ved properties of real computer systems. Except for measurement errors, the operational laws pre- sented in this paper apply with complete preci- sion to all ...
FUNDAMENTAL
LAWS OF COMPUTER
J.P.Buzen for Research in Computing Technology Harvard University Cambridge, Mass., 02138 and BGS Systems, Inc. Box 128 Lincoln, Mass., 01773
Center
ABSTRACT A number
device
between
utilization,
ious other
sis.
related
These
the operational
throughput,
space-time
factors
performance.
modeling
response
products
to computer
laws are obtained
method
The operational
nificantly
Table
of laws are derived which establish
relationships
of computer method,
approach,
ved properties
of real computer errors,
consists
Assume
of i second;
of 3 seconds.
Except laws pre-
response
than the average
response response
sented in this paper apply with complete
preci-
sion to all collections
data,
Appendix
and they are similar
to fundamental
in other areas of engineering
K e y Words:
Computer
throughput,
response
operational
method.
laws found
and applied
performance
Most analyses
models.
of computer
benchmarks
Benchmarks,
programs,
simulations,
cisely results nature
specified
of the workload
to be processing, load definition ly different
TIME
RR(Q=2)
B
C
8 2/3
7 1/3
C
B
A
6
6 2/3
B
A
C
6 2/3
6 2/3
Table
In the second same except
sensitive
benchmark to the
and slight
may sometimes
changes produce
average
than average
second
set of benchmark
mark contains
in the work-
response
higher
strong preference
that the system is assumed
line of Table
response
for FCFS,
presents
conclusions.
200
is reversed.
time for RR is 11%
time for FCFS.
results
is the
The
thus indicate
a
even though this bench-
the same set of jobs as the first.
As a point of interest,
significant-
i, everything
that the order of arrival
In this case,
under a pre-
However,
i
of real
or trace driven
system behavior
can be surprisingly
RESPONSE
FCFS
system performance
programs
workload.
re-
preference
A
or probabilistic
are most useful when it is neces-
sary to determine
AVERAGE
FIRST SECOND THIRD
systems,
which may consist
synthetic
the benchmark
a definite
WORKLOAD
evaluation~
time, queueing
Thus
(see
for RR.
science.
i. INTRODUCTION
rely on either
A for details).
suits in line one indicate
time for
at approximate-
time for RR with a quantum of two seconds
of observational
and
The first
Note that the average
for FCFS is 18% higher
of
the order of arri-
val is "ABC" with all jobs arriving ly the same time.
for
that the workload
the three jobs in the case where
to obser-
algorithms
Job A, with a duration
line of Table i gives the average
stochastic
that an ana-
(RR) and "first
Job B, with a duration
Job C, with a duration
sig-
Suppose
scheduling
of three jobs:
7 seconds;
system analy-
the operational
example.
"round robin"
a central processor.
system
systems.
simplified
the crux of the problem with
come first served"(FCFS)
and var-
by using
and directly
i illustrates
lyst is comparing
is based on a set of concepts
naturally
for measurement
a highly
time,
which differs
from the conventional
that correspond
SYSTEM PERFOP~IANCE
yet another
the third line of Table
arrival
sequence
in which
i the
two scheduling average
response
Although plified, real.
algorithms
produce
exactly
and observed
time•
the example
in Table
1 is highly
the dangers which it illustrates
In particular,
specification detail:
benchmark
as a result,
the analyst
where his knowledge
is usually
to confusing
ly equivalent
require
imprecise. where
studies produce
only a single
results
in areas
reason
enable
directly with situations ledge of the workload
the analyst
to deal
Suppose,
case of Table l, that the workload
is known
represented
is uncertain.
with a probabilistic
lyst may be able to formulate tion i n t e r m s suppose
that it is reasonable
the Law of
Theorem~and
intervals)
imply
model,
all cases.
for constructing
This is the major
simulation
programs
structing
that run for long periods
benchmarks
The assertion
the ana-
to assume
that a particular
confidence
level--
enough"--
may be satisfactory
tain performance
solu-
in ~raetiee
assertions
if the experiment
evaluation
are clearly
oriented
to meet all
computer
the
"CBA" will occur 10% of the time,
and the
with sets of data that have been collected
sequence
"BAC" will occur
sequence
as follows:
FCFS response
.60 x 8~- + .I0 x 6
intervals
can be regarded
times can be computed Expected
rect measurement
In
of time.
interested
response
Such individuals
of actual
2 + .30 x 6~
quantities could,
the internal
C. Formulas
the analyst
can compute
-- FCFS response
that RR response
time.
The analyst
benchmark
tests
fail to confirm his prediction;
indeed,
he fully
expects response
to find RR response
higher
of existing
sets
the effect
that certain
to the system or the workload quantities
such as
throughput and response time. Note that empirically mance analysts
than FCFS
time for 10% of the arrival
determined.
data.
would have on measured
is not troubl-
ed by the fact that individual
but which
that can be used to verify
that predict
modifications
that -- "on the
time will be 10% higher
data to other
empirically
consistency
of measurement
Thus,
are basically
that were not measured
in principleabe
B. Relationships
= 7.07
by difinite
expressions which re-
time
RR response time i 2 2 = .60 x 7~ + .i0 x 6~ + .30 x 6~
during
These analysts
per-
usually deal
in:
late existing measurement
Expected
average"
systems
A. Precise mathematical
= 7.80
cer-
but such
"ABC" will occur 60% of the time,
30% of the time.
analysts.
re-
is run "long
for resolving
issues,
insufficient
the needs of empirically
that the
theoretical
at some desired
sequence
and the expected
of
OBJECTIVES
sequence
the arrival
that
time and con-
formance
this case,
pre-
of simulated
sult can he validated
For example,
as a random variable,
the
that
average values will be reason-
3. OPERATIONAL
can be
a meaningful
of random variables.
In such cases,
(e.g.,
run for long periods
to con-
If the uncer-
the order of arrival
In prac-
in the
sist of Jobs A, B, and C, but the order in which the jobs will arrive
sequence.
of
real time.
where only partial know-
exists.
by the im-
to include many three
theory
the Ergodic
in
consisted
in a single benchmark.
dicted and observed
different
MODELS
models
arrival expect
from probability
ably close in almost
2. PROBABILISTIC
tainty concerning
three-job
theory of confidence
final conclusions.
Probabilistic
that the benchmark
Large Numbers,
This in seeming-
times discussed
paragraph were magnified
assumption
job sequences
is often compel-
between predicted
response
tice one would normally
in complete
decisions
situations
benchmark
plicit
discrepancies
average
the preceding
sim-
are very
evaluations
of the system workload
led to make subtle but critical
turn leads
The potential
the same
lationships
sequences.
values
computer interested
between random variables
of random variabies;
terested
201
oriented
are not primarily
in relationships
perforin re-
or expected
r a t h e ~ they are in-
between measured
quan-
tities and quantities measured.
which can~in principle,
The remainder
be
Throughput
of this paper will deve-
lop a theory of such relationships. This theory is based on the concept tional variables. mal objects
Operational
defined
of opera-
variables
so that they correspond
systems.
that will be derived
Some of the results
using operational
variables
have direct
parts which can be expressed variables
and probabilistic
is important
to notice
are fundamentally
that operational
different
results:
that is, operational precision
vational
data,
whereas
apply precisely
results
probabilistic
(with probability
one)
case where
the length of the observation infinity.
al
results
artificial
independent
service
and identically
and inter-arrival
equilibrium,
specific
that are commonly babilistic
operational
to broad classes results
are similar mental
as universal
laws of computer in certain
statements etc.)
tion,
in this
provide
They
and electrical
to the funda-
of these similarities
engineering.
of other
a precise
applicable
Before proceeding is useful
to consider
wSth the main analysis, two specific
questions
are typical of a large class of problems arise in the study of computer The two questions, the throughput problem,
problem
it
possible
to each question.
in the problem
conform
to Furare
to the
statements,
regard-
characteristics
For future reference,
to the throughput
to
problem
these
note that is 51 jobs
to the response
time
is 5 seconds.
which
5. THE OPERATIONAL The solutions
system behavior.
problem and the response
appear
or
under considera-
it is entirely
answer
level
METHOD
that
which will be referred
are defined
factors which would
per hour and the solution
TWO BASIC PROBLEMS
size,
system overhead,
to all systems which
the solution 4.
of memory
operating
In addi-
the answers which can be obtained
systems may have.
of the presentation.
or the utilization
in the system.
less of any other particular
will be noted in the course
is insufficient
posed in each case.
to the basic problems
specifications
Some
that the information
is made in the problem
I/O devices
Nevertheless,
thermore,
laws found in such fields as basic mechanics,
thermodynamics,
number
What is the ave-
questions
there is no discussion
tion.
and funda-
system performance. respects
a
time for the
of most modern day com-
of CPU utilization
be relevant
the
of the system
service
statements
no mention
of multiprogramming,
the
sys-
Suppose
to the system disk per
it might appear
the specific
of any other
they apply without
that will be derived
the complexity
For example,
statistical
of systems,
a time sharing
terminals.
think time is 15 seconds,
of accesses
in the problem
to answer
assumptions
forms,
of
time of the system?
systems,
provided
in the case of pro-
Because
paper can be regarded mental
puter
distributed
distributional
required
results.
restriction
times,
Consider
25 active
system disk is 40 milliseconds.
is that operation-
mathematical
per pro-
What is the throughput
is 7, the utilization
rage response
can be derived without making any of
the seemingly (e.g.,
interaction
Considering
difference
a page
and the average number
disk is .35, and the average
in the
approaches important
that the average
only
limiting
that the
of back-
to process
(i.e.~ reads and writes)
Time Problem:
average number
of obserresults
the average amount
time required
tem which contains
results
interval
Another
Reponse
it
apply with
to all collections
store
the system?
from probabilistic
complete
is .85,
gram is equal to 6000.
counter-
However,
(i.e.,
of page transfers
in terms of random models.
Suppose
of the backing
fault is 10 milliseconds,
of real
a batch processing
the utilization
ing store service
natur-
properties
Consider
system with a paged virtual memory.
paging device)
are for-
ally and d i r e c t l y to observed
Problem:
presented
in Section
tained by using the operational
to as
system analysis.
time
definitions
as follows:
tended
and working
to reflect
gaged in empirical performance.
202
This method
method of computer
is based on a set of
assumptions
the viewpoint studies
4 were ob-
that are in-
of individuals
of computer
As stated in Section
system
3, such
en-
individuals
usually deal with collections
measurement
data that are obtained
system behavior
during
In operational
terms,
specific
ed an observation
interval,
that are measured vation interval variables.
tional method
involve
The operational
variable
as follows: (i)
( actually
providing
ing the observation
interval.
ed that the processors
the opera-
system are numbered
the following
i
service)
dur-
It is assum-
and devices in the
in some fashion
the term "processor-or-device
three steps:
so that
i" refers
to a unique entity.
A. Define a set of operational that correspond manner
completions
of time that device-or-processor
is busy
to as operational
which utilize
generally
B(i) ~ Amount
during an obser-
are referred
of job
X = J/T
is call-
and the quantities
or computed
Analyses
X is expressed
of time.
of time during
data is collected
(i.e.~number
per unit time).
by monitoring
intervals
an interval
which such measurement
X = Throughput
of
in a direct
to the intuitive
interest
(e.g.,
lization,
U(i) = Utilization
variables and natural
concepts
throughput,
of
C(i) = Total number
variables
characterize
single observation
during a
i during
completed the obser-
interval.
S(i) = Average
service
time for processor-or-de-
vice i. S(i) is expressed
as follows:
S(i) = B(i)/C(i)
relationships
such as those described
sub-headings
vation
requests
interval.
C. Apply these mathematical to problems
among
which
system behavior
U(i)
(2)
of service
by processor-or-device relationships
i.
as follows:
U(i) = B(i)/T
device uti-
etc.).
B. Derive mathematical operational
of processor-or-device
is expressed
A, B and C of Section
D(i) = Average
in
number
or-device
3.
(3)
of requests
i per job. D(i)
for processoris expressed
as follows: The next section of this paper defines of operational formulate
variables
a set
that can be used
a number of performance
D(i) = C(i)/J
to
evaluation 7. THE THROUGHPUT
problems. 6. OPERATIONAL
set of symbols
will be used throughout discussion.
stood in the context described
of this
Throughput
should be under-
of the operational
the time interval
the system was observed ments were taken).
method
service
The
and measure-
(i) through
U(i) S(i)'D(i)
all other time
interval.
during
the obDespite
A job is a basic
unit of work and may refer to a program,
Throughput
a job step, a job, or an interaction,
result.
depending
device per
(5)
is immediate
from
(4) since B(i) T
=
J/T
=
X
C(i) B(i)
Law represents
a useful
to the throughput
in this discussion.
J C(i)
of the proof,
Note that it can be used
ed earlier
203
Law
the trivial nature
tain the answer
on the system being studied.
of average
for any i.
of the Throughput
quantities.
of jobs completed
servation
proof
equatio~
that
is equal to device util-
Symbolically,
U(i) S(i), D(i)
X
the time units in which T is expressed
J = Number
Throughput
by the product
time and the average number of requests
interval during which
It is assumed
are also used to express
to
basic result.
job for that device.
T = Length of the observation
dependent
Law:
ization divided
above.
(i.e.,
the following
it is possible
and definitions
the remainder
All definitions
LAW
Given the above definitions,
VARIABLES
derive The following
(4)
the
and powerful to directly
ob-
problem present-
As specified
in
the problem u(i)
statement,
8. THE UTILIZATION
.85
=
Note that the Throughput
S(i) = i0 milliseconds D(i) = Thus,
any device
6000
or processor
the law were applied
X = .85/(6000 X i0) jobs per millisecond
X =
Since throughput
that it is independent
of a large number
tors which must normally lyses.
These
programming;
factors
be specified
include:
the service
device and processor
tion
of fac-
time distribution
gardless
Equation
For this reason,
the Throughput
ed as a universal mance.
A probabilistic
version
put Law, which is valid central Buzen
server model,
interval.
re-
system perfor-
Law bears certain
(a) for an interval
it travels
1 d =~.a
t
to all sysperformance.
Probabilistie
of the Utilization
In order
at a constant
of time
(t), the
previously
regarded
in the above relationship. to consider
ween distance,
a basic model
such systems operate.
in which
for describing
that will be employed
[8], and Muntz
Note that it
blocked)
probabi-
process
and time is signi-
than the argument Law.
bet-
However,
of this paper will demonstrate laws of greater mathematical
that
interest
also exist.
204
with it.
in think state
enters
active).
Response that
system state and the
instant that it next leaves
system state. leaves
a proeessing
has
The in-
(i.e.,
the instant
that an interactive
is said to occur.
the next
and each terminal
(i.e.,
process
[7],
of this model
There are a fixed num-
time between
process
in this discussion
associated
is either
or system state
(i.e., completes
required
1.
the
The model and
[i0], Kleinrock
The essence
terminals,
process
time is the elapsed an interactive
the relationship
in Figure
interactive
teractive
[9].
to the
it is first neces-
sary to provide
one
them as expected values
acceleration,
the Throughput
few sections
in a
[4].
techniques
systems,
manner
ber of interactive
true t h a t the mathematical
to derive
ficantly more complex
operational
Equality
[2] and,
TIME LAWS
to apply operational
is illustrated
as operational
since no underlying
It is, of course, required
by Buzen
are based on the work of Scherr
listic model is involved.
to derive
versions
of interactive
terminology
(d) is given by:
between this result and ip . Law is that dlstance~ acceleration
of random variables
argument
Equality
Law,
is applicable
analysis
if
2
and time are most natually
is not natural
some observation
law of computer
similarities
One point of similarity
variables
during
can
of a set
tems and is a u n i v e r s a l
Moore the Throughput
This equation
consistency
very general c o n t e x ~ by Chang and Lavenburg
of the
For example,
a body starts at rest and accelerates
distance
data collected
have been derived
of the Through-
to in this discus-
was derived earlier by
laws of motion.
acceleration
(7)
Equality.
the internal
9 RESPONSE
The Throughput
(6) that
As in the case of the Throughput
the Utilization
factors.
in the context
(7) will be referred
of empirical
[2].
to elementary
used in equa-
(5) and
U(j) S(j),D(j)
=
be used to verify
Law can be regard-
law of computer
(i) has a
of the device
from equations
sion as the Utilization
equilibrium;
Law is valid
of the status of these other
in equation
for each
of CPU and I/0 processing
the Throughput
(5), it follows
U(i) S(i)-D(i)
the fact that
within a single program is or is not permitted. In other words,
as defined
in other ana-
the system is or is not in statistical the fact that overlap
j with j # i, one
(6)
unique value regardless
Law is
the degree of multi-
in the system;
to device
to
If, for example,
U(j) S(j) 'D(j)
= 51 jobs per hour
aspect of the Throughput
Law can be'applied
in a system.
would obtain
= 3600000 X .85/(6000 X i0) jobs per hour
An important
EQUALITY
Each time
system state
request),
an interaction
amount
of think time
per transition
from think state to system Think State
Interactive Terminals
plying
the same reasoning
state).
Ap-
used to define
R, N
Z System State
1 Z z(k) J" k=l
=
System By combining N
these definitions
sented previously,
several
ponse time in interactive General
Response
Model
context,
it is possible
group of operational
to
Proof:
assumed
J = Number
terminals.
that N is constant
out the observation
observation
interval.
is simply a specific the definition
response
system is expressed
The derivation
k-th terminal)
during
as
ji j
Z
(i0)
of equation
(i0) utilizes
(the process spends in
associated
with the
think state,
plus uhe
total time that the k-th interactive process
the
in
This definition instance
S(i).D(i) U(i)
active process
through-
of the observation
of
interval.
That is
of J given previousz(k) + r(k) = T for all k
r(k) = Total
time that the k-th interactive
process
Thus, N
response
time
(i.e.,
interaction).
Since R is the average
Dividing
of system state time necessary
to generate computed
one interaction,
(by this
and then dividing
quantity by the total number during
and
both sides by J and using equations
(9) to simplify JIz + R = J
the
total time spent in system state
completed
(12)
(8)
the left hand side,
R can
by first obtaining
all processes)
actions
N
Z z(k) + E r(k) = N-T k=l k=l
average
of time in system state per
amount
(ii)
spends in system state.
R = Average amount
spends
system state, must be equal to the length
ly.
be
can be derived.
the fact that the total time that the k-th inter-
It is
interval.
of interactions
res-
vari-
ables. N = Number of interactive
systems
concerning
follows:
R = N define an additional
results
Time Law: The average
time of an interactive i. Basic Interactive
Given this general
with the set pre-
J
Response Time
Fig.
(9)
By equation J T
of inter-
Combining
the observa-
N T (13)
J
(i) and the Throughput
Law,
U(i) S(i),D(i) equations
(14) (13) and
(14),
tion interval. ~ Z Z + R = N S(i)'D(i) Thus,
i R = ~
z(k) = Total
J
R is given by
The proof
N Z r(k) k=l
(8)
A useful
jobs leave
state during
the observation
Z = .Average think time
(i.e.,
simplication
large with respect
of times that interactive
think state and enter
by subtracting
of the General
Time Law arises if it is assumed
spends in think state.
jr = Total number
is completed
jf 7 z
from both sides.
time that the k-th interactive
process
(15)
u(i)
system
interval.
205
that J is very
This will occur if the
observation
interval
is substantially
the average
response
time of the system.
cases,
average
to N.
JI/J~l
Response
since ~JI-J I ~ N.
longer
than
In such
present It is then approximately
true that
at the start of the interval
that arrived
at time zero).
the following
R =
N S(i)'D(i) U(i)
Z
(16)
(i.e.,
jobs
For each job, define
function:
f(k,t)
= Amount
of memory allocated
to
the k-th job at time t. Equation
(16) will be referred
t_ic Response
to as the Asympto-
Time Law, or simply,
the Response
Let
Time
y(k) = Space-time
Law. If it is assumed N, the Asymptotic
this paper. seconds,
that J is large with respect
Response
solve the response Recall
D(i)
to
defined
Time Law can be used to
time problem
stated earlier
that N = 25 terminals,
= 7 requests,
S(i) = .040 seconds.
Z = 15
U(i) = .35, and
Y = Average
Thus,
time
space-time
product
tic Response
versions
Time Law have been derived
investigators
for particular
models
systems
[7,8,10].
Response
Time Law for more general
models
Derivations
are presented
Muntz and Wong mary objects
pressed
processes,
Space-time
program performance to programs
in virtual
ly, a program's its execution amount
memory
M = Average
space-time
time multiplied
of memory allocated
cution.
Since space-time
and throughput
(i.e.,
allo-
variable
m(t)
is
f (k, t)
(19)
of memory
M is defined
By equations
(19) and
as follows: T ~ m(t) 0
is equal
T
M=~
at
(20)
(20),
charges
Essential-
i n use during
interval.
i
to
A
:
Z
O
k=l
f (k,t) dt
(21)
by the average Reversing
to it during its exeproducts,
the order of integration
and applying
response
equations
(17) and
and summation,
(18),
are all used as indicators
of system performance, mine the manner
amount
M = ~i
accounting
product
Z k=l
the observation
of random
systems.
in use
the pri-
are often used to evaluate
andassign
(18)
as follows:
re(t) =
[i] and by
PRODUCT LAWS
products
job), Y is
A
variables. i0. SPACE-TIME
of memory
interactive
values
(i.e.~ space-
A Z k= I y(k)
defined
and results were ex-
in terms of expected
per completed
The operational
of the Asymptotic
In all these cases,
product
cated to some job ) at time t.
by several
of analysis were random variables
and stochastic
= Amount
of interactive
by Boyse and Warn
[9].
m(t)
of the Asympto-
(17)
as follows:
i Y = ~
= 5 seconds
probabilistic
y(k) is
as follows:
R = 25 x .040 x 7/.25 - 15
Specific
for the k-th job.
variable
T y(k) -- f f(k,t)dt 0
in
defined
time,
product
The operational
it is interesting
in which
M = Y.J/T
to exa-
these quantities
are
Applying
equation
(22)
(i)
related. Begin by considering which may be operating or a batch processing variable arrive terval.
an arbitrary in either
mode.
Assume
X = M/Y Equation
Let the operational
A be defined as the number
at the systemlduring
system
of jobs which
the observation
that A includes
(23)
an interactive
in-
(23)~
Product
throughput
is equal to average
use divided
those jobs
206
which will be referred
Space-Time
Throughput
by average
Law,
amount
space-time
to as the
states
that the
of memory
product.
in
A similar
relationship
time product asymptotic tions
between average
and average
response
case can be obtained
(23),
(16), and
(5).
space-
average
time in t h e
by combining
response
Define equa-
time
(R) times one unit of space.
the operational
variable
L as followsf
L = Average number of jobs present
This yields
the system during
in
the observation
interval. R = N'Y/M - Z
Equation
(24)
(24) will be referred
Time Product
Response
variables
correspond
tospace-time
class~
are expressed quantities
service
correspond
to event and some
integrals.
interval,
Within
are expressed
completion" EVENT COUNTS
each
are expressed
into equation
space units,
and simplifying,
certain
(25) is the operational
is usually
referred
of course,
that equation
' TIi.iE ~ DURATIONS
to as Little's
rather
SPACE-TIME INTEGRALS
throughput
than arrival
assumed
theory
formula.
(25) is expressed
variables
(25) involves rather
counterpart
from queueing
lues of random variables.
basis.
the
(25)
of operational
on a "per
(23), cancelling
L = X.R
of the well known result
on a "per job" basis,
are expressed
Substituting
Equation
for the
certain quantities
on a "per unit time" basis,
and certain quantities
Note
to time durations,
certain quantities
entire observation
M = L x one unit of space
of the
in this paper.
variables
some correspond
paragraph,
Y = R x one unit of space.
the basic structure defined
that some operational counts,
to as the S_pace-
Time Law.
Table 2 illustrates operational
From the previous
that Note,
in terms
than expected
va-
Note also that equation (i.e.,
rate
departure
(i.e.~ % ).
that the system being
rate)
If it is
studied
is in equi-
librium in the sense that the same number of proJ A J' C(i)
ENTIRE INTERVAL
~
T B(i) r(k)
i
z (k)
T
grams are present
y(k)= ~f (k, t) dt
be equal T
I
PER UNIT TIME
J
X=~i
the observation
!U(i)=B$i)
special
N
i
A
i X R=7 ~ r(k)
D(i)=Cj (-i)
Y = 7
k=l PER SERVICE COMPLETION
can be deduced
k=l
of Little's
ments[5],[6]. objective
formula
However,
previous
_B(i) s(1)-c(i)
and included
•
Table 2
Operational
additional
12. RELATIONSHIP
Variables
system.
that each job has unit size
M will be equal
the average number of jobs present
to
TO PROBABILISTIC
(response
times one unit of space.
time)
it is useful
all
assumptions
RESULTS
results presented
in this
properties
for the
systems to
207
time.
which can be expressed models
to consider
analyses
of specific
analyses•
are concerned
behavior
between Essential-
with the
sequences which
follow during well defined Most results
in
and random variables,
the relationship
and operational
ly, operational
is
Thus Y is equal
terms of probabilistic
probabilistic
in the
in this case that y(k)
equal to the time-in-system k-th job
space),
(i.e.,
number of occupied memory units or, equi-
It also follows
and consequently
additional
FOrmULA
one unit of memory
the average
results
arguments.
paper have counterparts If it is assumed
valently,
argu-
was to derive
required
Since most operational ii. LITTLE'S
occupies
analysis
to operational
in all cases the ultimate
of the analysis
analyses
of deri-
have included
in terms of random variables,
k=l
i
counter-
as a
out that a number
steps that are very similar
N
rate will
case.
vations y(k)
and the end of
the departure
rate and a direct
formula
It should be pointed
O
PER JOB
to the arrival
part to Little's
M = l ~ m ( t ) dt
at the beginning
interval,
in this paper
intervals involve
of
opera-
tional variables
that represent
over such time intervals, these results
averages
behavior
difference
and conventional
between
(e.g.,
periodic
se-
care must be taken to avoid
behavior
that is incompatible
tical equilibrium).
a given set of average values
(lhis is the principal tional analyses
results
and as a consequence
apply to all possible
quences which produce
computed
opera-
parts exist, generality
deterministic
Thus,
operational
and simpler
strictly
with statis-
even when direct
counter-
laws benefit
from greater
interpretation
and applica-
tion.
analyses). Probabilistic (i.e.,
results
also apply to an ensemble
a set) of possible
a probabilistic of statistical
result
behavior
then the Ergodic
rem and the Law of Large Numbers
averages
defined
sequences,
riables
values
(i.e.,
babilistic
and which involve
laws that apply
to almost
ensemble
members
of infinite-time
pro-
the implications
possible
device
behavior
standpoint,
of these infinite-time
behavior
quest
se-
A. The number sing element
of arrivals
sing element
at each proces-
is unbounded.
ber of requests
pending
sense,
and can be assumed
to be
queues develop
the observation
interval.
processing arrival
element,
Thus,
rates must be equal;
the number
of departures
each processing
element
which involve
properties
that do
A and B above,
only constants
and
to probabilistic
volve constants
and expected values
results which in-
made in the case of the corresponding
is at least moderately
long,
and can be safely
apply
on a "per job" basis the operational
to all variables
(e.g.,
X,Y, and R).
definitions
presented
are collected
during
that
the observation
interval
and that "per job" and "per service averages
interval
by simple division.
are computed
This assumption
to conform well to existing performance
com-
at the end of the
measurement
practice
in
and evaluation
field.
of random
assumptions
if the ob-
pletion"
Note additional
of S(i).
interval
the computer
variables. In many cases,
at
Fortuantely,
appears
and average values,
will correspond
complete
interval will contribute
in this paper are based on the assumption
to
it will be true that operational
not contradict
in a a ser-
This will artificially
measurements
results which are based on assumptions
Likewise,
raise the value
In effect,
at
can be assumed
re-
to B(i) but not to C(i).
defined
equivalently,
be equal. In general,
reduced.
The same considerations
and
and arrivals
If a service
disregarded.
for each
the departure
completed
at the start of the ob-
these effects will be negligible
during
variables.
that average
the value of S(i) will,
be artificially
servation
zero. B. No unbounded
interval.
the end of the observation
for each proces-
to avoid
to total device busy
vice request which is only partially
Thus the num-
at the start of the interval
is insignificant
interval,
states.
operational
of requests
complete
and
understood
(3) states
time is equal
the observation
servation
are significant:
consistent
of operational
equation
is partially
laws themselves
of certain
misinterpretation
service
during
only two
of end
inter-
and terminal
should be carefully
For example,
of the
initial
time divided by the number
From the operational
quences
of the observation
of operational
for all possible
definitions
and
opera-
the problem
A c t u a l l ~ end effects do not in any way im-
However,
sequences.
properties
points
since these laws are internally valid
to operational
all
concerns
the state of the system at the ini-
pair the validity
to
Thus,
only constants
values will correspond
associated
val).
random va-
averages).
(i.e.,
tial and terminal
se-
correspond
of the associated
the ensemble
effects
for
laws which are based on equilibrium
assumptions expected
One question which arises when applying
the operational
and will
CONSIDERATIONS
tional laws in practice
Theo-
imply that,
for each of the individual
quences will be identical the expected
13. PRACTICAL
If
is based on the assumption
equilibrium,
almost all behavior
sequences.
that the above
to the problem
must be
probabilistic
208
meters
comments
of estimating
actually
intrinsic
on the basis of observations
pertain
system para-
taken during
finite intervals
of time.
both operational
and probabilistic
it is independent actually
This problem
arises
analyses,
of the analysis method
in
constant
and
For example,
that is
techniques
employed.
A somewhat
problem
of disk and drum subsystems
arises
which untilize
of equation service
rota-
space-time
(3) to such devices
times.
latency will not appear in S(i). this phenomenon
does not affect
the operational
laws derived
it does illustrate
Once again, the validity
in this paper,
of but
verify
the internal
definitions.
stant,
then the Space-Time
measurements
collected
numerical
values
algebraic
equations
solution
by
wise,
the Space-Time
plies
that average
generated
through
models
methods;
(e.g., device
in terms of other variables
In a somewhat
service
since,
to which response interactive system.
Like-
Time Law imby
value of
terminals
increases
in the Response
time,
time will increase are added
to
Once such systems reach cer-
see Scherr
in the number
(N) leave all other depen-
of this approach
and Muntz and Wong
in
which are easier
different
context,
can be helpful
Time Laws unchanged. to estimating
[i0], Kleinrock[7],
resMoore[8]
[9].
15. ACKNOWLEDGEMENTS
to
Many of t h e c o n c e p t s
by equation
(7),
the Utiliza-
in locating
presented
to calculate
for each device
values
The
16. REFERENCES
i.0, when
Equality,
also makes
i. Boyse,
in the system.
J. W., and Warn,
forward model
These
tion~
can then be substitu-
2. Buzen,
Comp.
Surveys
J. P.
in this paper can also be used
such as throughput these estimates, which variables
ACM, New York, 3. Buzen,
that certain modifications
and response
it is necessary
time.
indicators
J.P.
of the IEEE 4. Chang,
To make
209
(June 1975),
predicpp.73-93.
of system bottlenecks
network model= on Sys. Perf.
Proc.
Eval.
ACM
(April 1971)
pp. 82-103.
63,6
architecture~
(June 1975),
A., and Lavenburg,
dent servers.
A straight-
performance
I/O subsystem
closed queueing
to understand
will change and which will remain
7,2
Analysis
SIGOPS Workshop
time.
D.R.
for computer
using a queueing
and lower bounds on response
to a system will have on performance
by
the work of R. R. Muntz and L. Kleinrock.
throughput
the effect
P. J. Denning,
The author was also influenced
ted into other laws to obtain upper bounds on
The laws derived
conversations
the author and D. B. Rubin,
and E. Gelenbe.
the maximum utilization
and processor
maximum utilization
in this paper
during
must
have the largest value of U(i).
with the Utilization
between
bottle-
the device or pro-
fact that no value of U(i) can exceed
to estimate
Response
time will decrease
dent variables
time)
cessor with the largest value of S(i).R(i)
it possible
Law
by 25%.
Time Laws can also be used to esti-
were refined and sharpened
tion Equality
combined
Product
response
terminals
ponse
the explicit
measure.
necessarily
Throughput
con-
The laws can also be used to express
an unknown variable
necks
Product
will increase
of interactive
For examples
programs;
of systems
(i.e., M) approximately
tain levels of saturation,
of: performance
by simulation
derived
of probabilistic
equilibrium.
a time-sharing
they can be used to
empirical
in use
that throughput
The Response
in this paper are direct-
consistency
the average
load on the system to keep the
when additional
to all systems,
that
N'Y/Z.
14. APPLICATIONS Since the laws derived
installa-
to assume
an amount equal to 20% of the original
the need to fully understand
of all operational
the working
(i.e., Y) by 20%, and if there
of memory
implies
out.
restructuring
to reduce
at a particular
amount
mate the extent
ly applicable
that program
will also reduce
product
is a sufficient
Thus,
are carried
If it is reasonable
this modification
delay due to seek and rotational
the implications
suppose
make it possible
tion by 20%.
times which are approximate-
ly equal to average data transfer the additional
in the case
sensing and block multiplexing[3].
The application yields average
the modifications
set size of all programs
different
tional position
after
networks Oper.
Res.
S.S.
pp.871-879. Work rates in
with general 22
Proc.
(1974),
indepenpp.838-847.
5. Jewell, 15
W.S.
(1967),
6. Maxwell, L=%W.
A simple proof of: L=%W~
W.L.
On the generality
Oper.
7. Kleinrock,
Oper. Res.
APPENDIX
Res.
L.
18 (1970),
Proc,
of the equation
results
IFIP Congress
Operational
for time
C.
Univ.
1968,
C(i) =
Network models
systems.
TR-71-1,
R.R.,
Annual Princeton Sciences I0. Scherr, Systems.
J.
M.I.T.
Press,
Proc.
Eighth
A
of jobs
M = Average
ANALYSIS
times of individual time
(completion
and A-3 present
It is assumed
presented
that the quantum
robin scheduling
algorithm
= Amount
(or interactions)
number
r(k) = Total
in Table i.
of memory
terminals.
time.
time in system state for k-th inter-
S(i) = Average
is 2 seconds.
(k-th terminal).
service
AVERAGE
FCFS
RR
U(i) = Utilization
7 8 ii
ii 3 8
X = Throughput.
8 2/3
Table A-I Workload
y = Average
= ABC
time for device
Z = Average z(k) = Total
FCFS
RR
3 4 Ii
6 3 ii
AVERAGE
6
B A C
AVERAGE
Workload
= CBA
RR
1 8 ii
i" ii 8
Table A-3 = BAC
6 2/3
210
(k-th terminal).
Operational
Asymptotic
Response
Response
Workload
think time.
Principal
General
6 2/3
FCFS
6 2/3
Table A-2
per job.
for k-th job.
time in think state for k-th interactive
process
C B A
i.
product
product
i.
interval.
of device
space-time
y(k) = Space-time
7 1/3
the
in use at time t.
response
active process
size in the round
in use during
interval.
of interactive
Average
for
the
interval.
amount
T = Length of observation
A B C
completed
of jobs in system during
of memory
N = Number
response
of the entire benchmark
each of the three workloads
m(t)
the completion
jobs and the average
time)
to the k-th job
system state. L = Average
observation Tables A-l, A-2,
per job for
at time t. J = Number
Computer
(1967).
BENCHMARK
by
i.
observation APPENDIX
completed
during the observation interval. j'= Number of transitions from think state to
pp.348-3521
of Time Shared
of requests
= Amount of memory allocated
properties
on Information
(March 1974),
device f(k,t)
i is busy.
requests
i.
D(i) = Average number
Eng.,
(April 1971).
Asymptotic
Conference
An Analysis
scale time
Industrial
network models,
and Systems A.
Dept.
Ann Arbor
and Wong,
of closed queueing
for large
at system.
of time device
Number of service device
of Michigan,
9. Muntz,
Variables
B(i) = Amount
pp. 838-845.
sharing
GLOSSARY
A = Number of arrivals
pp.172-174.
Certain analytic
shared processors~
8. Moore,
B
pp.l109-1116.
Response
Laws
Time Law - Eq.
Time Law - Eq.
Time Law - Eq.
(16)
(I0)
(16)
Space-Time
Product
Response
Space~Time
Product
Throughput
Time Law - Eq.
Throughput
Law - Eq.
Utilization
Equality
(5) - Eq.
(7)
Law - Eq.
(24)
(23)