Digital PDFs

XX-5F607-77

November 1975

14 pages

Original

1.1MB

Document:	Bell Strecker What we learned fm PDP-11
Order Number:	XX-5F607-77
Revision:
Pages:	14
Original Filename:	Bell_Strecker_What_we%20_learned_fm_PDP-11c%207511.pdf

OCR Text

ABSTRACT
Gordon Bell,
November 8,

William
1975

Il.

Strecker

COMPUTER STRUCTURES:
WHAT HAVE WE LEARNED FROM THE PDP-ll?
Over

the

FDP-11’S
six
year
life
20,000
specimens
have been
built
based on 10 species
(models).
was a design
goal,
Al though
range
the
actual
it was unquantified;
has
exceeded
expectations
range
(5OO:l
in memory
size
and system
The range has stressed
the
price].
mini (mall
computer
baa ic
architecture
along all dimensions.
The marn PM.5 structure,
i.e.
the
has
been
adopted
as a de
UNIBUS,
facto
standard
of
interconnection
and minicomputer
many micro
for
The
architectural
systems.
experience
gained
in the design
and
use of the PDP-11 will
be described
terms
Of
its
environment
in
(initial
goals
and
constraints,
the
technology,
and
organization
that
designs,
builds
and
distributes
the machine).

behave
in
a particular
Where
does it get inputs?
does
it
formulate
and
problems?

about

1.0

one might
think
that
computer
architecture
is the sole
determinant
of
a machine,
it
is
merely
the
focal
point
for
a
specification.
A computer
is
a
product
Of
its total
environment.
Thus
to
fully
understand
the
POP-11,
it
is
necessary
to
understand
its environment.
Org.
shows
Figure
the
various
groups
(factors)
affectrng
a
indicate
the
COmQUte r . The lines
flow
of
primary
information
for
product
functional
behavior
and for
QrOdUC t
specifications.
The
physical
flow of
goods
is
nearly
the
same 1 ines,
but
more
along
starting
with
direct:
aQQ1 ied
technology
semiconductor
(e.g.,
manufacturers),
going
through
computer
manufacturing,
and finally
to
the
service
personnel
before
turned
over
to
being
the
Einal
user.

The

as they

affect

development
is
the
of
makes

The
final

user,
who
OUtQUt.

receives

the

BACKGROUND

nature
of
engineer
ing
be goal oriented--the
much
no exception,
with
on deliverable
products.
Hence,
it is difficult
to plan
for
lifetime.
a
long
and extensive
more
Nevertheless,
the 11 evolved
rapidly
and over a wider
range than
we expected,
placing
unusual
stress
on even a carefully
planned
Systemhas evolved
under
The 11 family
market
and
implementation
group
pressure
to build
new machines.
In
been
this
the
plannrng
has
way
with
asynchronous
and
diffuse,
A
distributed
development
.
decentralized
organization
provides
checks and balances
since
it
is not
point,
all under a single
control
of
of ten
at
the
expense
the
compatibility.
Usually,
hardware
has been designed,
and the
software
1.5 modified
to
provide
compatibility.

technology--it
is
to
understand
the
components
that
are
available
to build
from, as they directly
affect
the
resultant
designs.

nature
that

It is
projects
11 is
pressure

basic

The
organrzatron--what
fundamentai
organrzatron

DEC
The
rest
of
the
includes
organization--this
assoc ia ted
applications
groups
with
market
groups # sales,
service
and manufacturing.

2.0

important

There are an ever increasing
number
of
groups
who feel
compel led to
them
control
all products
bringing
all
common
norm :
the
government
(“5) , testing
groups
such
as Underwriters
Laboratory,
and the
voluntary
standards
groups
such
as
ANSI
and
CBEWA. Nearly
all these
some
groups
affect
the
design
in
or another
(e.g.
by reguir
ing
way
time).

Although

parts,
are:

HOW
solve

Note,
that
if
we
assume
that
a
QrOduc t is
done sequentially,
and
each stage has a gestation
time
of
about
two
years,
it
takes roughly
basic
eight
years
for an idea from
research
to
finally
appear
at the
user’s
site.
Other
organizations
affect
design :
ala0
the
competitors
establish
a
(they
level
and determine
the
deaign
product
life):
and government
IsI
and standards.

TNTRODUCTTON

The relevant
the desrgn

way?

the
it

138

the

plannrnq,
the
Indepenaen
t of
the
machrne
has been very successful
in
and
wrth
the
the
marketplace,
written
for it.
sys terns
prbgrams
In the DaDer (Bell
et al,
1970)
we
with
market
f irs’t
concerned
are
Features
and
use.
acceptance
carried
to other
designs
are also
a
contributes
to
measure
of how
it
and
are
of
structures
computer
secondary
importance.

made to increase
its address
space.
In retrospect,
rt
IS clear
that
since
memory pr Ices decline
at 26%
to 413 per
and
many
users
year,
tend
to
constant
dollar
buy
systems,
then every
two
or
three
year 5 another
bit
is requrred
for
the
physical
address
space.
-

successful
in
The PDP-11 has been
with
over 20,000
the
marketplace
computers
in use
(1970-1975).
It
unclear
how rigid
a test
(aside
i:orn the marketplace)
we have given
since
a large
and
the
design
and
sales
aggressive
marketing
coupled
with
sof twace
organization,
architectural
to
cover
inconsistencies
and omissions,
can
There
was
save almost
any design.
difficulty
in teaching
the machine
this
required
a
to
new
users:
effort.
On the other
large
sales
various
machine
and operating
hand,
still
are to
capabilities
systems
be

2.1

Lack of
stacks
(weakness
3) has
been
solved,
uniquely,
with
the
auto-increment/auto-decrement
addressing
mechanism .
Stacks
are
used extensively
in some languages,
and generally
by most programs.

used.

GOALS AND CONSTRAlNTS

Weakness
2 of not enough
registers
was
solved
by
providing
eight
16-bit
registers:
subsequently
six
more
32-bit
registers
were added
for floating
point
arithmetrc.
The
number
of
registers
has
proven
adequate
.
More
registers
would
just
increase
the context
switching
time,
and
also
perhaps
the
programming
time
by
posrng
the
allocation
dilemma
for
a compiler
or a programmer.

Weakness
4, limited
interrupts
and
slow
con tex t switching
has
been
generally
solved
by the
11 UNTEIJS
vectors
which
direct
interrupts
when a request
occurs
from a given
I/O device.

- 1970

1970)
The
paper
(6ell
et
al,
beg inning
described
the
design,
with weaknesses
of minicomputers
to
remedy other
goals
and constraints.
These will
be described
tr iefly
in
provide
a
this
section,
to
framework,
but most
discussion
of
aspects
of
the
the
individual
machine
will
be described
later.

Byte
hand1 ing
(weakness
5)
was
provided
by direct
byte
addressing.
iiead-only
memory (weakness
6)
can
used
directly
without
special
programming
since
all
procedures
tend to be pure
(and reentrant)
and
can be programmed
to
be
recursive
(or multiply
reentrant).
Read-only
memories
are used
extensively
for
bootstrap
loaders,
debugging
programs,
and
normal
now provide
console
functions
(in
program)
using
a standard
terminal,
be

address
Weakness
1, that of limited
for
capability,
was
solved
its
rmmediate
future,
but not with
the
have
been.
finesse
it
might
this
has
been
a costly
Indeed,
oversight
in redundant
development
and sales.

Very

elementary

(weakness

by a better
but
so
far,
have not been

can
is only one mistake
that
in a computer
design
that
be
made
recover
from--not
is difficult
to
providing
enough
add:zis
bits
for
addressing
memory
memory
followed
management _
The
PDP-11
the unbroken
tradition
of
nearly
computer.
Of course,
every
known
rule
of
there
is
a fundamental
other )
computer
(and
per haps
designs
which
helps
to
alleviate
well-designed
this
problem:
any
machine
can be evolved
through
at
least
one
major
change.
It
is
extremely
embarrassing
that
the
11
with
memory
had
to
be
evolved
management
only
two years
after
the
the
paper
was
written
outlining
goal of providing
increased
address
space.
All predecessor
DEC designs
have suffered
the same problem,
and
only
the PDP-10 evolved
over a ten
year period
before
a change
was
There

processing
I/O
is partially
provided
interrupt
structure,
I/O processors
per se
implemented.

Weakness
8 suggested
that
we must
have a family.
Users would
like
to
move
about
over a rangyof
models.
High programming
costs
(Leakness
9)
should
be
addressed
because
users
program
in machine
language.
Here
we
have
no
suggest
data
to
improvement.
A
reasonable
comparison
would
be
programming
costs
on an
11
versus
other
machines.
We
built
more complex
systems
(e.g.,
operating
sys terns,
computers)
with
11 than
with
the
simpler
structures
(e.g.
PDP-8
or
15).
Also,
some
systems
programmrng
is
done
usrnq
higher
level
languages.

139

Another
constcalnt
lenq th
had
to
eight
bits.
While
expensive
within

lnvestmenc

the

was
this

has

DEC because

in 12,

and

least
two
organizations
have
machines
with similar
bus
and
ISP
structures
(use
of
address
modes,
behavror
of
registers
as
program
counter
and stack);
and a
third
organization
hdS
offered
a
plug-replacement
system for sale.

word
of
been

be in multiples

mdde

of our
36

bit

effect
has
probably
been
war thwhile
.
The
is
quite
notron
of
word
length
meaninqless
in machines
like
the 11
and
the
IBM 360 because
data-types
varying
lengths,
and
are
of
instructions
tend
to
be
in
multiples
of 16 bits.
However,
the
addressing
of
memory for floating
point
is inconsistent.
sys

terns,

the

net

been
The
UNIBUS
structure
has
accepted
by
many designers
as the
PMS
This
structure.
is
interconnect
ion
scheme
especially
used
in
microprocessor
designs.
as part
of
the
Also,
UNIBUS
design,
notion
of
the
I/O
data
and/or
control
mwerng
registers
into
the
memory
address
space
has
been
used often
in the
microprocessor
since
it
designs
eliminates
instructions
in the ISP
and requires
no
control
to
extra
the I/O section.

Structural
flexibility
(modularity1
was
an
impor tan t
goal.
This
succeeded
beyond expectations,
and
is
discussed
extensively
in
the
part
on PMS,
in
par titular
the
UNIBIJS section.

Finally,
we were concerned
in
1970
that
be
there
would
offsprings--cledrly
no
problem:
there
have
been
about
ten
implementations.
In
fact,
the
family
is
large
enough
to suggest
need
of family
planning.

There was not an explicit
goal
of
microprogrammed
implementation.
Since large
read-only
memories
were
available
at
the time of the
not
Model
20
implementation,
microprogramming
was
not
used.
Unfortunately,
all
subseouen
t
machines
have been microproqrammed
but
with some additional
difficulty
and expense
because
the
initial
design
had
poorly
allocated
but
more
opcodes,
impor tan t the
condition
codes
behavior
was
over
spacified.

3.0

The computers
we build
are strongly
influenced
by the basic
electronic
case
of
technology.
In
the
information
computers
, electronic
processing
technology
evolution
has
been
used
to
mark
the
four
generations.

was also
stated
that seems
to have
been missed.
The initial
handbook
was
terse
and as such
the
machine
saleable
to
those
was only
who
really
understood
computers.
It is
not clear
what the distribution
of
first
users
was, but probably
all
had previous
computing
experience
.
A large
number
of
machines
were
sold
to
extremely
knowledgeable
users
the
universities
and
laborato:?es.
The second
handbook
cdme out
in
1972
and helped
the
learning
problem
somewhat,
but
it
is
still
not clear
whether
a user
with
no
previous
computer
experience
can determine
how to use
a machine
from the
information
in
the
handbooks.
Fortunately,
two
comw.8 ter
science
tax tbooks
(E&house,
75;
and Stone
and
Siewiorek,
75)
have
been
written
baaed
on
the
11 to assist
in the
learning
problem.
Understandability
to
be
a

TECHNOLOGY

goal,

3.1

Effects
Memory
De8igns

computer

beqdn

1969

Core

memory

(program)

eventual

FEATURES THAT HAVE MIGRATED TO
OTHER COMPUTERS AND OFFSPRINGS

A suggested
test
was the features

(Hell
that

et al
1970)
have migrated

into

designs.

We.

competitive

reluctant

use

VariOUS

1dedS.

140

with

the
primary
for
with
an
memory
trend
memory.

toward

A comparatively
smdll
number
of
high
registers
for
speed
general
processor
state
(i.e.
registers),
with a trend
toward
register
larger,
higher
speed
files
at
lower
cost.
Note,
only
16
word
read-write
memories
were
avdilableat
design
time.

Unavailability
of
speed
read-only

have

not
fully
permitted
this
test
because
some basic
features
are
hence,
non-DEC
desrgners
patented:
are

.%&conductor
PDP-11
Model

set ies
design
the Model 20.
Subsequently,
3
models
were
introduced
as minimum
cost,
best
cost/performance,
and
maw imum
The memory
performance
machines.
several
technology
in 1969
formed
constraints:
The PDP-11

semiconductor
2.2

Of
The

large,
h iqh
memories,

microprogrammed
permitting
a
the design
of the
approach
to
for ca
Note, not
control
part.
read-only
memory
was
paper,
al though
slow,
unavailable
read-only
MOS was available
for
character
generators.
These constraint5
desrgn
following
attitudes:

established

principles

It should
be asynchronous
and
accepting
various
capable
of
conf iqurations
of
memories
in
sire
and speed.

to
should
be expandable,
take
advantage
of an eventually
laraer
number
of resisters
for
data-types
and
improve
more
context
switching
time.
Also,
would
Fermi t
registers
more
memory
to
eventually
mapping
machine
and
provide
a virtual
protected
multiprogramming.
It

It could
be relatively
complex,
so that
an eventual
microcode
approach
could
be
used
to
advantage
.
New
data-types
added
to
the
could
be
set
to
increase
rnstruction
though
performance
even
they
added complexity.

The UNIBUS width
would
be
relatively
wide,
to get as much
performance
as possible,
since
LSI
was not
yet available
to
encode functions.

3.2

Variations
In
PDP-11
Through
Technology

Models

Improve

performance
through
force
with
faster
The 11/45 and
11/70
memories.
bipolar
fast
MOS
uses
and
memory.
brute

Microprogramming
(see
improve
performance
more
complex
ISP
floating
point).

Mu1 tiple
copies
of
processor
state
(context)
to improve
time
to switch
context
among various
runnlnq
programs.

Additional
additional
floatrnq

point

by
one

another.

Improve
performance
by mapping
multiple
programs
Into
the same
physical
memory,
giving
each
program
a
virtual
machine.
Providing
the last
two points
requires
;I signif
icant
increase
in
the
number
of
reqlsters
(i.e.
at
least
64 word fast
memory arrays).

4.0

THE ORGANIZATION

OF PEOPLE

The

two usual
problems
of
design
:
inexperience
and
“second-systemitis”.
The
first
problem
is
simply
a resources
problem.
Are
there
people
available?
their
What
are
backgrounds?
Can a small
grow
work
effectively
on architectural
definitions?
Perhaps
most
important
is the principle,
that no
matter
who is
the
architect,
the
design
must
be clearly
understood
by at least
one person.
are

below)
to
through
a
(i.e.,

registers
data-types--i.e.,
arithmetic.

from

reliabilrty
(protectrngj

Three
types
of
design
are
based
both on the technology
and the cost
and
performance
considerations.
The
nature
of
this
tradeoff
is
shown in Figure
DS. Note,
that one
starts
at
0 cost and performance,
proceeds
to add cost,
to achieve
a
base
of
Iminimum
level
functionality).
At
this
point,
certain
minimum goals
are met:
for
the computer,
it
is
that
simply
there
is
program
counter,
and the
simplest
arithmetic
operations
can
be carried
out.
It is easy to show
(baaed on t:iwTurinq
machine)
that
only
a
Instructions
are
required,
and
from
these,
any
program
can be written.
From this
minimal
point,
performance
increases
very
rapidly
in a step
fashion
(to be described
later I for
quite
sometime
(due
to
fixed
overhead
of memories,
cabinets,
powar,
etc.)
to
a
point
of
inflection
where the cost-effective
solution
is found.
At this
point,
performance
continues
to
increase
until
another
point
where
the
performance
maxim ixed .
is
Increas rng
the
size
imp1 ies
physical
constraints
are
exceeded,
and
the
machine
becomes
unbuildable,
and
the
performance
can
go
to
0.
There is a general
tendency
of all designers
to
“n+l”
(i.e.,
incrementally
add
to
the
design
forever)
. No design
is
so
complete
, that
a redesign
can’ t
improve
it.

Semiconductor
memory (read-only
and
resd-wr i te)
were
used to tradeoff
coat
performance
across
a
reasonably
wide
range
of models.
Various
techniques
baaed
on
semiconductors
are
used
in
the
tradeoff
to provide
the
range
.
These include:
1.

the

the
and

Improve
lsolatlng
program

for

Second-systemitis
Of
def ininq
of
past

141

is the phenomenon
a system on the basis
sya tern
history.

the
Invar Iably,
problems..
past
unbulldable.

4.1

PDP-11

Some of
initially

system
. border

solves
ing

for securrty)
project
started
up,
design
and
planning
were
disarray,
since
Data
General
hi:
been
formed and was competing
with
the PDP-8 usrng
a very small
16-bit
computer .
Al though
the
Product
Line
former
Manager,
d
engineer
the
(NM)
for
PDP-8,
had
the
responsibility
this
time,
the
new
project
manager
was
mathematrcian/programmer
followeda
by
another
manager
(RC)
who had
managed the PDP-8.
Work
proceeded
for several
months based on the DCM
and
with
a
design
review
at
Carnegie-Mellon
University
in late
1969.
The DCM review
took
only
a
few minutes.
Aside from a general
dullness
and a feeling
that
it
was
too little
too late
to compete.
It
was
difficult
to
program
(especially
compilers).
by
However,
it’s
benchmark
results
were
good.
(We believe
it had been
tuned
to
the
benchmarks,
hence
couldn’
t do other
problems
very
well.)
One of
the
designers
(HM)
brought
along
the
kernel
of
an
which turned
out to be
alternative,
the
PDP-11.
We worked
on
the
design
all weekend,
recommending
a
switch
to the basic
11 design.

all

the

Exper xnce

the

PDP-11 architecture
was
carr red
out
by
at
Carnegie-Mellon
University
(HM with
CBI . Two of the useful
ideas:
the
UNIBUS,
and the
use
of
general
registers
in
a substantrally
more
fashion
general
(e.g.
as stack
came
out
of earlier
work
pointers)
(CB) at CMU and was described
in
STRUCTURES
COMPUTER
(Bell
and
During
the detailed
Newell,
1971).
design
amelioration,
2 persons
(HM,
responsible
for
and RC) were
the
specification.
Although
the architectural
activity
Of
the II/20
proceeded
In parallel
with
the implementation,
there
was
less
interaction
than in previous
designs
where
the
first
DEC
and
architecture
implementation
out
by
were
carried
the
same
As a result,
person.
a slight
penalty
was
paid
to
build
subsequent
designs,
especially
vis
a vi5 microprogramming.

this
point,
there
were
reviews
to
ameliorate
the design,
and each
suggestion,
in effect,
amounted
to
an n+l ;
the
implementation
was
proceeding
in
parallel
and
(JOI
since
the
logic
design
was
conventional
, it was difficult
to
the
tradeoff
extensions.
Also,
design
was constrained
with
boards
and
ideas
held over from the DCM.
(The only
safe
way to
design
a
range
is
simultaneously
do both
During
high and low end designs.)
the summer at DEC, we tried
to free
increased
space,
and
oP code
(nil’edl
the UNIBUS bandwidth
(with
an extra
set of address
lines),
and
outlined
alternative
models.
At

As the various
models began
to be
outside
built
the
original
POP-11/20
nearly
all
group,
architectural
control
(XC)
disappeared,
and
the
architecture
by more
people,
and
was
managed
de5ign
resided
with no one
person!
A similar
loss of control
occurred
in the design
of
the
per iphecals
after
the basic design.
The
first
designs
for
16-bit
came
from a group placed
computers
PDP-15
management
under
the
marketing
person,
with
engineeri:;
background).
It was called
PDP-X,
and did
include
a range.
As a
range
architecture,
it
was better
thought
out than the later
PDP-11,
but
didn’t
have
the
innovative
Unfortunately,
this group
aspects.
was
intimidating,
and some members
lacked
credibility.
The group also
managed to convince
management
that
machine
was potent ially
as
the
the
PDP-10
(which
it
complex
as
no one wan ted
wasn’t)
;
since
another
large
computer
disconnected
from
the
main
business,
it was a
sure
suicide.
The
(marketing)
management
had little
understanding
of the machine.
Since
the
people
involved
in
the
design
were
apparently
simultaneously
descgnrng
the PDP-X was not of
Data
General,
foremost
rmpor tance.
As

the

GCM

FDP-X

(for

pro]ect

@esk

folded

Calculator

The
advent
of
large,
read-only
made possible
the various
memories,
follow-on
designs
to
the
ll/ZO.
Figure
“Models’
sketches
the cost
of various
models versus
time,
with
cons is tent
oerformance.
lines
of
design
This very clearly
shows&the
The
11/40
5 tyles
(ideologies)
.
the
design
was started
right
after
11/20,
al though
it was the last
to
market
low
and
come
on the
(the
high
ends
had
higher
priority
to
they
into
product ion
as
get
Both
the
extended
the
market).
11/04
and 11/45 design
groups
went
through
extensrve
buy
in processes,
first
as they came into
the 11 by
alternative
designs.
In
proposrng
11/45,
a larger,
the case of the
11-like
18-bit
machine
was proposed
by the 15 group:
and
later,
the
LINC
engineerrng
group proposed
an
alternative
design
which was subset
at the symbolic
program
compa table
level.
As the
groups
constdered

and
the
Machine

142

computer
rnterconnectron
standard.
Al though
it
has been difficult
to
fully
specrfy
the UNIBUS such
that
one
can
be certain
that a grven
system
will
work
electrically
and
without
missed data,
specification
is the key to the UNTBUS. The bus
behavior
specification
is
a yet
unsolved
problem
in dealing
with
complex1 ty--the
best
descriptions
are based on behavror
(i.e.,
timing
diagrams).

the
software
ramificatrons,
buy-in
Frgure
Models shows the
was rap&d.
mrnrmum cost-orrented
group has two
lower
cost
successors
providing
and the
higher
performance)
(yet
have
same cost wrth the ability
to
laroer
memories
and perform
better.
thesecame
from
a
Note,
both of
strategy
to
the
LSI-11.
backup
read-only
These-come
from
larger
increased
memor ies,
and
implement
understanding
of how
to
the 11.
The 11/7O is, of course,
follow
on to extend
the
of the 11/45.

5.0

There are also
problems
with
the
deEi9n
of
the
UNIBUS.
Al though
parity
was assigned
as two
of
the
bits
on the bus
(parity
and parity
is
available),
it
has
not
been
widely
used.
Memory
parity
was
implemented
directly
in the memory,
since
checking
required
additional
time.
Memory
and UNIBUS parity
is
a
good
example
of
nature
of
engineering
optimization.
The
tradeoff
is
one
of
cost
and
decreased
performance
versus
decreased
service
cost
and more
data integrity
for the
user:
The
engineer
is
usually
measured
on
production
cost goals,
thus
parity
transmission
and
checking
are
clearly
a capability
to be
omitted
from design . ..especially
in view of
loet
performance.
The
internal
Field
Service
organization
has been
unable
to guantify
the increase
in
service
cost savings
due to shorter
MTTR by
better
isolation.
fault
Similarly,
many
of
the transient
errors
which parity
detects
can
be
detected
and corrected
by software
device
driver8
and
backup
procedures
without
parity.
With
lower
cost for logic
and
increased
r88pOnEibility
to include
(scope)
warranty
C08tS
a8 part
of
the
product
deeign
cO8t
forces
much
more
checking
into
the design.

a natural
performance

PMS STRUCTURE

give
an
section,
we
the
evolution
of the
of
its
PMS
PDP-I1
in
terms
of
structure,
and compare
it
with
expectatrons
(Bell
et
al.
1970).
the UNIBUS
aspects
include
:
The
performance:
structure;
UNIBUS
architectural
use
for diagnostics:
required:
and
control
and multi-processor
multi-computer
computer
structures.
In

this

overview

5.1

The UNIBUS - The Center
Pm structure

Of The

the UNIBUS has
behaved
In general,
expectations,
acting
as a
beyond
standard
for intercommunication
of
Several
hundred
types
peripherals.
of memor les
and peripherals
have
to it.
It has been
been
attached
PMS interconnect
ion
the principle
of MD-PC and oeriuherals
for
media
systems
in the range- 3K dollars
to
1OOK dollars
(1975).
For larger
supplementary
buees
for
sys terns
PC-Mp
and HP-MS traffic
have been
added.
For
very
small
sys terns,
like
the
LST-11,
a narrower
bus
((j-bus)
has
been
designed.

The
interlocked
nature
of
the
transfers
AS such that there
is a
deadlock
when
two compu thr s are
joined
together
using
the UNIBUS
window.
With the window a computer
can
map another
computer’s
address
space into
its own address
space in
a
true
mu1 tiprocessor
fashion.
De8dlock
occurs
when
the
two
computers
simultaneously
attempt
to
accees
the
other’s
addresses
through
each window.
A request
to
the window is in
progress
on
one
UNIBUS , and at
the
same time
a
request
to the other
UNfBUS
is
in
progress
on the requestee’s
UNIBUS,
hence
neither
can
be
request
answered,
causing
a deadlock.
Cne
or
both
requests
are
aborted
and
the
deadlock
is
broken
by having
the UNIBUS time out since
this
is
eguivalen
t
to
a
non-existent
address
(e.g.,
a memory).
In
this
the
sys tern
recovers
and
way
requests
can be reissued
(which may
cause deadlock)
. The UNIBUS window
is conf rned to
appi ica trons
where
there
likely
to
be a low
deadlock’sate.

The UNIBIJS by being a standard
ha8
provided
us with a PMS architecture
for
easily
configuring
systems;
other
organization
can
also
any
interface
build
components
which
the bus... clearly
ideal
for buyers.
make good
Good bU88eS
Istandards)
neighbors
( in
terms
of
engineering),
since
people
can
concentrate
delrign
in
a
fasE:on.
structured
Indeed,
the
UNIBUS
has
created
a complete
secondary
industry
dealing
in
alternatrve
sources
of supply
for
memorres
and peripherals.
Outside
of
IBM
the
360
r/o
Hultiplexor/Selector
bus,
the
UNIBUS
1s the
most
widely
used

143

5.2

UNIBUS
Optimality

and

Performance

5.3

always
more
Al though
we
want
performance
on one hand,
there
is
pressure
to have
lower
an equal
Since
cost
and peformance
cost.
are almost
totally
correlated
the
two goals
perfectly
conflict.
The
UNIBUS has turned
out to be optimum
over
wide
dynamic
range
of
product:,
(argued
below) . However,
the
lowest
site
system,
the
trbus
has
been
introduced,
which
about
l/2
the number of
contains
and at
the
conductors:
largest
systems,
the
data
path width
for
the processor
and memory
has
been
increased
to
32-bits
for
added
performance
although
the UNIBUS
is
still
used
for communication
with
most I/O controllers.

Evolution
Of Models:
Versus
Actual

Predicted

The or iq inal
prediction
(Bell
et
al,
1970)
was
that
models
with
increased
performance
would
evolve
using:
increased
width
for
path
data:
multi-processors:
and
wpar a ted
structures
for
bus
control
and data
transfers
to
secondary
and
tertiary
memory.
Nearly
all of these forms
have been
used,
though
not
exactly
as
predicted.
this
points
to
(Again.
lack
of
architectural
overall
planning
versus
our willingness
and
belief
that
the
suggest ions
and
plans
for the evolution
must
come
from the implementation
groups.1
In the earlier
11/45,
a separate
bus
was added for direct
access of
either
bipolar
(300ns)
or fast
MOS
(400ns)
memory.
In general,
it was
assumed that these
memories
would
be small,
and the user would move
the important
part of his algorithm
to
the
for
direct
fast
memory
11/45
provided
3
execution.
The
second
LJNIHUS
direct
for
transmission
of information
to
the
PC
fart
memory
without
interference.
The
rr/rs
also
increased
performance
by adding
a
second
autonomous
data
operation
Point
unit
called
the
Floating
Processor
(actually
not
boti
processor)
.
In
this
way,
integer
and
floating
point
computation
could
proceed
concurrently.

interconnection
schemes
Since all
are highly
constrained,
it is clear
future
lower
and
higher
that
systems cannot
be accomplished
from
a single
design
unless
a very
low
cost,
high
performance
communication
media (e.g.
optical)
is found.
The optimaiity
of the UNIBUS comes
about
because
memory size
(number
of address
bits)
and
traffic
I/O
correlated
with
the processor
are
Amdahl’s
rule-of-thumb
speed.
for
IBM computers
(including
the 360)
one byte of memory is required
instruction/set
and one bit of
E
I/O
is
required
for
each
executed.
For
our
instruction
applications,
we believe
there
is
more
computation
required
for each
memory word, because
of
the
bias
toward
control
and
scientific
there
has been
applications.
Also,
less
use
of
complex
instructions
typical
of
the
IBM
computers.
Hence, we assume one byte of memory
required
for
is
each
two
executed,
instructions
and assume
one byte
of I/O is an upper
bound
real
time
applications)
for
(for
each instruction
executed.
In
the
FDP-11,
an
average
instruction
accesses
three
to
five
bytes
of
memory, and with one byte of io, up
to six bytes of memory are accessed
instruction/set.
for
each
Therefore,
a bus which can
support
two megabytejsec
traffic
permits
instruction
execution
rates
of
.33
to
1 5 mega instruction/set.
This
imputes
to meory sizes
of
.16
to
.2s
megabytes:
the
maximum
allowable
memory
is
.3
to
,256
megabytes.
6y using
a cache memory
with
a processor,
the
effective
memory
processor
rate
can
be
increased
to
further
balance
the
Alternatively,
processor.
faster
floating
point
operations
will
or inq
the
balance
to be more like
the
IBM
data,
requiring
more
memory.

The 11/70,
a cache based processor.
is
a logical
extension
of using
fast,
local
memories,
but
without
need
for
expert
movement of data.
It
has
a memory
path
width
of
32-bits,
and
the
control
portion
I/O
transfers
and data portion
of
have
been
separated
as originally
suggested.
The
performance
limitation
of
UNIEIIJS are
the
removed,
since
the second Mp system
up to
permits
data
transfers
of
that
five megabytes/set
(2.5 times
that
a
of
the
UNIBUS) .
Note,
peripheral
memory
map control
is
needed
since
Mp address
space (two
meaabvtes)
exceeds
the UNIBUS.
In
this
- way,
direct
memory
access
devices
on the UNIEUS transfer
data
into
a mapped portion
of the larger
address
space.

5.4

Multi-processor
Structures

Computer

Although
it is not surprising
that
multi-processors
have not been used
specialized
except
on a highly
basis,
it
is
depressing.
In
Computer
Structures
(Bell
and
out
3n
Newell,
711 we carried
144

the
IBM
analysis
of
360,
multi-processor
predicating
a
The range
of
performance
aeslgn.
the
PDP-11
models
is
covered
by
with
the
substantially
worse than
compe t I t Lve
al though
the
360,
environment
of the two companies
is
substantially
different.
For the
smaller
models
appear
to
360,
perform
worse
than the technology
would
predict.
The
reasons
why
mu1 t iprocessors
have
not
materialized
may be:
1.

2
3
40

Planning
and
technology
are
Within
DEC, not
asynchronous.
all products
are
planned
and
built
at
a particular
time,
hence,
it is difficult
to
get
the
one
right
time
when
a
multiprocessor
would be better
than
an existing
Uniprocessor
together
with
one
or
two
additional
new processors.

Multiprocessors

PRICE/ SYS
PEW*

er1ce

3
3.23
3.47
3.35

PfQP*
1
.58
.Q8
.49

system

1.47

1.35

.66
-61
.6

Price/

only

System,assumrnq

l/3

The most
extensive
mu1 tiprocessor
structure,
has
been
C.mmP,
described
elsewhere.
Hopefully,
convincing
arguments
will
be
forthcoming
about
the effectiveness
of
multiprocessors
from this
work
in
order
to
establish
these
structures
on an applied
basis.

6.0

THE TSP

Determining
an ISP
is
a design
problem.
The initial
11 design
was
based substantially
on benchmarks ,
and
as previously
indicated
this
approach
yielded
a predecessor
(not
buil tl
that though
performing
best
the
six
benchmarks,
was
dOlfficu1
t
to program
for
other
applications.

Incremental
market
demands
specific
new machines.
require
products,
a
By having
more
company
can
better
track
specific
compe t i tor s
by
unipracessors.

Existent

PC.
PRICE
1
1.23

The second
type of structure
g iven
in
Figure
MP is
a convent ional
multiprocessor
using multiple
port
memories.
A
number
of
these
systems
have
been
installed
and
operate
guite
effectively,
however,
they
have
only
been
used
for
specialized
applications.

available
more
new processors
than
already..

*PC cost
* Total

1
1.85
2.4
2.25

multiple
srngle

From these results
we would
expect
to
use
up to three
processors,
to
give
the performance
of a model 40.
More
processors,
while
increasing
the
performance,
are
less
cost-effective.
This
basic
structure
has
been
applied
on a
production
basis
in the GT4X series
of graphics
processors.
In
this
scheme,
a second P.display
is added
to the UNIBUS for
display
picture
maintenance.

build
a better
We can always
special
processor.
single,
This
design
philosophy
stems
from
local
optimization
of the
and
designed
object,
ignores
global
costs
of
spares,
reliability
and
training,
the
of
user
ability
the
to
adjust
dynamically
a
configuration
to his load.
are
for
build

PC.
PERF

HP
.6
1.15
1.42

Pc.cost

The market
doesn ’ t demand them.
Ano the r
deadlock
: how can the
market
demand them,
since
the
market
doesn ’ t even know that
could
such a structure
exist?
not
blessed
the
IBM has
yet
concept.

There
designs
we can

5.4.1

IPC

The basic
nature
of engineering
is to be conservative.
this
is
a classical
deadlock
situation:
learn
how to program
we cannot
until
multiprocessors
such
exist:
a system
canot
systems
be built
before
programs
are
ready.

following
results
for
11/05 processors
sharing
UNIBUS:

6.1

Figure
MP gives
some
of
the
mu1 t lprocessor
sys terns
that
have
been built
on the 11 base.
The top
most structure
has been built
using
11/05 processors,
because
of
but
improper
arbitration
in
the
processor,
the performance
expected
based
on memory contention
didn’t
materialize.
expect
We would
the

General

ISP

Design

Problems

The guiding
principles
for
design
in
have
general,
especially
difficult
because:
1.

145

ISP
been

The range
of
machines
argues
for different
encoding
over
the
At
the
smallest
range.
byte-or
ien ted
a
systems,
approach
with
small
addresses

larger
is
optrmum,
whereas
Lmplementatlons
require
more
addresses
operations.
larger
and encoding
efficiency
can
be
traded
off to garn performance.

variable,
artistrc.
to
have
operations
data-type
integers)
operations

The 11 has
turned
out
to be
(and
hopefully
applied
effective)
over a range of
500
in
sys tern
price
($500
to
$250,000)
and memory
size
(8k
bytes
to 4 megabytes).
The 360
varied
over
a
by comparison
range:
similar
from 4k bytes
to 4 megabytes.
2.

for

data-types
Selection
totally
application.
considers
words
to
be - integers,
yet
doesn ' t have
a full
set
of
operations
for
the byte;
nor
are
the byte and word
the
ops
this
same.
to
BY adhering
pc inc iple , the
compiler
and
human
code
generators
are
greatly
aided.

At
a
style

given
time,
a certain
math ine
of
ISP
is used
because
of the rapidly
varying
technology.
For example,
three
address
machines
were initially
used
to
minimize
processor
the
state
(at
expense
of
encoding
efficiency),
and stack
machines
have never
been
used
extensively
due
to
memory
time
access
and
control
complexity.
In
fact,
we can
observe
that
machines
have
over
time
evolved
to include
virtually
all
important
operations
on
useful
data-types.

We would therefore
ask that
the
elegant,
where
machine
appear
elegance
is a combined
qua1 i ty
of instruction
formats
relating
to
significance,
mnemonic
operator/data-type
completeness
and
and
orthogonality,
addressing
consistency.
BY
general
having
completely
registers)
facilities
(e.g.,
and
which
are
not
context
dependent
assists
in minimizing
instruction
the
number
of
and greatly
aids
in
types,

use
The
machine
vat ies
over
time.
In the case of DEC, the
initial
users
were
sophisticated
and could
utilize
the
power
at
the
machine
level.
language
The
II
provided
more
fully
general
registers
and
was
unique
in the
minicomputer
marketplace,
which
at
the
time consisted
laroely
of 1 or 2 accumulator
machines
with
0 or
1 index registers.
the typical
Also,
m in icompu ter
operation
codes
were
small.
the 11 extended
data-typing
to
the
byte and to reals.
by the
extension
of the
auto-indexing
mode,
the
string
was
conveniently
programmed,
and
the same mechanism
provided
for
stack data-structures.

increasing

understandability
u5efulness).

The
machine
is
applied
into
widely
different
markets.
Initially
the 11 was
used
at
the
machine
language
level.
The
user
base
broadened
by
applications
with substantially
higher
level
languages.
These
languages
initially
were
the
based
scientific
reoister
transfer
languages
such
FORTRAN, CEC’S
BASIC,
FOCAL:
but
the
machine
eventually
began
to
be
applied
in
the
commercral
marketplace
fo.- the
COBOL,
RPG ,
DIBOL.
and
BASIC-PLUS
languages
which
provided
string
and
decimal
data-types.
The crlterla
In

lnstructlon

for

and borders
on the
Ideal
goals
ace thus
a complete
set
of
for
a q iven basic
(e.g.
--completeness,
and
would
be
the
same
varying
length
--orthogonality.
of the data-trues
is
a function
oi‘
the
That
is, the 11
both
bytes
and
full

a capability
set IS hlqhly

146

the
(and

Techniques
for generating
code
the human and compiler
vary
by
widely.
more
With
the
11.
addressing
modes
are provided
The a
than
any other
computer.
modes
for
source
and
destination
with
dyad ic
operators
provide
what amounts
to
64 possible
instructions:
the Program
and by associating
Counter
and
Stack
Pointer
registers
with the modes, even
more data accessing
methods
are
provided.
For
example,
18
MOVE
instruction
forms of the
can
be seen (Bell
et al,
1971)
as the machine
is
used
as a
two-address,
general
registers
and
stack
machine
program
forms.
(The
price
for
this
generality
is extra
bits).
In
general,
the
machine
has been
general
used
mostly
as
a
register
machine.

very
Basic design
can take the
be highly
general
form
or
and design
decisions
specific,
be
bound
in
some
can
combination
of
microcode
oc
macrocode
with no good criteria
for tradeoff.

Problems
Machine

Several

problems
have
machine
has been

basic

In
Range

Extendlnq

6.2

arisen

The

systems
provide
functions
get additional
segments).

the

7.0

extended:

extension
design
did
free
not
leave
enough
opcode
space for extending
the machine
to increase
the data-types.
The

opera tron-code
initial

SUMMARY

This
paper
has
re-examined
the
PDP-11
and
compared
it
with the
initial
goals
and constraints.
With
hindsight,
we now clearly
see what
the problemswith
the initial
design
were.
Design
faults
occurred
not
through
ignorance,
but because
the
design
was started
too late.
As we
continue
to
evolve
and improve
the
PDP-11 over the next five
years,
it

problem--the

the
11/45
Was
At
the
t lme
was
designed
(FPP
added) ,
several
extension
schemes
were
escape
mode to
examined :
an
the
floating
point
add
operations:
bringing
the
11
convent
ional
more
back
to
a
register
machine
by
general
reducing
the modes and finally,
data
by
adding
a
the
typing
mode which
could
be
global
switched
to
seiec t floating
point
( Lns tead
of
byte
operations).

Will

observe,
is use.

indeed
be interesting
however, the ultimate

to
test

HTELIOGRAPHY
Ames, G.T., Drongowski,
Fuller,
S.H.
Emulating
the
PDP-11/40:
Proc.
COMPCON (19;5,.

the
addressinq
Extend ina
VNTBUS limits
the
range-- the
physical
memory
to
262,144
(la-bits).
the
bytes
the1nll/70,
implementation
of
physical
address
was
th;
4
megabytes
by
extended
to
a UNZBUS map so that
providing
devices
in a 262K UNIBIJS space
could
transfer
into
the
4
space
by
mapping
megabyte
registers.

P.J.
and
the Nova on
case
study.

Bell,
G., Cady, R., McFarland,
H.,
Delagi,
E., O’Loughlin,
J., Noonan,
6. ,
and
Wulf,
w.
A
new
architecture
of minicomputers-the
DEC PDP-11.
Proc.
SJCC (1970)
Vo136,
~~-657-675.
Bell,

C.G.

Computer

and

Structures.

(1971)
Bell,
J.R.
ACM (June
370-372.

While
limits

the
physical
address
are acceptable
for both
the
UNIGUS and larger
sys terns,
for
a single
the
address
program
is still
confined
to an
instantaneous
space of 16 bits,
the user vlc tual
address.

Threaded
1973) Vol

Eckhouse,

Sys terns :

programming

Prentice-Hall,

The main method of dealing
with
relatively
small
addresses
is
via process-oriented
operating
sys terns
that
handle
large
numbers
of smaller
tasks.
This
a
trend
in
operating
is
especially
for
process
systems,
control
and
transaction
processing.
It also enforces
a
structuring
discipline
in the
organization.
(user)
program
operating
The
RSX
series
systems
are organized
this way,
the
need
for
large
and
addresses
except
for
problems
where large
arrays
are accessed
is minimized.

Fusfeld,
progress
Review

Newell,
McGraw

Hill

code.
No.

6, PP

16,

CONM

R.H.
Minicomputer
organization
and
(PDP-11).
(1975)

R. The
technological
function.
Technology
(Feb.
1973) FP. 29-38

McWilliams,
T.,
Sherwood,
W.,
Fuller,
S.,
PDP-11 Implementation
using
the Intel
3000 microprocessor
chips.
Submitted
to NCC (May 1976)
O’Loughlin,
J.F.
Microprogramming
a
fixed
architecture
machine.
Microprogramming
and
Sys terns
Architecture
Infotech
State
of the
Art Report
23.
~~205-224
Ornstein,

1972?

(page

28)

Stone,
H.S.
and Siewiorek,
D.P.
Introduction
to
computer
organization
and data
structures:
PDP-11
Edition.
McGraw-Hill,
(19751

memory
management
The initial
to extend
the virtual
proposal
predicated
on
memory
was
than
dynamrc,
rather
static
assignment
of
memory
segment
current
registers.
In
the
scheme,
the
memory management
are usually
address
registers
considered
to be static
for
a
task
(al though
some
operating

Turn,
R. Computers
Columbia
University
Gulf,
W.A., Bell,
multi-mini-processor.

147

in the
1980’s.
Press 1974.
C.G.,

C.mmp:

FJCC (1972)

I
BASIC REtC;
ADVANCED
DEVELOPMENT

APPLIED
TECHNOLOGY
(E.G. SEMICOt0JCTCRS)

IMPLEMENTATION

I
I
APPLICATIONS
(HARDWARE/
SOFTWARE)

MKT/SALES
4

ARCHITECTURE

Competitors

Governments,
standards,/
testing,
professional
societies

op.

SYS.

LANGUAGES

FlGURE.ORt.

flow

of information

STRUCTURE

(specifications,

OF ORGANIZATION

AFFECTING

ideas,

etc.)

A COMPUTER

DESIGN

USEH

Fig.

DESlcll

STYLES

(IDEOLOGIES)

111 TERIIS

OF COST

AND

PERFORMANCE

Performance

it
/
/

maximum

/
/

maximum
casteffective

/
/
/
/
/
/
1

minimum
level of
functionalit)

.-------

incremented
fixed 8 variable

costs

/
/
/

/
unbuildable

149

Physical
Cons?raints

Price

A--

Fig.

HODELS

PDP-I

I MODELS

PRICE

vE9SIJ;

T’rlE

\llTtI

LINES

OF COttSTA;IT

i’CEt0QtVA’CE

PC
I

PC...
I

a. MUM-PC strucfure

using a sir&

Pdisplry

PC
I

Mp...
I

hip...

used in CT4X

sericr; al~ernr~~wly

P spcciahrcd

(e.g.. FIT)

usq

h1p(CO:lS)
3

muttiporl

KT...
I

KMS...

PC spcrirlvcd

Mp.

S encral:crosspom~:
Pd ii 0: I5 ;‘I I ;4Oj C’
16x16
3
E

d. C.mmp Cblll rnulri-mini.pr@:cssor

Figure bit’

WK...

Unibos.

c. hlultiprnccssor

ET...
I

hlul~i~l’roccssor

Compwcr

SICnibui)

slru;wre.

Compu~cr

S~rw~ures

151

Im~tenren~crj

uiln;

PDP.t t

Ki...

ys...