and programme testing are

very much simplified if a calculation is reduced to standard mathe-

matical processes for which library routines exist. Even if this

39
represents an inefficient
'
' way of solving a problem, the use of library
routines often costs less in the long run than the development of
special routines which must be discarded when the calculation is

complete.
A programme for a 'single' problem is like a temporary piece of
laboratory apparatus, roughly built but adequate for its job. Since
it will probably be used only by its creator and discarded when the
job is finished, there is little point in spending a lot of time on
unnecessary refinements of design.

4.2. Repetitive problems

Most engineering calculations are not, however, the complicated
single problems of research. Much of the work carried out in design
offices involves relatively short stereotyped calculations which occur

at intervals, the same operations being carried out each time on

different sets of numbers. This type of work is particularly common in
firms whose products, whether steam turbines or shell roofs, are
designed individually to meet particular specifications.
The development of a new design often involves analysing a series
of trial designs, modifying each one in the light of previous results,
and the ease with which such analyses can be carried out may deter-
mine the extent to which the final product approaches the best
possible design. One problem of this type is the design of steel frame-
works for buildings and bridges. Here the aim is to find the lightest
and cheapest structure which will support the applied loads safely,
and as long as steel structures continue to be built designers will
have to carry out stress-analysis calculations, continually repeating
numerical operations which are essentially the same.
A computer is extremely suitable for this type of calculation.
Once a general or 'standard' programme has been constructed for
a particular class of problem, each individual calculation merely
involves supplying the computer with the appropriate data. No
furtherprogramming is required, and the machine can in fact be
operated by semi-skilled labour.

4.3. Factors in the design of standard programmes

Just as in the case of a single problem, the designer of a standard
'
'

programme tries to achieve his objective with the least possible effort.
There are, however, several additional factors which he must consider.
One important difference springs from the fact that a standard
programme, like a library routine, is intended to be of permanent

40
 : ; ; '

value. Any time spent in making it more efficient is therefore a

capital investment which will pay dividends each time the pro-
gramme is used. Even though many standard programmes are
designed for relatively simple problems and take up only a few
minutes of machine time, a few seconds saved may represent a con-
siderable gain over a long period. If a programme is really going
to be of general service, it is worth spending a considerable amount of
time on making it as perfect as possible.
Unlike a programme for a single problem, a standard programme
is always designed to cover a whole class of problems, and this means
that the programmer looks for mathematical techniques which are
as general as possible. With a digital computer it is very often easier
to carry out an operation which in some cases may be unnecessary,
than to test whether it is possible to use a short cut, while many
of the intuitive judgements ordinarily made by the human computer
are almost impossible to programme. Thus it often happens that the
'best' way of solving a problem on an automatic machineis one

which would be quite impracticable on a desk computer, so that the

programmer has to adopt new ideas as to which methods are easy '

and which are 'difficult'.

A standard programme is essentially a production tool, which may
be used by many different operators, some of whom may not be
skilled programmers. It must therefore be made foolproof as regards
operation and data preparation, and all 'special cases' must be
foreseen. It is also most important that the programme should be
designed to minimize human error. Such errors are all too common,
and are often more difficult to detect than in hand-computing work.
They may occur during either
(a) the preparation of the initial data
(b) the operation of the programme
(c) the interpretation of the results.
Errors during data preparation may be minimized by arranging
the programme so that the machine takes in data in the form which
is most natural to the user. Small preliminary calculations, involving

scaling, evaluation of coefficients, etc., are sufficiently simple to

invite careless treatment, with the result that the machine may get
supplied with the wrong data. For this reason it is foolish to pro-
gramme the difficult parts of a calculation and neglect the easier
ones. For similar reasons the programme should print the final
answers in the most convenient form for the user, and not leave the
last few steps to the human computer.

41
 There is another reason why it is good policy to feed the basic
parameters of a problem directly into a computer. The first part of
many calculations consists of evaluating various coefficients from
the initial information. These coefficients are often far more numerous
than the original data, so that if they are calculated by hand outside
the machine many more quantities will eventually have to be pre-
pared for input. Whether input is by punched cards or paper tape,
the smaller the volume of material the less will be the chance of
mistakes in punching.
The
attitude which a programmer adopts towards machine results
may vary between blind faith and complete scepticism, according to
the standard of reliability of his computer. Many routine engineering
calculations are so short in terms of machine time that it is usually
simplest to omit checking sequences, repeating the whole calculation
if errors are suspected. On longer programmes of half-an-hour's
duration or more it is a good idea to incorporate checks at various
points, but most computers are sufficiently reliable to render complex
checking sequences unjustified.
It is hardly necessary to say that a standard programme should
never invite human error by requiring extensive manipulation of the
machine's controls. Experience has shown that it is perfectly possible
for a semi-skilled machine operator to produce satisfactory results,
provided that the computer is reliable. An operator having no
knowledge of the construction of a programme is naturally at a
disadvantage if suspect results appear, as he may be unable to tell
whether the fault lies in the machine, the programme, or the punching
of the data.
The standard programmes were often merely direct transla-
first

tions intomachine code of operations previously carried out by hand.

They were therefore 'small' calculations from the computer point
of view, and required only a few minutes of machine time. It seems
likely, however, that as engineers realize the power of automatic

computers the size of standard programmes will increase, until calcu-

lations which would be quite impracticable to carry out by hand
become 'standard practice'.
A standard programme can be compared with a machine tool
designed for mass-production work. Just as a machine tool is re-

designed in the fight of experience, so a standard programme is often

gradually modified as experience of using it accumulates.

42
4.4. Problems which have been solved on computers
This section does not attempt to describe all the problems which

have so far been tackled. It merely gives examples of the kind of

work which is being done at present on digital machines.

4.41. Problems in linear algebra

Many engineering calculations are concerned with linear systems,
built up from elements which obey Hooke's or Ohm's law, and in
such systems there are usually two kinds of variable which are
related in a linear manner. In an elastic structure, for instance, the
displacements are linear functions of the applied loads, while in an
electric network currents are similarly related to applied voltages.
If such a calculation involves finding the static response to a given
applied stimulus, it can always be formulated as a set of simultaneous
linear equations, while if the general static response to any arbitrary
applied stimulus is required, the solution involves the inversion of
a matrix. In the same way, the common problem of finding natural
frequencies becomes from the mathematical point of view the extrac-
tion of the latent roots of a matrix. Solving such a problem on a
computer merely involves calling in the appropriate library routine.
In recent years a great deal of effort has been spent on developing
techniques for dealing with large sets of equations, so that the pro-
grammer can consider his problem solved if he can reduce it to one
of the three standard matrix processes mentioned above. The Man-
chester computer, for example, which is not a particularly fast
machine, has a library routine which solves n simultaneous equations
(n<64) in approximately n 2 /Q sec, the inversion of the associated
matrix taking about three times as long. Other machines have
routines of comparable speed and capacity. Larger sets of equations
do not often occur, but when they do it is usually possible to break the
problem down into several smaller problems of more manageable size.
The fact that a linear problem can be reduced to a standard
mathematical form does not mean that a computer is only of use in
solving the equations. If the problem is a single one it may be best
'
'

to work out the coefficients by hand, but in a repetitive problem it is

most desirable to use the machine to set up the equations as well as
to solve them. A set of 50 simultaneous equations, for instance, may
have up to 2500 non-zero coefficients, and if all these are calculated
by hand it will be a major task merely to check the punching. The re-
marks on the avoidance of human error in §4.3 are very relevant here.
43
 :

The programmer's main problem in dealing with repetitive linear

problems is therefore the design of routines for translating the data
of a physical system into a set of equations. One example of such
routines occurs in a standard programme for elastic structural analysis
developed on the Manchester computer. In this programme, the
machine is first supplied with the dimensions and section constants
of the various members of a framework, from which it calculates the
stiffnesses of the individualmembers. The data tape then specifies the
applied loads and the way in which the members are connected
together, and the programme uses this information, together with
the stiffnesses previously calculated, to set up the load-displacement
equations for the complete structure. After a library routine has
solved these equations for the displacements, the internal forces and
moments in the members are calculated and printed out in order.
The programme can be applied to any rigid- jointed plane framework,
the complete analysis of a frame with 10 joints taking about 8 min.
A similar technique has been applied to problems of electric network
analysis.
Another problem which was solved on the Manchester computer
by reducing it to a set of simultaneous linear equations was the

determination of the elastic stress distribution in a transversely loaded

square floor supported by edge-beams, for various symmetric and
unsymmetric conditions of loading and support. Mathematically
the problem one of solving the bi-harmonic equation in a square
is

region, with appropriate boundary conditions at the edges, where the

floor-slab is attached to the supporting joists. The method adopted
was to replace the deflexion of the slab by the deflexions at the nodes
of a 7 x 7 square mesh, using finite difference techniques to set up the
equations relating the displacements of the mesh points to the applied
loads. Here again the problem was the setting up of the 49 simultaneous
equations; the final programme did this automatically, as well as
calculating the values of the bending and twisting moments at
the mesh points. Computing times for the various parts of the
calculation were

Reading the values of beam/slab stiffness for the four

edge-beams, the loads at the mesh points, and set-

ting up the 49 difference equations 50 sec.

Solving the equations and printing the displacements

at the mesh points 5 min.
Calculating the bending and twisting moments 4 min.
(mostly printing)
44
On one occasion 25 different cases were analysed in 4§ hr. It is

interesting to note that the formation of the equations inside the

machine took considerably less time than would have been required
for reading the equations from tape.
Another common problem for which several standard programmes
exist is the determination of natural frequencies of torsional and
transverse vibration in turbine shafts and other power transmission
systems. Although this can be reduced to a latent-root problem, the
special characteristics of such systems allow other methods to be
used. These are based on determining the exciting force needed to
maintain vibration at a number of frequencies, the natural frequencies
corresponding to the zeros of the forcing term. With a computer it is

a very simple matter to carry out such a calculation for a large

number of frequencies, thus determining the response of a system
over the whole working range.

4.42. Linear programming

Problems of analysis can often be reduced to the solution of sets
of linear equations, but problems of design give rise to a different type
of calculation.
The technical problem of satisfying a certain set of design require-
ments can usually be met in an infinite number of different ways, and
it is the designer's job to find the solution which, in some sense or

other, is the best one. If the design problem is a linear one, it can
'
'

sometimes be reduced to the mathematical problem of determining

the values of a set of variables which minimize a known linear
function, while satisfying a set of linear inequalities. The linear
function may represent the cost of an article, for instance, while the
inequalities represent certain specifications which it must satisfy.
In designing a steel structure to support a given loading system, for
example, the cost of the structure may be considered to be roughly
proportional to its weight. If the design is based on the criterion of
plastic collapse, the fact that the structure must not fail in any one
of all the possible collapse mechanisms defines a set of linear in-
equalities satisfied by the fully plastic moments of
which must be
the members. The designer's task is to assign dimensions to the
various members in such a way that these conditions are satisfied,
while at the same time minimizing the total weight and cost. Such
a problem is called a 'linear programming' problem the word —
'programming' being used here not in the sense of a computer
programme, but rather with the meaning 'planning'.
45
 In all but the most trivial cases the exact solution of a linear
programming problem is only practicable on a computer, and most
of the numerical techniques which have so far been developed have
been designed for machine use. Although in many cases common-
sense may give a solution which is very near to the correct one, a
saving of a small percentage may yet represent quite a considerable
amount in a large-scale design, so that digital computers may have
a useful part to play in this type of work. Linear programming
techniques may also be used to solve problems arising in the economic
planning of other human activities such as transport, marketing, etc.

4.43. Ordinary differential equations

Computers have been applied successfully to many problems
involving the numerical integration of differential equations. By
introducing extra variables, where necessary, any set of ordinary
simultaneous differential equations can be reduced to a set of n first-

order equations of the form

§-/,(**;*) (i.j-1,2 n).

If the values of the variables x t are all known for one value of t, the
equations can be integrated forward (or backward) in by anyt

standard numerical step-by-step process. The Runge-Kutta method

is perhaps the one most commonly met within computer programmes,
as it does not require a series of starting values and allows the step-
length to be changed without difficulty, but other methods have also
been used successfully.
As mentioned previously, the programmer normally has a library
routine available for carrying out the integration. In most computer
libraries this is designed to advance the integration by one step each
time it is called in, and a master routine is required for controlling
the length and number of the steps. The programmer also has to
provide sub-routines for calculating the appropriate //s at each stage.
In many cases the equations to be integrated are equations of
motion. Problems of this type which have been solved on computers
include trajectory calculations for electrons, projectiles and the
satellites of binary In each case the use of a computer allows
stars.
the calculation of any number of trajectories, the operator merely
having to provide the machine with the appropriate series of starting
conditions. A slightly different type of problem recently solved at
Manchester involved the calculation of surge voltages in a transformer

46
winding. By treating the physical system as an inductance- capaci-
tance network, the problem was reduced to the forward integration
of a set of forty ordinary differential equations of the type described
above.
In cases where the boundary conditions are given at two ends of an
interval a process of trial and error is normally required. Starting
values which are not known are assumed, and the integration carried
forward to the other end of the interval. The remaining boundary
conditions are then compared with the values calculated, and an
adjustment made to the assumed initial conditions. The process is
continued until satisfactory agreement is reached at both ends of the
interval. If the functions ft are linear in the x's, an alternative
approach is to replace all differential coefficients by differences. This
has the effect of reducing the problem to the solution of a set of
simultaneous linear algebraic equations.

4.44. Partial differential equations

Partial differential equations with two independent variables may
be treated in a number of ways, according to the equations and the
boundary conditions. Problems requiring a solution of Laplace's,
Poisson's or the bi-harmonic equation, in which boundary values are
specified on the edge of a closed region, are often treated by reducing
all differential coefficients to differences, and solving the resulting

algebraic equations by relaxation or by direct means. An example

of this approach is the floor-slab problem mentioned in §4.41. For
hyperbolic equations such as the wave equation, on the other hand,
it is more usual to remove only one independent variable by using

finite differences, integrating the resulting set of ordinary differential

equations in the manner described above.
Equations with more than two independent variables, such as
arisein unsteady two- or three-dimensional heat or fluid flow
problems, can be treated by similar means. In general, they require
a very fast machine with a large amount of storage.

4.45. Problems in non-linear algebra

Non-linear problems are in general more difficult than linear ones,
and the computers are correspondingly greater.
possibilities of digital
Many non-linear problems can be solved by iterative means, each
iteration being a linear calculation. An example of this technique is
an extension of the structural analysis programme described in
§4.41 to cover the non-linear stability effects of axial forces in

47
members. A preliminary linear elastic analysis calculates approxi-
mate values for the axial forces, and these are then used to modify the
bending stiffnesses of the members, in accordance with well-known
theory. The calculation is then repeated with these modified stiff-
nesses, the cycle being repeated until a sufficiently accurate answer
has been obtained.
Another interesting non-linear problem is the design of optical
lenses. Ray-tracing through an optical system is merely a matter of
coordinate geometry, normal refraction laws being used to find the
change of direction at each surface, and with an automatic computer
it is possible to trace the path of a single ray through a complete lens

system in a fraction of a second. In one design programme at present

in use, a series of rays are traced, and the information gained is used
to evaluate a 'performance function', which measures the amount of
distortion, coma, etc., of the complete system. The routines for this
part of the calculation appear as sub-routines in a larger programme,
which systematically tries to minimize the performance function by
altering the design constants of the various lenses. With this pro-
gramme a two -hour run achieved considerable improvement in a
design of the 'Tessar' type which has been in use for many years.
Work is at present in progress on similar programmes for the design
of magnetic lenses for focusing electron beams.

4.46. Function tabulation

Computers rarely use tables of functions themselves, since they
can so easily be programmed to generate most of the functions they
may require, but they are nevertheless extremely suitable for pre-
paring tables. Perhaps the most famous set produced on a machine
are the Harvard tables of Bessel functions, but many other useful
tables have been produced elsewhere. By incorporating column and
page lay-out in the output routine of a programme it is possible to
produce tables with the minimum of human intervention and error.

4.47. Non-mathematical problems

One important application of digital computers is to commercial
data-processing and accountancy calculations. This type of work
requires a great deal of input and output with very little intervening
arithmetic, so that the type of machine designed for scientific calcu-
lationsis not particularly suitable. Specially designed machines are,

however, playing an increasingly important part in mechanizing

office procedure.

48
 On the more frivolous side, machines have been programmed to
play chess and draughts, and to make opening They
bids at bridge.
have not, however, so far evinced any marked superiority over human
beings at these games.

4.5. Digital computers and mathematical methods

It will be realized that the problems facing the programmer are
quite different from those facing the user of a desk machine. The
human computer is always concerned with minimizing the total
amount of arithmetic he has to do, and to do this he draws exten-
sively on intuition and experience. The programmer, on the other
hand, tries to reduce his calculation to a simple logical process, and
makes as much use as he can of standard mathematical techniques.
The difference in the two approaches is particularly marked in
engineering problems, where mathematical analysis is often developed
solely as a basis for subsequent numerical work. The more often a
problem has been solved on a desk machine, the more certain it is
that the methods developed need re-examining before translation
into a computer programme. An example of this may be found in
the mathematical technique of 'relaxation' and the comparable
methods of moment-distribution in structural analysis. Both these
' '

techniques were developed because classical methods of formulating

linear problems led to large sets of simultaneous equations, the direct
solution of which was tedious by hand. Indeed, one might say that
most of the methods developed for elastic structural analysis are
really devices for avoiding the formal solution of large sets of linear
equations. It is rather strange to find that with an automatic com-
puter available one often actually tries to reduce a problem to this
form.
When a problem is prepared for hand computing, the mathematical
treatment often includes analytical transformations which, while
they may be formally elegant, result in the calculation being carried
out with numbers which have no direct physical significance. The
programmer, on the other hand, tends to keep close to the basic
equations of a physical system, even if this leads to a slightly longer
calculation. Mathematical analysis developed as a prelude to a com-
puter programme may thus help to throw fresh light on the nature
of a problem, while the possibility of using a machine to obtain
rapid exact solutions has already stimulated fresh theoretical work
on problems previously treated empirically.

49
4.6. The human factor
The engineer who wishes to use a digital computer, but who has
not sufficient numerical work to justify the purchase of a machine,
has two courses open to him. He can either learn to programme
himself, hiring time as required on one of the machines now available,
or he can employ a commercial computing organization to solve his
problems for him. Learning to programme may take him a couple
of months, but if the work which he requires can be done by an
existing standard programme a few days' training in data preparation
and machine operation is all that is required. Hire charges for large
machines range from £20 to £50 per hour. This may appear high, but
in fact it represents a considerable economy in many problems. Once
a programme has been completed, the speed of a computer is from
100 to 500 times that of a human being equipped with a desk
calculating machine.
From the remarks made in §4.5 it is apparent that the programmer
may more than the mere ability to
require on occasion considerably
translate mathematical processes into machine code. To take full
advantage of his machine he needs sufficient vision and mathematical
skill to design new methods of solving old problems, while if he is

dealing with engineering calculations he also needs a sound back-

ground of engineering knowledge. At present there is a serious
shortage of engineers who have had a mathematical training suitable
for programming work.
It is often assumed that an automatic computer is only useful in
solving large problems. This may be true in the case of 'single'
calculations, but it is doubtful if the assumption holds for problems
which can be treated by standard programmes. In many cases a
decision as to which is the best method of attack depends not so
'
'

much on the relative merits of hand and automatic machines as on

the relative abilities of the human beings available to use them. The
consequences of human incompetence are amplified rather than con-
cealed by the speed of an electronic machine, and it is vitally im-
portant that the human organization built round a computer should
be able to make the most of its potentialities. A machine may be
able to do in an hour what would take a human being six months,
but the value of what is done in that hour remains entirely a human
responsibility.

50
4.7.Future prospects
The early machines were relatively slow, used large quantities of
equipment, and were often unreliable. Fortunately, however, com-
puter development followed a pattern familiar in other branches of
engineering.Speed and reliability improved, while cost, size and
power consumption decreased. The open racks and trailing wires of
the laboratory-built machines gradually gave way to the enamelled
cabinets and neatly arranged control panels of their factory-built
successors. Many different types of computer are now on the market,
ranging from very fast machines with large storage capacities designed
for general scientific calculations, to simple computers with rudi-
mentary programming facilities intended for use in accountancy and
other clerical work.
While the engineering side of computers has progressed, mathe-
maticians have been exploring the potentialities of the new machines,
and developing the art of programming. What was at first an abstruse
specialism only practised by a few people has become a subject in
which regular courses are held at many universities. Programming
itself has become easier; this is partly because of accumulated

experience, and partly because modern machines offer many facilities

to the programmer which their predecessors lacked. On some com-
puters learning to programme now takes only a few weeks, as com-
pared with the three or four months necessary on earlier machines.
The computers of the future will doubtless be better than those
available at present, but the changes will probably be ones of detailed
design rather than of fundamental principle. Magnetic cores are
already coming to be regarded as the standard elements for the high-
speed stores of large machines, while transistors are also likely to
play an important part in future designs.
Developments are more likely to be spectacular in the field of
applications. On the research side there are many problems, parti-
cularly in the field of partial differential equations, which tax severely
the speed and capacity of the largest machines yet built. At a lower
level, the development of standard programmes similar to those
described in this chapter may lead to radical changes in the routine
calculations carried out in industry. Another new and interesting
application of computers is and com-
in the control of industrial
mercial processes. The idea of an automatic factory directly con-
trolled by a special-purpose computer no longer belongs exclusively
to the realms of science fiction, while the mechanization of clerical

51
work in large offices has created ademand for medium-sized com-
puters, usually with complex input-output equipment.
The digital computer is now firmly established among the many
other tools available to the engineer, and it is natural that this fact
should be reflected in his education. This does not mean that every
student should be taught how to programme —programming is after
all merely a technique which can be learnt fairly quickly when
required. It is far more important that he should know something of
the mathematical techniques most useful in computer work, so that
later on he will be in a position to judge whether his own particular
problems are suitable for machine solution.

52
 :

REFERENCES FOR FURTHER READING

The following are merely a selection from the literature available

General principles

Bowden, B. V. (Ed.). Faster than Thought (Pitman, 1953). A popular

account of digital computers, including a general survey of their
present and future applications.
Hartree, D. R. Calculating Instruments and Machines (Cambridge
University Press, 1950). A general introduction also covering
certain types of analogue computer.

Booth, A. D. and Booth, K. H. V. Automatic Digital Calculators

(Butterworth, 1953).
Wilkes, M. V. Automatic Digital Computers (Methuen, 1956).
These two books are more technical and are largely concerned with
details of machine design. A few applications are described.

Applications to particular problems

Brooker, R. A. Application of digital computing techniques to physics.

Brit. J. Appl. Phys. vol. 4, p. 321 (November 1953).
Hunt, P. M. The digital computer in aircraft structural analysis.
Aeronaut. Engng, vol. 28, p. 70 (March 1956).
Live sle y R , . K
The application of a digital computer to some problems
.

of structural analysis. Struct. Engr, vol. 34, p. 1 (January 1956).

Locke, W. N. and Booth, A. D. The Machine Translation of Languages
(John Wiley, New York, and Chapman and Hall, London, 1955).
Convention on Digital Computer techniques, London, April 1956.
Proc. Instn Elect. Engrs, vol. 103, part B, Supplement no. 1. The
section entitled 'Engineering and Scientific Applications' contains
nine papers describing the application of computers to electrical
engineering problems.

53
 o{
e&*