See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/258705139

Stochastic Petri Nets -An Introduction to the Theory

Book · November 2013

CITATIONS READS

182 4,495

2 authors:

Falko Bause Pieter Kritzinger

Technische Universität Dortmund University of Cape Town
133 PUBLICATIONS 1,443 CITATIONS 81 PUBLICATIONS 982 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Falko Bause on 04 June 2014.

The user has requested enhancement of the downloaded file.

Stochastic Petri Nets
- An Introduction to the Theory -

Falko Bause
Informatik IV
Universität Dortmund
D-44221 DORTMUND
Germany
e-mail: bause@ls4.cs.uni-dortmund.de

Pieter S Kritzinger
Data Network Architectures Laboratory
Department of Computer Science
University of Cape Town
Private Bag, RONDEBOSCH
7700 South Africa
e-mail: psk@cs.uct.ac.za

c Bause and Kritzinger, 2002.

2
 5

Preface
Any developer of discrete event systems knows that the most important quality of
the final system is that it be functionally correct by exhibiting certain functional,
or qualitative properties decided upon as being important. Once assured that the
system behaves correctly, it is also important that it is efficient in that its running
cost is minimal or that it executes in optimum time or whatever performance
measure is chosen. While functional correctness is taken for granted, the latter
quantitative properties will often decide the success, or otherwise, of the system.
Ideally the developer must be able to specify, design and implement his system
and test it for both functional correctness and performance using only one for-
malism. No such formalism exists as yet. In recent years the graphical version
of the Specification and Description Language (SDL) has become very popular
for the specification, design and partial implementation of discrete systems. The
ability to test for functional correctness of systems specified in SDL is, however,
limited to time consuming simulative executions of the specification and perfor-
mance analysis is not directly possible. Petri nets, although graphical in format
are somewhat tedious for specifying large complex systems but, on the other
hand were developed exactly to test discrete, distributed systems for functional
correctness. With a Petri net specification one can test, e.g., for deadlock, live-
ness and boundedness of the specified system. Petri nets in their various formats,
have been studied extensively since first proposed by Carl Adam Petri in 1962
[133] and several algorithms exist to determine the functional properties of nets.
Another paradigm which is aimed at testing for functional correctness is that of
process algebras or calculi for communicating systems.
The major drawback of Petri nets, as originally proposed and process algebras
(amongst others) is that quantitative analyses are not catered for. As a conse-
quence, the developer who needs to know about these properties in his system
has to devise a different model of the system which, apart from the overhead con-
cerned provides no guarantee of consistency across the different models. Because
of the latter, computer scientists during the last decade added time, in various
forms, to ordinary Petri nets to create Stochastic Petri nets (SPNs) and General-
ized Stochastic Petri nets (GSPNs) for performance modelling and a great deal
of theory has developed around Stochastic Petri nets as these are generically
known.
Another aspect which also contributed significantly to the development of Stochas-
tic Petri nets is the fact that their performance analysis is based upon Markov the-
ory. Since the description of a Markov process is cumbersome, abstract models
have been devised for their specification. Of these, queueing networks (QNs) was
originally the most popular, especially since the analysis of a large class of QNs
(product-form QNs) can be done very efficiently. QNs cannot, however, describe
system behaviours like blocking and forking and with the growing importance
of distributed systems this inability to describe synchronisation naturally turned
the focus to Petri nets as well.
6 PREFACE

Stochastic Petri nets are therefore a natural development from the original Petri
nets because of
• the advantage of their graphical format for system design and specification
• the possibility and existing rich theory for functional analysis with Petri
nets
• the facility to describe synchronisation, and
• the natural way in which time can be added to determine quantitative prop-
erties of the specified system.
The disappointing thing about Stochastic Petri nets is that the integration of time
changes the behaviour of the Petri net significantly. So properties proven for the
Petri net might not hold for the corresponding time-augmented Petri net. E.g., a
live Petri net might become deadlocked or a non-live Petri net might become live.
We will see that the analysis techniques developed for Petri nets are not always
applicable to SPNs. But there are ways around this, as we shall see in this book.
Also, using Stochastic Petri nets to specify the sharing of resources controlled by
specific scheduling strategies is very cumbersome. So the pendulum has swung
back, in a sense, that we introduce certain concepts from queueing theory when
presenting Queueing Petri nets (QPNs) which offer the benefits of both worlds,
Petri nets and Queueing networks.
This book itself arose out of a desire by the authors to collect all one needs to
understand Stochastic Petri net theory in one volume. It is in three parts. The
first part is on stochastic theory leading to introductory queueing theory and
simple queues. In Part I we emphasise Markovian theory, because where general
queueing theory fails, Markovian analysis can often still be useful.
Part II is about Petri nets, starting with ordinary Petri nets and ending with
Coloured Petri nets. Ordinary and Coloured Petri nets do not involve time and
were developed to test the functionality of concurrent systems. In this part of
the book we give an overview of the most important analysis techniques paying
particular attention to the validation of those properties which are essential for
Stochastic Petri nets.
Our emphasis in Part III is on those Stochastic Petri net models which can be
analysed by Markovian techniques. The intention of this book is not to give an
overview of several or all Stochastic Petri net models appearing in the litera-
ture, but to stress a combined view of functional and performance analysis in the
context of some Stochastic Petri net models.
We hope that by reading this book, the reader will become as excited as we are
about the subject of Stochastic Petri nets and the many unsolved problems arising
from the increasing demands for correctness and performance when specifying
discrete event systems.
Falko Bause and Pieter Kritzinger
Dortmund, Germany
Cape Town, South Africa
1995.
 7

Preface to the Second Edition

A great deal of progress has been made in the analysis of Petri nets and Stochastic
Petri nets since the first edition of this book appeared over 6 years ago. Amongst
others, partial state space exploration methods have been proposed and state-
based analysis techniques exploiting the structure of the system being modeled
are now known. In the case of Queueing Petri nets and stochastic Petri nets in
general, results and algorithms based on product-form solutions have been intro-
duced.
The results are that nets with up to 50 million states can now be analysed on
ordinary computing equipment. In order to guide the reader to these results we
have added several links in this edition to the relevant literature and updated the
Further Reading sections at the end of each chapter as starting points for more
detailed information.
Naturally, we were tempted to include the new material mentioned in the book.
That would have however, detracted from the focus and advantage of this text: A
concise introduction to both the functional and performance aspects of Petri nets
without emphasising the one or the other.
Falko Bause and Pieter Kritzinger
Dortmund, Germany
Cape Town, South Africa
2002.
8

For Heinz Beilner, our friend and mentor, without

whom this book would never have been written.
 9

Contents

Preface 5

Preface to the Second Edition 7

Contents 9

I STOCHASTIC THEORY 11

1 Random Variables 13
1.1 Probability Theory Refresher . . . . . . . . . . . . . . . . . 13
1.2 Discrete Random Variables . . . . . . . . . . . . . . . . . . 16
1.3 Continuous Random Variables . . . . . . . . . . . . . . . . 18
1.4 Moments of a Random Variable . . . . . . . . . . . . . . . 19
1.5 Joint Distributions of Random Variables . . . . . . . . . . . 20
1.6 Stochastic Processes . . . . . . . . . . . . . . . . . . . . . 21

2 Markov Processes 23
2.1 Discrete Time Markov Chains . . . . . . . . . . . . . . . . 25
2.1.1 Steady State Distribution . . . . . . . . . . . . . . . . . 31
2.1.2 Absorbing Chains and Transient Behaviour . . . . . . . 35
2.2 Semi-Markov Processes . . . . . . . . . . . . . . . . . . . 40
2.2.1 Formal Model of a Semi-Markov Process . . . . . . . . 41
2.2.2 Interval Transition Probabilities . . . . . . . . . . . . . 42
2.2.3 Steady State Behaviour . . . . . . . . . . . . . . . . . . 44
2.3 Continuous Time Markov Chains . . . . . . . . . . . . . . 46
2.3.1 Steady State Distribution . . . . . . . . . . . . . . . . . 51
2.4 Embedded Markov Chains . . . . . . . . . . . . . . . . . . 53

3 General Queueing Systems 55

3.1 Little’s Law . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2 Birth-Death Processes . . . . . . . . . . . . . . . . . . . . 61
3.3 Poisson Process . . . . . . . . . . . . . . . . . . . . . . . . 63
3.4 M/M/1 Queue . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.5 M/M/m Queue . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6 Queues with Processor Sharing Scheduling Strategy . . . . 69
3.7 Queues with Infinite Servers . . . . . . . . . . . . . . . . . 69
3.8 Queues with Priority Service . . . . . . . . . . . . . . . . . 70

4 Further Reading 72
10 Contents

II PETRI NETS 75

5 Place-Transition Nets 77
5.1 Structure of Place-Transition Nets . . . . . . . . . . . . . . 81
5.2 Dynamic Behaviour of Place-Transition Nets . . . . . . . . 84
5.3 Properties of Place-Transition Nets . . . . . . . . . . . . . 86
5.4 Analysis of Place-Transition Nets . . . . . . . . . . . . . . 90
5.4.1 Analysis of the Reachability Set . . . . . . . . . . . . . 90
5.4.2 Invariant Analysis . . . . . . . . . . . . . . . . . . . . 95
5.4.3 Analysis of Net Classes . . . . . . . . . . . . . . . . . 101
Analysis of State Machines . . . . . . . . . . . . . . . . 101
Analysis of Marked Graphs . . . . . . . . . . . . . . . 102
Analysis of EFC-nets . . . . . . . . . . . . . . . . . . . 105
5.4.4 Reduction and Synthesis Analysis . . . . . . . . . . . . 110
5.5 Further Remarks on Petri Nets . . . . . . . . . . . . . . . . 112

6 Coloured Petri Nets 116

7 Further Reading 125

III TIME-AUGMENTED PETRI NETS 127

8 Stochastic Petri Nets 131

9 Generalized Stochastic Petri Nets 139

9.1 Quantitative Analysis of GSPNs . . . . . . . . . . . . . . . 141
9.2 Qualitative Analysis of GSPNs . . . . . . . . . . . . . . . . 148
9.2.1 Qualitative Analysis of EFC-GSPNs . . . . . . . . . . . 155
9.3 Further Remarks on GSPNs . . . . . . . . . . . . . . . . . 158

10 Queueing Petri Nets 162

10.1 Quantitative Analysis of QPNs . . . . . . . . . . . . . . . . 164
10.2 Qualitative Analysis of QPNs . . . . . . . . . . . . . . . . 169
10.2.1 Qualitative Analysis of EFC-QPNs . . . . . . . . . . . 170
10.3 Some Remarks on Quantitative Analysis . . . . . . . . . . . 174

11 Further Reading 176

12 Application Examples 179

12.1 Resource Sharing . . . . . . . . . . . . . . . . . . . . . . . 179
12.2 Node of a DQDB network . . . . . . . . . . . . . . . . . . 180

13 Solutions to Selected Exercises 185

Bibliography 193
Index 209
 Part I

STOCHASTIC THEORY
 13

1 Random Variables

Much of the world around us is not very deterministic although it may not be
apparent at first glance. Consider a computer, for instance, which given the same
input values, will always give the same output. While a computer program is
processing incoming data however, it is often not possible to predict from one
moment to the next

– what input values will arrive for processing, or

– the time sequence in which they will arrive.

Think of the node of a computer network to understand this. Although the set of
messages which may arrive at the node is finite and known, we cannot tell for
certain from instant to instant which messages will arrive from where. Moreover,
the network software is likely to be using the same processor(s) at the node as
the operating system. When the process executing the network software will be
interrupted and by which process cannot be said for certain. All of which makes
it impossible to tell for certain what will happen next. We say the process just
described is stochastic.
The term stochastic has an exact mathematical meaning and there is a vast theory
developed to predict the behaviour of stochastic processes. This part of the book
gives only a basic introduction to that theory. Goodman [86] provides a more
thorough introduction to the subject while an extensive treatment can be found
in Howard [90].

1.1 Probability Theory Refresher

In order to understand stochastic theory, one needs to know some fundamen-

tal concepts of probability theory. This section provides such a basic introduc-
tion. For students wishing a more fundamental introduction, there are many good
books on probability theory, such as those of Feller [75] and Ross [152].
The first concept in probability theory that we need to know is that of an exhaus-
tive set of events which is a set of events whose union forms the sample space
S of all possible outcomes of an experiment. The sample space when we roll an
ordinary die, consists of 6 events for instance.
If two events A and B are such that

A ∩ B = ∅ (the empty set)

14 1 Random Variables

then the two events are said to be mutually exclusive or disjoint. This leads to
the concept of mutually exclusive exhaustive events {A1 ,A2 , . . . ,An } which are
events such that

Ai Aj = Ai ∩ Aj = ∅ for all i 6= j
A1 ∪ A2 ∪ . . . ∪ An = S (1.1)

The next important concept is that of conditional probability. The conditional

probability of the event A, given that the event B occurred (denoted as P [A|B])
is defined as

P [AB]
P [A|B] :=
P [B]

whenever P [B] 6= 0.
The statistical independence of events can be defined as follows. Two events A
and B are said to be statistically independent iff

P [AB] = P [A]P [B]. (1.2)

For three statistically independent events A, B, C each pair of events must satisfy
Eq. (1.2) as well as

P [ABC] = P [A]P [B]P [C]

and so on for n events requiring the n-fold factoring of the probability expres-
sion as well as the (n − 1)-fold factorings all the way down to all the pairwise
factorings. Moreover, for two independent events A and B

P [A|B] = P [A]

which merely states that the knowledge of the occurrence of an event B does not
affect the probability of the occurrence of the independent event A in any way
and vice-versa.
We also need to know the theorem of total probability for our basic understanding
of probability theory.

Theorem 1.1 Theorem of Total Probability. Consider an event B and a set of

mutually exclusive exhaustive events {A1 ,A2 , . . . ,An }. If the event B is to occur
it must occur in conjunction with exactly one of the mutually exhaustive events
Ai . That is
n
X
P [B] = P [Ai B]
i=1
1.1 Probability Theory Refresher 15

From the definition of conditional probability we may always write

P [Ai B] = P [Ai |B]P [B]

= P [B|Ai ]P [Ai ]

which leads to the second form of the theorem of total probability

n
X
P [B] = P [B|Ai ]P [Ai ]
i=1

The last equation suggests that to find the probability of some complex event B,
one simplifies the event by conditioning it on some event Ai in such a way that
computing the probability of event B given event Ai is less complex and then to
multiply by the probability of the conditional event Ai to yield the joint proba-
bility P [Ai B]. Having done this for a set of mutually exclusive exhaustive events
{Ai } we may then sum these probabilities to find the probability of the event B.
If we need to simplify the analysis even further, we may condition event B on
more than one event and then uncondition each of these events by multiplying
by the probability of the appropriate condition and then sum all possible forms
of all conditions.
The final bit of probability theory that we are certain to come across in our study
of stochastic systems is Bayes’ theorem.

Theorem 1.2 Bayes’ theorem. Let {Ai } be a set of mutually exclusive and ex-
haustive events. Then
P [B|Ai ]P [Ai ]
P [Ai |B] = Pn (1.3)
j=1 P [B|Aj ]P [Aj ]

Bayes’ theorem allows us to compute the probability of one event conditioned

on a second by calculating the probability of the second conditioned on the first
and other terms.

Exercise 1.1 If there are n people present in a room, what is the probability
that at least two of them have the same birthday? How large may n be for this
probability to be less than 0.5?

Exercise 1.2 A student writes a multiple-choice examination where each ques-

tion has exactly m possible answers. Assume that a student knows the correct
answer to a proportion p of all the questions; if he does not know the correct
answer, he makes a random guess. Suppose that the student got the answer to a
particular question wrong. What is the probability that he was guessing?
16 1 Random Variables

1.2 Discrete Random Variables

We call a variable random and denote it χ if we cannot tell for certain what its
value will be. Examples of such random variables are the temperature outside on
any particular day, the number of customers in a supermarket checkout line or
the number of messages arriving at a network node.
A random variable is said to be discrete if the set of possible values of χ is
countable (but not necessarily finite). Since we do not know for certain what
value it will have, we say that it will have value x with a probability pχ (x). That
is
pχ (x) = P [χ = x]. (1.4)
In this formula, x can be any real number and 0 ≤ pχ (x) ≤ 1 for all values of x.
pχ (x) is called the probability mass function of χ.
Suppose that χ can take on the values x1 ,x2 ,x3 ,x4 or x5 with probability p1 ,p2 ,p3 ,p4
and p5 respectively. Clearly,
5
X
pi = 1
i=1

The following random variables are important for our studies.

Definition 1.1 Bernoulli variable. A Bernoulli random variable χ takes on values
of 0 and 1. The event {χ = 1} is called a success and occurs with probability p.
The event {χ = 0} is called a failure and occurs with probability q = 1 − p.
Suppose an experiment consists of spinning an unbiased coin. If we spin the coin
n times we say that we have performed n trials and if the outcome is either 0 or
1, true or false, we refer to that as a Bernoulli trial.
Definition 1.2 Binomial variable. The probability mass function of a binomial
random variable χ that yields k successes in n independent Bernoulli trials is
defined by
 
n
pχ (k) = pk q n−k
k

Definition 1.3 Geometric variable. Suppose it took N Bernoulli trials to obtain

the first success. The variable N is said to be geometric and its probability mass
function is given by
pN (k) = q k−1 p (1.5)

Another way of describing a random variable χ which takes values from an or-
dered set is to give a formula for the probability that it will take on values of
xi which are less than or equal to some value a. This leads to the following
important definition.
1.2 Discrete Random Variables 17

Definition 1.4 The cumulative distribution function of a random variable χ is

the function X
Fχ (a) = pχ (x)
x≤a

defined for all real variables a.

Another way of denoting the cumulative distribution often encountered in the

literature is

Fχ (x) = P [χ ≤ x]

Again, it should be evident that the values of the function Fχ are between 0 and
1. Using Fχ we can calculate

P [a < χ ≤ b] = Fχ (b) − Fχ (a) (1.6)

This follows easily from the fact that

{χ ≤ b} = {χ ≤ a} ∪ {a < χ ≤ b}

so that
Fχ (b) = Fχ (a) + P [a < χ ≤ b]
and the equation in (1.6) follows.

Exercise 1.3 If the probabilities of a male or female offspring are both 0.5, find
the probability of a family of five children being all male.

Exercise 1.4 A person has 18 bank-notes which includes 4 counterfeits in his

purse. If he pays for an item with 2 bank-notes selected randomly from his purse,
what are the probabilities that

1. both notes are genuine;

2. one of the notes is a counterfeit;

3. both notes are counterfeits?

Exercise 1.5 An Ethernet local network has k stations always ready to transmit.
A station transmits successfully if no other station attempts to transmit at the
same time as itself. If each station attempts to transmit with probability p, what
is the probability that some station will be successful?
18 1 Random Variables

1.3 Continuous Random Variables

In Sec.1.2 we obtained the distribution function of the discrete random variable

by summing values of the mass function. When we now replace summation by
integration, we obtain the notion of a continuous random variable.
A random variable χ is said to be continuous if there exists a nonnegative func-
tion fχ (x) such that the cumulative distribution function Fχ (x) can be calculated
from
Z a
Fχ (a) = fχ (x)dx (1.7)
−∞

and frequently defined by the expression

Fχ (a) = P [χ ≤ a]

The function fχ (x) is called the probability density function of the random vari-
able χ. Again, because we are concerned with probabilities, we must have the
condition
Z ∞
fχ (x)dx = 1 (1.8)
−∞

Also, analogous to Eq. (1.6) we can calculate the probability that a random vari-
able χ lies in the interval (a,b) from
Z b
P [a < χ < b] = fχ (x)dx (1.9)
a

The density function fχ (x) does not have to be continuous, but the distribution
function Fχ (x) is automatically continuous. This implies

P [χ = x] = 0 (1.10)

for any value of x, so that the events

{a ≤ χ < b} {a < χ ≤ b} {a < χ < b}

all have the same probability given by the integral in (1.9).

It should be clear that we can compute the density function from the distribution
function from
d
fχ (x) = Fχ (x) (1.11)
dx
The probability density function we will meet over and over again is the negative
exponential density function given by

fχ (x) = λe−λx (1.12)

1.4 Moments of a Random Variable 19

The constant λ is called the parameter of the distribution and the function is
undefined for x < 0.
The corresponding cumulative distribution function is easily calculated to be
Z a
F χ(a) = λe−λx dx = 1 − e−λa (1.13)
0

for a ≥ 0, and F (a) = 0 if a < 0. Note also that lima→∞ F (a) = 1 as it should
be, since it is certain that 0 ≤ x < ∞.

Exercise 1.6 Find the probabilities that a random variable having an exponen-
tial distribution with parameter λ = 10 assumes a value between 0 and 3, a
value greater than 5, and a value between 9 and 13.

Exercise 1.7 The life expectancy of a certain kind of lightbulb is a random vari-
able with an exponential distribution and a mean life of 100 hours. Find the
probability that the lightbulb will exceed its expected lifetime.

1.4 Moments of a Random Variable

In most cases we are not interested in the specific distribution function of a ran-
dom variable, but only in some characteristic values, the moments. The mean or
average value of a real positive random variable χ(t) is used very often and it is
more frequently referred to as the expectation of that variable. We write E[χ] or
χ for that value and it is defined by
Z ∞
E[χ] = tfχ (t)dt
0

Note that we integrate only over the interval t ∈ [0,∞) since the independent
variable will always be time in our discussions.
We will see later on that we will need to know the expectation of the power of
a random variable as well. The expected value of the nth power of a random
variable is referred to as its nth moment. Thus the more general nth moment (the
mean is just the first moment) is given by
Z ∞
n
E[χ ] = tn fχ (t)dt
0

Furthermore, the nth central moment of a random variable is defined to be

Z ∞
n
(χ − χ) = (t − χ)n fχ (t)dt
0

The second central moment is used very often and is referred to as the variance,
usually denoted by σχ2 and defined as before by
20 1 Random Variables

σχ2 = (χ − χ)2
= χ2 − (χ)2

The square root σχ of the variance is referred to as the standard deviation. The
ratio of the standard deviation to the mean of a random variable is called the
coefficient of variation denoted by

σχ
Cχ =
χ

Exercise 1.8 Referring to Ex. 1.5 (page 17), compute the mean number of colli-
sions to be expected by any one of the k stations before a successful transmission.

Exercise 1.9 Compute the mean and coefficient of variation of a random vari-
able which is exponentially distributed with parameter λ.

1.5 Joint Distributions of Random Variables

Use the symbol Rn to denote the set of all n-tuples of real numbers. Let χ1 ,χ2 , . . . ,χn
be random variables. These random variables are said to have a joint discrete
distribution if there exists a nonnegative function p(x1 ,x2 , . . . ,xn ) of n real vari-
ables that has the value 0 except at a countable set of points in Rn , such that

P [χ1 = x1 ,χ2 = x2 , . . . ,χn = xn ] = p(x1 ,x2 , . . . ,xn )

for all points (x1 ,x2 , . . . ,xn ) in Rn . Obviously we must have that
X
p(x) = 1
x∈Rn

Similarly we say that the collection χ1 ,χ2 , . . . ,χn of random variables has a
joint continuous distribution if there exists a nonnegative integrable function
f (x1 ,x2 , . . . ,xn ) of n real variables that satisfies

Z a1 Z an
P [χ1 ≤ a1 , . . . ,χn ≤ an ] = ... f (x1 ,x2 , . . . ,xn )dx1 . . . dxn
−∞ −∞

for all choices of upper limits a1 , . . . ,an . The function f is called the joint prob-
ability density function of the random variables χ1 ,χ2 , . . . ,χn and as in the
discrete case, we must have

Z ∞ Z ∞
... f (x1 ,x2 , . . . ,xn )dx1 . . . dxn = 1
−∞ −∞
1.6 Stochastic Processes 21

If we know the joint distribution f of χ1 ,χ2 , . . . ,χn , we can obtain the distribu-
tion of any one, say χm , of the random variables by simply integrating over all
values of the remaining random variables. That is, fm (x) =

Z ∞ Z ∞
... f (x1 , . . . ,xm−1 ,x,xm+1 , . . . ,xn )dx1 . . . dxm−1 dxm+1 . . . dxn
−∞ −∞

is the probability density function of χm . The same holds true for the discrete
case where we would sum, rather than integrate, over all possible values of the
other variables.

1.6 Stochastic Processes

In the previous sections we frequently referred to a random variable χ taking on

a value x. Nowhere did we make any mention of time, or, in other words, when
χ took on what value or how that varied with time. Should we do that, we have
what is known as a stochastic process. Mathematically then, a stochastic process
is a family of random variables {χ(t)} defined over the same probability space.
Put differently, the values (also called states) that members of the family χ(t)
can take on all belong to the same set called the state space of χ(t).
Examples of stochastic processes are the number of persons on the beach as
a function of the time of the day or the number of processes executing on a
computer as a function of time. You will come to suspect already that if we can
describe the latter mathematically we have made great progress at predicting the
behaviour of the computer.
The classification of stochastic processes (some people also call them random
processes) depends on three things: the state space; the nature of the time pa-
rameter and the statistical dependencies among the random variables χ(t) for
different values of the time parameter.

Definition 1.5 If the values x = (x1 ,x2 , . . .) in the state space of χ(t) are finite
or countable, then we have a discrete-state process, also called a chain. The state
space for a chain is usually the set of integers {0,1,2, . . .}. If the permitted values
in the state space may range over a finite or infinite continuous interval, then
we say that we have a continuous-state process. The theory of continuous-state
stochastic processes is not easy and we will only be considering discrete-state
processes in this book.

Definition 1.6 If the times t = (t1 ,t2 , . . . ,tn ) at which we observe the value of
χ(t) are finite or countable, then we say that we have a discrete-time process;
if these times may, however, occur anywhere within a set of finite intervals or
an infinite interval of time, then we say that we have a continuous-time process.
When time is discrete we write χn rather than χ(t) and refer to a stochastic
sequence rather than a stochastic process.
22 1 Random Variables

Definition 1.7 Consider the joint distribution function (refer Sec. 1.5) of all the
random variables X = {χ(t1 ),χ(t2 ), . . .} given by

FX (x; t) = P [χ(t1 ) ≤ x1 , . . . ,χ(tn ) ≤ xn ] (1.14)

for all x = (x1 ,x2 , . . . ,xn ), t = (t1 ,t2 , . . . ,tn ) and all n. Then the nature of
FX (x; t) is the third quantity which determines the class of a stochastic process.

In this book we will consider only the class of stochastic processes known as
Markov processes.
 23

2 Markov Processes

In 1907 a Russian mathematician, A.A. Markov, described a class of stochastic

processes whose conditional probability density function is such that

P [χ(t) = x|χ(tn ) = xn ,χ(tn−1 ) = xn−1 , . . . ,χ(t0 ) = x0 ] (2.1)

= P [χ(t) = x|χ(tn ) = xn ], t > tn > tn−1 > . . . > t0

The above condition is known as the Markov property. A Markov process is a

stochastic process {χ(t), t ∈ T } for which this property holds. We will assume
that T = [0,∞) in our discussions and write S = {xi = i; i ∈ N0 }1 , the state
space of the process.
The intuitive explanation of the Markov property is to say that the future of the
process, from time tn onwards, is determined only by the present state. However
the process may have evolved to its present state χ(tn ), it does not influence
the future. However, even if the history of the system up to the present state
does influence the future behaviour, we may still be able to satisfy the Markov
assumption by a change in the state structure. Let us assume, for example, that
the next state depends on the last two states of our N state Markov chain (MC)2 .
Then we could define a new Markov process with N 2 states, where each state in
the new process would consist of successive pairs of states in the old process. In
this way the Markov property of Eq. (2.1) would still be satisfied, albeit at the
expense of considerable increase in computational complexity. Any dependence
of future behaviour on a finite number of historical steps can, at least in theory,
be treated in the same way.

Definition 2.1 Homogeneous Markov Processes. A Markov process {χ(t)} is

said to be homogeneous or stationary if the following condition holds

P [χ(t + s) = x|χ(tn + s) = xn ] = P [χ(t) = x|χ(tn ) = xn ] (2.2)

The equation expresses that a homogeneous Markov process is invariant to shifts

in time.
Throughout our discussions we shall use as an example of a Markov process
a surfer which goes from beach to beach in some random way as surfers tend
to do. We shall describe the state of this Markov process, xi , i = 1,2, . . . ,N
by the number i of the particular beach the surfer is on. In fact, for notational
convenience, we shall use throughout the integer i to denote the state xi of a
Markov process.
1
N denotes the set of positive integers and N0 additionally includes the 0.
2
A Markov process with a discrete state space is also called a Markov chain.
24 2 Markov Processes

In the case of a homogeneous Markov process, the particular instant tn in Eq. (2.2)
does not matter either so that the future of the process is completely determined
by the knowledge of the present state. In other words,

pij (t − tn ) := P [χ(t) = j|χ(tn ) = i] (2.3)

In fact, worse than that, an important implication is that the distribution of the
sojourn time in any state must be memoryless. Our surfer does not know how
long he has been at this beach! If you think about it, if the future evolution de-
pends on the present state only, it cannot depend on the amount of time spend in
the current state either.
When time is continuous, there is only one probability distribution fχ (y) of the
time y spent in a state which satisfies the property

P [χ ≥ y + s|χ ≥ s] = P [χ ≥ y]

and that is the negative exponential function

fχ (y) = λe−λy , y ≥ 0 (2.4)

In other words, the sojourn times in a Continuous Time Markov Chain (CTMC)
have an exponential probability distribution function. We will prove this fact in
Sec. 2.3 on page 46. Not surprisingly, we will meet the exponential distribution
many times in our discussions.
Similarly, for a Discrete Time Markov Chain (DTMC), the sojourn time η in a
state must be a geometrically distributed random variable (cf. Eq. (1.5))

pη (n) = P [η = n] = q n−1 (1 − q), n = 1,2,3, . . . ; 0 ≤ q < 1. (2.5)

with cumulative distribution function Fη (n)

n
X
Fη (n) = pη (k)
k=1

Note that when a process has an interarrival time distribution given by Fη (n) it
is said to be a Bernoulli arrival process. Moreover, let η = nδ for n an integer
and δ the basic unit of time. Then the mean time is given by
∞
X δ
δ kpη (k) = (2.6)
k=1
(1 − q)

from which the mean arrival rate is (1 − q)/δ.

In order to decide whether a particular process is a Markov process, it suffices to
check whether the distribution of sojourn times is either exponential or geometric
and whether the probabilities of going from one state to another only depend on
the state the process is leaving and on the destination state.
2.1 Discrete Time Markov Chains 25

Exercise 2.1 The weather bureau in an European country decided to improve its
record for weather prediction. This is made a little easier by the fact there are
never two sunny days in a row. If it is a sunny day however, the next day is just
as likely to be rainy as it is likely to be just grey and dull. If it is not a sunny day,
there is an even chance that the weather will be the same the next day. If there is
a change from a rainy or dull day, there is only a 50 percent chance that the next
day will be sunny.
1. Is the stochastic process we have just described Markovian?

2. If it is only approximately Markovian, what can one do to improve the ap-

proximation?

2.1 Discrete Time Markov Chains

In this section we concern ourselves with the case where the time spent in a
Markov state has a discrete distribution whence we have a Discrete Time Markov
Chain (DTMC).

Definition 2.2 The stochastic sequence {χn |n = 0,1,2, . . .} is a DTMC pro-

vided that

P [χn+1 = xn+1 |χn = xn ,χn−1 = xn−1 , . . . ,χ0 = x0 ] = (2.7)

P [χn+1 = xn+1 |χn = xn ]

for n ∈ N.

The expression on the right-hand side of this equation is the one-step transition
probability of the process and it denotes the probability that the process goes
from state xn to state xn+1 when the time (or index) parameter is increased from
n to n + 1. That is, using the indices for notating the states,

pij (n,n + 1) = P [χn+1 = j|χn = i]

The more general form of the sth step transition probabilities is given by

pij (n,s) = P [χs = j|χn = i]

which gives the probability that the system will be in state j at step s, given that
it was in state i at step n where s ≥ n.
Note that the probabilities pij (n,s) must satisfy the following requirements:

0 < pij (n,s) ≤ 1, i,j = 1,2, . . . ,N ; n,s = 0,1,2, . . .

X
pij (n,s) = 1, i = 1,2, . . . ,N ; n,s = 0,1,2, . . .
j∈S
26 2 Markov Processes

The probability of going from state i to state j is the probability of somehow get-
ting from i at time n to some intermediate state k at some time r and from there
to state j. The events {χr = k|χn = i} and {χs = j|χr = k} are independent,
so that using this and the fact that from the Markov property,

P [χs = j|χr = k,χn = i] = P [χs = j|χr = k]

we can write recursively over all possible intermediate states k

X
pij (n,s) = P [χr = k|χn = i]P [χs = j|χr = k]
k∈S
X
= pik (n,r)pkj (r,s) (2.8)
k

for n ≤ r ≤ s. Eq. (2.8) is known as the Chapman-Kolmogorov equation for

DTMC.
If the DTMC is homogeneous (cf. Eq. (2.2)) which will be the case in all of our
discussions, the probability of various states m steps into the future depends only
upon m and not upon the current time n; so that we may simplify the notation
and write

pij (m) = pij (n,n + m) = P [χn+m = j|χn = i]

for all m ∈ N. From the Markov property we can establish the following recur-
sive equation for calculating pij (m)
X
pij (m) = pik (m − 1)pkj (1), m = 2,3, . . . (2.9)
k

We can write Eq. (2.9) in matrix form by defining matrix P = [pij ], where
pij := pij (1), so that

P (m) = P (m−1) P (2.10)

where

P (0) = I

the identity matrix. Note that

P (1) = P (0) P = IP
P (2) = P (1) P = P 2
P (3) = P (2) P = P 3

and in general

P (m) = P m , m = 0,1,2, . . . (2.11)

2.1 Discrete Time Markov Chains 27

This equation enables us to compute the m-step transition probabilities from the
one-step transition probabilities.
(m)
Next we consider a very important quantity, the probability πj of finding our
DTMC in state j at the mth step:
(m)
πj = P [χm = j] (2.12)

How can we calculate these probabilities?

If we write
(m)
pij = pij (m) = P [χm = j|χ0 = i]

for the m-th step transition probability where we have assumed, without loss of
generality, that we entered state i at time 0, then multiplying both sides of this
(0)
equation by πi = P [χ0 = i] (cf. definition in Eq. (2.12)), summing over all
states and applying theorem of Total Probability (cf. page 14), we obtain
X (m) X
P [χ0 = i]pij = P [χ0 = i]P [χm = j|χ0 = i]
i i
X (0) (m)
πi pij = P [χm = j]
i
(m)
= πj

or, alternatively
(m) X (0) (m)
πj = πi pij (2.13)
i

That is, the state probabilities at time m can be determined by multiplying the
multistep transition probabilities by the probability of starting in each of the
states and summing over all states.
The row vector formed by the state probabilities at time m is called the state
probability vector Π(m) . That is,
(m) (m) (m)
Π(m) = (π0 ,π1 ,π2 , . . .)

With this definition, Eq. (2.13) can be written in matrix form as follows

Π(m) = Π(0) P (m) , m = 0,1,2, . . .

or from Eq. (2.11)

Π(m) = Π(0) P m , m = 0,1,2, . . . (2.14)

28 2 Markov Processes
WAIKIKI
2
3
10
1
7 2
10
1 3
2 10 3
2
10

1
IPANEMA
1
2

CLIFTON

Figure 2.1 A Markov chain.

Example 2.1 Consider the simple discrete time MC in Fig.2.1 which illustrates
the behaviour of our surfer. This diagramme is also called the state transition
diagramme of the DTMC. Every instant a unit of time elapses the surfer decides
to do something. When at the Clifton, he decides to go to Waikiki with probability
1
2 or may decide to go to Ipanema with the same probability (our surfer happens
to be very affluent). When in Ipanema he may in fact decide to stay there with
probability 12 at the end of a time period. With our beaches numbered as shown,
we have

0 0.5 0.5
 

P =  0.3 0 0.7 
0.2 0.3 0.5

Assume that our surfer starts off at Clifton (beach 1). In other words the initial
distribution is Π(0) = (1,0,0). From Clifton he can go to Ipanema or Waikiki
with equal probability, i.e.,

0 0.5 0.5
 

Π(1) = (1,0,0)  0.3 0 0.7  = (0,0.5,0.5)

0.2 0.3 0.5

from Eq. (2.14) and so on.

As we will see later, the vector Π(m) of state probabilities tends to a limit for
m → ∞. Even more, one can show that for specific DTMCs the effect of Π(0) on
the vector Π(m) completely vanishes. For our surfer that means, e.g., even if we
do not know at which beach he started the probability of finding him at a specific
beach after a long time is nearly constant. This phenomenon does not hold for
all DTMCs. Consider, e.g., the DTMC of Fig. 2.2. If the process starts in state 0
it stays there forever. But starting in state 3 there is a chance that the process gets
absorbed in state 5. Clearly, the probability Π(m) is not independent of the initial
2.1 Discrete Time Markov Chains 29

p
2 3
q p

0 p q 1
1 5

q
p
1 4
q

Figure 2.2 A simple Markov chain.

distribution. This effect or to be more precise the absence of such effects can
be verified by investigating the structure of the state transition diagramme. E.g.,
from state 0 or 5 of the DTMC given in Fig. 2.2 no other state can be reached,
thus intuitively explaining the described effect.
Next we consider a classification of Markov states based on the structure of the
state transition diagramme.
Consider states i,j ∈ S. If there is a path from i to j, i.e., there exists an integer
n (which may depend on i and j) such that

pij (n) > 0

then we write i * j.
Two states i and j are said to communicate, written i j, if there is a path from
state i to state j and vice versa.
Let C[i] = {j|i j; j ∈ S},∀i ∈ S. We call C[i] the class of state i.

Example 2.2 Consider the simple MC in Fig. 2.2. In that figure, C[0] = {0},C[5] =
{5},C[1] = {1,2,3,4}.

Definition 2.3 A MC is said to be irreducible if every state communicates with

every other state.

An irreducible MC clearly has only one class of states, i.e. C[i] = C[j] ∀i,j ∈ S.
The MC of Fig. 2.2 is reducible since 0 1 is for instance not true.
Let C denote any class of state and C be the set of Markov states not in the class
C.

Definition 2.4 A class C is said to be closed if no single-step transition is pos-

sible from any state in C to any state in C. If C consists of a single state, say i,
then i is called an absorbing state. A necessary and sufficient condition for i to
be an absorbing state is that pii = 1.
30 2 Markov Processes

Since the latter implies pij = 0 for i 6= j, an absorbing state does not communi-
cate with any other state.
The MC of Fig. 2.2 has two absorbing states, 0 and 5.
Definition 2.5 A class C is said to be transient if there is a path out of C. That
is, if ∃i ∈ C and k ∈ C such that pik > 0. The individual states in a transient
class are themselves said to be transient.
States 1, 2, 3 and 4 in the MC of Fig. 2.2 are all transient.
Definition 2.6 A MC is said to be absorbing if every state in it is either absorbing
or transient.
Finally we define an ergodic class.
Definition 2.7 A class C is said to be ergodic if every path which starts in C
remains in C. That is
X
pij = 1, ∀i ∈ C
j∈C

The individual states in an ergodic class are called ergodic. An irreducible MC

consists of a single ergodic class, i.e. C[i] = S,∀i ∈ S.
(m)
Next write fj for the probability of a Markov process leaving a state j and first
returning to the same state j in m steps. Clearly the probability of ever returning
to state j is given by

∞
X (m)
fj = fj
m=1

We now classify the states j of a MC depending on the value fj of the state. Not
surprisingly, if fj = 1 we say the state is said to be recurrent; if a return is not
certain, that is fj < 1, then state j is said to be transient. Furthermore, if our
MC can return to state j only at steps η,2η,3η, . . ., where η ≥ 2 is the largest
such integer, then state j is said to be periodic with period η. If such an integer
number η does not exist, then the state j is said to be aperiodic.
(m)
Knowing the probability fj of returning to state j in m steps, we can now
define another interesting quantity, the mean recurrence time Mj of state j.
∞
X (m)
Mj = mfj (2.15)
m=1

The mean recurrence time is thus the average number of steps needed to return
to state j for the first time after leaving it.
We can further describe a state j to be recurrent null if Mj = ∞, whereas it
is recurrent nonnull if Mj < ∞. Note that an irreducible MC can only have
recurrent null states if the number of states is infinite.
With all this in mind, we can now state the following important result[108] with-
out proof:
2.1 Discrete Time Markov Chains 31

Theorem 2.1 The states of an irreducible DTMC are all of the same type; thus
they can be either
• all transient,
• all recurrent nonnull, or
• all recurrent null.
Moreover, if periodic, then all states have the same period η.

Exercise 2.2 Assume that we don’t know for certain where our surfer has started.
An oracle tells us that he might have started at Clifton with a chance of 19%, at
Waikiki with 26% and at Ipanema, the beach he likes most, with a chance of 55%.
What is our vector π (0) now? Calculate π (1) , π (2) , π (3) .

2.1.1 Steady State Distribution

The most interesting DTMCs for performance evaluation are those whose state
(m)
probability distribution πj does not change when m → ∞ or to put it dif-
ferently, a probability distribution πj defined on the DTMC states j is said to
(m)
be stationary (or have reached a steady state distribution) if πj = πj when
(0)
πj = πj , that is, once a distribution πj has been attained, it does not change in
the future (with m).

Definition 2.8 Define the steady state probability distribution

{πj ; j ∈ S} of a DTMC by
(m)
πj = lim πj
m→∞

We are after the steady state probability distribution {πj } of being in state j at
some arbitrary point in the future. Clearly, if we know this, we can say a great
deal about the system modelled by the MC. When the DTMC is irreducible,
aperiodic and homogeneous the following theorem [108] helps us out.

Theorem 2.2 In an irreducible and aperiodic homogeneous MC the limiting

probabilities πj always exist and are independent of the initial state probabil-
ity distribution. Moreover, either
1. all states are transient or all states are recurrent null. In both cases πj =
0 ∀ j and there exists no steady state distribution, or
2. all states are recurrent nonnull and then πj > 0 ∀ j, in which case the set
{πj } is a steady state probability distribution and
1
πj = (2.16)
Mj
32 2 Markov Processes

In this case the quantities πj are uniquely determined through the following
equations
X
πi = 1 (2.17)
i
X
πi pij = πj (2.18)
i

where Mj is defined in Eq. (2.15).

A recurrent nonnull DTMC is also referred to as an ergodic MC and all the states
in such a chain are ergodic.
From our definitions in the previous section, we know that a finite MC is ergodic
if from any state it is possible to reach any other state and belong to a single
class.
The limiting probabilities πj of an ergodic DTMC are referred to as equilibrium
or steady state probabilities. It should be clear that if we observe an ergodic MC
for a fixed time T , that the average sojourn time τ i spent in state i by the DTMC
during T can be computed from
τ i = πi T (2.19)
Another quantity which will be useful to us is the average time υij spent by the
DTMC in state i between two successive visits to state j in steady state. This
quantity is also known as the visit ratio or mean number of visits, and can be
computed from
πi
υij = (2.20)
πj
Referring to Eq. (2.18), it is more convenient to find that equation expressed in
matrix notation. In order to do this we define the probability vector Π as
Π = [π0 ,π1 ,π2 , . . .]
so that we may now rewrite Eq. (2.18) as
Π = ΠP (2.21)
Note that Eq. (2.21) follows directly from the equation Π(m) = Π(m−1) P by
taking the limit as m → ∞. The following example illustrates that no unique
steady state distribution exists for a periodic MC.
Example 2.3 Consider the periodic MC illustrated in Fig.2.3 and let Π(0) =
(1,0,0). Then
Π(1) = Π(0) P = (0,1,0)
Π(2) = Π(1) P = (0,0,1)
Π(3) = Π(2) P = (1,0,0)
Π(4) = Π(3) P = (0,1,0)
...
2.1 Discrete Time Markov Chains 33

1 1

2 3
1

Figure 2.3 A simple periodic Markov chain.

0.75 0.25
1 1
2 3 4 5
1 1

Figure 2.4 A simple reducible Markov chain.

Clearly the limit Π = limm→∞ Π(m) does not exist. Similarly the MC must be
irreducible for a unique solution to exist as the following example illustrates.

Example 2.4 Consider the reducible MC illustrated in Fig.2.4 and let Π(0) =
(1,0,0,0,0). Then
 
0 0 0.75 0.25 0

 0 0 1 0 0 

(1) (0) 
Π =Π 0 1 0 0 0  = (0,0,0.75,0.25,0)


0 0 0 0 1 
 

0 0 0 1 0

Π(2) = Π(1) P = (0,0.75,0,0,0.25)

Π(3) = Π(2) P = (0,0,0.75,0.25,0)
Π(4) = Π(3) P = (0,0.75,0,0,0.25)
...

Again there is no limit Π = limm→∞ Π(m) .

So far we have not said anything about the size of the state space of our DTMC.
When the state space in our examples is finite we speak of a finite MC. The
states of a finite aperiodic irreducible DTMC are always ergodic.
In the following example we determine the steady state distribution for our surfer
example in Fig. 2.1.

Example 2.5 Using Eq. (2.21) we can write the following set of linear equations:
34 2 Markov Processes

1−p
1 2
1

p 1−q

3
q

Figure 2.5 The homogeneous discrete time MC of the exercise

π1 = 0π1 + 0.3π2 + 0.2π3

π2 = 0.5π1 + 0π2 + 0.3π3 (2.22)
π3 = 0.5π1 + 0.7π2 + 0.5π3
Note that in Eqs. (2.22) above, the last equation is a linear combination of
the second and the first, indicating that there is a linear dependence among
them. There will always be a linear dependence amongst the set of equations
in Eq. (2.21) and it is the reason why we have to use the additional Eq. (2.17)
to derive a solution. Using the latter equation and any two of the equations in
(2.22) we obtain approximately
π1 = 0.188
π2 = 0.260
π3 = 0.552
So our surfer is most likely to be found on the beach at Ipanema (probability
1
.552) and in fact he returns every 0.552 or 1.81 days to that beach.

Exercise 2.3 Consider the stochastic process described in Exercise 2.1. Let C,R
and D represent a sunny, rainy and dull day respectively and in this way define a
new stochastic process with 9 states. Determine the new transition probabilities.
Consider this process to be a discrete time MC and find the probability of two
dull days following upon one another.
Exercise 2.4 Consider the homogeneous MC illustrated in Fig. 2.5.
1. Give the probability matrix P for the chain.
2. Under what conditions will the chain be irreducible and aperiodic, if at all?
3. Solve for the steady state probability vector Π.
4. What is the mean recurrence time of state 3?
5. For what values of p and q will π1 = π2 = π3 ?
2.1 Discrete Time Markov Chains 35

2.1.2 Absorbing Chains and Transient Behaviour

When using MCs to model real systems it is often very useful to know the num-
ber of steps (or, equivalently, the time) spent in the transient states before reach-
ing an absorbing state. Think of executing a multi-layer network protocol: The
time spent by processes executing the protocol in one layer (transient states)
before going to the the next layer (absorbing state) is one example of such an
application. The absorbing MC illustrated in Fig. 2.6 consisting of a set St of nt
transient states and a set Sa of na absorbing states, illustrates what we have in
mind.

1 1

2 2

nt na

St Sa

Figure 2.6 An absorbing MC.

We begin our analysis by numbering the states in the MC such that the na ab-
sorbing states occur first and writing the transition probability matrix P as
 
I 0
P = (2.23)
R Q

Once in an absorbing state the process remains there, so I is the identity matrix
with all elements pii = 1, 1 ≤ i ≤ na . R is an nt × na matrix describing
the movement from the transient to the absorbing states, and Q is a nt × nt
36 2 Markov Processes

matrix describing the movement amongst transient states. Since it is not possible
to move from the absorbing to the transient states 0 is the na × nt zero matrix.
Since the formula for matrix multiplication also applies to matrices written in
block form, we can calculate the powers of P in terms of the matrices R and Q:
 
2 I 0
P =
R + QR Q2

and  
3 I 0
P =
R + QR + Q2 R Q3
or in general  
I 0
Pn =
N n R Qn
where N n = I + Q + Q2 + . . . + Qn−1 = ni=1 Qi−1 .
P

We can now state the following fundamental result for an absorbing MC:

Theorem 2.3 When n → ∞, then Qn → 0 and N n → (I − Q)−1 . In particular,

the matrix I - Q is invertible.

We will not prove the theorem (for a proof see [86]) but, from Eq. (2.14) above
and the knowledge that for a transient state the steady state probability πj = 0,
the first part of the result is easy to accept intuitively.
Write

N = [nij ] = (I − Q)−1

N is called the fundamental matrix of the MC.

It follows from the last theorem that
 
n I 0
lim P =
n→∞ NR 0

For absorbing chains, the only interesting starting states are the transient ones.
Assume that we start with an initial state i ∈ St . For each state j ∈ St , define
the random variable υij to be the number of visits to state j before an absorbing
state is reached. Define υij = 1 when i = j.
We know from Th. 2.2 that υij < ∞ for any transient state j, and that υij has
finite expectation. Assuming these properties we can now prove the following
theorem:

Theorem 2.4 For every pair of transient states i,j

E[υij ] = nij

where N = [nij ] is the fundamental matrix as before.

2.1 Discrete Time Markov Chains 37

Proof. Suppose that we move from starting state i to state k in the first step. If k
is an absorbing state, we can never get to state j. If k is a transient state, we are
in the same situation as before with starting state k instead. Using the Markov
property,
X
E[υij ] = δij + qik E[υkj ]
k∈St

The term δij is the Kronecker delta function with value 1 if i = j and 0 otherwise
and it counts the initial visit to state j in case the starting state is j. Denote by M
the matrix whose i,j-entry is E[υij ] for all i,j ∈ St . Then the last equation can
obviously be written

M = I + QM

so that M = (I − Q)−1 = N . t
u

Referring to Fig. 2.6, starting off in some state i ∈ St , the total number of steps
(transitions) before reaching an absorbing state is clearly the sum of times we
visit every state in St before absorption. Denote this random variable by υi and
its expected value E[υi ] by τi .

Theorem 2.5
X
τi = nij i ∈ St (2.24)
j∈St

and τi < ∞.

Proof. Since the expectation of the sum is the sum of the expectations the latter
result follows from the previous theorem. t
u

Write ~τ = (τ1 ,τ2 , . . . ,τnt ). Then

~τ = N e> (2.25)

where e is the row vector with 1’s in every position.

What about the expected number of steps needed to reach an absorbing state,
given that we will start in any of the transient states with probability distribution
~r = (r1 ,r2 , . . . ,rnt )? This quantity we will denote by the scalar value τ and refer
to it simply as the mean time before absorption.

Theorem 2.6 Let τ be the mean time before absorption of a DTMC with tran-
sient state set St = {1,2, . . . ,nt } and initial probability distribution ~r = (r1 ,r2 , . . . ,rnt ).
Then

τ = ~r N e> (2.26)
38 2 Markov Processes

Proof. Since ri is the probability of a starting state i ∈ St , and τi the expected

value of the number of steps to reach an absorbing state from that state i, the
result follows. t
u

Furthermore, define σi2 as the variance of the time υi before absorption starting
in state i ∈ St and σ~2 = (σi2 , . . . ,σn2 t ) as the vector of the variance of these
times. Let τ~2 = (E[v12 ], . . . ,E[vn2 t ]). Then it can be shown (cf. [103], page 49)
that

Theorem 2.7

σ~2 = (2N − I)~τ − τ~2 2 (2.27)

Proof. From Sec. 1.4 we know that

σi2 = E[υi2 ] − τi2 (2.28)

Using the same argument as in Th. 2.4 we write for

X X
E[υi2 ] = qik · 1 + qik E[(υi + 1)2 ]
k∈Sa k∈St
X X
= 1+ qik E[υi2 ] + 2 qik E[υi ]
k∈St k∈St
X  
= 1+ qik E[υi2 ] + 2E[υi ]
k∈St

Using vector and matrix notation, we can thus write

 
E[υi2 ] = e> + Q E[υi2 ] + 2~τ

which reduces to

E[υi2 ] = N (e> + 2Q~τ )

= ~τ + 2(N − I)~τ

by substituting for N Q and N e> from Eq. (2.26). The result follows. t
u

In theory we therefore have the formulas in Eqs. (2.26) and (2.27) to compute
the mean and variance respectively for the time before absorption. In practice,
computing the matrix N = (I − Q)−1 for a MC with a large state space is
no mean task. Courtois and Semal[56] fortunately have devised a method of
computing τ and σ 2 from P. We next describe their technique without proving
the results. Proofs can be found in [103].
To start off, we define the augmented transition probability matrix
!
Q (I − Q)e>
(2.29)
~r 0
2.1 Discrete Time Markov Chains 39

Note that (I − Q)e> is the vector of transition probabilities from the states in St
to a new state say, a ∈ Sa , and ~r and Q are the same as before.
The clever idea is that, assuming irreducibility and aperiodicity, the Markov pro-
cess defined by the matrix in Eq. (2.29), has the same behaviour as a new pro-
cess with the state a designated as an absorbing state provided one assumes that
whenever the latter chain reaches the absorbing state a, it is restarted with the
initial vector ~r and this is done infinitely many times. The new, absorbing MC is
described by the matrix

!
Q (I − Q)e>
(2.30)
0 1

Again, the ergodic behaviour of the process described by Eq. (2.29) describes the
behaviour of the absorbing chain of Eq. (2.30) over an infinite number of runs,
each started with the initial distribution vector ~r.

Theorem 2.8 If τ is the mean time before absorption of the chain

!
Q (I − Q)e>
0 1

when started with the initial distribution ~r, then

1
τ= − 1,
πa

where πa is the last component of the steady state distribution of the DTMC
described by the matrix !
Q (I − Q)e>
~r 0

The proof of this theorem can be found in [56]. Intuitively, 1/πa is the mean
time between two visits to the last state a of the Markov process described by the
matrix in Eq. (2.29) (cf. Th. 2.2) and each time the system is in this last state, one
further step is needed to restart the absorbing chain with the initial distribution ~r.
A similar result exists for the variance σ 2 of the time before absorption.

Theorem 2.9 If σ 2 is the variance of the time before absorption of the chain
!
Q (I − Q)e>
0 1

when started with the initial distribution ~r, then

σ 2 = 2τ τ 0 − τ − τ 2
40 2 Markov Processes

where τ is as before and τ 0 is given by

1
τ0 = −1
πa0
and πa0 is the last component of the steady state distribution of the DTMC de-
scribed by the matrix !
Q (I − Q)e>
~r0 0
with
1
~r0 = (π1 ,π2 , . . . ,πnt )
1 − πa
where πi is the steady state distribution of the MC given in Th. 2.8.
This concludes our study of discrete time MCs.
Exercise 2.5 Two gamblers are betting on the outcome of an unlimited sequence
of coin tosses. The first gambler always bets heads, which appears with proba-
bility p, 0 < p < 1 on every toss. The second gambler always bets tails, which
appears with probability q = 1 − p. They start with a total of C chips between
them. Whenever one gambler wins he has to give the other one chip. The game
stops when one gambler runs out of chips (is ruined). Assume the gamblers start
with C = 3 chips between them.
1. Determine the probability, in terms of p and q, that a gambler is ruined?
2. How long will the game last if the first gambler starts with 1 coin?
Exercise 2.6 Write a program to simulate the game described in the previous
exercise for a sufficient number of coin tosses. Use values of p = 0.2,0.4, . . . ,0.8.
Compare your simulation results with the theoretical answers in the previous
exercise. Assume p = 0.6.
1. Determine the theoretical mean time of a game. Compare your answer with
your simulation results.
2. Determine the theoretical variance of the duration of a game. Compare
your answer with your simulation results.

2.2 Semi-Markov Processes

In our discussions thus far the Markov process had the property that a transition
was made at every time instant. That transition may well return the process to
the same state, but a transition occurred nevertheless.
We now turn our attention to a more general class of processes where the time
between transitions may be several unit time intervals, and where this time can
depend on the particular transition being made. This process is no longer strictly
Markovian, however, it retains enough of the Markovian properties to deserve
the name semi-Markov process[90].
2.2 Semi-Markov Processes 41

2.2.1 Formal Model of a Semi-Markov Process

A semi-Markov process can be thought of as a process whose successive state

occupancies are governed by the transition probabilities of a Markov process,
but which spends time in any state described by an integer-valued random vari-
able that depends on the state currently occupied and on the state to which the
next transition will be made. At the transition instants, the semi-Markov pro-
cess behaves just like a Markov process. We call this process the embedded
Markov process and denote the single step state transitional probabilities by
pij , i,j = 1,2, . . . ,N . For simplicity we assume that the process has a finite state
space of size N .
The times at which transitions occur are now, however, governed by a different
probability mechanism. Before making the transition from state i to the next state
j, the process “sojourns”3 for a time τij in state i before proceeding to state j.
In our discrete time case, these sojourn times τij with expected value τij are
positive, integer-valued random variables, each governed by a probability mass
function sij (m), m = 1,2, . . . called the sojourn time mass function for a tran-
sition from state i to state j. That is,

P [τij = m] = sij (m), m = 1,2,3, . . . ; i,j = 1,2, . . . ,N

∞
X
τij = msij (m)
m=1

One distribution sij (m) we are familiar with is the geometric distribution

(1 − a)am−1 , m = 1,2,3, . . .

We assume that the mean of the sojourn time is finite and that the sojourn times
are at least one time unit in duration, i.e.,

sij (0) = 0

In other words, to fully describe a discrete-time semi-Markov process, we must

specify N 2 holding time mass functions, in addition to the one step transition
probabilities.
Observe that we can consider the discrete-time Markov process we discussed in
the previous section to be the discrete-time semi-Markov process for which

1,if m = 1,
sij (m) = ∀ i,j = 1,2, . . . ,N
0,if m = 2,. . .

that is, all sojourn times are exactly one time unit in length.
We next define the waiting time τi with expected value τi as the time spent in state
i, i = 1,2, . . . ,N irrespective of the successor state and we define the probability
mass function of this waiting time as
3
Some authors refer to this time as the “holding time”.
42 2 Markov Processes

N
X
P [τi = m] = pij sij (m)
j=1
N
X
τi = pij τij
j=1

That is, the probability that the system will spend m time units in state i if we do
not know its successor state, is the probability that it will spend m time units in
state i if its successor state is j, multiplied by by the probability that its successor
state will indeed be j and summed over all possible successor states.

2.2.2 Interval Transition Probabilities

As in the DTMC case, we next set out to compute the n−step transition proba-
bilities, which we denote φij (n), for the semi-Markov case. That is, how can a
process that started by entering state i at time 0 be in state j at time n?
One way this can happen is for i and j to be the same state and for the pro-
cess never to have left state i throughout the period (0,n). This requires that the
process makes its first transition, to any other state, only after time n. That is

δij P [τi > n]

where

1, if i = j,
δij =
0, otherwise

It is not difficult to see that the latter term can be written as

 
n X
X N
δij 1 − pij sij (m) (2.31)
m=0 j=1

 
Let W (n) = {δij 1 − nm=0 N j=1 pij sij (m) } be the matrix of these ele-
P P

ments.
Every other way to get from state i to j in the interval (0,n) would mean the
process made its first transition from state i to some other state k at a time m,
and then by a succession of such transitions to state j at time n. Note that we
have to consider all intermediate times m, 0 < m ≤ n and intermediate states
k ∈ S, S the Markov state space. In other words
n X
X N
pik P [τik = m]φkj (n − m) (2.32)
m=0 k=1

where i,j = 1,2, . . . ,N ; n = 0,1,2, . . . or

2.2 Semi-Markov Processes 43

n X
X N
pik sik (m)φkj (n − m) (2.33)
m=0 k=1

Define the congruent matrix multiplication C = A◦B of two matrices A = {aij }

and B = {bij } by C = {aij bij }, for all i,j. Furthermore, write

Φ(n) = {φij (n)}

n
X
S(n) = {1 − sij (m)}
m=0

With these definitions, Eq. (2.33) becomes

n
X
(P ◦ S(m))Φ(n − m) n = 0,1,2, . . .
m=0

and
n
X
Φ(n) = W (n) + (P ◦ S(m))Φ(n − m) n = 0,1,2, . . . (2.34)
m=0

Φ(n) is called the interval transition probability matrix for the semi-Markov pro-
cess in the interval (0,n) and clearly

Φ(0) = I

Eq. (2.34) provides a convenient recursive basis for computing Φ(n) for any
semi-Markov process. The quantities P and S(m) come directly from the defi-
nition of the process.

Example 2.6 Consider again our typical surfer of Fig. 2.1 on page 28, but let
us now give the example a semi-Markovian flavour. The nature of the surfer now
dictates that the length of time he will stay on a particular beach will depend on
both which beach he is on and where he intends to go next. The (sojourn) time
τij is thus the length of time spent surfing on beach i with the intention of going
to beach j (where j = i is certainly possible). The lifeguards on each beach
have been keeping record of our surfer and have come up with the following
probability mass functions describing the surfer’s behaviour:
   m−1    m−1    m−1
1 2 1 3 1 2
s11 (m) = s12 (m) = s13 (m) =
3 3 4 4 3 3
   m−1    m−1    m−1
1 1 1 7 2 3
s21 (m) = s22 (m) = s23 (m) =
2 2 8 8 5 5
   m−1    m−1    m−1
1 3 1 2 1 1
s31 (m) = s23 (m) = s33 (m) =
4 4 3 3 2 2

for m = 1,2,3, . . .
44 2 Markov Processes

Solution. Consider the general geometric distribution function (cf. Sec. 1.2)
p(n) = (1 − a)a(n−1) . The first moment or mean n can be calculated in the fol-
lowing way:

∞
X
n= n(1 − a)a(n−1)
n=0
P d d P
Using the property that da = da we write
∞
X d n
n = (1 − a) a
n=0
da
d 1
= (1 − a)
da 1 − a
1
=
1−a
so that the first moment τ ij in the example is given by

τ 11 = 3; τ 12 = 4; τ 13 = 3;
τ 21 = 2; τ 22 = 8; τ 23 = 2.5;
τ 31 = 4; τ 32 = 3; τ 33 = 2;

Clearly beach 2 (Waikiki) is the most popular with our surfer, since he spends 8
units of time on the average surfing there and then immediately returning to it.
The mean time he spends on that beach, irrespective of where he might go next
is given by

τ2 = p21 τ 21 + p22 τ 22 + p23 τ 23

= 0.3 × 2 + 0 × 8 + 0.7 × 2.5
= 2.35

Exercise 2.7 In Ex. 2.6, compute the mean time the surfer spends on Ipanema,
assuming that he has arrived there from anywhere else but Waikiki.

2.2.3 Steady State Behaviour

We now set out to find the limiting behaviour of the interval transition prob-
abilities over long intervals. It is important to note that the MC structure of a
semi-Markov process is the same as that of its embedded Markov process. There-
fore the interval transition probabilities of a semi-Markov process can exhibit a
unique limiting behaviour only within the chain of the embedded Markov pro-
cess.
We begin by defining a limiting interval transition probability matrix Φ for our
process by
2.2 Semi-Markov Processes 45

Φ = lim Φ(n)
n→∞
with elements φij . However, in steady state, the limiting interval transition prob-
abilities φij do not depend on the starting state i and we therefore write φij = φj .
Define a vector ϕ = (φ0 ,φ1 , . . . ,φN ) as the vector of probabilities φj that the
semi-Markov process is in state j as time n → ∞ and let Π = {π0 ,π1 , . . . ,πN }
be the steady state probability vector of the equivalent embedded MC [cf. Eq. (2.21)].
One can prove (see e.g., Howard[90]) that
πj τ j
φj = PN (2.35)
i=1 πi τ i
or
1
ϕ = ΠM
τ
where we have written
N
X
τ= πj τ j
j=1

and M is the diagonal matrix [τ j ] of mean waiting times.

Although the proof requires transform analyses which are beyond the scope of
this book, Eq. (2.35) is intuitively somewhat obvious in the following way: The
probability φj of finding the semi-Markov process in state j is the product of the
average time τ j spent in that state, multiplied by the probability πj of being in
that state in the embedded MC and normalised to the mean total time τ spent in
all of the N possible states.
Note that Eq. (2.35) can also be written as
τj
φj = PN πi
i=1 πj τ i
τj
= PN
i=1 υij τ i
where υij , given by Eq. (2.20), is the visit ratio defined on page 32.
Example 2.7 Suppose we want to know the steady state probability of finding
our surfer on Waikiki beach. This we can do by applying Eq. (2.35). In Ex. 2.5
we determined that
Π = (0.188,0.260,0.552)
so that
π2 τ 2
φ2 =
π1 τ 1 + π2 τ 2 + π 3 τ 3
0.260 × 2.35
=
0.188 × 3.5 + 0.260 × 2.35 + 0.552 × 2.7
= 0.22
There is thus a 22 percent chance of finding him on Waikiki or beach number 2.
46 2 Markov Processes

Exercise 2.8 A car rental company has determined that when a car is rented in
Town 1 there is a 0.8 probability that it will be returned to the same town and
a 0.2 probability that it will be returned to Town 2. When rented in Town 2,
there is a 0.7 probability that the car will be returned to Town 2, otherwise it
is returned to Town 1. From its records, the company determined that the rental
period probability mass functions are:
   m−1    m−1
1 2 1 5
s11 (m) = s12 (m) = m = 1,2,3, . . .
3 3 6 6
   m−1   m−1
1 3 1 11

s21 (m) = s22 (m) = m = 1,2,3, . . .
4 4 12 12

What percentage of the time does a car spend in Town 2?

2.3 Continuous Time Markov Chains

In Sec. 2.1 we described how our surfer had to decide at regular, equal intervals
of time whether to leave or whether to stay on the beach where he is. If we now
allow him to decide at an arbitrary time which beach to go to next, we have the
continuous-time version of that example.
The Continuous Time Markov Chain (CTMC) version of the Markov property,
Eq. (2.1), is given by

Definition 2.9 The stochastic process {χ(t),t ≥ 0} is a CTMC provided

P [χ(t) = x|χ(tn ) = xn ,χ(tn−1 ) = xn−1 , . . . ,χ(t0 ) = x0 ] (2.36)

= P [χ(t) = x|χ(tn ) = xn ]

for any sequence t0 ,t1 , . . . ,tn such that t0 < t1 < . . . < tn and xk ∈ S where S
is the (discrete) state space of the process.

The right-hand side of the above equation is the transition probability of the
CTMC and we write

pij (t,s) = P [χ(s) = xj |χ(t) = xi ]

to identify the probability that the process will be in state xj at time s, given that
it is in state xi at time t ≤ s. Since we are still considering discrete state Markov
processes (chains) we will continue to use i ∈ N rather than xi to denote a state
of our Markov processes.
Note that we need to define

1, if i = j
pij (t,t) = (2.37)
0, otherwise
2.3 Continuous Time Markov Chains 47

to establish the fact that the process may not leave immediately to another state.
We already mentioned in Sec. 2 on page 24 that the time a Markov process
spends in any state has to be memoryless. In the case of a DTMC this means that
the chain must have geometrically distributed state sojourn times while a CTMC
must have exponentially distributed sojourn times. This is such an important
property that we include a proof from Kleinrock[108] of it here. The proof is
also instructive in itself.
Let yi be a random variable which describes the time spent in state i. The Markov
property specifies that we may not remember how long we have been in state i
which means that the remaining sojourn time in i may only depend upon i. As-
sume that this remaining time t has distribution h(t). Then our Markov property
insists that

P [yi > s + t|yi > s] = h(t)

i.e. h(t) is independent of s. Using conditional probabilities, we may write

P [yi > s + t ∩ yi > s]

P [yi > s + t|yi > s] =
P [yi > s]
P [yi > s + t]
=
[yi > s]

where the last step follows from the fact that the event yi > s + t implies that
yi > s. Rewriting the last equation we obtain

P [yi > s + t] = P [yi > s]h(t) (2.38)

Setting s = 0 and since we know P [yi > 0] = 1 we have

P [yi > t] = h(t)

which we then substitute in Eq. (2.38) to obtain the relation

P [yi > s + t] = P [yi > s]P [yi > t] (2.39)

for all s,t ≥ 0. All that remains is to show that the only continuous distribution
function which satisfies Eq. (2.39) is the negative exponential distribution. Write
fi (t) for the corresponding density function. Then we can write

d d
P [yi > t] = (1 − P [yi ≤ t])
dt dt
= −fi (t) (2.40)

Using the latter result and differentiating Eq. (2.39) with respect to s we obtain

dP [yi > s + t]
= −fi (s)P [yi > t]
ds
48 2 Markov Processes

Dividing both sides by P [yi > t] and setting s = 0 we obtain the differential
equation
dP [yi > t]
= −fi (0)ds
P [yi > t]
which by using (2.40) gives the following density function
fi (t) = fi (0)e−fi (0)t
for all t ≥ 0. Setting λ = fi (0) we are back to fi (x) = λe−λx which we had
before in Eq. (2.4) on page 24. 2
In other words, if a stochastic process has the Markovian property, the time spent
in a state will have a negative exponential distribution. This may seem rather
restrictive since many real systems do not have exponential time distributions.
In 1955 Cox [57] showed however, that a random variable χ whose distribution
Fχ (x) has a rational Laplace transform can be represented as a series-parallel
network of stages as illustrated in Fig. 2.7 where the time distribution of each
stage has a different negative exponential distribution. In that diagramme stage
j, j = 2, . . . ,k follows stage j − 1 with probability αj or does not follow with
probability 1 − αj .
The first and second moments of this distribution are given by [102]

k i
X X 1
E[χ] = δi
i=1 j=1
µj
  2 
k i i
X X 1 X 1 
E[χ2 ] = δi  +

µ2 µ

i=1 j=1 j j=1 j

where
i−1
Y
δi = (1 − αi ) αj
j=1

α1 α2
µ1 µ2 ... µk

1 − α1 1 − α2 1 − αk−1
...

Figure 2.7 The Coxian stage-type distribution.

If we now encode the stage number in which the process finds itself in the state
descriptor of our stochastic process, it is clear that the process will again be
Markovian.
2.3 Continuous Time Markov Chains 49

In fact one can approximate almost any distribution as closely as one wishes. The
drawback is that the size of the corresponding state space increases by a factor
k.
The obvious goal of CTMC analysis is to find the probability that the process
will be in state i at a given time t or alternatively, its steady state distribution.
In the continuous time case the solution is different from that for discrete time
in Th. 2.2, although the arguments are analogous. Let us now find that solution.
First of all define the following time-dependent transition probability

pij (s,t) = P [χ(t) = j|χ(s) = i]

where the process is in state j at time t given that it was in state i at time s; s ≤ t.
Using the same arguments as in Sec. 2.1 on page 25 that the process must pass
at some intermediate time u through some intermediate state k in order to get
from state i to state j we can arrive at the Chapman-Kolmogorov equation for
continuous-time MCs:
X
pij (s,t) = pik (s,u)pkj (u,t) (2.41)
k

where i,j = 0,1,2, . . . Using the notation

H(s,t) = (pij (s,t))

we can write Eq. (2.41) as

H(s,t) = H(s,u)H(u,t), s≤u≤t

where from Eq. (2.37)

H(t,t) = I

for I the identity matrix.

In Sec. 2.1 the solution Π we wanted was given by Π = ΠP with P the one-step
transition probability matrix. For the CTMC the one-step transition probabilities
are replaced by the infinitesimal rates which are in turn given in terms of the
time derivative of pij (s,t) as t → s.
In order to determine how this comes about we first of all write from Eq. (2.41)
X
pij (s,t + ∆t) = pik (s,t)pkj (t,t + ∆t)
k

Subtracting pij (s,t) from both sides and dividing by ∆t we obtain

pij (s,t + ∆t) − pij (s,t) X pkj (t,t + ∆t) − pkj (t,t)
 
= pik (s,t)
∆t k
∆t

since pij (s,t) = k pik (s,t)pkj (t,t). ∆t → 0 yields the following alternate form
P

of the forward Chapman-Kolmogorov equations for continuous-time MCs

50 2 Markov Processes

∂H(s,t)
= H(s,t)Q(t), s≤t (2.42)
∂t
where we have written the matrix Q(t) for

H(t,t + ∆t) − I
Q(t) = lim
∆t→0 ∆t
whose elements are defined as follows
pii (t,t + ∆t) − 1
qii (t) = lim
∆t→0 ∆t
pij (t,t + ∆t)
qij (t) = lim , i 6= j
∆t→0 ∆t
The intuitive interpretation of the quantities qii (t) and qij (t) is that when our
process is in state i at time t, then the probability that a transition occurs to any
state other than i is given by −qii (t)∆t + o(∆t). Thus, −qii (t) is the rate at
which the process leaves state i and similarly the probability of a transition to
state j in time interval ∆t is qij (t)∆t + o(∆t) or rate qij (t). Since we know that
P
j pij (s,t) = 1 it follows that
X
qij (t) = 0, ∀ i (2.43)
j

Using the same arguments as above we could also derive the backward Chapman-
Kolmogorov equations

∂H(s,t)
= −Q(s)H(s,t), s≤t
∂s
Note that in terms of the individual matrix elements we can write the forward
Chapman-Kolmogorov equations as

∂pij (s,t) X
= qjj (t)pij (s,t) + qkj (t)pik (s,t) (2.44)
∂t k6=j

The backward Chapman-Kolmogorov equations can be written as

∂pij (s,t) X
= −qii (s)pij (s,t) − qik (s)pkj (s,t) (2.45)
∂s k6=i

From these equations we can solve for H(s,t), but what we really need to know
is the probability πj (t) that the Markov process will be in state j at time t. Define
the vector of state probabilities

Π(t) = (π0 (t),π1 (t),π2 (t), . . .)

2.3 Continuous Time Markov Chains 51

If we are given the initial state distribution Π(0) then we can solve for the time-
dependent state probabilities in a way similar to that in Sec. 2.1.1 from

Π(t) = Π(0)H(0,t)

Differentiating both sides of the last equation we obtain after applying Eq. (2.42)
dΠ(t) dH(0,t)
= Π(0)
dt dt
= Π(0)H(0,t)Q(t)
= Π(t)Q(t) (2.46)

Writing the latter matrix equation in terms of its elements, we obtain

dπj (t) X
= qjj (t)πj (t) + qkj (t)πk (t) (2.47)
dt k6=j

Note the similarity between Eqs. (2.44) and (2.47). The former equation de-
scribes the probability that the process is in state j at time t given that it was
in state i at time s.
We now know that we can solve for the time-dependent probability distributions
Π(t) from Eq. (2.47). The explicit solution of this system of differential equa-
tions is given by
Rt
Q(u)du
Π(t) = Π(0)e 0

which is difficult to solve in most interesting cases and numerical methods have
to be used. As in the case of the DTMC we therefore fall back on the steady state
distribution when t → ∞ while remembering that there may indeed be cases
where the time-dependent solution is very useful.
Before concluding this section we would like to note that concepts such as
reducibility, transient and absorbing states and chains which we discussed for
DTMCs apply to CTMCs as well and we will not repeat those discussions here.

2.3.1 Steady State Distribution

Consider the case where our CTMC is homogeneous. If so, we can drop the
dependence on time and adopt the following notation

pij (t) = pij (s,s + t)

qij = qij (t), i,j = 1,2, . . .
H(t) = H(s,s + t) = (pij (t))
Q = Q(t) = (qij )

The forward and backward Chapman-Kolmogorov equations now become, re-

spectively,
52 2 Markov Processes

dpij (t) X
= qjj pij (t) + qkj pik (t) (2.48)
dt k6=j
and
dpij (t) X
− = −qii pij (t) − qik pkj (t)
dt k6=i
or for the state probabilities themselves
dπj (t) X
= qjj πj (t) + qkj πk (t) (2.49)
dt k6=j
When our CTMC chain is irreducible, homogeneous and all states are recurrent
nonnull, it can be shown (cf. [108]) that the following limit exists and is inde-
pendent of the initial state distribution. That is,
lim πj (t) = πj
t→∞
The vector Π = (π0 ,π1 ,π2 , . . .) is the steady state probability distribution we are
after. As was the case for a DTMC, the MC is said to be ergodic.
This limiting distribution is given uniquely by the solution of the following sys-
tem of linear equations
X
− qjj πj + qkj πk = 0 (2.50)
k6=j
X
πj = 1
j
where the latter equation again follows from the rules for probability conserva-
tion. In matrix form Eq. (2.50) may be expressed as
ΠQ = 0 (2.51)
This latter equation is the CTMC equivalent of the very important Eq. (2.21) on
page 32 for the DTMC case. In the latter case P was the matrix of transition
probabilities, whereas the so-called infinitesimal generator Q is a matrix of
transition rates.
Eq. (2.50) are also known as the global balance equations of the MC. If in
addition the following property holds, namely
qij πi = qji πj ∀i,j (2.52)
then the MC is said to satisfy local balance and the solution to Eq. (2.52) is also
a solution to Eq. (2.50). Note that the reverse is not necessarily true.
Exercise 2.9 A computing center has only one technician and 3 computers. Each
computer randomly breaks down at an exponential rate of twice a day, and the
technician fixes them at an exponential rate of 3 computers per day. Whenever
all three computers are broken, the users help the technician and 4 computers
are fixed per day. Answer the following:
1. On the average how many computers are out of order?
2. At what mean rate are the computers breaking down?
2.4 Embedded Markov Chains 53

2.4 Embedded Markov Chains

In the case of a continuous time Markov process it is again possible to have an

embedded DTMC and arrive at a solution for the steady state probability distri-
bution in the way described in Sec. 2.2. The steady state transition probabilities

pij = P [χn+1 = j|χn = i]

can be calculated as a function of the transition rates qij . In the case of a CTMC
the sojourn times τij ; i,j = 1,2, . . . ,N (cf. page 48) are all exponentially dis-
tributed with parameter qij .
That is,

P [τij ≤ x] = 1 − e−qij x (2.53)

or P [τij > x] = e−qij x (2.54)

The probability pij is the same as the probability that τij is the smallest of the
sojourn times τik , k 6= i. Denote this greatest lower bound by τglb . Then

[
P [τglb ≤ x] = P [ {τik ≤ x}]
k6=i
k6=j
\
= 1 − P[ {τik > x}]
k6=i
k6=j
P
− k6=i qik x
or Fτglb (x) = 1 − e k6=j

from Eq. (2.54). The corresponding density function fτglb (x) is thus
P
X − k6=i qik x
fτglb (x) = − qik e k6=j

k6=i
k6=j

However,
X
−qii = qik
k6=i
X
or − (qii + qij ) = − qik
k6=i
k6=j

so that

fτglb (x) = −(qii + qij )e(qii +qij )x

It follows that
54 2 Markov Processes
Z ∞
pij = fτglb (x)P [τij ≤ x]dx
0
Z ∞
= − (1 − e−qij x )(qii + qij )e(qii +qij )x dx
0
qij
=
−qii

The steady state probabilities π are then computed in a manner similar to that
described in Sec. 2.2.
This completes our discussion of discrete-state Markov processes.

Exercise 2.10 Consider the car rental exercise of Ex. 2.8 except that now you are
given that cars with destination Town 2 are rented at the average rate of 2 per
day in Town 1 and 14 per day in Town 2. Cars destined for Town 1 are rented at
the average rate of 6 per day in Town 2 and 8 per day in Town 1. Again calculate
the percentage of the time a car spends in Town 2?
 55

3 General Queueing Systems

With the basic theory of stochastic processes behind us we can now proceed to
apply that theory to model the performance of systems. Think about a typical
computer system, be it a multiprocessor, a computer network or a database sys-
tem. In all cases we have entities (processes, packets, transactions) requiring ser-
vice from some resource or set of resources (central processors, nodal processor,
file server). In doing so, these entities usually have to wait their turn before be-
ing served in some order. Even humans are regrettably only too familiar with the
world we have just described: it is known as queueing and a vast body of theory
known as queueing theory has developed around the subject over the years.
What then is the link between queueing theory and stochastic processes? As you
know from experience it is not always possible to predict the length of the waiting
line at the post office or bank. We know that the length of that line is a stochastic
process as is, for instance, the number of packets in the input buffer at a network
node at any time. It would thus make sense to have an understanding of basic
queueing theory in our study of performance modelling.
We focus our attention on the general queueing system illustrated in Fig. 3.1. In
queueing theory terminology the entities requiring service are called customers
and they are said to wait in a waiting line to be served by one or more servers
using some scheduling strategy such as First-Come-First-Served (FCFS). The
queue may have a finite capacity while the customers in turn may come from a
finite population.
In order to describe and analyse a queue we tag each customer with a subscript
n and denote the customer itself by Cn . So Cn is the nth customer to enter the
(queueing) system. As suggested already, we define the random process N (t) as

N (t) := number of customers in the system at time t.

Another stochastic process is the unfinished work U (t) that exists in the system
at time t. When U (t) > 0, then the system is said to be busy, and when U (t) = 0

arrivals departures

WAITING LINE

QUEUE

Figure 3.1 A queue.

56 3 General Queueing Systems

sn

Cn−1 Cn Cn+1 Cn+2

wn xn xn+1 xn+2

SERVER
time

Cn Cn+1 Cn+2

WAITING
τn τn+1 τn+2
LINE
time
tn+1 tn+2

Cn Cn+1 Cn+2

Figure 3.2 Time sequence diagramme describing the queueing process.

it is said to be idle. These stochastic variables are displayed in the time sequence
diagramme in Fig. 3.2 where a customer Cn ,Cn+1 , . . . arrives at the queue, waits
for a time wn before proceeding to the server as soon as the previous customer
Cn−1 ,Cn , . . . has completed his service (FCFS service) and eventually leaves
after a service time of duration xn .
More formally then, we define the arrival time of customer Cn to the queueing
system as the instant

τn := arrival time for Cn

and define the time between the arrivals of Cn−1 and Cn (the interarrival time)
as

tn := interarrival time between Cn−1 and Cn

= τn − τn−1

We assume moreover that the interarrival times are drawn from a continuous
distribution A(t), so that

P [tn ≤ t] = A(t)

independent of n. The average interarrival time is a quantity of special signifi-

cance in queueing theory and it is given the symbol 1/λ. That is
Z ∞
1
= tdA(t).
λ 0
 57

Note that the average arrival rate is given by λ. Similarly, we define the service
time for Cn as

xn := service time for Cn

and write B(x) for the distribution of that time. That is

P [xn ≤ x] = B(x)

and write µ for the average service rate with

Z ∞
1
= xdB(x)
µ 0

for the average service time which are also quantities we will come across very
frequently.
The sequences {tn } and {xn } are the independent variables of the queueing
system. The quantity of most interest to us is the waiting time which we define
as

wn := time taken by Cn waiting in line.

The total time spent in the system by Cn , which we will call the queueing time1
is thus given by

sn := waiting time plus service time for Cn

= wn + xn (3.1)

Thus we have defined the interarrival time, the waiting time, the service time and
the queueing time for our customer Cn . Before concluding this section we would
like to introduce a very convenient shorthand notation that is in common use for
describing a queue. The reader will suspect already that the mathematical notion
of a queue is described by the arrival distribution A(t) and the service distribution
B(x). Other determining attributes are the number m of servers and the system
capacity K, the total number of customers, including those in service, the queue
can accommodate. Another attribute is the scheduling strategy at the server while
it may also be necessary to specify the size of the customer population if this is
not infinite.
The shorthand way of describing the abovementioned attributes of a queue, known
as Kendall’s notation, is to write A/B/m/K/Z/Sched to identify the arrival pro-
cess/service process/number of servers/ maximum capacity/user population/scheduling
strategy of a queue. In the description of the arrival or service process the fol-
lowing conventions are used:
1
In this book, the terms “waiting time” and “queueing time” will refer to the time waiting in line
and the total time spent waiting in line and being served respectively. The meaning of these is
not consistent throughout the queueing theory literature and the wise reader will make sure he
knows which meaning the particular author uses.
58 3 General Queueing Systems

G: general distribution

M: exponential distribution

C k: k-phase Coxian distribution

D: deterministic distribution

Thus the queueing system M/G/1///FCFS has exponentially distributed interar-

rival times, a general service time distribution, a single server, infinite capacity,
infinite user population and uses the FCFS scheduling strategy. Usually if a field
is empty the extra slashes are left out, e.g. M/G/1/FCFS.

Exercise 3.1 Think of any freeway in your city. Can that be considered as a
queueing system? What would the customers be? And the server(s)?

Exercise 3.2 Give the shorthand notation for a queue with fixed interarrival
times and a 4-phase Coxian type service time distribution where the capacity
of the queue is K.

3.1 Little’s Law

Most of us would suspect that the faster customers arrive at a queue and the
heavier their service requirements, the longer they will spend in the queueing
system. This is indeed the case and the rule which describes this relationship
is known as Little’s Law (or theorem). A theoretical proof of this law can be
found in the original work of Little [115] and several other papers [68, 100, 170,
180]. We will use operational (i.e., “how it works”) arguments to derive the same
result.
Suppose we observe the queueing process described by Fig. 3.2 for an interval
of time (0,t), at the beginning of which there are no customers and count the
number a(t) of customers which arrive during this time. That is

a(t) := number of arrivals in (0,t)

Simultaneously we count the number d(t) of departures during the same interval.
Thus

d(t) := number of departures in (0,t)

A typical behaviour for each of these two functions is illustrated in Fig. 3.3.
The number N (t) of customers in the system at time t is then clearly

N (t) = a(t) − d(t)

3.1 Little’s Law 59

13
12
11 N (t)
Number of
10 Customers
9 a(t)
8
7
6
5
4 d(t)
3
2
1
TIME t

Figure 3.3 Arrival and departure functions at a typical queue.

The area from 0 to time t between the curves a(t) and d(t) in Fig. 3.3 is the
integral over time of the quantity N (t) and obviously represents the time spent
by all the customers in the system (measured in units of customer-seconds) up
to point t. Denote this area by I(t). Moreover let λt be defined as the average
arrival rate of customers during the interval (0,t); that is,

a(t)
λt :=
t
If we define Tt as the average queueing time (remember, waiting time plus ser-
vice time) per customer during the interval (0,t), then clearly

I(t)
Tt =
a(t)
Finally define Nt as the average number of customers in the system during the
interval (0,t), which can obviously be computed from

I(t)
Nt =
t
a(t) I(t)
=
t a(t)
= λt Tt
60 3 General Queueing Systems

1
..
.

Central Computer

Figure 3.4 Simplified computer system used in the exercise

Assuming that the limits

λ = lim λt
t→∞
T = lim Tt
t→∞
exist, then so will the limit N for Nt the average number of customers in the
system. The mean arrival rate λ we have seen before and T is merely the average
value of the time sn in the system defined in Eq. (3.1) on page 57. This leads us
to the following important result
Theorem 3.1 Little’s Law. The average number N of customers in a queueing
system with arbitrary service and arrival distribution, with average arrival rate
λ and average time in the system T , is given by
N = λT (3.2)
Note that in the derivation described earlier in this section we made no assump-
tions about the distributions of the service and interarrival times nor did we spec-
ify the service discipline or number of servers. Little’s law thus applies regard-
less and can be applied to any system with average arrival rate λ and average
time T in the system. Moreover, if we know any two of the three unknowns in
the equation we can always compute the third.
There are also extensions to Little’s law dealing with the calculation of moments
of the random variables [46, 116, 120].
Exercise 3.3 Consider the computer system in Fig. 3.4. Assume that the think
time at each of the M terminals has a mean of 15 seconds and that, on the aver-
age, half of them are busy “thinking”. What is the delay in the central computer
when there are 10 terminals in use?
3.2 Birth-Death Processes 61

3.2 Birth-Death Processes

A very useful class of Markov chain is the birth-death process. Birth-death pro-
cesses may be either discrete- or continuous-time processes in which the defin-
ing condition is that state transitions take place between state k and its nearest
neighbours k − 1,k and k + 1 only. In fact, as the name indicates, the birth-death
process is appropriate for modelling changes in the size of a population. When
the population is of size k we will say that the system is in state k. Moreover
a transition from state k to k + 1 will signify a “birth” within the population,
whereas a transition from k to k − 1 will denote a “death” in the population.2
Although the time parameter may be discrete as well (as the state space) for
the birth-death process, the continuous-time case is much more interesting and
the one we will analyse. We begin by introducing the notion of a birth rate λk ,
which describes the rate at which births occur when the population is of size k,
and similarly we define a death rate µk which is the rate at which deaths occur
when the population is of size k. Note that these rates depend only on k and are
independent of time so that our Markov chain is thus homogeneous. In terms of
our notation in the previous chapter, we have

λk , if j = k + 1




µk , if j = k − 1
qkj =

 −(λ k + µk ), if j = k

0, otherwise

Note that the fact that qkk = −(λk + µk ) above, follows directly from Eq. (2.43)
on page 50. Note the assumptions that λk > 0 for all k, that µ0 = 0 (no deaths
are possible in a population of size 0), and that µk > 0 for all k > 0.
Using the notation developed in the previous chapter we can write the infinitesi-
mal generator Q for the general homogeneous birth-death process as
 
−λ0 λ0 0 0 0

 µ1 −(λ1 + µ1 ) λ1 0 0 

0 µ2 −(λ2 + µ2 ) λ2 0
 
Q= 
0 0 µ3 −(λ3 + µ3 ) λ3 ... 
 

..
 
.

Note that the sum of all elements in each row of the matrix Q is zero as it al-
ways will be. Compare this with the sum of all row elements of the transition
probability matrix P in Sec. 2.1 which will always be 1.
The state probabilities πk (t), k = 0,1, . . . at time t can thus be found by solving
the set of differential equations equivalent to Eq. (2.47)
2
The astute reader will already suspect that this process resembles a queue where a “birth”
signifies an arrival to and a “death” a departure from the queue. That this is indeed the case for
a queue with very specific interarrival time and service time distributions (the M/M/1 queue)
will be discussed at the end of this chapter.
62 3 General Queueing Systems

λ0 λ1 λ2 λk−1 λk λk+1

0 1 2 ... k-1 k k+1 ...

µ1 µ2 µ3 µk µk+1 µk+2

Figure 3.5 State-transition diagramme of the birth-death process.

dπk (t)
= −(λk + µk )πk (t) + λk−1 πk−1 (t) + µk+1 πk+1 (t),k ≥ 1 (3.3)
dt
dπ0 (t)
= −λ0 π0 (t) + µ1 π1 (t) (3.4)
dt
These equations can be solved for the general transient case, but that is not easy.
Instead, we consider the steady state probabilities πk , k = 0,1, . . .. In that case
the above equations are a particular case of Eq. (2.51) and can be written

− (λk + µk )πk + λk−1 πk−1 + µk+1 πk+1 = 0, k≥1 (3.5)

−λ0 π0 + µ1 π1 = 0 (3.6)

Eqs. (3.5) and (3.6) are known as the global balance equations. Before pro-
ceeding to solve these equations we digress to Fig. 3.5 which illustrates the state
transition diagramme of the birth-death process. Concentrating on state k we ob-
serve that one may enter it only from the state k − 1 or from the state k + 1 and
similarly one leaves state k only to enter state k − 1 or state k + 1. From that
diagramme it is clear why we refer to the process we described as the nearest-
neighbour, birth-death process.
The clever thing is to note that we can obtain Eqs. (3.5) – (3.6) directly from the
state transition diagramme in Fig. 3.5 by equating the rates of flow into and out
of each state k, k = 0,1, . . . Indeed, the flow into state k comes from state k − 1
(with probability πk−1 ) at rate λk−1 , and from state k + 1 at rate µk+1 . The total
flow into state k is the sum of the rates weighted by the probabilities of the origin
states, or

flow into state k = λk−1 πk−1 + µk+1 πk+1

The flow out of state k is the product of the probability of state k and the sum of
the rates out of state k. That is,

flow out of state k = (λk + µk )πk

Using the fact that under steady state conditions the flow out of state k equals the
flow into state k and Eqs. (3.5) – (3.6) follow. We can solve the latter equations
from the fact that
λ0
π1 = π0
µ1
3.3 Poisson Process 63

and recursively from there on

k−1
Y λj
πk = π0 (3.7)
j=0
µj+1
P
for all k = 1,2, . . . Imposing the normalising condition k πk = 1 we finally
obtain
 −1
X k−1
Y λj
π 0 = 1 + 
k6=0 j=0
µj+1

Substituting into Eq. (3.7) we obtain

! −1
k−1 X k−1
Y λi Y λj
πk = 1 +  (3.8)
i=0
µi+1 k6=0 j=0
µj+1

Note that for this limiting distribution to exist we must have

X k−1
Y λj
<∞
k6=0 j=0
µj+1

and we call the latter the condition of ergodicity (i.e., that a steady state solution
exists) for the birth-death process.

Exercise 3.4 Consider a birth-death system with the following birth and death
coefficients

λk = (k + 2)λ k = 0,1,2, . . .
µk = kµ k = 1,2,3, . . .

1. Find πk and give the condition of ergodicity.

2. Determine the average number of customers in the system.

3.3 Poisson Process

We mentioned in the last section that to find the transient solution for Eqs. (3.3) –
(3.4) is not very easy in general. In fact, more than that, the birth-death process in
the general case is not all that interesting either. A special case which is, however,
of central importance in queueing theory, is that for a pure birth process in which
we assume that µk = 0 for all k. Moreover, to simplify the problem even further,
we will assume that λk = λ for all k = 0,1,2, . . . Substituting this into Eqs. (3.3)
– (3.4) we have
64 3 General Queueing Systems

dπk (t)
= −λπk (t) + λπk−1 (t), k≥1 (3.9)
dt
dπ0 (t)
= −λπ0 (t) (3.10)
dt
Assuming that the process starts out from state 0 at time t = 0, or

1, k = 0
πk (0) =
6 0
0, k =

then solving for π0 (t) from Eq. (3.10) we have

π0 (t) = e−λt

Inserting this result into Eq. (3.9) for k = 1 results in

dπ1 (t)
= −λπ1 (t) + λe−λt
dt
with solution

π1 (t) = λte−λt

Solving recursively, we finally obtain as a solution to Eqs. (3.9)-(3.10):

(λt)k −λt
πk (t) = e ; k ≥ 0,t ≥ 0 (3.11)
k!
The latter is the well-known Poisson distribution which is central to our stud-
ies of queueing theory. Eq. (3.11) gives the probability of k arrivals (births) in
the interval (0,t) when the average arrival rate is λ. The Poisson process is im-
portant because it resembles or models many real-life processes very well. So,
for instance, it was observed that the number of telephone calls arriving from
subscribers at a central exchange is modelled very well by the Poisson process.
The same is true for packets arriving at a node in a packet switching computer
network.
There is another equally important reason, however, why the Poisson process is
so popular: Consider the random variable t̃ which represents the time between the
arrivals of a process which has a Poisson arrival distribution. As we did before,
we denote the probability distribution and probability density functions of these
times by A(t) and a(t) respectively. From the definition a(t)∆t + o(∆t) is the
probability that the next arrival occurs at least t and at most (t + ∆t) seconds
after the event of the last arrival. Furthermore,

A(t) = P [t̃ ≤ t]
= 1 − P [t̃ > t]

But P [t̃ > t] is the probability that there was no arrival in the interval (0,t), that
is π0 (t) for the Poisson distribution. Therefore, we have
3.4 M/M/1 Queue 65

A(t) = 1 − π0 (t)

and so from Eq. (3.11) we obtain, for the Poisson distribution, that the interarrival
time distribution is described by

A(t) = 1 − e−λt , t ≥ 0

which is of course the negative exponential distribution!

What we have just shown is that when the arrival process has a Poisson distri-
bution, the interarrival times are exponentially distributed, and we have proven
already that the exponential distribution has the remarkable memoryless property
which was necessary for our Markov condition.

Exercise 3.5 Let X be the time interval required to observe k Poisson arrivals.
Consider the two events: exactly k − 1 arrivals occur in the interval (0,t − ∆t)
and the event that exactly one arrival occurs in the interval (t − ∆t,t). By letting
∆t → 0 show that
λ(λx)k−1 −λx
fX (x) = e x≥0
(k − 1)!

3.4 M/M/1 Queue

The last special case of Eqs. (3.3) – (3.4) we will consider is a birth-death process
in which all the birth rates are the same and equal to λ for k ≥ 0 and all death
coefficients are equal to µ for k ≥ 1. This birth-death process with constant
coefficients is the simplest model of a queue, namely the well-known M/M/1
queue. The arrival process we know now from the previous section is Poisson
with exponential interarrival times. The service time distribution we define to be
exponential as well, or

B(x) = 1 − e−µx , x ≥ 0

Clearly, if we consider the number N (t) of customers in the queue at time t as

the stochastic variable in this case and assume furthermore that the interarrival
times and service times are mutually independent as well as independent from
the number of customers in the system, our process is Markovian. The steady
state distribution is given by Eq. (3.7) or, writing ρ for λ/µ, by

πk = (1 − ρ)ρk (3.12)

since
∞
!−1
X
k
π0 = 1+ ρ
k=1
= 1−ρ (3.13)
66 3 General Queueing Systems

where we obviously require 0 ≤ ρ < 1 for this solution to exist (cf. condition
of ergodicity on page 63). The ratio ρ is known as the utilisation of the system.
Note for instance that π0 or the probability that the system will be idle is given
by

π0 = 1 − ρ

or, conversely the probability that the system will be busy is given by ρ so that
“utilisation” seems an appropriate name for this quantity.
Another important measure of any queueing system is the average number N of
customers in the system. Since we know πk we can now easily compute that
quantity from
∞
X
N = kπk
k=0
∞
X
= (1 − ρ) kρk (3.14)
k=0

There is a very useful trick one uses to evaluate Eq. (3.14), namely the fact that
for an absolutely convergent series of functions, the derivative of the sum of
terms is the sum of the derivative of the individual terms. Applying this trick we
obtain from Eq. (3.14)
∞
X ∂
N = (1 − ρ)ρ ρk
k=0
∂ρ
∞
∂ X
= (1 − ρ)ρ ρk
∂ρ k=0
∂ 1
= (1 − ρ)ρ
∂ρ 1 − ρ
ρ
=
1−ρ

Applying Little’s law, Eq. (3.2), we can write for T the average time in the M/M/1
queueing system

N
T =
λ
1
=
µ−λ

This dependence of average time in the M/M/1 queue on the utilisation ρ is

illustrated in Fig. 3.6. Note that when ρ = 0 the average time in the system is
just 1/µ, the average service time expected by a customer as we would suspect.
3.4 M/M/1 Queue 67

15 .
...
...
..
..
...
14 ..
...
...
....
13 ..
....
..
.

...
...
12 ...
...
....
.
..
11 ..
...
....
..
10 ...
..
..
...
.
..
9 ..
..
...
.
...
8 ...
RESPONSE ...
....
...
TIME, T 7 ...
...
....
..
..
...
6 ..
...
.
..
..
5 ...
...
...
...
..
4 .
...
...
...
..
3 .....
....
......
.....
.....
......
2 ..
........
.........
.
...
..

..........
.............
..
..................
.
1 .
.........................
.........................................
.................................................................
..
...
...
...
...
.

0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
UTILISATION, ρ
Figure 3.6 Delay time as a function of utilisation in an M/M/1 queue.

The behaviour illustrated in Fig. 3.6 drawn for µ = 2.0, is rather dramatic. As
ρ → 1 the average delay T → ∞. This type of behaviour with respect to ρ as it
approaches 1 is characteristic of almost any queueing system one encounters. It
is also the explanation of why many a computer system manager has lost his job:
The system works well with increasing load in that response times increase only
slowly until, all of a sudden, the knee in Fig. 3.6 is reached and response times
become unacceptably long almost overnight!

Exercise 3.6 Consider an M/M/1 queue where there is no room for waiting cus-
tomers. We can define this as a birth-death process with coefficients
 
λ, k = 0 µ, k = 1
λk = µk =
6 0
0, k = 6 1
0, k =
1. Give the differential equations for πk (t) (k = 0,1).
2. Solve these equations in terms of π0 and π1 .

Exercise 3.7 At an automatic carwash there is room for only C waiting cars.
Dirty cars arrive at the rate of 6 per hour and are washed at an exponential rate
of 3 per hour. If there are more than two cars waiting the chances are 50 percent
that the arriving car will try to find another carwash.
68 3 General Queueing Systems

1. Set up the global balance equations for the system.

2. Calculate the average number of cars at the carwash at any moment.

3. If a car wash costs ECU10, what is the income of the carwash per hour?

3.5 M/M/m Queue

An interesting generalisation of the M/M/1 queue is the case of m servers rather

than a single one. This queue is illustrated in Fig. 3.7 and the CTMC with the
number of customers in the system as the state variable is illustrated in Fig. 3.8.

..
.

Figure 3.7 The M/M/m queueing system.

λ λ λ λ λ λ

0 1 2 ... m-1 m m+1 ...

µ 2µ 3µ mµ mµ mµ

Figure 3.8 State-transition diagramme of the M/M/m queueing process.

From Fig. 3.8 we can easily set up the global balance equations:

µπ1 = λπ0
(i + 1)µπ(i+1) + λπ(i−1) = (λ + iµ)πi , i < m
mµπ(i+1) + λπ(i−1) = (λ + mµ)πi , i ≥ m

So that
  k k
 π0 λ m
k! , k<m

mµ
πk =  k m
 π0 λ
 m
, k≥m
mµ m!
3.6 Queues with Processor Sharing Scheduling Strategy 69

with
m−1
!−1
X (mρ)k (mρ)m 1
π0 = +
k=0
k! m! 1 − ρ

λ
where ρ = mµ . Solving we obtain for the average number N of customers in the
system
∞
X (mρ)m π0
N= kπk = mρ + ρ
k=0
m! (1 − ρ)2

Exercise 3.8 Show that the average time in an M/M/m system is less than in m,
M/M/1 queues.

Exercise 3.9 Consider the freeway mentioned in Ex. 3.1. Assume that cars have
an interarrival rate of 3 per second and that the 4 lane freeway can accommodate
6 cars passing a particular point per second.
1. What is the utilisation of the freeway?
2. Assuming the same arrival rate, determine what happens if one of the lanes
is suddenly closed.

3.6 Queues with Processor Sharing Scheduling Strategy

In practice the processor scheduler of a computer system often uses a principle

known as time-slicing, whereby all processes requiring service are served for a
fixed length of time (the time-slice) in a round-robin fashion. A customer whose
service is completed leaves immediately. This scheduling strategy is approxi-
mated in theory by a Processor Sharing (or PS for short) server.
All customers have an identical mean service requirement µ1 and all n customers
in the queue are served simultaneously. Each customer is served at a rate n1 µ. It is
not difficult to see that, if we represent the state of the process by the number of
customers present, that the corresponding CTMC is similar to that for the M/M/1
queue illustrated in Fig. 3.5. The rate of a service completion is given by n n1 µ.
Note that the minimum of a set of exponentially distributed random variables
is again exponentially distributed (cf. Sec. 2.4). The steady state probability
distribution is thus given by the identical Eqs. (3.13) and (3.12).

3.7 Queues with Infinite Servers

A variation of the PS scheduling strategy is that called an Infinite Server (IS)

where the processor rate is independent of the number of customers present,
70 3 General Queueing Systems

yet all are served simultaneously with the same rate µ. Thus µk = kµ for an IS
queue. As for the PS scheduling strategy there is no waiting time in the queue and
each customer is merely delayed for a random length of time with distribution
that of the service time. An Infinite Server can also be viewed as a queue with
an infinite number of servers so that a free server can be dedicated to each newly
arriving customer. An example of such a queue is, e.g., a terminal pool.
Queues with the PS and IS service disciplines belong to a larger family of queues
know as being of the BCMP-type[15]. Queueing networks using these servers
can be analysed very efficiently, since the steady state probability distribution of
the entire network can be represented by a (possibly normalised) product of the
steady state probability distributions of the queues considered in isolation. Such
queueing networks are also known as having product-form.

3.8 Queues with Priority Service

Up to this point in our discussions the stochastic behaviour of all our customers
was identical and independent. There are practical situations however where we
may want to distinguish between different classes of customers[102]. This may
be the case because different classes of customers may have different service re-
quirements or because the server may wish to give different priorities to different
classes of customers or both. In this last section on queueing we will investigate
priority service.
When one has priority service it is obviously necessary to define exactly what the
priority rules for the different classes of customers are. One may decide for in-
stance that an arriving priority customer will pre-empt a lower priority customer
in service. In the latter case it is then necessary to make a distinction between
the case where the service of the lower priority customer is resumed where it had
left off on pre-emption, or whether the service will be repeated from scratch.

Example 3.1 Consider an M/M/1 queue with two classes of customers whose
service follow a negative exponential distribution with parameters µ1 and µ2 ,
and with arrival rates λ1 and λ2 respectively. We assume the n1 customers of
class 1 have priority and that the n2 customers of class 2, whose service is in-
terrupted, will resume their service. Resuming service or restarting the service
is equivalent in this case, however, since the service times are exponentially dis-
tributed (the memoryless property, cf., page 24).
Since it is clear that priority customers are not affected by customers of class 2,
we have
λ1
π1 (n1 ) = ρn1 1 (1 − ρ1 ) with ρ1 =
µ1

as was the case for the M/M/1 queue.

3.8 Queues with Priority Service 71

Write π(n1 ,n2 ) for the steady state probability distribution of the number of class
1 and class 2 customers in the system. In this case the global balance equations
are

(λ1 + λ2 + µ1 )π(n1 ,n2 ) = λ1 π(n1 − 1,n2 ) + µ1 π(n1 + 1,n2 )

+λ2 π(n1 ,n2 − 1), n1 > 0,n2 > 0,
(λ1 + λ2 + µ1 )π(n1 ,0) = λ1 π(n1 − 1,0)
+µ1 π(n1 + 1,n2 ), n1 > 0,n2 = 0,
(λ1 + λ2 + µ2 )π(0,n2 ) = µ1 π(1,n2 ) + λ2 π(0,n2 − 1)
+µ2 π(0,n2 + 1), n1 = 0,n2 > 0,
(λ1 + λ2 )π(0,0) = µ1 π(1,0) + µ2 π(0,1)

It is not easy to solve these balance equations and we will not describe how to
do so. Instead we simply quote the results (see, e.g., [77]). The mean number N1
and N2 of class 1 and 2 customers, respectively, are given by
ρ1
N1 =
1 − ρ1
ρ2 + N 1 ρ2
N2 =
1 − ρ 1 − ρ2

As one might expect π(0,0) = 1 − ρ1 − ρ2 as in the case of the M/M/1 queue.

Exercise 3.10 Explain how one can solve a queue using semi-Markov analysis?

Exercise 3.11

(a) Consider a queue with K different classes of customers arriving. The arrival
process for each customer class i,i = 1, . . . ,K is specified by a Poisson
process with parameter λi . The service distribution of customers of class i
is Coxian with mean value µ1i .
Calculate the steady state probability distribution of the number of cus-
tomers in the system. I.e., compute P [(n1 , . . . ,nK )], the steady state prob-
ability that ni customers of class i (i = 1, . . . ,K) are in the queue. Assume
a queue with a PS service scheduling strategy.

(b) Consider the same queue as in part (a), but now with an Infinite Server strat-
egy. Determine P [(n1 , . . . ,nK )] again. Also compute the mean number of
customers in the queue.
72

4 Further Reading

There is a huge selection of books on the subjects of with stochastic theory and
especially Markov processes and queueing theory. The following subjective se-
lection of books is thus not an exhaustive list. With the basics already covered in
this book, the reader can choose the books with the style and notation (s)he likes.
E.g., an introduction to stochastic processes is given in [53] and an old, but still
excellent introduction to the theory of single queues can be found in the book by
Leonard Kleinrock [108]. Results on priority queues is given in [92]. All these
references also discuss queues with general arrival and service time distributions.
In this book we only touched on the theory of queues and focused on models
where single customers arrive at and leave a queue. Queues with batch (or bulk)
arrival and service rates are found in [49].
The analysis of networks of queues is also a vast subject, in particular the so-
called product-form networks, which can be analysed very efficiently. The first
product form networks were analysed by Jackson [91] in 1957 and Gordon/Newell
[87] in 1967. Before that the output process of a single queue, for example the
M/M/m queue described above, was found to have the identical Poisson de-
parture process as the input process. In other words, the inter-departure times
are exponentially distributed with the same parameter λ as the input Poisson
process.[48, 102, 151].
Similar results hold for all M/M-queues with work-conserving1 scheduling dis-
ciplines which are independent of the service time requirements [102]. A feed-
forward network of such queues is easy to analyse since the output process is
Poisson with the same parameter as the input process. Each queue can be con-
sidered in isolation and the steady-state distribution of the overall network is
given by the product of the steady-state distributions of each individual queue.
Unfortunately the above is not true for a network of queues with feedback since
the output process is no longer Poisson, as successive inter-departure times are no
longer identically distributed because of the (partial) dependencies of the input
and output processes. In open queueing networks one can show that the output
from the network is still Poisson, but that in closed networks none of the flows
between queues is Poisson [102].
Surprisingly, one can nevertheless solve such networks of queues by analysing
each queue in isolation. In a seminal paper published in 1975 by F. Baskett, K.
M. Chandy, R. R. Muntz and F. G. Palacios, established a product form solution
for queueing networks, called BCMP-networks, which have a mixture of four
different type of stations [15]. Based on the product-form property very efficient
1
A scheduling discipline is work-conserving when no overhead or loss of work is incurred by
scheduling.
 73

algorithms can be designed to compute the steady state probability distribution

and performance metrics of these BCMP-networks. The best known of such al-
gorithms are the convolution algorithm and the mean value analysis for closed
nets [102].
Mean Value Analysis (MVA) is based on Little’s theorem and the arrival theorem
for closed product-form queueing networks. The arrival theorem states that a job
entering a queue, observes the steady-state distribution for the network with one
customer removed. In subsequent years significant research went into extending
the class of queueing networks for which a product-form solution of the steady-
state distribution can be found, e.g. [14]. For further reading we recommend [42].
Further information on queues and queueing networks can be found in the World
Wide Web at URL http://liinwww.ira.uka.de/bibliography/
Distributed/QLD/index.html. Here the reader finds a bibliography on
queueing systems and telecommunication, which is part of the collection of com-
puter science bibliographies.
Additional addresses offering search facilities (also for informa-
tion on other scientific computer relevant subjects) are http:
//liinwww.ira.uka.de/bibliography/waisbib.htmland
http://citeseer.ist.psu.edu/.
This concludes our introduction to stochastic theory.
74 4 Further Reading
 Part II

PETRI NETS
 77

5 Place-Transition Nets

Petri nets were invented in 1962 by Carl Adam Petri [133] and are a formalism
for the description of concurrency and synchronisation inherent in modern dis-
tributed systems. Since first described by Petri, many variations of his original
nets have been defined such as Coloured Petri nets, which we shall discuss in
Ch. 6. In this chapter we shall define Place-Transition nets, also referred to as
“ordinary” Petri nets. Throughout this book we shall use the term “Petri nets” to
refer to any type of net which is derived from that originally defined by Petri.
Apart from modelling and formally analysing modern distributed systems, Petri
nets also provide a convenient graphical representation of the system being mod-
elled. The graphical representation of a Place-Transition net comprises the fol-
lowing components:

places, drawn by circles. Places model conditions or objects, e.g., a program

variable. Places might contain

tokens, drawn by black dots. Tokens represent the specific value of the condition
or object, e.g., the value of a program variable.

transitions, drawn by rectangles. Transitions model activities which change the

values of conditions and objects.

arcs, specifying the interconnection of places and transitions thus indicating

which objects are changed by a certain activity.

Place-Transition nets are bipartite graphs, i.e. we may only connect a place to a
transition or vice versa, but we must not connect two places or transitions. This
also would not make sense, if we keep the interpretation of the components of a
Place-Transition net in mind.
Consider Fig. 5.1. p1 is a place marked with one token and is connected to tran-
sition t1 . Because of the direction of the arc connecting p1 and t1 we will call
p1 an input place of t1 and t1 accordingly an output transition of p1 . Places and
transitions might have several input/output elements. E.g., place p2 has three in-
put transitions: t1 , t3 , t5 , while t4 has two output places: p4 and p5 . The latter is
marked with two tokens.
p6 and t7 are not interconnected to any other element. Such elements will be
called isolated places or transitions respectively. As expected, isolated elements
of a Place-Transition net do not influence the rest of the net and therefore we can
neglect them. Until now we have only described the static structure of a Petri net.
Its dynamic behaviour or the modification of the condition/object values, can be
expressed via the following rules:
78 5 Place-Transition Nets

t1
p1
p4

t3

p2 t5
t4
p5
p3 t2

t6
p6 t7

Figure 5.1 A Place-Transition net

Enabling of a transition:
A transition is enabled if all its input places are marked at least with one to-
ken. For example transitions t1 and t4 in the Place-Transition net of Fig. 5.1
are enabled.

Firing of an enabled transition:

An enabled transition may fire. If a transition fires, it destroys one token on
each of its input places and creates one token on each of its output places.
E.g., if the enabled transition t1 fires, it destroys the token on p1 and creates
one token on p2 yielding the Place-Transition net given in Fig. 5.2. Note
that a transition need not fire.

Apart from modeling the behaviour of complex systems, Place-Transition nets

were originally designed to assist the analysis of such systems. In order to un-
derstand how this might be done, consider the following example.
Our example system consists of a sender and a receiver communicating with
each other via two unidirectional channels. Since we are interested in an imple-
mentation of this system, we do not wish to waste hardware resources. So we
want to implement channels which need to hold at most one message each. With
that restriction, sender and receiver have to obey a certain communication proto-
col. We first of all write down an implementation of the sender and receiver in a
Pascal-like notation.
 79

t1
p1
p4

t3

p2 t5
t4
p5
p3 t2

t6
p6 t7

Figure 5.2 Marking of the Place-Transition net after firing of t1

Sender:
while true do
begin
prepare message
send message to receiver via channel1
read acknowledgement of receiver on channel2
interpret acknowledgement
end
Receiver:
while true do
begin
read message from sender on channel1
interpret message
prepare acknowledgement
send acknowledgement to sender via channel2
end
Given this description of our system, we have to prove that it will operate as
intended. Particularly, this means that neither sender nor receiver send a mes-
sage/acknowledgement via the corresponding channel before a previous mes-
sage/acknowledgement has been read by the communication partner.
How can a Place-Transition net help us to prove that our system is correct? All
the above statements describe activities of the sender and receiver. Activities can
be modelled by transitions. So with each statement we associate a transition as
follows:
80 5 Place-Transition Nets

Sender:
while true do
begin
t1 prepare message
t2 send message to receiver via channel1
t3 read acknowledgement of receiver on channel2
t4 interpret acknowledgement
end
Receiver:
while true do
begin
t5 read message from sender on channel1
t6 interpret message
t7 prepare acknowledgement
t8 send acknowledgement to sender via channel2
end
Which objects do these transitions modify? There are several objects involved,
e.g., the message, the acknowledgement, the state of a channel, the states of
sender and receiver. Since we are not concerned with the content of the mes-
sages, we will neglect such information, thus, e.g., the internal state of a chan-
nel is given by the number of messages/acknowledgements currently present.
The Place-Transition net illustrated in Fig. 5.3 represents our system, where the
places p1 to p8 represent the internal state of sender and receiver respectively.
p9 represents the channel from sender to receiver and p10 the channel in reverse
direction. Using this Place-Transition net model, we can get a better insight into
the functionality of our system.
So far we presented an informal introduction of Petri nets. In the following sec-
tions we will introduce Place-Transition nets formally.

Exercise 5.1 Consider the Place-Transition net of Fig. 5.1.

1. Name the input and output places of t5 , t6 , t7 and the input and output
transitions of p3 , p4 , p5 , p6 .

2. How many tokens are on p4 ?

3. Transitions t3 and t5 are enabled. Is that correct?

4. What happens after firing t2 . Is t4 still enabled?

Such a situation is called a conflict, since firing one transition disables
another transition.

5. Count the number of all tokens in the net before and after firing of t4 . This
illustrates, that a transition does not merely move tokens around, but really
changes the value of objects by destroying and creating tokens.
5.1 Structure of Place-Transition Nets 81

p1
p5
t1 t5

p2 p6
p9
t6
t2

p3 p7

t3
t7

p4 p8
p10
t4
t8

Figure 5.3 Place-Transition net model of the sender/receiver system

Exercise 5.2 Play the token game for the Place-Transition net of Fig. 5.3 by
putting tokens onto the places and “move” them around according to the en-
abling and firing rules given before. Is our implementation correct?

5.1 Structure of Place-Transition Nets

A Place-Transition net is a bipartite directed graph composed of places, drawn

as circles, and transitions, drawn as rectangles. As usual, a directed graph is
formally given by the description of its elements (here separated into two sets,
since we have got two kinds of nodes) and functions or matrices specifying their
interconnection. Here we employ a functional notation.

Definition 5.1 (Place-Transition net) A Place-Transition net

(P-T net) is a 5-tuple P N = (P,T,I − ,I + ,M0 ) where

• P = {p1 , . . . ,pn } is a finite and non-empty set of places,

• T = {t1 , . . . ,tm } is a finite and non-empty set of transitions,

82 5 Place-Transition Nets

2 p2
3
p1
t1
t3
t5
t2
2
p4
2
t4 p3
t6

Figure 5.4 Place-Transition net

• P ∩ T = ∅,

• I − ,I + : P × T → N0 are the backward and forward incidence functions,

respectively

• M0 : P → N0 is the initial marking.

If we refer to the structure only, a Place-Transition net can also be viewed as a 4-

tuple P N = (P,T,I − ,I + ) omitting the initial marking M0 . Some authors refer
to the unmarked net P N = (P,T,I − ,I + ) as the Place-Transition net and call a
marked net P N = (P,T,I − ,I + ,M0 ) a system.
The functions I − and I + specify the connection between places and transitions.
If I − (p,t) > 0, an arc leads from place p to transition t; therefore I − is called the
backward incidence function of transition t. I − maps into the natural numbers
thus attaching a weight to the arc leading from p to t. In the graphical represen-
tation of a Place-Transition net this arc weight is written near the corresponding
arc. It is convention that an arc weight of 1 is not shown explicitly. Arc weights
specify that a transition is enabled only when at least as many tokens as given by
the arc weight are located on that place. Firing will destroy exactly this number
of tokens from p. Similarly I + (p,t) specifies the number of tokens created on
place p in case of firing t.

Example 5.1 The formal definition of the Place-Transition net depicted in Fig. 5.4
is given as follows:
P N = (P,T,I − ,I + ,M0 ) where

• P = {p1 ,p2 ,p3 ,p4 }

• T = {t1 ,t2 ,t3 ,t4 ,t5 ,t6 }

5.1 Structure of Place-Transition Nets 83

• I − (p1 ,t1 ) = 2, I − (p2 ,t2 ) = 1, I − (p2 ,t3 ) = 1, I − (p3 ,t4 ) = 2, I − (p4 ,t5 ) =
1, I − (p4 ,t6 ) = 1. All other values of function I − are zero.

• I + (p2 ,t1 ) = 3, I + (p4 ,t2 ) = 1, I + (p3 ,t3 ) = 2, I + (p4 ,t4 ) = 1, I + (p1 ,t5 ) =
1, I + (p4 ,t6 ) = 1. All other values of function I + are zero.

 1 if p = p1
• M0 (p) = 2 if p = p2 ∀p ∈ P.
0 otherwise


The input and output elements of a place and a transition can be defined formally
as follows.

Definition 5.2 Let P N = (P,T,I − ,I + ,M0 ) be a Place-Transition net.

• Input places of transition t: •t := {p ∈ P |I − (p,t) > 0},

• Output places of transition t: t• := {p ∈ P |I + (p,t) > 0},

• Input transitions of place p: •p := {t ∈ T |I + (p,t) > 0} and

• Output transitions of place p: p• := {t ∈ T |I − (p,t) > 0},

extension to sets X ⊆ P ∪ T is defined as •X = •x,

S
the usual
S x∈X
X• = x∈X x•.

The sets •p, • t and p • ,t• are also respectively referred to as the preset and
postset of the place p and transition t respectively.
For the Place-Transition net of Fig. 5.4 we get, e.g., •p1 = {t5 },
p2 • = {t2 ,t3 }, • {t2 ,t3 } = {p2 },{p2 ,t2 ,p4 }• = {t2 ,t3 ,t5 ,t6 ,p4 }.
A Place-Transition net can also be viewed as a usual directed graph and a flow
relation defined as follows:

Definition 5.3 Let P N = (P,T,I − ,I + ,M0 ) be a Place-Transition net.

1. F ⊆ (P × T ) ∪ (T × P ) given by F := {(x,y)|x,y ∈ P ∪ T : x ∈ •y} is

called the flow relation of P N .

Let F ∗ denote the reflexive and transitive closure of F , i.e.

∀x,y,z ∈ P ∪ T :

(a) (x,x) ∈ F ∗
(b) (x,y) ∈ F =⇒ (x,y) ∈ F ∗
(c) (x,y) ∈ F ∗ and (y,z) ∈ F ∗ =⇒ (x,z) ∈ F ∗

2. P N is weakly connected iff 1 ∀x,y ∈ P ∪ T : xF ∗ y or yF ∗ x

1
iff stands for “if and only if”.
84 5 Place-Transition Nets

3. P N is strongly connected iff ∀x,y ∈ P ∪ T : xF ∗ y and yF ∗ x

For the Place-Transition net of Fig. 5.4 the following, e.g., holds:

p1 F t1 ,p2 F ∗ p2 ,t1 F ∗ t3 ,t3 F ∗ p1 ,t1 F ∗ p1

Exercise 5.3 Give the formal definition of the Place-Transition net given in Fig. 5.1.

Exercise 5.4 Determine the following quantities for the Place-Transition net of
Fig. 5.1.

1. •t1 , •t5 , t2 •, •t6 , •p3 , p4 •, •p2 , p2 •.

2. •{t1 ,t2 ,t3 }, •{t1 ,t2 ,t6 }, •{t1 ,t2 ,t7 }, •{t1 ,t2 ,t4 }, •{p2 ,p3 },
{p2 ,p3 }•, {p2 ,p3 ,p5 }•, {p2 ,p6 }•, {t2 ,t3 }•, {t4 ,t3 }•, •{t1 ,p2 }.

3. •T , T •, •P , P •, •(T ∪ P ), (T ∪ P )•.

Exercise 5.5 Prove that the Place-Transition nets of Fig. 5.1 and 5.4 are strongly
or weakly connected.

5.2 Dynamic Behaviour of Place-Transition Nets

In the last section we defined the structure of a Place-Transition net. The dynamic
behaviour is determined by enabling and firing of transitions as given as follows.

Definition 5.4 Let P N = (P,T,I − ,I + ,M0 ) be a Place-Transition net.

1. A marking of a Place-Transition net is a function M : P 7→ N0 , where

M (p) denotes the number of tokens in p.

2. A set P̃ ⊆ P is marked at a marking M , iff ∃p ∈ P̃ : M (p) > 0; otherwise

P̃ is unmarked or empty at M .

3. A transition t ∈ T is enabled at M , denoted by M [t >, iff M (p) ≥

I − (p,t),∀p ∈ P .

4. A transition t ∈ T , enabled at marking M , may fire yielding a new marking

M 0 where

M 0 (p) = M (p) − I − (p,t) + I + (p,t),∀p ∈ P,

denoted by M [t > M 0 . We say M 0 is directly reachable from M and write

M → M 0 . Let →∗ be the reflexive and transitive closure of →. A marking
M 0 is reachable from M, iff M →∗ M 0 .
5.2 Dynamic Behaviour of Place-Transition Nets 85

5. A firing sequence (occurrence sequence) of P N is a finite sequence of

transitions σ = t1 . . . tn ,n ≥ 0 such that there are markings M1 , . . . ,Mn+1
satisfying Mi [ti > Mi+1 ,∀i = 1, . . . ,n. A shorthand notation for this case
is M1 [σ > and M1 [σ > Mn+1 respectively. The empty firing sequence is
denoted by ε and M [ε > M always holds.

The initial marking is the starting point of the dynamic behaviour of the Place-
Transition net and is therefore singled out and denoted M0 . Since transitions are
only enabled if their input places are marked with at least as many tokens as
specified by the backward incidence function (cf. Def. 5.4(3)), we are especially
interested in marked (sub)sets of places.
The expression M 0 (p) = M (p) − I − (p,t) + I + (p,t) in Def. 5.4(4) corresponds
to our intuitive understanding of the firing process. If transition t fires, it destroys
I − (p,t) tokens at place p and creates additional I + (p,t) tokens on p. Note that the
place p is the same in both expressions I − (p,t) and I + (p,t). In most cases one
of these expressions is 0, but there might exist some cases in which a transition
can destroy and create tokens at the same place. For instance, if transition t6 in
Fig. 5.4 fires at some marking in which it is enabled, the number of tokens on
place p4 remains unchanged. Such connection structures are called self-loops.

Example 5.2 Consider P N , the Place-Transition net illustrated in Fig 5.4.

A marking of PN is, e.g., M (p1 ) = 100,M (p2 ) = 1,M (p3 ) = 0,
M (p4 ) = 109 . Note that a marking is only a mapping and need not be reachable
from the initial marking. Of course, in analysis of Place-Transition nets we are
particularly interested in reachable markings.
P̃1 := {p2 } and P̃2 := {p1 ,p2 } are marked at M0 . t2 is enabled at M0 , so it is
correct to write M0 [t2 >. After firing t2 a new marking M 0 is reached, denoted
by M0 [t2 > M 0 where M 0 (p1 ) = M0 (p1 ) = 1, M 0 (p2 ) = M0 (p2 ) − 1 = 1,
M 0 (p3 ) = M0 (p3 ) = 0, M 0 (p4 ) = M0 (p4 ) + 1 = 1. Furthermore σ = t2 t5 t1 is
a firing sequence of PN, since M0 [σ >.

Exercise 5.6 Use the Place-Transition net in Fig 5.4 for the following exercises.
1. Determine all sets of places marked at M0 .

2. Prove or disprove:
M0 [t1 > ,M0 [t2 > ,M0 [t4 > ,M0 [t6 > ,M0 [t2 t5 t3 t4 > ,
M0 [t3 t4 t5 t1 > ,M0 [ε >.

3. Calculate the following markings M :

• M0 [t2 > M ,
• M0 [t3 t4 > M ,
• M0 [t2 t5 t1 > M ,
• M0 [t2 t6 t6 t6 t5 t1 > M ,
• M0 [ε > M .
86 5 Place-Transition Nets

4. Prove or disprove:

• M0 →∗ M , where M is given by M0 [t3 t2 > M .

• M0 →∗ M , where M is given by M0 [t3 t2 t4 t5 > M .
• M0 →∗ M , where M is given by
M (p1 ) = 0,M (p2 ) = 2,M (p3 ) = 2,M (p4 ) = 0.
• M0 →∗ M , where M is given by
M (p1 ) = 1,M (p2 ) = 3,M (p3 ) = 0,M (p4 ) = 0.
• M0 →∗ M , where M is given by M (p) = 0,∀p ∈ P .

5.3 Properties of Place-Transition Nets

Now that we have described the dynamic behaviour of Place-Transition nets we

turn our interest to the properties of such nets and how one might verify them.
Obviously we are only interested in properties concerning reachable markings.
One property to verify for our sender-receiver system in Chapter 5 was to en-
sure that the places representing the channels contain at most one token in every
reachable marking.
Another property concerns the firing of transitions. Since transitions model ac-
tivities, it is often desirable that no marking can be reached, such that a transition
is never enabled again. If such situations are excluded the Place-Transition net is
called “live”.
Last but not least, it is also desirable that when we view the set of reachable
markings as a graph, that it contains exactly one strongly connected subset of
markings. We know from our Markovian theory, that this is necessary for the
existence of a steady state distribution. Fig. 5.5 depicts two possible structures
of the reachability set drawn as a graph, where the subsets Ij are assumed to be
strongly connected. The left part of Fig. 5.5 contains two strongly connected sub-
sets. If we assume that this graph represents the state space of a Markov process,
we would have no steady state distribution (cf. Ex. 2.4 on page 33). Concern-
ing the right-hand part of Fig. 5.5, a steady state distribution exists assuming a
finite reachability set. Note that this situation is characterised by some marking
M which is reachable from all other markings.
All these remarks give rise to the following definition.

Definition 5.5 Let P N = (P,T,I − ,I + ,M0 ) be a Place-Transition net.

1. The reachability set of PN is defined by R(P N ) := {M |M0 →∗ M }.

If P N denotes an unmarked Place-Transition net or if we want to con-
sider parts of the reachability set, the set of reachable markings for a given
marking M̃ will be denoted by R(P N,M̃ ) := {M |M̃ →∗ M }. Thus for a
marked Place-Transition net we have R(P N ) = R(P N,M0 ).
5.3 Properties of Place-Transition Nets 87

M0 M0

M M
M0
M0

I1 I2

I0

Figure 5.5 Two possible structures of the reachability set

2. P N is a bounded Place-Transition net, iff ∀p ∈ P : ∃k ∈ N0 : ∀M ∈

R(P N ) : M (p) ≤ k.
P N is safe, iff ∀p ∈ P : ∀M ∈ R(P N ) : M (p) ≤ 1.
3. A transition t ∈ T is live, iff ∀M ∈ R(P N ) : ∃M 0 ∈ R(P N ) :
M →∗ M 0 and M 0 [t >.
P N is live, iff all transitions are live, i.e. ∀t ∈ T,M ∈ R(P N ) : ∃M 0 ∈
R(P N ) : M →∗ M 0 and M 0 [t >.
4. A marking M ∈ R(P N ) is a home state, iff ∀M 0 ∈ R(P N ) : M 0 →∗ M .
Safeness is of interest if the places of a Place-Transition net represent conditions.
In that case the presence or absence of a token models that the condition is or
is not satisfied. When modelling conditions, we are only interested in safe nets,
since two or more tokens on a place make no sense. Historically, one of the
first Petri nets, Condition-Event nets, were especially designed for this purpose.
Condition-Event nets do not allow multiple arc weights and transitions are only
enabled if their output places are empty, thus avoiding markings with more than
one token on a place. The situation that output places of an enabled transition
(with respect to Def. 5.4) are marked, is called contact.
We will use a vector notation to describe markings in order to simplify our no-
tation. If the set of places P is given by {p1 , . . . ,pn } the function M : P 7→ N0
can also be viewed as a vector M := (M (p1 ), . . . ,M (pn ))T . For notational con-
venience we will represent markings also by row vectors.
Example 5.3 Consider the net in Fig. 5.3 with the following reachability set
R(P N ) = {(1,0,0,0,1,0,0,0,0,0),(0,1,0,0,1,0,0,0,0,0),
(0,0,1,0,1,0,0,0,1,0),(0,0,1,0,0,1,0,0,0,0),
(0,0,1,0,0,0,1,0,0,0),(0,0,1,0,0,0,0,1,0,0),
(0,0,1,0,1,0,0,0,0,1),(0,0,0,1,0,1,0,0,0,0)}
88 5 Place-Transition Nets

left subnet right subnet

t3
p1

t2
t1 p3 p4

p2
t4

Figure 5.6 Place-Transition net not being live and bounded

left subnet right subnet

t1 p3

p2
p1 t3 t4

t2 p4

Figure 5.7 Further Place-Transition net not being live and bounded

From this reachability set it is easy to derive that the Place-Transition net is
bounded and in fact safe (M (p) ≤ 1,∀p ∈ P,M ∈ R(P N )) and that M is a
home state ∀M ∈ R(P N ). Furthermore the Place-Transition net is live.
Liveness and boundedness impose a restriction on the structure of a Place-Transition
net. Suppose transition t1 of the Place-Transition net in Fig. 5.1 fires emptying
place p1 . Since place p1 has no input transitions it will remain empty in all fur-
ther reachable markings, implying that t1 is not live. In the same Place-Transition
net, transition t6 has no input places. Thus it is always enabled, so that no upper
bound for the number of tokens can be found for place p5 . This example shows
that if the Petri net has source (or sink) elements, it can not be both live and
bounded. The reader should convince himself that also sink elements prevent
live- and boundedness.
5.3 Properties of Place-Transition Nets 89

Figures 5.6 and 5.7 illustrate that liveness and boundedness can not hold both
even in weakly, but not strongly, connected Place-Transition nets. Considering
Fig. 5.6 we can find two strongly connected subnets which are connected via the
arc from t1 to p3 . If we assume that the left subnet is live, then the right subnet is
obviously not bounded. On the other hand, if we assume that the right subnet is
bounded, the left subnet can not be live. So in summary the net of Fig. 5.6 can not
be both live and bounded for any initial marking. The same can be concluded for
the Place-Transition net of Fig. 5.7 following similar arguments. As the reader
might expect the following holds.

Theorem 5.1 ([34]) Let PN be weakly connected.

PN is live and bounded =⇒ PN is strongly connected.

In other words strong connectedness is a necessary condition for a PN to be live

and bounded. Since we are only interested in Place-Transition nets for which
these properties hold, we will only consider strongly connected Place-Transition
nets in the following.
Note that if a Place-Transition net is not weakly connected it may still be live and
bounded. In that case the Place-Transition net consists of more than one isolated
strongly connected Place-Transition nets which can be analysed separately.

Exercise 5.7 1. Determine R(P N ) for the Place-Transition net displayed in

Figures 5.1 and 5.4. Prove or disprove the following conjectures for both
Place-Transition nets:

• PN is bounded.
• PN is live.

2. Consider the Place-Transition net in Fig. 5.4 and change the arc weight of
the arc leading from t1 to p2 to 2, i.e. define I + (p2 ,t1 ) = 2.
Prove or disprove the following conjectures:

• PN is bounded.
• PN is live.
• M0 is a home state of PN.

Exercise 5.8

• Find further examples of weakly connected Place-Transition nets which are

live, but not bounded.

• Find further examples of weakly connected Place-Transition nets which are

not live, but bounded.

• Disprove:
PN strongly connected =⇒ PN is live and bounded.
90 5 Place-Transition Nets

5.4 Analysis of Place-Transition Nets

Verifying that a given Place-Transition net satisfies certain properties is typical

of the type of analysis we may wish to do. E.g., we want to show that the Place-
Transition net is live.
The most common way to analyse a Place-Transition net is to analyse its reach-
ability set, since all properties are defined from this reachability set.

5.4.1 Analysis of the Reachability Set

The reachability set of a Petri net, cf. Ex. 5.3 on page 87, is often drawn as a tree,
where the nodes of the tree are markings of the Place-Transition net. Two nodes
M and M 0 are connected with an directed arc iff M [t > M 0 for some t ∈ T .
This arc is also labelled with t ∈ T .
The reachability tree can be generated starting with the initial marking of the
Place-Transition net and adding directly reachable markings as leaves. Next we
proceed with these new markings and determine their directly reachable mark-
ings. These markings now become the new leaves of the already generated part
of the reachability tree etc. If we reach a marking previously explored we need
not continue building the tree any further at that node.

Example 5.4 The reachability tree of the net in Fig. 5.8 is shown in Fig. 5.9.

Reachability trees can be transformed directly into graphs by removing multiple

nodes and connecting the nodes appropriately. Such a graph is called a reacha-
bility graph. The corresponding reachability graph of the reachability tree dis-
played in Fig 5.9 is shown in Fig. 5.10.
Unfortunately the method of creating the reachability tree by simulating the to-
ken game fails in the case of unbounded nets like the one depicted in Fig. 5.11
(see Exercise 5.9). In order to cope with infinite reachability sets a special symbol

t1 p2 t3

p1

t2 p3 t4

Figure 5.8 A Place-Transition net

5.4 Analysis of Place-Transition Nets 91

(1,0,0)
t1 t2

(0,1,0) (0,0,1)
t3 t4
(1,0,0) (1,0,0)

Figure 5.9 Its reachability tree

t3 t4
(1,0,0)
t1 t2

(0,1,0) (0,0,1)

Figure 5.10 The corresponding reachability graph

t1

p1 2
p2

t2

Figure 5.11 Place-Transition net with an infinite reachability set

ω is introduced for unbounded nets to represent the marking of an unbounded

place of the Place-Transition net. ω can be referred to as infinity.
The arithmetic rules for ω are: ∀a ∈ N0 : ω + a = ω,ω − a = ω,
a < ω,ω = ω.
Definition 5.6 A marking M covers a marking M 0 , denoted by M ≥ M 0 , iff
M (p) ≥ M 0 (p),∀p ∈ P.
The following algorithm leads to a finite representation of the reachability tree
even for unbounded nets. Since the marking of unbounded places is represented
by ω, which is larger than every natural number, this representation of the reach-
ability tree is also called a coverability tree. The terminal nodes of the tree are
called leaves. A node is called an inner node, if it is not a leaf of the tree.
Algorithm 5.1 (cf. [132])
92 5 Place-Transition Nets

t4

p2

t2
t3
p1

t5
p3

t1

p4

Figure 5.12 Unbounded Place-Transition net

X := {M0 }; /* M0 is the root of the coverability tree */

while X 6= ∅ do
begin
Choose x ∈ X.
∀t ∈ T : x[t > do
create a new node x0 given by x[t > x0 and connect x and x0
by a directed arc labelled with t.
Check ∀p ∈ P :
If there exists a node y on the path from M0 to x0 with
y ≤ x0 and y(p) < x0 (p) then set x0 (p) := ω.
X := {x|x is a leaf of the coverability tree generated so far,
in x at least one transition is enabled
and there is no inner node y with y = x }
end

The reason for setting x0 (p) = ω is that x0 can be reached from y by a firing
sequence σ ∈ T ∗ , i.e. y[σ > x0 . Because x0 ≥ y x0 [σ > also holds, leading to a
marking x00 with x0 [σ > x00 where x00 (p) > x0 (p) > y(p). This shows that place
p is unbounded, which justifies the use of ω. The coverability tree coincides with
the “normal” reachability tree if the net is bounded.

Example 5.5 The coverability tree of the Place-Transition net of Fig. 5.12 is
given in Fig. 5.13. Note that the algorithm does not necessarily lead to a unique
coverability tree for a given Place-Transition net, because of the arbitrary choice
of x ∈ X.
5.4 Analysis of Place-Transition Nets 93

(1,0,0,0)
t1 t2

(0,0,1,1) (0,1,0,0)

t3 t5 t4

(0,1,0,1) (1,0,0,0) (1,0,0,0)

t4

(1,0,0,ω )
t1 t2

(0,0,1,ω ) (0,1,0,ω )

t5 t3

(1,0,0,ω ) (0,1,0,ω )

t4

(1,0,0,ω )

Figure 5.13 Coverability tree of the unbounded Place-Transition net

The coverability tree can be used for the analysis of Place-Transition nets. The
following theorem is obviously true.
Theorem 5.2 PN is bounded iff no node in its coverability tree contains the sym-
bol ω.
In the case of bounded Place-Transition nets, liveness can also be checked by
investigating the coverability tree. The idea is to transform the coverability tree
into the corresponding coverability graph, similar to the transformation of the
reachability tree into a reachability graph.
Theorem 5.3 Let PN be bounded. PN is live iff all transitions appear as a label
in all terminal strongly connected components of the coverability graph.
A strongly connected component is terminal if no arcs leave that component. In
terms of Markov chains such a subset of markings (states) is called closed (cf.
Def. 2.4).
Definition 5.7 R̃ ⊆ R(P N ) is a terminal strongly connected component of
R(P N,M0 ) iff ∀M ∈ R̃,M 0 ∈ R(P N ) :
(M →∗ M 0 =⇒ M 0 →∗ M and M 0 ∈ R̃).
94 5 Place-Transition Nets

p1 t1 p2 t2
3 2
(1,0)

t1

p1 t1 p2 t2 (1,ω)
3
t1 t2

(1,ω) (1,ω)

Figure 5.14 Two Place-Transition nets and their coverability trees

Note that, checking the existence of home states is also straightforward for bounded
nets.

Theorem 5.4 Let PN be bounded. A home state exists in a Place-Transition net

iff its coverability graph contains exactly one terminal strongly connected com-
ponent.

For unbounded Place-Transition nets the information given by their coverability

trees is not always sufficient to decide whether home states do exist. Consider the
Place-Transition nets of Fig. 5.14, which both have the same coverability tree. In
the first net the marking (1,1) is a home state, since by firing t2 the token count
on place p2 can always be reduced to one. This is not possible in the second
Place-Transition net, which therefore has no home states.
As you may have noticed this kind of Place-Transition net analysis will lead to
large coverability trees/graphs even for fairly simple Place-Transition nets. Even
modern computers can not cope with the problem of generating the coverability
tree of the Place-Transition net depicted in Fig. 5.15. This problem is also known
as the “state space explosion problem”, since the state space or the coverability
tree becomes intractably large.
In the following subsections we will learn about analysis techniques which ad-
dress this problem.

Exercise 5.9 Determine the reachability set of the unbounded Place-Transition

net given in Fig. 5.11.

Exercise 5.10
1. Use algorithm 5.1 to determine the coverability tree of the Place-Transition
nets of Fig. 5.1 and 5.4.
2. Determine the coverability trees of the Place-Transition nets given in Fig. 5.16
([16]).
5.4 Analysis of Place-Transition Nets 95

t1

p1 p2

10
1010 tokens t2

Figure 5.15 Place-Transition net with a huge reachability set

2 2 2 2
t1 p1 t2 t1 p1 t2

Figure 5.16 Two Place-Transition nets with the same coverability tree

Exercise 5.11 Show that theorem 5.3 does not hold for unbounded nets?

5.4.2 Invariant Analysis

We have seen before that a marking can be represented in either functional or in

vector notation. Analogously the incidence functions I − and I + can be notated
as matrices. These matrices (C − and C + ) are called the incidence matrices and
are defined as follows.

Definition 5.8 P N = (P,T,I − ,I + ,M0 ).

The backward incidence matrix C − = (c− n×m
ij ) ∈ N0 is defined by
− −
cij := I (pi ,tj ),∀pi ∈ P,tj ∈ T ,
the forward incidence matrix C + = (c+ ij ) ∈ N0
n×m
is defined by
+
cij := I + (pi ,tj ),∀pi ∈ P,tj ∈ T ,
and the incidence matrix of P N is defined as C := C + − C − .

Enabling and firing of a transition can now be expressed in terms of these inci-
dence matrices. A transition ti ∈ T is enabled in a marking M , iff M ≥ C − ei ,
1 ,0, . . . ,0)T .
where ei is the i-th unit vector (0, . . . ,0, |{z}
i
If an enabled transition ti ∈ T fires in marking M the successor marking M 0 is
given by
96 5 Place-Transition Nets

s t
p1 p4

t1 t3
n

p2 n p5
p3 n

t2 t4

Figure 5.17 M0 = (s,0,n,t,0)T and I − (p3 ,t3 ) = I + (p3 ,t4 ) = n

M 0 = M + Cei (5.1)
Eq. (5.1) states that the firing process can be described by adding the vector
Cei to the given marking M yielding the successor marking M 0 . Cei is the i-th
column of the incidence matrix C specifying the influence of transition ti ∈ T
on its input and output places. Since markings can be calculated in that way by
the addition of vectors, Place-Transition nets can be viewed as vector addition
systems with the restriction that only enabled transitions may fire.
Consider σ = tk1 . . . tkj ,j ∈ N0 a firing sequence with
M0 [tk1 > . . . Mj−1 [tkj > Mj . Marking Mj can be calculated as follows. For
each reachable marking the following equation holds:
Mi = Mi−1 + Ceki , i = 1, . . . ,j
Substituting the right hand side of the equation for i = r into the equation for
i = r + 1 yields
j
X
Mj = M0 + C eki
i=1

The vector f := ji=1 eki is called a firing vector and is of dimension |T |. A

P

component fi of f characterises the number of occurrences of transition ti in σ.

Example 5.6 The Petri net of Fig. 5.17 has the following incidence matrices:
   
1 0 0 0 0 1 0 0

 0 1 0 0 


 1 0 0 0 

− +
C = 1 0 n 0 C = 0 1 0 n 
   

0 0 1 0 0 0 0 1 
   
  
0 0 0 1 0 0 1 0
5.4 Analysis of Place-Transition Nets 97

 
−1 +1 0 0

 +1 −1 0 0 

C= −1 +1 −n +n 
 
0 0 −1 +1 
 

0 0 +1 −1

At the initial marking t1 is enabled and the resultant marking after firing t1 is
given by
     
s −1 s−1
1
 
 0   1   1 
0      
M 0 = M0 + C 
 
= n + −1 = n−1
    
 0   
t 0 t
     
0
     
0 0 0

Assume s ≥ 3 and n ≥ 3, then σ = t1 t1 t1 t2 is a firing sequence with f =

(3,1,0,0)T its corresponding firing vector. The marking M 00 reached after the
occurrence of σ is given by
       
s −1 1 s−2
3
 
 0   1   −1   2 
1        
M 00 = M0 + C 
 
= n  + 3 −1 + 1 = n−2
      
 0   
t 0 0 t
       
0
       
0 0 0 0

Since reaching a marking implies the existence of an appropriate firing sequence

the following theorem obviously holds.

Theorem 5.5

∀M ∈ R(P N,M0 ) : ∃f ∈ Nm
0 : M = M0 + Cf (5.2)

For verifying invariant properties of all reachable markings we can exploit Th. 5.5
as follows. If v ∈ Zn is2 multiplied with equation 5.2 we get

v T M = v T M0 + v T Cf, ∀M ∈ R(P N,M0 ). (5.3)

Of special interest are those vectors v ∈ Zn satisfying v T C = 0, since Eq. (5.3)
becomes v T M = v T M0 in those cases. Since v T and M0 are known, (5.3) then
establishes a condition on all reachable markings M .

Definition 5.9 v ∈ Zn ,v 6= 0, is a P-invariant iff v T C = 0.

Theorem 5.6 If v ∈ Zn is a P-invariant then

∀M ∈ R(P N,M0 ) : v T M = v T M0 .
2
Z denotes the set of integers and N the set of positive integers.
98 5 Place-Transition Nets

Example 5.7 (Readers/Writers-Problem) Consider a system

with several processes reading or writing a common file. Readers never modify
the file, while writers do modify it. So it is obvious that several readers might
simultaneously access the file, but for writing operations all other writers and
readers must be excluded from accessing the file.
A possible solution might be to introduce an integer semaphore S, initialised with
n, with parameterised P- and V-operations, specifying the value to be subtracted.
So our solution might look in pseudo-code:
Readers: Writers:
. . . (before reading) . . . (before writing)
P(S,1); P(S,n);
read file; write file;
V(S,1); V(S,n);
. . . (after reading) . . . (after writing)
The problem is now to verify that this solution is correct, e.g., respects the mutual
exclusion condition. Let s denote the number of readers and t the number of
writers.
A possible Petri net model of this system is the one of Fig. 5.17, where the places
p1 and p2 specify the number of readers currently not reading and reading re-
spectively. The places p4 and p5 have a similar function for modeling the state
of the writers, and place p3 represents the semaphore.
Solving the system of linear equations v T C = 0 (cf. Ex. 5.6) yields, e.g., the
following P-invariants:

v1 = (1,1,0,0,0)T ,v2 = (0,0,0,1,1)T ,v3 = (0,1,1,0,n)T

Substituting these solutions for v T M = v T M0 we get

M (p1 ) + M (p2 ) = s
M (p4 ) + M (p5 ) = t
M (p2 ) + M (p3 ) + nM (p5 ) = n,∀M ∈ R(P N ).

The first two equations simply express that the number of readers and writers is
constant. The last equation allows the following conclusions:
∀M ∈ R(P N ) we have

a) M (p2 ) ≥ 1 =⇒ M (p5 ) = 0

b) M (p5 ) ≥ 1 =⇒ M (p2 ) = 0

c) M (p5 ) ≤ 1

In terms of the readers/writers system this means:

• if a reader is reading, no writer is writing (a);

• if a writer is writing, no reader is reading (b);

5.4 Analysis of Place-Transition Nets 99

• and finally that at most one writer is writing (c).

Furthermore we can conclude that M (p2 ) ≤ n showing that we should choose
n ≥ s to allow all readers to operate in parallel. Since we have proved that
our Petri net model has the desired properties, we know the same for our system
provided the Petri net reflects the modelled system behaviour correctly.
We have seen that Th. 5.6 assists us in the analysis of Petri nets. For live Place-
Transition nets the converse of Th. 5.6 also holds.
Theorem 5.7 Let PN be a live Place-Transition net and v ∈ Zn ,v 6= 0. If
v T M = v T M0 ,∀M ∈ R(P N,M0 ) then v is a P-invariant.
Proof. Since PN is live we have ∀ti ∈ T : ∃M,M 0 ∈ R(P N,M0 ) : M [ti > M 0
and M 0 = M + Cei giving v T M 0 = v T M + v T Cei . Since v T M = v T M0 this
yields v T Cei = 0. Since this holds for all transitions we get ∀ti ∈ T : v T Cei =
0 and thus v T C = 0, showing that v is a P-invariant. t
u

P-invariants are sometimes called S-invariants, because some authors denote the
set of places by S after the German word “Stellen” (≡ “Places”).
P-invariants can be employed for checking the boundedness property.
Definition 5.10 PN is covered by positive P-invariants iff
∀pi ∈ P : ∃ P-invariant v ∈ Zn with v ≥ 0 and vi > 0.
Theorem 5.8 PN is covered by positive P-invariants =⇒ PN is bounded.
Proof. PN is covered by positive P-invariants =⇒ ∃ P-invariant v ∈ Nn0 : vi >
0,∀i ∈ {1, . . . ,n}. With v T M = v T M0 = const.,∀M ∈ R(P N,M0 ), we get
vi M (pi ) ≤ v T M0 ,∀pi ∈ P , since M (p) ≥ 0,∀M ∈ R(P N,M0 ).
T
vi > 0 yields M (pi ) ≤ v vM i
0
,∀pi ∈ P proving that the net is bounded. t
u

If the net is not covered by positive P-invariants, but there are some positive
P-invariants, then at least these covered places are bounded.
Corollary 5.1 If there exists a P-invariant v ∈ Zn : v ≥ 0,vi > 0 then pi is
bounded, i.e. ∃k ∈ N : ∀M ∈ R(P N,M0 ) : M (pi ) ≤ k.
Proof. The proof follows that of Th. 5.8. t
u

Another invariant which is useful in Place-Transition net analysis is called the

T-invariant. Suppose we have a marking M and after firing of some transitions
we want to reach M again. So let f ∈ Nm 0 be the corresponding firing vector.
Since we started in M and reached M again the following equation holds (cf.
Eq. (5.2))
M = M + Cf
which implies Cf = 0. In order to return to a marking of the Place-Transition net
it is necessary that the firing vector satisfies Cf = 0. This leads to the following
definition.
100 5 Place-Transition Nets

Definition 5.11 w ∈ Zm ,w 6= 0, is a T-invariant iff Cw = 0.

Definition 5.12 PN is covered by positive T-invariants iff

∀ti ∈ T : ∃ T-invariant w ∈ Zm with w ≥ 0 and wi > 0.

Covering by T-invariants is a necessary condition for a Place-Transition net to be

bounded and live.

Theorem 5.9 PN is bounded and live =⇒ PN is covered by positive T-invariants.

Proof. Since PN is bounded its reachability set is finite.

Let R̃ ⊆ R(P N,M0 ) be a terminal strongly connected subset of R(P N,M0 ).
Choose an arbitrary marking M ∈ R̃. M can always be reached after some
transition firings, since R̃ is strongly connected.
Since PN is live there ∃f ∈ Nm 0 : fi > 0,∀i ∈ {1, . . . ,m} : M = M + Cf ,
which implies Cf = 0.
Thus f is a positive T-invariant covering all transitions in the Place-Transition
net. t
u

Corollary 5.2 Let PN be bounded.

If ti ∈ T is live then ∃ T-invariant w ∈ Zm : w ≥ 0,wi > 0.

Proof. The proof is similar to that of Th. 5.9. t

u

Analysing a Place-Transition net by calculating its invariants is in most cases

more time and space efficient than inspecting the reachability set, since the com-
plexity of this kind of analysis depends only on the number of places and tran-
sitions and not on the size of the reachability set. In [117] an algorithm is de-
scribed for determining the invariants of a Place-Transition net. This algorithm
is e.g. part of the GreatSPN-tool [50] and the QPN-tool (cf. [29]) as well.
A Place-Transition net can also be analysed very efficiently if its structure is
restricted. Imposing restrictions on the structure of Place-Transition nets leads to
several classes of Place-Transition nets, which will be our next topic.

Exercise 5.12 Determine the incidence matrices of the Place-Transition net given
in Fig. 5.4.
From these prove if M0 ≥ C − e1 and M0 ≥ C − e2 holds. What exactly does that
mean?
Calculate: M1 = M0 + Ce2 and M2 = M1 + Ce5

Exercise 5.13 Prove, that if v1 and v2 are P-invariants then

• v1 + v2
• rv1 ,r ∈ Z
are also P-invariants.

Exercise 5.14 Disprove the converse of Theorem 5.8.

5.4 Analysis of Place-Transition Nets 101

5.4.3 Analysis of Net Classes

The analysis of net classes has been a topic for Petri net researchers since the
0 70s. In this section we give an overview on the most important results, paying

particular attention to Extended Free Choice (EFC)-nets.

Definition 5.13 Let P N = (P,T,I − ,I + ,M0 ) be a Place-Transition net with
I − (p,t),I + (p,t) ∈ {0,1},∀p ∈ P,t ∈ T . P N is called a
1. State Machine (SM) iff ∀t ∈ T : | • t| = |t • | ≤ 1.3
2. Marked Graph (MG) iff ∀p ∈ P : | • p| = |p • | ≤ 1
3. Free Choice net (FC-net), iff
∀p,p0 ∈ P : p 6= p0 =⇒ (p • ∩p0 • = ∅ or |p • | = |p0 • | ≤ 1).
4. Extended Free Choice net (EFC-net), iff
∀p,p0 ∈ P : p • ∩p0 • = ∅ or p• = p0 •.
5. Simple net (SPL-net), iff
∀p,p0 ∈ P : p 6= p0 =⇒ (p • ∩p0 • = ∅ or |p • | ≤ 1 or |p0 • | ≤ 1).
6. Extended Simple net (ESPL-net), iff
∀p,p0 ∈ P : p • ∩p0 • = ∅ or p• ⊆ p0 • or p0 • ⊆ p•.

Fig. 5.19 provides a description of these different classes of nets. For these net
classes the following relations hold (cf. Fig. 5.18):
Marked Graphs ⊂ FC-nets, State Machines ⊂ FC-nets ⊂ EFC-nets ⊂ ESPL-nets
and FC-nets ⊂ SPL-nets ⊂ ESPL-nets.
We first of all investigate State Machines, which can be analysed very easily.

Analysis of State Machines

All transitions of a State Machine have at most one input and one output place.
Since we are interested in live and bounded nets, Th. 5.1 tells us that we only
have to consider strongly connected State Machines. Fig. 5.20 depicts such a
State Machine.
If a State Machine is strongly connected every transition has exactly one input
and one output place. A token in a strongly connected State Machine can be
“moved” from one place to every other place of the net, since this token directly
“enables” all output transitions of the place it is currently located. This observa-
tion yields the following characterisation of liveness.
Theorem 5.10 (Liveness in State Machines [38]) Let PN be a State Machine
with | • t| = |t • | = 1,∀t ∈ T .
PN is live iff PN is strongly connected and M0 6= 0.
3
|x| denotes the number of elements in the set x.
102 5 Place-Transition Nets

ESPL

EFC

FC

SPL
MG SM

Figure 5.18 Relationship of net classes

Firing of transitions in a strongly connected State Machine obviously does not

change the number of tokens in the Place-Transition net and thus the net is
bounded.

Theorem 5.11 (Boundedness in State Machines [38]) Let PN be a live State

Machine with | • t| = |t • | = 1,∀t ∈ T .
Then PN is bounded.

The former argument for liveness also implies, that a token can always return to
a place it started from. So we can always reach the initial marking.

Theorem 5.12 (Home states in State Machines) If PN is a live and bounded

State Machine then M0 is a home state.

Th. 5.12 directly implies that all reachable markings of a live and bounded State
Machine are home states.
We have seen that State Machines are very easy to analyse. Let us investigate the
next net class.

Analysis of Marked Graphs

Similar to State Machines we only consider strongly connected Marked Graphs.

This implies that each place has exactly one input and one output transition. A
typical example of a Marked Graph is shown in Fig. 5.21.
Playing the token game for this Place-Transition net, one realises that, e.g., the
token on p5 seems to cycle around. Note that we have avoided speaking so far
5.4 Analysis of Place-Transition Nets 103

allowed not allowed

State
Machines

Marked

Graphs

FC−nets

EFC−nets

SPL−nets

ESPL−nets

Figure 5.19 Illustrating different net classes

104 5 Place-Transition Nets

t1 p2 t3

p1

t2 p3 t4

Figure 5.20 Example of a State Machine

p1 p3 p5 p7

t5

t1 t2 t3

t4

p2 p4 p6

Figure 5.21 An example of a Marked Graph

about tokens moving in a Place-Transition net, since a transition might destroy

more or less tokens than it creates. But in this context it makes sense. So the
token on p5 is first fired on p6 than on p7 and afterwards returns to p5 . If we
modify the initial marking just by removing the token on p5 , we also see that the
net is not live. It seems that these cycles play an important role in the context of
Marked Graphs. So let us define them:

Definition 5.14 A cycle is a sequence x0 , . . . ,xk where xi ∈ P ∪ T,0 ≤ i ≤ k

and xi+1 ∈ xi • ,∀i ∈ {0, . . . ,k − 1} and x0 = xk .
A cycle is called simple iff ∀xi ,xj ∈ x0 , . . . ,xk : i 6= j =⇒
xi 6= xj or (i = 0 and j = k).

A simple cycle is a cycle where each element appears once, except for the first
and last element. The Marked Graph of Fig. 5.21, e.g., contains amongst others
the following simple cycles:
p1 ,t1 ,p2 ,t2 ,p1 and p5 ,t3 ,p6 ,t4 ,p7 ,t5 ,p5 .
The following result should now be no surprise for us.
5.4 Analysis of Place-Transition Nets 105

Theorem 5.13 (Liveness in Marked Graphs [150]) Let PN be a Marked Graph

with | • p| = |p • | = 1.
PN is live iff every simple cycle contains a place with at least one token at M0 .

Determining boundedness of a Marked Graph can also be based on the concept

of simple cycles.

Definition 5.15 PN is covered by simple cycles iff ∀x ∈ P ∪ T : ∃ a simple

cycle x0 , . . . ,xk with x = xi for some i ∈ {0, . . . ,k}.

Theorem 5.14 (Boundedness in Marked Graphs [150]) A live Marked Graph

PN is bounded iff PN is covered by simple cycles.

Similar to our discussion concerning State Machines Fig. 5.21 shows that M0 is
a home state. After firing of every transition M0 is reached again. In fact Marked
Graphs unveil the same property as State Machines, i.e.

Theorem 5.15 (Home states in Marked Graphs [125]) If PN is a live and bounded
Marked Graph then M0 is a home state.

So far we have seen that the two net classes, State Machines and Marked Graphs,
are very simple to analyse. In the next subsection we draw attention to EFC-nets,
since they are a superset of FC-nets and we can establish some useful analysis
techniques for this “extended” class directly.

Analysis of EFC-nets

According to the definition, EFC-nets unveil the possibilities of connecting tran-

sitions to their input places as shown in Fig. 5.22.

t0

Figure 5.22 Possible connections between transitions and their input places in EFC-
nets

With that the following is obvious: If a transition t of an EFC-net is enabled

then all transitions which are in structural conflict, i.e. all t0 ∈ (•t)•, are enabled
as well. So amongst conflicting transitions, we can choose the transition to fire
freely.
For the analysis of FC- and EFC-nets we need to understand the terms “dead-
lock” and “trap”.
106 5 Place-Transition Nets

t1 p2
t3

p1

t2 p3

Figure 5.23 non-live FC-net

Definition 5.16
1. P 0 ⊆ P,P 0 6= ∅ is a deadlock4 iff •P 0 ⊆ P 0 •.
2. P 0 ⊆ P,P 0 6= ∅ is a trap iff P 0 • ⊆ •P 0 .
If a deadlock empties, this subset will remain empty. On the other hand once a
trap is marked, it will remain marked.
Consider the FC-net in Fig. 5.23, which has the deadlocks {p1 ,p2 },
{p1 ,p3 } and {p1 ,p2 ,p3 }, which is also the only trap of the net. At the initial
marking all deadlocks are marked. Firing t1 now empties the deadlock {p1 ,p3 }
and thus this deadlock will remain empty. Since {p1 ,p3 } is empty all transitions
in {p1 ,p3 }• are not enabled and are therefore not live. An analogous argument
holds for the deadlock {p1 ,p2 } if t2 fires first at the initial marking.
A natural idea is to avoid such situations by ensuring that a deadlock can never
be emptied. This is, e.g., the case if the deadlock contains a marked trap, because
a trap remains marked in all further reachable markings. For EFC-nets this idea
yields a characterising condition for liveness:

Theorem 5.16 (Liveness in EFC-nets [34, 38]) An EFC-net is live iff every dead-
lock contains a marked trap at M0 .

This condition is often called the deadlock/trap-property (dt-property).

Proof.
We will only prove sufficiency.
Assume the EFC-net is not live.
Then ∃t ∈ T : M ∈ R(P N,M0 ) : ∀M 0 ∈ R(P N,M ) : ¬M 0 [t >, i.e. we
can find a transition t and a marking M , such that t is not enabled in all further
reachable markings. The salient property of EFC-nets is now that there must be
a fixed place which is and remains empty, i.e.
4
Other authors refer to deadlocks as siphons or tubes.
5.4 Analysis of Place-Transition Nets 107

∃p ∈ •t : M 0 (p) = 0,∀M 0 ∈ R(P N,M ).

The existence of such a place p is ensured by the fact that all transitions of (•t)•
are also not live. Furthermore since p remains empty all its input transitions are
not live, i.e.
∀t̃ ∈ •p : ∀M 0 ∈ R(P N,M ) : ¬M 0 [t̃ >.
Let Pempty denote the set of empty places in R(P N,M ), i.e.
∀p ∈ Pempty ,M 0 ∈ R(P N,M ) : M 0 (p) = 0
and let Tdead be the set of transitions which are not live in R(P N,M ), i.e.
∀t ∈ Tdead ,M 0 ∈ R(P N,M ) : ¬M 0 [t >.
This yields •Pempty ⊆ Tdead and because of the EFC-net structure we get
Pempty • = Tdead . Thus •Pempty ⊆ Pempty • and Pempty is an empty deadlock,
which contains no marked trap. t
u

In [104, 107] efficient polynomial-time algorithms are described for determining

the dt-property for FC-nets.
Checking boundedness of EFC-nets is somewhat more complicated. In the fol-
lowing we present only the main theorem without further discussion. Before we
can establish this theorem we need the following definitions.

Definition 5.17 Let P N = (P,T,I − ,I + ,M0 ) and

P N 0 = (P 0 ,T 0 ,I 0− ,I 0+ ,M00 ) be two Place-Transition nets and X ⊆ P ∪ T .

1. P N 0 is a subnet of P N iff
P 0 ⊆ P,T 0 ⊆ T and
I 0− (p,t) = I − (p,t),I 0+ (p,t) = I + (p,t), ∀p ∈ P 0 ,t ∈ T 0 , and
M00 (p) = M0 (p),∀p ∈ P .

2. P N 0 is a subnet of P N generated by X iff

P 0 = (X ∩ P ) ∪ •(X ∩ T ) ∪ (X ∩ T ) • ,
T 0 = (X ∩ T ) ∪ •(X ∩ P ) ∪ (X ∩ P ) • , and
I 0− (p,t) = I − (p,t),I 0+ (p,t) = I + (p,t), ∀p ∈ P 0 ,t ∈ T 0 , and
M00 (p) = M0 (p),∀p ∈ P .

Definition 5.18

1. P N 0 = (P 0 ,T 0 ,I 0− ,I 0+ ,M00 ) is a P-component5 of
P N = (P,T,I − ,I + ,M0 ) iff P N 0 is a subnet of P N generated by P 0 ⊆ P
and ∀t0 ∈ T 0 : | • t0 ∩ P 0 | ≤ 1 and |t0 • ∩P 0 | ≤ 1, where the •-notation is
with respect to P N 0 .

2. P N is covered by P-components iff ∀x ∈ P ∪ T : ∃ a P-component P N 0

of P N with x ∈ P 0 ∪ T 0 .

Theorem 5.17 (Boundedness in EFC-nets [35]) A live EFC-net PN is bounded

iff PN is covered by strongly connected P-components.
5
A P-component is related to a State Machine.
108 5 Place-Transition Nets

Figure 5.24 Live and bounded Place-Transition net with no home states

Research efforts to determine the existence of home states for EFC-nets have
only been undertaken in the recent past. Some authors conjectured that all bounded
and live Place-Transition nets have home states, e.g. [124]. This conjecture was
proven to be false in 1984, where E. Best and K. Voss [39] published a bounded
and live ESPL-net without any home states (cf. Fig. 5.24). Since 1989 it is known
that the above-mentioned conjecture holds for EFC-nets.

Theorem 5.18 (Home states in EFC-nets [36, 178]) If PN is a live and bounded
EFC-net then its reachability set contains home states.

Furthermore we can characterise all home states of a live and bounded EFC-net
as follows.

Theorem 5.19 ([36]) Let PN be a live and bounded EFC-net. Then

M is a home state iff M marks all traps of PN.

Proof. We will only prove necessity.

Let P̃ ⊆ P be an arbitrary trap of PN and p ∈ P̃ . Since PN is live and bounded, it
is strongly connected (cf. Theorem 5.1) and ∃t ∈ •p. Since PN is live ∃M̃ ,M̃ 0 ∈
R(P N,M0 ) : M̃ [t > M̃ 0 . M̃ 0 is a marking which marks P̃ . Since M is a home
state, we have M̃ 0 →∗ M and thus M marks P̃ . t
u

This characterisation of home states can be exploited as follows.

5.4 Analysis of Place-Transition Nets 109

Theorem 5.20 ([36]) Let PN be a live and bounded EFC-net and σ ∈ T ∗ be a

firing sequence with ∀t ∈ T : #(σ,t) > 0.6 Then M given by M0 [σ > M is a
home state.
Proof. After the firing sequence σ has occurred, all traps are marked, since a trap
being marked remains marked. t
u

Th. 5.20 is important for analysing Stochastic Petri nets, discussed in Ch. 8, by
means of simulating the net. Once each transition has fired, all further reached
markings are recurrent. These are exactly the states (markings) we are inter-
ested in when analysing the steady state behaviour of the associated Markov
process (cf. Ch. 8). Hence EFC-nets are a convenient class of Place-Transition
nets, because there are characterising conditions for boundedness and liveness
and furthermore the existence of home states is guaranteed for bounded and live
EFC-nets.
For Place-Transition nets with more complex structures such theorems are not
known up to now. E.g. it is known that the dt-property is sufficient for ensuring
liveness in ESPL-nets, but not necessary (see [38]).
Another analysis technique which has evolved in the last decade to analyse also
more complex Place-Transition nets efficiently is the subject of the next section.
Exercise 5.15 Prove the relations between the net classes as given in Fig. 5.18.
Exercise 5.16 Let Q1 ,Q2 ⊆ P . Prove the following conjectures:
1. Q1 ,Q2 are deadlocks =⇒ Q1 ∪ Q2 is a deadlock.
2. Q1 ,Q2 are traps =⇒ Q1 ∪ Q2 is a trap.
Exercise 5.17 Prove or disprove:
1. If P N = (P,T,I − ,I + ,M0 ) is a live EFC-net then
P N = (P,T,I − ,I + ,M ) is live ∀M ≥ M0 .
2. If P N = (P,T,I − ,I + ,M0 ) is a live Place-Transition net then P N =
(P,T,I − ,I + ,M ) is live ∀M ≥ M0 .
Exercise 5.18 Show that the Place-Transition net of Fig. 5.24 is bounded and
live, but has no home states.
Exercise 5.19 Prove or disprove:
1. If PN is a live and bounded EFC-net then M0 is a home state.
2. Let PN be a live and bounded Place-Transition net. Then
M is a home state =⇒ M marks all traps of PN.
Exercise 5.20 Prove that if PN is a live and bounded Marked Graph or a live
and bounded State Machine then M0 marks all traps of PN.
6
#(σ,t) denotes the number of occurrences of t ∈ T in σ ∈ T ∗ .
110 5 Place-Transition Nets

p1 t p2

p0

Figure 5.25 Simple reduction rule

5.4.4 Reduction and Synthesis Analysis

Reduction analysis deals with the reduction of the Place-Transition net by re-
placing subnets of the net by less complex subnets such that several properties
remain invariant. Since the reduced Place-Transition net is less complex than the
original net, once sufficiently reduced, the analysis can be performed, e.g., on
the coverability/reachability tree of the reduced Place-Transition net. A simple
example of a reduction is shown in Fig. 5.25.
This particular reduction rule specifies that if two empty places (p1 ,p2 ) and one
transition (t) with | • p1 | = |p1 • | = | • p2 | = |p2 • | = | • t| = |t • | = 1 and
p1 • 3 t ∈ •p2 are given, these three elements may be reduced to one empty place
p0 with •p0 = •p1 and p0 • = p2 •. It is obvious that properties like boundedness
and liveness are not affected by this reduction (transformation) of the net.
Several authors have independently developed several sets of such reduction
rules, e.g. [31, 32, 62, 125, 163]. We only describe one example of a typical
reduction rule. The interested reader is referred to the literature.
This transformation from [31] removes redundant places by removing arcs from
transitions to the place. If all arcs are removed then the place can be removed
from the Place-Transition net. A place is redundant “when its marking is al-
ways sufficient for allowing firings of transitions connected to it” [31]. This is
expressed by the following definition.

Definition 5.19
Let P N = (P,T,I − ,I + ,M0 ) be a Place-Transition net. A place p ∈ P is re-
dundant iff there exists a subset Q ⊆ P \ {p} (possibly empty) and a weighting
function V : Q ∪ {p} 7→ N such that the following conditions hold:

1. ∃b ∈ N0 : V (p)M0 (p) − q∈Q V (q)M0 (q) = b,

P

i.e. the weighted marking of p is greater or equal to the weighted marking

of the places in Q.
5.4 Analysis of Place-Transition Nets 111

t1

p1

t2 p3

p2

t3

Figure 5.26 Portion of a Place-Transition net with a redundant place

2. ∀t ∈ T : V (p)I − (p,t) − q∈Q V (q)I − (q,t) ≤ b,

P

i.e. firing of a transition removes less or equal tokens on p than from the
places in Q with respect to the weight function.

3. ∀t ∈ T : ∃ct ∈ N0 :
V (p)(I + − I − )(p,t) − q∈Q V (q)(I + − I − )(q,t) = ct ,
P

i.e. with respect to the weight function there are a greater or equal number
of tokens placed onto p than on the places of Q after firing of t.

After having identified the redundant place the transformation can be performed
by removing all edges in •p and p•. For every transition, t ∈ T , an edge with
weight ct is added from t to p. If all the weights ct are zero then the place p
is isolated and can be removed. Simplification of redundant places preserves
properties like boundedness, liveness and the existence of home states (cf. [31]).

Example 5.8 Figure 5.26 shows a portion of a Petri net with a redundant place
p3 . Figure 5.27 shows the same portion of the Place-Transition net after the
transformation has been performed. In this example we have Q = {p1 ,p2 }; V (p) =
1,∀p ∈ Q ∪ {p3 }; b = 0; ct = 0,∀t ∈ T, and the place p3 has been removed.

Most of these transformation/reduction rules can be applied in the reverse direc-

tion, thus increasing the size of the Place-Transition net. This form of building a
net is called synthesis of Petri nets. If we start with a Place-Transition net which
embraces all positive properties like boundedness and liveness and apply these
synthesis rules, we always yield a Place-Transition net with the same properties
making the analysis of the target net unnecessary.
112 5 Place-Transition Nets

t1

p1

t2
p2

t3

Figure 5.27 Portion of the Place-Transition net after removing the redundant place

Reduction and synthesis techniques are analysis techniques which have not been
examined systematically in the literature. Several sets of rules have been pub-
lished, but it is, e.g., not known how powerful all these published rules are.
E.g. are all life and bounded simple-nets completely reducible to an atomic net
(P N = ({p},{t},I − ,I + ,M0 ) where I − (p,t) = I + (p,t) = 1 and M0 (p) > 0)?
Only for FC-nets this and similar questions have been considered, see e.g. [69,
70, 72]. Furthermore the effort of testing the application conditions is not always
considered in the literature. In [62] the computational complexity of several re-
duction and synthesis rules is examined.

5.5 Further Remarks on Petri Nets

This introduction covers only a part of current Petri net theory. We have con-
centrated on those properties of a Petri net which are also significant for the
performance analysis of the corresponding Stochastic Petri net.
Other properties which are of interest are e.g.

1. fairness, i.e.
∀t ∈ T : ∃kt ∈ N : ∀t0 ∈ T,∀σ ∈ T ∗ : M0 [σ >: #(σ,t0 ) − #(σ,t) ≤ kt ,
where #(σ,t) denotes the number of occurrences of t ∈ T in σ ∈ T ∗ .
Fairness means that each transition will eventually fire and can not be ex-
cluded from firing. The sender/receiver system given in Fig. 5.3 is fair,
whereas the Place-Transition net in Fig. 5.8 is not fair, since t2 is not in-
cluded in the infinite firing sequence (t1 t3 )+ .
5.5 Further Remarks on Petri Nets 113

2. reversibility, i.e. M0 is a home state of the Place-Transition net.

3. persistence, i.e. ∀ti ,tj ∈ T,ti 6= tj and ∀M ∈ R(P N,M0 ) :

M [ti > and M [tj >=⇒ M [ti tj >.
Persistence means that a transition can only be disabled by firing of itself.
Marked Graphs are e.g. a subclass of persistent nets.

4. properties of the net language where a language of a Place-Transition net

is given by {W (σ)|σ ∈ T ∗ : M0 [σ >} with
W : T 7→ “set of symbols (words)” and
W (tσ̃) := W (t)W (σ̃),∀t ∈ T,σ̃ ∈ T ∗ .

5. synchronic distance sd of two transitions ti ,tj ∈ T . This notion is related

to fairness and defined as follows:
sd(ti ,tj ) := sup{|#(σ,ti ) − #(σ,tj )| |M [σ > ,∀M ∈ R(P N,M0 )}.

Finally we turn to the question of whether an algorithm exists to decide whether

a Place-Transition net is live or not. This decidability problem remained open
for about 20 years until 1981.
In fact it was well known since the early days of Petri net theory that the above
mentioned problem is equivalent to the well-known reachability problem which
can be described as follows:
Given a Place-Transition net and a marking M , does an algorithm exists to decide
whether M ∈ R(P N,M0 )? As we learned in Ex. 5.11, the coverability tree is
not suitable for deciding both problems. In 1981 E.W. Mayr [118, 119] and in
1982 R. Kosaraju [109] published algorithms for the reachability problem, i.e.
reachability and liveness problem are both decidable. Furthermore it is known
that both problems are exponential-space hard with complexity exponential in
|P |,|T | and the flow relation |F | (cf. [119]).

Exercise 5.21 (cf. [146]) This exercise is dedicated to some real life problems in
modelling computer systems. It concerns the ’mutual exclusion’ problem, which
occurs if several parallel processes are sharing common resources. If one pro-
cess is, e.g., changing the internal state of a resource (consider writing a file),
then for all other processes access has to be denied, in order to keep the state of
the resource consistent.
A very common solution is to restrict simultaneous access to a resource by mod-
ifying the source code of each process and insert special lines of code.
E.g. a control variable is introduced, which represents the access status of a re-
source (free or occupied) and each process has to check this control variable
before using the resource. The section of code where the resource is used is re-
ferred to as the “critical section”. The crux of all such solutions is the atomicity
or non-atomicity of certain statements. In this exercise we will denote atomic
statements (operations) in pointed brackets, e.g. <x := y; >.
In the following, four algorithms are presented in pseudo-code, and you are
asked to prove or disprove the correctness of these algorithms using Petri nets.
114 5 Place-Transition Nets

For simplicity we only look at two processes P0 and P1 in the first three algo-
rithms and assume that the presented part of the code of each process is executed
infinitely often.

1. Algorithm:
Control variables:
var flag : array [0..1] of boolean;
turn : 0..1;
Initial value of turn is 0 and all flags are false

Program for process Pi :

<flag[i] := true;>
while <flag[j]> do /* j = i + 1 mod 2 */
if <turn = j> then
begin
<flag[i] := false;>
while <turn = j> do <nothing> enddo;
<flag[i] := true; >
endif;
enddo;
critical section;
<turn := j>;
<flag[i] := false;>

2. Algorithm:
Control variables:
var flag : array [0..1] of boolean;
turn : 0..1;
Initial value of turn is 0 and all flags are false

Program for process Pi :

<flag[i] := true;>
while <turn 6= i> do
while <flag[j]> do <nothing> enddo; /* j = i + 1 mod 2 */
<turn := i;>
enddo;
critical section;
<flag[i] := false;>

3. Algorithm:
This algorithm uses a binary semaphore sem. P(sem) decreases the value
5.5 Further Remarks on Petri Nets 115

of sem by 1, provided it is greater than 0, otherwise the process has to wait

until sem becomes > 0. V(sem) increases the value of sem by 1.
Control variables:
var sem : binary semaphore;
Initial value of sem is 1

Program for process Pi :

< P(sem); >

critical section;
< V(sem); >

4. Algorithm:
In this problem we have 5 processes P0 , . . . ,P4 sharing common resources.
The restriction we have is that if process Pi is in its critical section, then
(Pi−1 mod 5) and (Pi+1 mod 5) are both not able to be or to enter their
own critical section.
This algorithm uses an array of binary semaphores given by

Control variables:
var sem[0..4] : binary semaphore;
Initial value of all semaphores is 1

Program for process Pi :

< P(sem[i]); >

< P(sem[i+1 mod 5]); >
critical section;
< V(sem[i]); >
< V(sem[i+1 mod 5]); >
116

6 Coloured Petri Nets

You may have noticed that the graphical representation of Petri nets becomes
fairly complex if we try to model real life problems. The main reason is that we
only have one type of token. In our mutual exclusion example (see Exercise 5.21,
page 113) we had to represent each process by a separate subnet, although their
respective behaviour is identical. Furthermore we had to represent all values of a
variable by different places.
In this chapter we introduce Coloured Petri nets (CPNs) in which a type called
the colour is attached to a token. CPNs were first defined by K. Jensen (cf. [93]).
There are several other Petri net models which distinguishes between individual
tokens, e.g. [79, 99, 148], but since Queueing Petri nets (QPNs) [18], discussed
in Ch. 10, are based upon CPNs, we need to know only about the latter.
Fig 6.1 illustrates a part of a CPN in which place p1 is marked by a token of
colour a and p2 is marked by two tokens, one of colour b and one of colour c.
How can we describe the firing of transition t? E.g., we may require transition t
to be enabled in the marking shown and that it destroys the tokens a and b and
creates a token of colour d on p3 . Another possible behaviour of t may be that it
destroys tokens a and c and creates a token of colour e on p3 . These two different
ways in which t might fire are referred to as the different modes in which t fires.
In CPNs these modes of a transition are also described by colours, but now at-
tached to the transition. Let us denote the set consisting of all these colours by
C(t). So we say, e.g., that transition t fires with respect to colour x (characteris-
ing a certain “firing mode” of t) or with respect to colour y. With that in mind the
incidence functions of a CPN can be defined similar to the incidence functions
of a Place-Transition net (see Def. 5.1 on page 81). E.g., if transition t fires with
respect to colour x, tokens a and b will be destroyed and one token of colour d
will be created on p3 .
Before defining CPNs formally we need to know about multi-sets. Informally a

p1 a
t

p3

b
p2 c

Figure 6.1 A simple CPN

 117

a
a c
b b d
b d

Figure 6.2 A multi-set m

multi-set is the same as a set, except that individual elements may occur more
than once and for this reason they are sometimes also called bags. For instance
if we have a set {a,b,c} and add element b to the set we still have only the set
{a,b,c}. However, if we added b to the multi-set {a,b,c} we have the new multi-
set {a,b,b,c} with two occurrences of the element b.

Definition 6.1 (Multi-set) A multiset m, over a non-empty set S, is a function

m ∈ [S 7→ N0 ]. The non-negative integer m(s) ∈ N0 is the number of appear-
ances of the element s in the multi-set m.

Refer to Fig. 6.2. The definition of the multi-set m in that figure is given by
if s ∈ {a,d}

 2,
m(s) = 3, if s = b
1, if s = c


We denote the set of all finite multi-sets over S by SM S and define addition of
multi-sets and multiplication with ordinary numbers as follows:
Definition 6.2 ∀m1 ,m2 ∈ SM S and r ∈ R 1 define
(m1 + m2 )(s) := m1 (s) + m2 (s)
(rm1 )(s) := rm1 (s)
Now, remember that the incidence functions of a Place-Transition net reflect the
connection between places and transitions. That is, I − (p,t) describes the number
of tokens which are destroyed on place p in case of t fires. The corresponding
backward incidence function of a CPN is now an element of
I − (p,t) ∈ [C(t) 7→ C(p)M S ].
The forward incidence function I + (p,t) is obviously defined similarly.
Let us return to our example. Since t can fire in two different modes, the colour
set should consist of (at least) two elements, e.g. C(t) = {x,y}. The (backward)
incidence function I − (p1 ,t) ∈ [C(t) 7→ C(p1 )M S ] is given by I − (p1 ,t)(x) =
{a}, which is a multi-set, and I − (p1 ,t)(x)(a) = 1 denotes the number of ele-
ments of colour a in that multi-set.
We define a CPN formally as follows:
1
R denotes the set of real numbers.
118 6 Coloured Petri Nets

Definition 6.3 (CPN) A Coloured Petri net (CPN) is a 6-tuple

CPN=(P,T,C,I − ,I + ,M0 ), where
• P is a finite and non-empty set of places,

• T is a finite and non-empty set of transitions,

• P ∩ T = ∅,

• C is a colour function defined from P ∪ T into finite and non-empty sets,

• I − and I + are the backward and forward incidence functions defined on

P×T such that
I − (p,t),I + (p,t) ∈ [C(t) → C(p)M S ],∀ (p,t) ∈ P × T ,

• M0 is a function defined on P describing the initial marking such that

M0 (p) ∈ C(p)M S ,∀p ∈ P .

The preset and postset of a transition and place, respectively, of a CPN are de-
fined analogous to that for Place-Transition nets on page 83:

Definition 6.4
The preset •(p,c) of p ∈ P and c ∈ C(p) is
•(p,c) := {(t,c0 )|t ∈ T,c0 ∈ C(t) : I + (p,t)(c0 )(c) 6= 0}.
The preset •(t,c0 ) of t ∈ T and c0 ∈ C(t) is
•(t,c0 ) := {(p,c)|p ∈ P,c ∈ C(p) : I − (p,t)(c0 )(c) 6= 0}.
The postset (p,c)• of p ∈ P and c ∈ C(p) is
(p,c)• := {(t,c0 )|t ∈ T,c0 ∈ C(t) : I − (p,t)(c0 )(c) 6= 0}.
The postset (t,c0 )• of t ∈ T and c0 ∈ C(t) is
(t,c0 )• := {(p,c)|p ∈ P,c ∈ C(p) : I + (p,t)(c0 )(c) 6= 0}.

Next we can define the behaviour of a CPN, again very much the same way as
for Place-Transition nets.

Definition 6.5 (Enabled Transition) A transition t ∈ T is enabled in a marking

M w.r.t. a colour c0 ∈ C(t), denoted by M [(t,c0 ) >, iff M (p)(c) ≥ I − (p,t)(c0 )(c),∀p ∈
P,c ∈ C(p).
An enabled transition t ∈ T may furthermore fire in a marking M w.r.t. a colour
c0 ∈ C(t) yielding a new marking M 0 , denoted by M → M 0 or M [(t,c0 ) > M 0 ,
with
M 0 (p)(c) = M (p)(c) + I + (p,t)(c0 )(c) − I − (p,t)(c0 )(c),∀p ∈ P,c ∈ C(p).

Not surprisingly the various properties we defined for Place-Transition nets can
also be defined for CPNs where we now denote the reachability set by R(CP N ) :=
R(CP N,M0 ) := {M |M0 →∗ M } where →∗ is the reflexive and transitive clo-
sure of →.

Definition 6.6 Let CPN=(P,T,C,I − ,I + ,M0 ) be a Coloured Petri net.

 119

1. CPN is bounded iff R(CP N,M0 ) is finite.

2. CPN is live iff ∀M ∈ R(CP N,M0 ),t ∈ T,c ∈ C(t) : ∃M 0 : M →∗ M 0

and M 0 [(t,c) >.

3. A marking M 0 is called a home state iff

∀M ∈ R(CP N,M0 ) : M →∗ M 0 .

4. D ⊆ p∈P c∈C(p) (p,c) is called deadlock (trap) iff •D ⊆ D•

S S

(D• ⊆ •D).
D is marked at a marking M iff ∃(p,c) ∈ D : M (p)(c) > 0.

p1 p5

t1 t4

p2 p6

t2 t5

p3 bus access bus access p7

bus free

t3 t6

Figure 6.3 Place-Transition net model of a dual multiprocessor system

Example 6.1 Consider a dual processor system where each processor accesses
a common bus, but not simultaneously. Fig. 6.3 illustrates a Place-Transition net
model of the system. Places p1 ,p2 ,p3 represent the states of the first processor
while places p5 ,p6 ,p7 represent the states of the second one.
However, since the processors are identical, an obvious idea (now that we know
about CPNs) is to represent each of the two processors by a different colour
token. A CPN model of the same system is illustrated in Fig. 6.4. The formal
definition is given by CP N = (P,T,C,I − ,I + ,M0 ) where

1. P = {p1 ,p2 ,p3 ,p4 },

2. T = {t1 ,t2 ,t3 },

120 6 Coloured Petri Nets

3. C(p1 ) = C(p2 ) = C(p3 ) = {c1 ,c2 },C(p4 ) = {•},

C(t1 ) = C(t2 ) = C(t3 ) = {c01 ,c02 }.
4. For the definition of the backward and forward incidence functions it is
sufficient to specify only the functional non-zero values.
I − (p1 ,t1 )(c01 )(c1 ) = 1, I − (p1 ,t1 )(c02 )(c2 ) = 1, I − (p2 ,t2 )(c01 )(c1 ) = 1,
I (p2 ,t2 )(c02 )(c2 )
−
= 1, I (p4 ,t2 )(c01 )(•) = 1,
−
I − (p4 ,t2 )(c02 )(•) = 1,
I − (p3 ,t3 )(c01 )(c1 ) = 1, I − (p3 ,t3 )(c02 )(c2 ) = 1, I + (p2 ,t1 )(c01 )(c1 ) = 1,
I + (p2 ,t1 )(c02 )(c2 ) = 1, I + (p3 ,t2 )(c01 )(c1 ) = 1, I + (p3 ,t2 )(c02 )(c2 ) = 1,
I + (p1 ,t3 )(c01 )(c1 ) = 1, I + (p1 ,t3 )(c02 )(c2 ) = 1, I + (p4 ,t3 )(c01 )(•) = 1,
I + (p4 ,t3 )(c02 )(•) = 1,

5. M0 (p1 )(c1 ) = M0 (p1 )(c2 ) = M0 (p4 )(•) = 1.

c1
p1 c2

t1

p2

t2

p3 bus access
p4
bus free

t3

Figure 6.4 Coloured Petri net model of the same dual multiprocessor system

Now that we know about the structure, behaviour and properties of CPNs, we
next would like to know how to analyse them. Fortunately, every CPN can be
unfolded uniquely into a Place-Transition net so that all the concepts relating to
CPNs are consistent with those of Place-Transition nets.
A CPN=(P,T,C,I − ,I + ,M0 ) is unfolded into a Place-Transition net in the follow-
ing way:
1. ∀ p ∈ P,c ∈ C(p) create a place (p,c) of the Place-Transition net.
2. ∀ t ∈ T,c0 ∈ C(t) create a transition (t,c0 ) of the Place-Transition net.
 121

p1 I − (p1 ,t)
t

p3
I + (p3 ,t)

p2 I − (p2 ,t)

Figure 6.5 The graphical representation of a CPN

3. Define the incidence functions of the Place-Transition net as

I − ((p,c)(t,c0 )) := I − (p,t)(c0 )(c),
I + ((p,c)(t,c0 )) := I + (p,t)(c0 )(c).
4. The initial marking of the Place-Transition net is given by
M0 ((p,c)) := M0 (p)(c),∀ p ∈ P,c ∈ C(p)
The unfolded CPN is given by
(t,c0 ),I − ,I + ,M0 )
[ [ [ [
PN = ( (p,c),
p∈P c∈C(p) t∈T c0 ∈C(t)

It should be obvious that while the unfolding of a CPN yields a unique ordinary
Petri net the converse is not true. Folding (as it is termed) of a Place-Transition
net can be done in more than one way.
Example 6.2 As an example, the unfolded net of Fig 6.4 is displayed in Fig. 6.3.
It should thus be clear that Coloured Petri nets are only a means of simplifying
the graphical representation of a Petri net model of a complex system. We can
analyse such models by unfolding the CPN and performing the analysis on the
unfolded net.
Coloured Petri nets have undergone several revisions. Jensen[96, 97] uses ex-
pressions to specify the incidence functions and markings, but for invariant anal-
ysis a function representation as presented here is used. The expression represen-
tation is based on the meta language ML [122]. In the same book Jensen also
introduces guards which can be attached to the transitions of a CPN. A guard
determines a boolean value and the corresponding transition is only enabled if
this value is true.
For simplicity and to avoid confusion, we will describe the “older” graphical
representation presented by Jensen in [93, 94] and only touch on the graphical
representation of CPNs using ML constructs and not go into the formal details.
Fig. 6.5 illustrates a part of a CPN where the incidence functions
I − (p,t),I + (p,t) ∈ [C(t) 7→ C(p)M S ] are represented as “arc weights”. If we
can find mnemonics for these functions, this idea of representing a CPN graphi-
cally will be sufficient for a human reader. This is the original idea presented by
Jensen in [93].
122 6 Coloured Petri Nets

a
p1 x+y
b c t

p3
2x + y
p2 b
x
a

Figure 6.6 CPN with expressions attached to the arcs

Later, concepts of Predicate/Transition nets [79] were adopted for the graphical
representation of CPNs. In order to understand this we need to remember the
usual concept of binding all or a subset of all free variables in an expression
to values. Consider some variables x,y and z and an expression involving these
variables, e.g. 2x+y −z. If we bound these variables to values i,j,k respectively,
then the given expression evaluates to 2i + j − k. In most cases the incidence
function of a CPN can now be written using an expression. Consider e.g. the
CPN of Fig. 6.6, where we have attached an expression to each arc remembering
that the x and y are variables on a multi-set.
Thus the expression x + y states that in case transition t fires, a token of kind x
and a token of kind y is destroyed on p1 . To determine whether a transition t is
enabled and the result of it firing, the free variables of all expressions attached to
the input and output arcs of t must be bound to some values in order to evaluate
the expression. An arc expression must evaluate to a multi-set over the colour
set of the corresponding place. Thus possible evaluations of the arc expression
x + y are a + a,a + b,a + c,a + d etc. with C(p1 ) = {a,b,c,d}. Clearly, for all
arc expressions adjacent to a transition we have to bind identical variables to the
same value. E.g., if we bound the variable x to a and y to b, this implies that the
arc expressions x + y, x and 2x + y evaluate to a + b, a and 2a + b respectively.
After the evaluation of all arc expressions of a transition, enabling and firing of
that transition is similar to that of Place-Transition nets. E.g., if the arc expres-
sions x + y and x evaluate to a + b and a respectively, then t is enabled at a
marking M iff there is at least one token of colour a and at least one token of
colour b in the marking of p1 and at least one token of colour a in the marking
of p2 . This is the case in Fig 6.6 and if t fires it destroys the tokens a and b on
p1 as well as the token a on p2 and creates 2 tokens of colour a and one token of
colour b on p3 as prescribed by the arc expression 2x + y.
If we bound x to c and y to b, then t is not enabled since there is no token of
colour c in the marking of p2 depicted in Fig. 6.6.
If, depending on the binding, we want a transition to behave in more than one
way, we can describe this by employing syntactical constructs of programming
languages. Jensen in [96, 97] uses the programming language ML for this and
we offer the CPN in Fig. 6.7 as an example.
 123

In that example, C(p1 ) = {a,b,c},C(p2 ) = {a,b,c,d},C(p3 ) = {r,s} and if x is

bound to a and y to b then t is enabled and the arc expression “if . . .” evaluates
to 2r + s. Note that this arc expression always evaluates to a multi-set over the
colour set of p3 , as before. If x is bound to a and y to c then t is also enabled and
in that case, when it fires, no token is created on p3 as denoted by the expression
“empty”.

a
p1 x+y
b c t

p3
if (x = a)
b then if (y = b) then 2r + s
p2 x
a else empty
else 2s

Figure 6.7 CPN with code rather than functional description of the incidence functions

In [96] the reader will find a formal definition of CPNs and its graphical rep-
resentation which differs somewhat from that described above. In the following
we will use the definition of a CPN given in this chapter and use the concepts
concerning the graphical representation only to draw CPNs compactly.

a
p1 x
b t

r p3
if (x = a)
c then s
p2 if (x = a)
d else empty
then c
else d

Figure 6.8 C(p1 ) = {a,b,c},C(p2 ) = {b,c,d},C(p3 ) = {r,s}

Exercise 6.1 Fold the following sets of places and transitions of the Petri net in
Fig. 5.3, page 81:
{p9 ,p10 } and {pi ,pi+4 },{ti ,ti+4 },i ∈ {0, . . . ,4}.

Exercise 6.2

1. Give the formal definition (see Def. 6.3) of the CPN displayed in Fig. 6.8.
124 6 Coloured Petri Nets

2. Draw the reachability graph of this CPN.

3. Determine the unfolded Place-Transition net. What does the “empty”-expression

imply for the connection of transitions and places of the unfolded net?
 125

7 Further Reading

Further introductions on Petri nets are given in [132, 149, 150] and short intro-
ductory papers are [125, 130].
A common extension of Petri nets found in the literature is to impose priorities
on transitions and/or to introduce so-called inhibitor arcs. If two transitions of
different priority are enabled in the sense of ordinary Petri nets then only the
higher priority transition may fire. An inhibitor arc connects a place p to a transi-
tion and imposes an additional constraint on the enabling of that transition. The
transition is now enabled iff the enabling rule of ordinary Petri nets concerning
the usual arcs between places and that transition is satisfied and if place p is not
marked. In [132] these extensions are discussed, amongst others. Petri nets with
inhibitor arcs or priorities have the power of Turing machines and thus several
problems, e.g. reachability problem or determining liveness, become undecid-
able. This is the reason why most Petri net researchers do not consider these
extensions, although they might be necessary to model real life problems.
The analysis of net classes, especially of (Extended) Free Choice nets, is the
subject of [34, 36, 37, 38, 39, 60, 175] and several results are summarised in
[38, 125]. Analysis of Marked Graphs is considered in [78]. This class of nets
is also called “Synchronisationsgraphen” or T-graphs. Another name for State
Machines is S-graphs. (Extended) Simple nets, also called Asymmetric Choice
nets, are dealt with in [38].
A further net class with similar properties like EFC-nets are investigated in [173,
174]. This class is called Equal Conflict nets (EC-nets) and allows multiple arc
weights. Its definition is:
∀t,t0 ∈ T : (•t ∩ •t0 6= ∅ =⇒ I − (p,t) = I − (p,t0 ),∀p ∈ P ).
Reduction and synthesis analysis can be found in several papers. The following
papers are concerning reduction and synthesis analysis of Petri nets: [31, 32, 70,
80, 101, 126, 168, 171, 172, 176, 182].
A study on different efficient methods for the analysis of the reachability set
(symmetry method, stubborn set method, symbolic model checking, incomplete
reachability analysis) can be found in [145]. Progress has been made in the anal-
ysis of the reachability set exploiting dynamic priorities[21], structural informa-
tion for a Kronecker based representation of the state space (e.g. [106] (see also
page 175)) or Binary Decision Diagrams (e.g. [123]).
Further Petri net models introducing the concept of individual tokens are de-
scribed in [79, 99, 148]. The invariant analysis of CPNs was first discussed in
[93]. The latest version of CPNs is defined in [96, 97, 98] and it is shown how a
functional representation can be used for calculating the invariants of this CPN
version. A coloured version of the reduction rules of [31, 32] are defined in [88].
CPNs with an infinite number of colours also have the power of Turing machines
126 7 Further Reading

[131].
An international conference with the title “Application and Theory of Petri Nets”
started in 1980 and is held annually, e.g. [95, 3, 177, 58, 40, 12, 59, 64, 128, 54,
71]. Further publications on Petri nets can be found in “Advances in Petri Nets”
(Lecture Notes in Computer Science, Springer-Verlag, e.g. [44, 45, 153, 154,
155, 156, 157, 158, 159, 160, 161, 162, 41, 1, 67]).
The German “Gesellschaft für Informatik (GI)” has a “Special Interest Group
on Petri Nets and Related System Models”, which publishes a Petri net
newsletter to its members half-yearly. For membership contact Gesellschaft
für Informatik, email: gs@gi-ev.de (yearly costs: ≈ EUR 10 ) or visit
the URLs http://www.gi-ev.deand http://www.informatik.
uni-hamburg.de/TGI/GI-Fachgruppe0.0.1/resp. There is also a
Petri net mailing group (see http://www.informatik.uni-hamburg.
de/TGI/PetriNets/mailing-lists/). Additional information on
Petri nets can be accessed via URL http://www.informatik.
uni-hamburg.de/TGI/PetriNets/.
First proposals have been published defining a standard for high-level
Petri nets (High-level Petri Nets - Concepts, Definitions and Graphical No-
tation) within the International Organization for Standardisation (ISO). For
more information see http://www.informatik.uni-hamburg.de/
TGI/PetriNets/standardisation/.
 Part III

TIME-AUGMENTED PETRI
NETS
 129

Petri nets involve no notion of time, since it is not defined at what point in time
a transition will fire. Analysing the performance of a system with a Petri net
is thus not possible. Petri nets can only be used for qualitative analysis, i.e., to
analyse the functional or qualitative behaviour of a system. For performance or
quantitative analysis the temporal (time) behaviour of the system has to be part
of the Petri net description.
Since the early 70’s several suggestions on incorporating time into Petri nets have
been published. In general, there are two possible ways to do this:

1. Specifying sojourn times of tokens on places.

If tokens are fired onto a place p they become unavailable to all output
transitions of p for a certain time interval. Once this time has elapsed the
tokens become available and can be consumed by the output transitions of
p. Such Petri nets are known as Timed Places Petri nets (TPPNs).

2. Specifying a firing delay of enabled transitions.

Once a transition is enabled, it will fire after a certain time interval has
elapsed. These Petri nets are also called Timed Transitions Petri nets (TTPNs)
which, in turn, can be classified into two groups.

(a) Preselection models.

Once a transition is enabled, it selects all tokens it needs to fire on
its input places so that they are unavailable to any other transition of
the Petri net. The enabled transition waits until its firing time interval
has elapsed and then fires immediately. By firing, the transition de-
stroys all its reserved tokens and creates tokens onto its output places
according to the firing rule of a Petri net.
(b) Race models.
In these Petri nets tokens are not reserved by a transition. Once a tran-
sition is enabled, it waits for its firing time to elapse and than fires
immediately provided it is still enabled at that time. Since a faster
transition might disable other transitions, enabled transitions compete
or “race” for tokens; hence the name of this class of Petri net.

These time-augmented Petri nets, TPPN as well as TTPN models, can be clas-
sified further depending upon whether the times mentioned are deterministic or
stochastic. In the first case the class of such Petri nets is called “Timed Petri
Nets” and in the latter they are called “Stochastic Petri Nets”.1
A further classification depending upon the firing policy of transitions (resam-
pling, age memory, enabling memory) can be found in [4, 5], but this classifica-
tion is not relevant for exponentially distributed firing delays.
1
Unfortunately the name of this class coincides with the name of a member of the class: Molloy’s
SPNs [124].
130

Since our interest is in time-augmented Petri nets which can be analysed by

Markovian techniques, the next sections will consider those time-augmented
Petri nets where time is exponentially distributed.
We first of all introduce a wide-spread family of Stochastic Petri nets, not surpris-
ingly called “Stochastic Petri Nets” (SPNs) as well as “Generalized Stochastic
Petri Nets” (GSPNs) and then describe a different class called “Queueing Petri
Nets” (QPNs). QPNs combine the two ways of incorporating time by specifying
a sojourn time of tokens on places as well as a firing delay for enabled transitions.
All three classes are race models, i.e., enabled transitions compete for the tokens
on their input places. A preselection policy, as we will see later, can be repre-
sented using immediate transitions as defined for GSPNs and QPNs.
 131

8 Stochastic Petri Nets

The continuous-time Stochastic Petri net (M.K. Molloy [124], S. Natkin [127])
SP N = (P N,Λ) is formed from the Place-Transition net P N = (P,T,I − ,I + ,M0 )
by adding the set Λ = (λ1 , . . . ,λm ) to the definition. λi is the, possibly mark-
ing dependent, transition rate of transition ti . I.e., the firing time is exponentially
distributed and the distribution of the random variable χi of the firing time of
transition ti is given by
Fχi (x) = 1 − e−λi x
A typical example of a Stochastic Petri net (SPN) is illustrated in Fig. 8.1 (cf.
[124]).
In the SPN in the figure, transition t1 is enabled at M0 = (1,0,0,0,0). The time
elapsed until t1 fires is exponentially distributed with rate λ1 , i.e. the average
time for t1 to fire is λ11 . Once t1 has fired, using the firing rule of Place-Transition
nets, we obtain marking M1 = (0,1,1,0,0). At M1 , t2 and t3 are concurrently
enabled. That is, one of these two transitions will fire next after a certain time has
elapsed. If transition t2 fires first, the SPN changes to marking M2 = (0,0,1,1,0)
and if t3 fires before t2 , we get the marking M3 = (0,1,0,0,1). The next marking
thus depends on which transition “wins the race”. The probability that t2 fires
first is given by:

P [t2 fires first at M1 ] = P [χ2 < χ3 ]

p1

t1

p2 p3

t4 t2 t3

p4 p5

t5

Figure 8.1 A typical Stochastic Petri net

132 8 Stochastic Petri Nets
Z ∞ Z x
= ( λ2 e−λ2 y dy)λ3 e−λ3 x dx
0 0
Z ∞
= (1 − e−λ2 x )λ3 e−λ3 x dx
0
λ2
=
λ2 + λ3

and similarly,

λ3
P {t3 fires first at M1 } = .
λ2 + λ3

This shows that the probability of changing from marking M1 to some other
marking is independent of the time spent in M1 !
The sojourn time in M1 is given by the minimum of the independent, exponen-
tially distributed firing times of both transitions, namely:

P [min(χ2 ,χ3 ) ≤ x] = P [χ2 ≤ x or χ3 ≤ x]

= 1 − P [χ2 > x and χ3 > x]
= 1 − e−λ2 x e−λ3 x
= 1 − e−(λ2 +λ3 )x

Thus the sojourn time in M1 is also exponentially distributed with parameter

(λ2 + λ3 ). In combination with the fact that the probability of changing the state
is independent of the sojourn time, this implies that a SPN describes a Markov
process.
The rate of going from M1 to say e.g. M2 is then given, as usual, by:

λ2 1 λ2 1
= 1 = λ2
(λ2 + λ3 ) E[min(χ2 ,χ3 )] (λ2 + λ3 ) (λ2 +λ3 )

The performance or quantitative analysis of SPNs can be carried out straight-

forwardly by analysing the corresponding Markovian process. Consider Fig. 8.2
which depicts the reachability graph of the SPN of Fig. 8.1. Consider each mark-
ing in that reachability graph as a state of the corresponding Markov chain (MC)
and attach the firing rate λi of transition i as an arc label to each transition in the
Markov chain, giving rise to Fig. 8.3.
To be more precise: The Markov chain (cf. Sec. 2.3) of an SPN can be obtained
from the reachability graph of the associated Place-Transition net P N as follows:
The MC state space is the reachability set R(P N ) and the transition rate from
state Mi to state Mj is given by qij = λk , the possibly marking-dependent,
firing rate of transition tk from Mi to Mj . If several transitions lead from Mi to
Mj then qij is the sum of the rates of those transitions. Similarly qijP= 0 if no
transitions lead from Mi to Mj and qii is determined so as to satisfy j qij = 0
(cf. Eq. 2.43 on page 50).
 133

(1,0,0,0,0)
t1

(0,1,1,0,0)
t4 t2 t3
t5
(0,0,1,1,0) (0,1,0,0,1)
t2
t3 t4
(0,0,0,1,1)

Figure 8.2 Reachability graph of the SPN’s underlying Place-Transition net

(1,0,0,0,0)
λ1

(0,1,1,0,0)
λ4 λ2 λ3
λ5
(0,0,1,1,0) (0,1,0,0,1)
λ2
λ3 λ4
(0,0,0,1,1)

Figure 8.3 Markov chain of the SPN

The square matrix Q = (qij ) of order s = |R(P N )| is the matrix Q in Sec. 2.3.
As before, the steady state distribution π of the MC is obtained by solving the
linear equations
s
X
πQ = 0; πj = 1
i=1
From the vector π = (π1 ,π2 , . . . ,πs ) we can then compute the following perfor-
mance measures:

Probability of being in a subset of markings: Let B ⊆ R(P N ) constitute

the markings of interest in a particular Stochastic Petri net. Then the probability
of being in a state of the corresponding subset of the MC is given by:
X
P [B] = πi .
Mi ∈B
134 8 Stochastic Petri Nets

Mean number of tokens: Let B(pi ,n) be the subset of R(P N ) for which the
number of tokens in a place pi is n, i.e. B(pi ,n) = {M ∈ R(P N )|M (pi ) = n}.
Then the average number of tokens in place pi is given by:
∞
X
mi = (nP [B(pi ,n)]).
n=1

Probability of firing transition tj : Let ENj be the subset of R(P N ) in which

a given transition tj is enabled, i.e. ENj = {M ∈ R(P N )|M [tj >}. Then, the
probability rj that an observer, who looks randomly into the net, sees transition
tj firing next is given by:
X λj
rj = πi ( )
Mi ∈ENj
(−qii )

where (−qii ) is the sum of transition rates out of Mi .

Throughput at a transition tj : The throughput at a timed transition is given

by it’s mean number of firings at steady state:
X
dj = πi λ j
Mi ∈ENj

Example 8.1 Consider the Stochastic Petri net shown in Fig.8.1. That SPN dis-
plays sequential operation (t5 ,t1 ), parallel operation (t2 ,t3 ), forking (t1 ), join-
ing (t5 ), and conflict (t4 ,t5 ). Assume mean firing rates λ1 = 2, λ2 = 1, λ3 = 1,
λ4 = 3, and λ5 = 2. Starting with an initial marking of one token in place p1
and no tokens in the remaining places, the reachability set has five markings or
equivalently, there are five states in the Markov chain.
Solving the Markov chain we obtain the following steady-state marking proba-
bilities:
5
P [M1 ] = P [(1,0,0,0,0)] =
43
8
P [M2 ] = P [(0,1,1,0,0)] =
43
2
P [M3 ] = P [(0,0,1,1,0)] =
43
23
P [M4 ] = P [(0,1,0,0,1)] =
43
5
P [M5 ] = P [(0,0,0,1,1)] =
43
Using the marking probabilities and the number of tokens in each place in a
particular marking we can deduce the steady state probabilities of their being µi
tokens at place pi . This is known as the token probability density function.
 135

38 5
P [µ1 = 0] = 43 P [µ1 = 1] = 43

12 31
P [µ2 = 0] = 43 P [µ2 = 1] = 43

33 10
P [µ3 = 0] = 43 P [µ3 = 1] = 43

36 7
P [µ4 = 0] = 43 P [µ4 = 1] = 43

15 28
P [µ5 = 0] = 43 P [µ5 = 1] = 43

If we were to assume a different initial marking then we would obtain a different

Markov chain. The more tokens we have in the system the larger the Markov
chain. If we start with two tokens in place p1 we would find 14 states in the
reachability set. Similarly, if we started with 3 tokens in place p1 , we would find
30 states in the reachability set. This state space explosion is a familiar problem
in practice.

The most important aspect of SPNs is that the reachability graph of the underly-
ing Place-Transition net and the Markov chain are isomorphic. In other words,
the number of states/markings and the connection structure of both graphs are the
same. The Markov chain describes the “reachability graph” of the SPN. There-
fore all properties of the underlying Place-Transition net also hold for the SPN
and vice versa. For that reason qualitative (functional) analysis of an SPN can be
done by applying the algorithms and techniques for Place-Transition nets.
Although SPNs embrace these important features, they also exhibit significant
problems, which led to their modification to Generalized Stochastic Petri nets
(GSPNs). Before we introduce GSPNs we first point out the difficulties of SPNs.
Consider the SPN in Fig. 8.4 with firing rates Λ = (1.0,1.0,1.0,µ,µ,µ). The
corresponding Markov chain is shown in Fig. 8.5.

t3 t1 t6

p2 p1 p3 p4

t2 t4 t5

Figure 8.4 Simple SPN

We now want to analyse this continuous-time Markov chain for different values
of µ. The infinitesimal generator Q of the MC is given by
136 8 Stochastic Petri Nets

M1 = (1,0,0,0)
µ 1.0
1.0 1.0

M3 = (0,0,1,0) M2 = (0,1,0,0)

µ µ

M4 = (0,0,0,1)

Figure 8.5 Markov chain

−2.0 1.0 1.0 0

 
 1.0 −1.0 0 0 
Q= 
 µ 0 −2µ µ 
0 0 µ −µ

and we have to solve the equation

πQ = 0 (8.1)

yielding the steady-state distribution π = (P [M1 ],P [M2 ],P [M3 ],P [M4 ]).
A simple iterative method (power method) to solve Eq. (8.1) is:

π k (aQ + I) = π k+1 (8.2)

where π k is the approximate steady state distribution in the k-th step of the iter-
ation and a is a factor such that the sum of absolute values of all rows is less or
equal to 1.
For our example we choose a := 1.001 1
(max{2,2µ})−1 . In general a good choice
is1
1 |R(P N )|
a := (maxi=1 {|qii |})−1
1.001
Table 8.1 shows the number of iterations needed for several values of µ and
the corresponding result for π starting with the vector (0.25,0.25,0.25,0.25).
The iteration stops once kπ n+1 − π n k <  where kxk := ni=1 |xi | for x =
P

(x1 , . . . ,xn ) without taking the convergence speed into account for simplicity.
π iteration −πiexact
The relative error is given here by max4i=1 (100% × | i |).
π exact
i

1
For more information on the numerical analysis of Markov chains see [169].
 137

µ #iterations π iteration rel. error

1.0 1 (0.25000,0.25000,0.25000,0.25000) 0.00000
2.0 47 (0.33333,0.33333,0.16667,0.16667) 0.00064
10.0 122 (0.45455,0.45454,0.04546,0.04546) 0.00156
50.0 470 (0.49022,0.49017,0.00980,0.00980) 0.00545
100.0 861 (0.49510,0.49500,0.00495,0.00495) 0.01043
200.0 1574 (0.49761,0.49741,0.00249,0.00249) 0.02045
500.0 3464 (0.49925,0.49875,0.00100,0.00100) 0.05049
1000.0 6226 (0.50000,0.49900,0.00050,0.00050) 0.10060
10000.0 39155 (0.50495,0.49495,0.00005,0.00005) 1.00143
100000.0 161009 (0.55004,0.44995,0.00001,0.00001) 10.01045
200000.0 183245 (0.60010,0.39990,0.00000,0.00000) 20.02048
300000.0 153105 (0.65015,0.34985,0.00000,0.00000) 30.03040
400000.0 88950 (0.70020,0.29980,0.00000,0.00000) 40.04046
500000.0 59 (0.74997,0.25003,0.00000,0.00000) 160.13592
1000000.0 59 (0.74998,0.25001,0.00000,0.00000) 270.28565
10000000.0 59 (0.75000,0.25000,0.00000,0.00000) 2252.93736

Table 8.1 Iterative solution for different µ;  = 10−6

Table 8.1 shows that for increasing µ the number of iterations also increases im-
plying increased CPU-time to obtain the steady-state distribution π. Furthermore
the iteration becomes unstable for large values of µ, yielding incorrect results
µ µ 1 1
for π. Note that the solution for π is ( 2µ+2 , 2µ+2 , 2µ+2 , 2µ+2 ) and thus π tends to
( 12 , 12 ,0,0), if µ tends to infinity.
If µ = ∞ markings M3 and M4 will be left immediately by the Markov process
in Fig. 8.5. So the only significant markings for the steady-state distribution are
M1 and M2 . With this in mind we can reduce the Markov chain to that in Fig. 8.6
before determining the steady-state distribution. Solving this reduced Markov
chain iteratively will not lead to complications. Note that the size of the MC is
now smaller than the original one.
The last example illustrates that if the firing rates of transitions differ in orders of
magnitude, problems in the quantitative analysis of SPNs might occur. For this
reason SPNs were extended to the so-called Generalized Stochastic Petri nets
(GSPNs).
M1 = (1,0,0,0)
1.0
1.0

M2 = (0,1,0,0)

Figure 8.6 Reduced Markov chain

Exercise 8.1
138 8 Stochastic Petri Nets

1. Determine the infinitesimal generator Q of the MC described by the SPN of

example 8.1 and check the given steady state distribution.

2. Calculate the following quantities:

(a) mj ,∀j ∈ {1, . . . ,5}, i.e. the average number of tokens on all places at
steady state.
(b) rj ,∀j ∈ {1, . . . ,5}, i.e. the probability of firings of all transitions.

Exercise 8.2 Derive Eq.(8.2) from Eq.(8.1).

 139

9 Generalized Stochastic Petri Nets

Generalized Stochastic Petri nets (GSPNs)(M. Ajmone-Marson, G. Balbo, G.

Conte [6, 11, 7]) have two different classes of transitions: immediate transitions
and timed transitions. Once enabled, immediate transitions fire in zero time.
Timed transitions fire after a random, exponentially distributed enabling time as
in the case of SPNs.
In Fig. 9.1 the Producer-Consumer system is modelled by a GSPN. Timed transi-
tions are drawn by open bars and immediate transitions are drawn by solid bars.
Since we assume that producing and consuming an item takes much longer than
inserting and removing an item to or from the buffer, we decided to model the
activities of production and consumption by timed transitions ({t3 ,t6 }) and the
other activities by immediate transitions.
As always, several transitions may be simultaneously enabled at a marking M . If
the set of enabled transitions, denoted by ENT (M ), comprises only timed tran-
sitions, then the enabled timed transition ti ∈ ENT (M ) fires with probability
λi
P
j:tj ∈ENT (M ) λj

Producer Consumer

t1 t4

t3 t6

t2 t5

Legend:
timed transition
immediate transition

Figure 9.1 Producer-Consumer system with limited buffer capacity

140 9 Generalized Stochastic Petri Nets

exactly as before for SPNs.

If both timed and immediate transitions are enabled in the sense of Def. 5.4 (see
page 84), then by definition, only the immediate transitions are enabled in the
GSPN. Thus, in GSPNs, firing of immediate transitions has priority over the
firing of timed transitions. So

ENT (M ) := {t ∈ T |M [t > and if ∃t0 ∈ T2 : M [t0 > then t ∈ T2 }

where T2 denotes the set of immediate transitions. The definition of ENT (M )

reflects that if t is a timed transition then no immediate transition is enabled in
the sense of Def. 5.4.
When ENT (M ) has only one immediate transition than that transition fires with
probability 1.0. If ENT (M ) comprises several immediate transitions, we have
to specify the probability with which each immediate transition t ∈ ENT (M )
fires at M .
This probability distribution is called a random switch or a switching distribu-
tion. For specification of a probability distribution for concurrently enabled im-
mediate transitions, we actually need to know which transitions contribute to the
switching distribution. To get this information usually requires a pre-analysis of
the GSPN. This may be complex. Therefore it is often more suitable to define
firing weights from which the switching distribution can be determined.1 So, if
we attach firing weights wi and wj to two transitions ti and tj and only these
transitions are both enabled at some marking M , the probability of firing ti is
given by wiw+wi
j
.
The formal definition of a GSPN is as follows:
Definition 9.1 A GSPN (cf. [6, 11, 76, 7]) is a 4-tuple
GSP N = (P N,T1 ,T2 ,W ) where
• P N = (P,T,I − ,I + ,M0 ) is the underlying Place-Transition net

• T1 ⊆ T is the set of timed transitions, T1 6= ∅,

• T2 ⊂ T denotes the set of immediate transitions,

T1 ∩ T2 = ∅, T = T1 ∪ T2

• W = (w1 , . . . ,w|T | ) is an array whose entry2 wi ∈ R+

– is a (possibly marking dependent) rate of a negative exponential dis-

tribution specifying the firing delay, when transition ti is a timed tran-
sition, i.e. ti ∈ T1 or
1
In practice the specification of these firing weights often requires a form of pre-analysis, e.g.
by calculating so-called extended conflict sets [76]. Also confusion situations have to be con-
sidered. Confusion occurs if a conflict situation is resolved or produced by a third transition.
E.g., after firing t1 t2 in the Petri net of Fig. 8.1 we reach a marking in which t4 is enabled and
is not in conflict with t5 , since p5 is empty. But firing t3 leads to this conflict situation. When
defining firing weights we have to take this influence of “third party” transitions into account.
2
R+ denotes the set of positive real numbers.
9.1 Quantitative Analysis of GSPNs 141

– is a (possibly marking dependent) firing weight, when transition ti is

an immediate transition, i.e. ti ∈ T2 .

Note that if T2 = ∅ then the definition of a GSPN coincides with the definition
of an SPN, i.e. SPNs ⊂ GSPNs.

Example 9.1 The definition of the GSPN which models the Producer-Consumer
system is as follows: GSP N = (P N,T1 ,T2 ,W ) where

• P N is the underlying Place-Transition net. We omit the definition of that

Petri net, since one can obviously derive it from Fig. 9.1.

• T1 = {t3 ,t6 },T2 = {t1 ,t2 ,t4 ,t5 }

• W = (w1 ,w2 ,w3 ,w4 ,w5 ,w6 ), wi ∈ R+ .

How do we analyse a GSPN? Note that a GSPN does not directly describe a
continuous-time Markov process, since immediate transitions fire in zero time.
So the sojourn time in markings which enable immediate transitions is no longer
exponentially distributed. Such markings we will call vanishing, because if a
random observer looks at the stochastic process of a GSPN, he will never observe
such states, although the stochastic process sometimes visits them.
On the other hand, markings which enable timed transitions only will be ob-
served, since the stochastic process sojourns in such markings for an exponen-
tially distributed length of time. So these markings are not left immediately.
Therefore we will call them tangible. In the following we will refer to a marking
also as a state.

9.1 Quantitative Analysis of GSPNs

Throughout this section we will assume that the GSPN has a finite reachability
set. In order to analyse a GSPN we analyse the embedded Markov chain of the
corresponding stochastic process (cf. Sec. 2.2.3). Note that from the definition
of GSPNs we know that the probability of changing from one marking to an-
other is independent of the time spent in a marking. Thus a GSPN describes a
semi-Markov process. Imagine that the firing of transitions take place at points
d0 ,d1 , . . . ,dn , . . . in time. If a timed transition fires at, say, dk then dk+1 −dk > 0,
since timed transitions constitute tangible states in which the process sojourns for
a certain time interval. If an immediate transition is responsible for a state change
at time dk in the corresponding stochastic process, then we have dk = dk+1 .
So we can think of our stochastic process as a discrete-time Markov chain where
some points coincide in time. To specify this embedded, discrete-time Markov
chain, we have to determine the transition probability from marking or state Mk
to Mr . These probabilities are given by
142 9 Generalized Stochastic Petri Nets
P
i:ti ∈{t∈T |Mk [t>Mr }∩ENT (Mk ) wi
P [Mk → Mr ] = P (9.1)
j:tj ∈ENT (Mk ) wj

If {t ∈ T |Mk [t > Mr } = ∅ we have P [Mk → Mr ] = 0. Normally {t ∈ T |Mk [t > Mr }

contains at most one element, since otherwise “redundant” transitions t,t0 ∈ T
causing the same state change would be present, i.e. I − (p,t) = I − (p,t0 ) and
I + (p,t) = I + (p,t0 ),∀p ∈ P .
Note that these probabilities are independent of the time spent in marking Mk ,
which is why we are allowed to consider an embedded Markov chain. From the
embedded Markov chain, we can write the transition probability matrix as an
augmented matrix with respect to the sets of vanishing and tangible states:
 
C D
P := (9.2)
E F

where

C = (cij ), cij = P [Mi → Mj ; Mi ∈ V̂ ,Mj ∈ V̂ ],

D = (dij ), dij = P [Mi → Mj ; Mi ∈ V̂ ,Mj ∈ T̂ ],
E = (eij ), eij = P [Mi → Mj ; Mi ∈ T̂ ,Mj ∈ V̂ ],
F = (fij ), fij = P [Mi → Mj ; Mi ∈ T̂ ,Mj ∈ T̂ ].

and T̂ denotes the set of tangible states and V̂ the set of vanishing states, T̂ ∩
V̂ = ∅. C describes the transition probabilities between vanishing states and F
specifies the probabilities between tangible states.
The steady state distribution π̃ of the embedded Markov chain, provided it exists,
is given as usual by
X
π̃P = π̃ and π̃i = 1 (9.3)
Mi ∈T̂ ∪V̂

From this steady state distribution we can calculate the steady state distribution π
of the original stochastic process by weighting the probability π̃j with the portion
of time the process spends in marking Mj .
Obviously πj is 0 if Mj is a vanishing marking. For tangible markings πj can be
calculated as follows:
vsj := π̃π̃sj determines the mean number of visits (cf. Eq. (2.20) on page 32) to
state Ms between two consecutive visits to state Mj . Since ( k:tk ∈ENT (Ms ) wk )−1
P

is the mean time the process will spend in state Ms , the mean time between two
consecutive visits to Mj (mean cycle time) is given by:

1 X
wk )−1
X
π̃s × (
π̃j
Ms ∈T̂ k:t k ∈ENT (Ms )

On the other hand, the stochastic process will spend

9.1 Quantitative Analysis of GSPNs 143

t4

p2 M0 = (1,0,0)
t4 t5
t1 t1 t2
p1 t3

t2 p3 M1 = (0,1,0) M2 = (0,0,1)
t3
t5

Figure 9.2 A GSPN and its reachability graph

wk )−1
X
(
k:tk ∈ENT (Mj )

time units on the average in state Mj .

As stated before, the steady state probability πj is given by the fraction of time
the process spends in marking Mj (see also Eq. (2.35) on page 45). This portion
is characterised by the mean time spend in marking Mj divided by the mean cy-
cle time. Thus the steady state distribution π of the stochastic process described
by a GSPN is given by

wk )−1
X


 π̃j × (

 k:tk ∈ENT (Mj )
if Mj ∈ T̂


wk )−1
 X X
π̃s × (

πj = (9.4)

 Ms ∈T̂ k:tk ∈ENT (Ms )





if Mj ∈ V̂

 0

Example 9.2 Consider the GSPN displayed in Fig. 9.2 and define w1 = w2 = 1
and w3 = w4 = w5 = 3. The only vanishing marking is the initial marking M0 .
Matrix P is given by
 
1 1
  0 2 2
C D
P = = 1 0 0 
 
E F 1 1
2 2 0
P
and solving the global balance equations π̃P = π̃; i π̃i = 1 yields π̃ =
( 49 , 39 , 29 ). Thus the mean cycle time for tangible marking M1 is 1× 13 + 32 × 3+3
1
=
4 3 1 1 2
9 and for marking M2 this is 2 × 3 + 1 × 6 = 3 . So the steady state distribution
is given by
1 1
3 1
π0 = 0, π1 = 34 = , π2 = 62 = .
9
4 3
4
144 9 Generalized Stochastic Petri Nets

The main task in the calculation of the steady state distribution described above,
is to solve (9.3). The complexity of this calculation is determined by the number
of reachable markings, including the vanishing markings. From the description
of immediate transitions we know that the steady state probability of being in a
vanishing state is defined to be zero.
We can eliminate vanishing markings from the embedded Markov chain before
calculating the steady state distribution in a way similar to that in the example
of the SPN in Ch. 8, where the reachability set was reduced to two markings for
µ → ∞. This will lead to a reduction of the number of global balance equations
and a more efficient analysis.
To obtain a reduction of the state space we have to determine the transition prob-
abilities between tangible states. On leaving a tangible state Mi , the process
might visit one or more vanishing states, Mr ∈ V̂ before finally reaching a tan-
gible marking, say Mj . Thus the reduced embedded Markov chain is specified
by transition matrix P 0 :
P 0 := [p0ij ], p0ij := fij + eir P [Mr −→∗ Mj ],
X
(9.5)
Mr ∈V̂

where P [Mr −→∗ Mj ] := P[the stochastic process starts in state Mr

and reaches the tangible state Mj ,
and all states reached in between
are vanishing states], Mr ∈ V̂ ,Mj ∈ T̂ .

P [Mr −→∗ Mj ] can be obtained as follows. Define

h
P [Mr −→ Mj ] := P[the stochastic process starts in state Mr
and reaches the tangible state Mj after
exactly h steps, and all states reached in between
are vanishing states], Mr ∈ V̂ ,Mj ∈ T̂ .

This yields
∞
h
P [Mr −→∗ Mj ] =
X
P [Mr −→ Mj ].
h=1
From our discussions on Markov chains we know that, given the matrix P char-
acterising the single step probabilities between all states, the h-step probabilities
are given by P h . From (9.2) we know that the single step probabilities out of van-
ishing states are described by matrices C and D. Thus the h-step probabilities
h
P [Mr −→ Mj ] are given by C h−1 × D.
Note that D specifies the single step probabilities between vanishing and tangible
h
states, and that Mj is a tangible state. So the probability P [Mr −→ Mj ] is simply
given by the corresponding element of matrix C h−1 × D, namely (C h−1 × D)rj .
Referring to Eq. (9.5) we can now write matrix P 0 as follows
∞ ∞
P0 = F + E ×
X X
C h × D provided C h exists. (9.6)
h=0 h=0
9.1 Quantitative Analysis of GSPNs 145

The steady state distribution π̃ of the reduced embedded Markov chain is given
by
π̃ = π̃ × P 0 ;
X
π̃s = 1, (9.7)
Ms ∈T̂

if a unique solution for π̃ exists. The steady state distribution π of the Markovian
process is given by Eq. (9.4) for tangible markings.
A prerequisite for the existence of P 0 is the existence of ∞ h
P
h=0 C . How can we
characterise the existence of this limit?
P|V̂ |
Define ui := 1 − j=1 cij . ui > 0 if the sum in row i of matrix C is less
then 1 and ui = 0 if it is 1. With that we can separate the set of vanishing states
into two subsets J0 := {Mi |ui = 0} and J1 := {Mi |ui > 0}. If the stochastic
process is in a state Mi ∈ J0 , it can only transit to another vanishing state, being
in a state Mj ∈ J1 there is a positive probability to reach a tangible state. The
existence of the limit in Eq. (9.6) can be analysed by inspecting the structure of
the reachability set with respect to vanishing states. If the process can always
reach tangible states, then this limit exists.
The reader should refer to Sec. 2.1.2 on absorbing Markov chains and transient
behaviour where the matrix Q of transition probabilities between transient states
correspond to matrix C of the transition probabilities between vanishing states
in this section. First we define:

Definition 9.2 (A Trap.) (cf. [110]) C has no trap :⇐⇒

∀Mi ∈ J0 : ∃Mj ∈ J1 reachable from Mi ; otherwise C has a trap.

Theorem 9.1 (cf. [110]) C has no trap iff (I − C)−1 =

P∞ h
h=0 C exists.

Because C consists of the transition probabilities between vanishing states, we

speak of a timeless trap. If there is no timeless trap, we can determine matrix P 0 .
Note that a timeless trap can only exist if the GSPN is not live, since all timed
transitions are not live.
Another condition for the existence of a steady state distribution is the existence
of a unique solution of Eq. (9.7), i.e. in case of a finite reachability set, the GSPN
has to contain home states (cf. Sec. 5.3).

Example 9.3 Consider the GSPN of Fig. 9.3 with wi = i. The reachability set
comprises the following markings:
M0 = (1,1,1,0,1),M1 = (0,0,2,1,0),M2 = (0,1,2,0,0),
M3 = (1,0,1,1,1),M4 = (2,1,0,0,2),M5 = (2,0,0,1,2),
M6 = (0,0,2,0,1),M7 = (1,0,1,0,2),M8 = (2,0,0,0,3)
and the corresponding reachability graph is shown in Fig. 9.4.
The markings M0 ,M1 ,M2 ,M3 ,M4 ,M5 are vanishing and the markings M6 ,M7 ,M8
are tangible. The matrix P of the embedded Markov chain is
146 9 Generalized Stochastic Petri Nets

p1 p2

t1 t2

p3 p4 t4

t3
t5

p5

Figure 9.3 wi = i in the example

M0
t1 t2

t3 M6 M3 t4
t5
M1 t4 t3

t2 M2 M7

t5

M8
t1 t3
t4 M5

M4 t2

Figure 9.4 Reachability graph of the GSPN

2 1
 
0 0 0 3 0 0 3 0 0

 0 0 0 0 0 0 1 0 0 

0 1 0 0 0 0 0 0 0
 
 
0 0 0 0 0 0 0 1 0
 
   
C D  2 1

P = = 0 0 0 0 0 3 0 3 0 
E F  

 0 0 0 0 0 0 0 0 1 

4 5

 0 0 9 0 0 0 0 9 0 

 4 5 
 9 0 0 0 0 0 0 0 9

0 0 0 0 1 0 0 0 0

Since matrix
9.1 Quantitative Analysis of GSPNs 147
 2 
1 0 0 3 0 0

 0 1 0 0 0 0 

−1

0 1 1 0 0 0 
(I − C) = 
 
0 0 0 1 0 0

 
 2 
 0 0 0 0 1 3

0 0 0 0 0 1
we get
4 5
 
9 9 0
 
 
0 −1  4 8 5
P = F + E × (I − C) × D =  27

 27 9 

 
1 2
0 3 3
Solving
8
π̃P 0
X
= π̃ and π̃i = 1
i=6

1
yields π̃6 = 11 ,π̃7 = 15 25
44 ,π̃8 = 44 . Thus the mean number of visits is
4 4
v66 = 1,v67 = 15 ,v68 = 25 ,v76 = 15 3 25 5
4 ,v77 = 1,v78 = 5 ,v86 = 4 ,v87 = 3 ,v88 =
1.
The sojourn time ts (Mi ) for tangible marking Mi is given by
ts (M6 ) = ts (M7 ) = (4 + 5)−1 = 19 ,ts (M8 ) = 14 .
301
Let X̂ = π̃6 × ts (M6 ) + π̃7 × ts (M7 ) + π̃8 × ts (M8 ) = 1584 ,
then we get the following mean cycle times tc (Mi ) for tangible marking Mi :
tc (M6 ) = π̃16 X̂ = 301 1 301 1 301
144 ,tc (M7 ) = π̃7 X̂ = 540 ,tc (M8 ) = π̃8 X̂ = 900 ,
which gives the steady state distribution for tangible states:
π6 = ttsc (M
(M6 )
6)
16
= 301 , 60
π7 = 301 , 225
π8 = 301 .

The steady state distribution π of the GSPN can be employed for the calculation
of performance figures as shown in Ch. 8. Only the throughput at an immediate
transitions has to be calculated differently, since immediate transitions are only
enabled in vanishing states.
Let r = (ri ) be a vector, which determines for a vanishing state Mi the rate at
which the state is entered and left immediately at steady state. Since we know the
steady state distribution π of tangible states, we can calculate r as follows. Let
Ẽ be the matrix specifying the rates for a specific state change leaving a tangible
state Mi and entering a vanishing state Mj , i.e.
X
Ẽ = (ẽij ), ẽij = wk ; Mi ∈ T̂ ,Mj ∈ V̂
k:tk ∈{t∈T |Mi [t>Mj }∩ENT (Mi )

Then
r = π × Ẽ × (I − C)−1
The throughput at an immediate transition tj ∈ T2 is then given by
148 9 Generalized Stochastic Petri Nets
X wj
dj = ri P
k:tk ∈ENT (Mi ) wk
Mi ∈ENj ∩V̂

where ENj = {M ∈ R(P N )|M [tj >}.

Example 9.4 For the GSPN of Fig. 9.3 (cf. Ex. 9.3) we have
0 0 4 0 0 0
 

Ẽ =  4 0 0 0 0 0 
0 0 0 0 4 0
and thus
1
r = π × Ẽ × (I − C)−1 = (240,64,64,160,900,600)
301
The throughputs at transition t1 and t2 are now
240 1 900 1 380
d1 = × + × =
301 3 301 3 301
240 2 64 900 2 824
d2 = × + ×1+ × =
301 3 301 301 3 301
The throughput at the timed transition t5 can be calculated as for SPNs which
yields
380
d5 = (π6 + π7 ) × 5 =
301
and the throughput at transition t4 is λ4 = 4, since it is enabled in all tangible
markings.

9.2 Qualitative Analysis of GSPNs

In the previous section we saw that qualitative properties, like the existence of
home states, are essential preconditions for the existence of a steady state distri-
bution of the stochastic process represented by the GSPN.
For the qualitative analysis of a GSPN we want to exploit the underlying Place-
Transition net of the GSPN and use the algorithms presented in Sec. 5.4.
Our interest is therefore in the combined qualitative and quantitative analysis il-
lustrated in Fig. 9.5. First of all, while neglecting time, certain qualitative aspects,
such as boundedness and liveness of the Place-Transition net are fully analysed.
Then, if the Place-Transition net satisfies all the required qualitative properties,
it may be worthwhile to do a quantitative analysis. For this procedure to apply,
all qualitative features of the Place-Transition net have to remain valid after the
introduction of time. Unfortunately, this is not always the case.
Incorporating time, particularly introducing a priority relation on transitions,
changes the properties of the Place-Transition net significantly [17, 19, 20]. In
this section we investigate this further.
9.2 Qualitative Analysis of GSPNs 149

time-augmented Petri net

=
timing constraints + Petri net

?
qualitative analysis

?
- +

?
quantitative analysis

Figure 9.5 Principal analysis procedure for time-augmented Petri nets

In the following discussion we will denote the underlying Place-Transition net

of a GSPN by PN. Since firing of immediate transitions has priority on firing of
timed transitions the enabling rule in GSPNs is simply given by

Definition 9.3 A transition t ∈ T of a GSPN is enabled at M , denoted by

M [t >GSP N , iff M (p) ≥ I − (p,t),∀p ∈ P and
(∃t0 ∈ T2 : M (p) ≥ I − (p,t0 ),∀p ∈ P =⇒ t ∈ T2 ).

In other words, a transition is enabled iff the usual enabling rule for Place-Tran-
sition nets holds and if there exists another enabled immediate transition t0 , then
t must also be an immediate transition, since otherwise the firing of t0 would
have priority over the firing of t. Or to put it differently, t ∈ T of a GSPN is
enabled at M iff t ∈ ENT (M ), with ENT (M ) defined on page 140.
With respect to the enabling rule in GSPNs, all definitions concerning the dy-
namic behaviour of GSPNs are similar to those concerning Place-Transition
nets (cf. Def. 5.4 and 5.5). We will use subscripts P N and GSP N to distin-
guish between the same notions for the underlying Place-Transition net, e.g.
M [t >P N ,M [t >GSP N ,M [σ >P N , M →∗GSP N M 0 or R(GSP N,M0 )
etc.
We first of all investigate timeless traps in the context of the state space of a
GSPN. A timeless trap means that the GSPN can reach a marking whence only
immediate transitions can fire in the future.

Definition 9.4 A GSPN has a timeless trap, iff ∃M ∈ R(GSP N,M0 ) :

∀k ∈ N : ∃σ ∈ T ∗ ,|σ| ≥ k such that M [σ >GSP N and
∀σ̃ ∈ T ∗ : M [σ̃ >GSP N =⇒ σ̃ ∈ T2∗ .
150 9 Generalized Stochastic Petri Nets

t7

t4 t8
t1

t3 t6

t2 t5

Figure 9.6 A GSPN with a timeless trap

where |σ| denotes the length of σ. Since T is a finite set, the existence of a time-
less trap implies that some immediate transitions fire infinitely often. Fig. 9.6
depicts an example of a timeless trap which is caused by the firings of the imme-
diate transitions t7 and t8 . Note that this definition of a timeless trap corresponds
directly with that for a trap of the matrix C in Def. 9.2. Timeless traps are avoided
by the following condition.

Definition 9.5 Condition NoTT. 3

Condition NoTT :⇐⇒ ∀ T̃ ⊆ T2 : T̃ 6= ∅ =⇒ •T̃ 6= T̃ •.

Condition NoTT prevents the existence of a strongly connected subnet generated

by immediate transitions.
Let GSPN be a Generalized Stochastic Petri net with a finite reachability set.
Then condition NoTT implies that the GSPN has no timeless trap.

Theorem 9.2 Let GSPN have a finite reachability set.

Condition NoTT =⇒ GSPN has no timeless trap.

Proof.: Assume there is a timeless trap and let T̃ be the set of immediate tran-
sitions which fire infinitely often. Let P̃ be the set of places which
 are either

an input place for T̃ or an output place, but not both, i.e. P̃ = •T̃ \ T̃ • ∪
 
T̃ • \ • T̃ . If T̃ 6= ∅ we have P̃ 6= ∅ since otherwise, •T̃ = T̃ • contradicting
condition NoTT. There are two cases to consider

1. ∃ p ∈ •T̃ \ T̃ •.
Since p 6∈ T̃ • not all transitions of T̃ can fire infinitely often, contradicting
our assumption.
3
NoTT stands for No Timeless Trap.
9.2 Qualitative Analysis of GSPNs 151

Figure 9.7 GSPN not satisfying condition NoTT

2. ∃ p ∈ T̃ • \ • T̃ .
Since p 6∈ •T̃ and all t ∈ T̃ fire infinitely often, the GSPN has an infinite
reachability set, since the number of tokens on p is not bounded, which is
again a contradiction. t
u

Condition NoTT is not necessary to avoid timeless traps, as illustrated by the

GSPN in Fig. 9.7. From the proof of Theorem 9.2 it follows that the next theorem
holds.

Theorem 9.3 Condition NoTT =⇒ ∃h0 ∈ N : C h = 0, ∀h ≥ h0 .

where C is the matrix given on page 142. Remember that C h describes the h-step
transition probabilities between vanishing states. From the proof of Th. 9.2 we
know that the number of firings of immediate transitions is always bounded. The
theorem essentially tells us that when condition NoTT applies, the number of
steps between vanishing states do not exceed limit h0 . This implies that there are
no loops between vanishing states in the state space of the GSPN. This property
is particularly important when eliminating vanishing states while generating the
state space. It is only necessary to multiply the probabilities along every path,
starting at a particular tangible state Mi and leading to a particular tangible state
Mj , and sum all these path probabilities to determine the single step probability
between these two tangible states.
Condition NoTT can be verified very easily with the algorithm in Fig. 9.8.
Incorporating time also changes other properties of a Place-Transition net. E.g.
an unbounded Place-Transition net may yield a bounded GSPN (see Fig. 9.9).
Fortunately the reverse does not present a similar problem.

Theorem 9.4 PN bounded =⇒ GSPN bounded.

Proof. Obvious, since R(GSP N,M0 ) ⊆ R(P N,M0 ). t

u

Other properties unveil greater problems. E.g., there is no relationship between

the liveness of a GSPN and its underlying Place-Transition net PN. Fig. 9.10
shows a non-live GSPN, whose PN is live. The immediate transition t∗ is only
enabled at the initial marking which is the only marking where places p1 and p2
152 9 Generalized Stochastic Petri Nets

TH := T2
while •TH 6= TH • do
begin
PH := (•TH \ TH •) ∪ (TH • \ • TH )
∀t ∈ TH do
if (•t ∩ PH ) ∪ (t • ∩ PH ) 6= ∅
then TH := TH \ {t}
end
if TH = ∅
then Condition NoTT is satisfied
else Condition NoTT is not satisfied

Figure 9.8 Algorithm for checking condition NoTT

Figure 9.9 GSPN bounded, PN not bounded

p1 p2
t∗

Figure 9.10 PN live, GSPN not live

9.2 Qualitative Analysis of GSPNs 153

M0 M0
t∗ t∗

Figure 9.11 Structure of reachability graphs: PN (left); GSPN (right)

are marked simultaneously. The reachability graphs of PN and GSPN are shown
in Fig. 9.11.
Fig. 9.12, on the other hand, depicts a live GSPN with a non-live underlying Petri
net. After firing t2 and t3 the Place-Transition net is dead, but the firing sequence
t2 t3 can not occur in the GSPN, since the firing of t1 has priority on the firing of
t2 . The only conclusion we can establish for a GSPN whose underlying Place-
Transition net is live is the following:

Lemma 9.1 Let GSPN be a Generalized Stochastic Petri net whose underly-
ing Place-Transition net is live. Then ∀M ∈ R(GSP N,M0 ) : ∃t ∈ T :
M [t >GSP N .

Proof. Since the enabling rule of GSPNs is a restriction of the enabling rule of
Place-Transition nets, we know R(GSP N,M0 ) ⊆ R(P N,M0 ). Liveness of the
underlying Place-Transition net implies that ∀M ∈ R(GSP N,M0 ) : ∃t ∈ T :
M [t >. If t ∈ T2 then also M [t >GSP N holds and our proof is complete. If
t ∈ T1 then ¬M [t >GSP N will only hold if there is an immediate transition
which prevents the enabling of t, i.e. ∃t0 ∈ T2 : M [t0 >GSP N so that t0 is
enabled at M according to the enabling rules of GSPNs. t
u

So in such GSPNs there is always some enabled transition, although there might
be no live transition (see Fig. 9.13). Lemma 9.1 only implies that total deadlocks
cannot occur.
Furthermore, although the PN contains home states, this may not be true for the
GSPN as shown in Fig. 9.14. The reachability graph of the GSPN, see Fig. 9.15,
now contains two terminal strongly connected components (cf. Def. 5.7 and
Th. 5.4 on page 94).
All these examples illustrate that the analysis procedure in Fig. 9.5 has to be
modified or possibly completely changed to usefully combine qualitative and
quantitative analysis. What possibilities do we have to cope with this basic prob-
lem? One way is to modify existing Petri net algorithms to render them suitable
for the analysis of time-augmented nets. This may not be so useful, as Petri net
theory and existing software tools could then no longer be used directly. Further-
more, all future research results in the Petri net area would have to be adapted to
time-augmented Petri nets.
154 9 Generalized Stochastic Petri Nets

t1 t3

t2

Figure 9.12 GSPN live, PN not live

Figure 9.13 GSPN with no live transitions

Another way is to determine suitable restrictions for an integration of time, such

that the results of a qualitative analysis remain valid for a quantitative analysis
as well. The latter idea leads to the analysis procedure illustrated in Fig. 9.16.
The great benefit of this is that the standard theory of Petri nets remains unaf-
fected. Finding such restrictions for general net structures is difficult. We restrict
ourselves to EFC-nets. Another reason for this restriction is that this class of
Petri nets has been studied exhaustively, and many qualitative properties can be
characterised by conditions which can be tested efficiently (cf. Sec. 5.4.3).
9.2 Qualitative Analysis of GSPNs 155

t1 t2
p1

p2
p5 p8

p3
t3 t4

p6 p9

p4
t5 t6

p7 p10

Figure 9.14 GSPN with no home states

t3 t4 t3 t4
t1 t2 t1 t2
t5 t6
t2 t1 t2 t1
M0 M0
t2 t5 t6 t5 t6
t1
t6 t5
t4 t3 t4 t3

t1 t2 t1 t2
t6 t5
t6 t5

Figure 9.15 Reachability graphs: PN (left); GSPN (right)

9.2.1 Qualitative Analysis of EFC-GSPNs

Definition 9.6 A GSPN is an EFC-GSPN iff its underlying Place-Transition net

is an EFC-net.

Since EFC-nets have the structures shown in Fig. 5.22 on page 105, it seems
obvious to insist that conflicts may only occur between transitions of the same
kind. When this condition, which we shall call EQUAL-Conflict, holds for an
EFC-GSPN, all positive properties of the underlying PN remain valid.
156 9 Generalized Stochastic Petri Nets

time-augmented Petri net

=
timing constraints + Petri net

?
qualitative analysis

? 
?
additional restrictions - +

?
quantitative analysis

Figure 9.16 Modified analysis procedure for time-augmented Petri nets

Definition 9.7 A GSP N = (P N,T1 ,T2 ,W ) satisfies condition EQUAL-Conflict,

iff ∀t,t0 ∈ T : •t ∩ •t0 6= ∅ =⇒ {t,t0 } ⊆ T1 or {t,t0 } ⊆ T2 .

We will show that this condition ensures the absence of timeless traps and that
liveness and the existence of home states extend to the GSPN.

Theorem 9.5 ([17, 19, 20]) If we are given an EFC-GSPN whose underlying
Place-Transition net is live and bounded, the following holds:

1. Condition EQUAL-Conflict =⇒ GSPN has no timeless trap.

2. Condition EQUAL-Conflict ⇐⇒ GSPN live.

3. Condition EQUAL-Conflict =⇒ GSPN has home states.

Proof. Since the proof of part 3 is lengthy we will only prove the first two parts.
Proof of part 1: Assume the GSPN has a timeless trap, i.e.
∃M ∈ R(GSP N,M0 ) :
∀k ∈ N : ∃σ ∈ T ∗ ,|σ| ≥ k such that M [σ >GSP N and
∀σ̃ ∈ T ∗ : M [σ̃ >GSP N =⇒ σ̃ ∈ T2∗ .
Since PN is bounded Th. 9.4 ensures boundedness of the GSPN. Thus R(GSP N,M0 )
is finite and there exists a terminal strongly connected component (cf. Def. 5.7)
CM ⊆ R(GSP N,M ). Define T̃ := {t ∈ T |∃M 0 ∈ CM : M 0 [t >GSP N } and
P̃ := •T̃ .
9.2 Qualitative Analysis of GSPNs 157

T̃ ,P̃ T \ T̃ ,P \ P̃

ti tj
a: 6 3

pr j

pk j j pl
>

b: tr
6


Figure 9.17 Illustration of part 1 of the proof

T̃ can be referred to as the set of live transitions with regard to CM . Because of

our assumption T̃ ⊆ T2 holds. Since the Place-Transition net is live and bounded,
it is strongly connected according to Th. 5.1. So one of the following two cases
holds (see Fig. 9.17):

a: ∃pr ∈ P̃ : pr • ⊇ {ti ,tj } and ti ∈ T̃ ,tj ∈ T \ T̃ .

The assumption of a timeless trap yields ti ∈ T2 . If M 0 [ti >GSP N for some
marking M 0 ∈ CM , the EFC-net structure and condition EQUAL-Conflict
imply M 0 [tj >GSP N contradicting tj ∈ / T̃ .

b: ∃tr ∈ T̃ : tr • ⊇ {pk ,pl } and pk ∈ P̃ ,pl ∈ P \ P̃ .

Since tr ∈ T̃ , tr fires infinitely often with regard to CM and our assump-
tion of a timeless trap. Because pl ∈ / P̃ = •T̃ , place pl is not bounded,
contradicting the boundedness of the GSPN.

Thus our assumption is not valid, which completes the proof of part 1.

Proof of part 2:

if: Assume the GSPN is not live.

Then ∃t ∈ T : ∃M ∈ R(GSP N,M0 ) : ∀M 0 ∈ R(GSP N,M ) : ¬M 0 [t >GSP N .
Since the net is an EFC-GSPN there exists a place p ∈ •t which is empty
in all markings of R(GSP N,M ). Note that from part 1 we know that the
GSPN has no timeless trap. This further implies that all input transitions of
that place are not live with regard to R(GSP N,M ), i.e.

∀t ∈ •p : ¬M 0 [t >GSP N ,∀M 0 ∈ R(GSP N,M ). (9.8)

Define Tdead := {t ∈ T |¬M 0 [t >GSP N ,∀M 0 ∈ R(GSP N,M )} and

Pempty := {p ∈ P |M 0 (p) = 0,∀M 0 ∈ R(GSP N,M )}.
Because of the EFC-net structure we have Pempty • = Tdead and further-
more •Pempty ⊆ Tdead from (9.8). Thus we have •Pempty ⊆ Pempty • and
Pempty is an empty deadlock, which contradicts the dt-property, since the
Place-Transition net is a live EFC-net (cf. theorem 5.16 on page 106).
158 9 Generalized Stochastic Petri Nets

only if: Assume that condition EQUAL-Conflict does not hold,

i.e. ∃t,t0 ∈ T : •t ∩ •t0 6= ∅ and {t,t0 } 6⊆ T1 and {t,t0 } 6⊆ T2 .
Since the GSPN is an EFC-GSPN, obviously t and t0 can not be both live,
contradicting the liveness of the GSPN. t
u

Note that in part 3 of Th. 9.5, because of the EFC-net structure, boundedness and
liveness of the Place-Transition net implies the existence of home states for the
Place-Transition net as well. [19, 20] shows that Th. 9.5 also holds for EC-nets
(cf. page 125) and extends the results to general net structures and multiple levels
of priorities for immediate transitions.

9.3 Further Remarks on GSPNs

The difficulties with the qualitative analysis of GSPNs apply to nearly any kind of
time-augmented Petri net. This is an intrinsic problem which cannot be avoided.
Since immediate transitions have priority over timed transitions, GSPNs are
equivalent to Turing machines and thus several properties, such as liveness, are
undecidable for general GSPNs (cf. [132]).
There are yet further disadvantages concerning the modelling of systems using
GSPNs.
GSPNs, like Place-Transition nets, do not distinguish between individual to-
kens so that the graphical representation very soon becomes complex. Defining
a coloured version of GSPNs (Coloured GSPNs, CGSPNs) can be done eas-
ily by folding places and transitions of the same kind. The firing rates or firing
weights are then associated with the specific colour of a transition. Analysis of
such a CGSPN can, e.g., be done by unfolding the net and analysing the GSPN
as usual.

Definition 9.8 A Coloured GSPN (CGSPN) is a 4-tuple

CGSP N = (CP N,T1 ,T2 ,W ) where

• CP N = (P,T,C,I − ,I + ,M0 ) is the underlying Coloured Petri net.

• T1 ⊆ T is the set of timed transitions, T1 6= ∅,

• T2 ⊂ T is the set of immediate transitions, T1 ∩ T2 = ∅,

T = T1 ∪ T2 ,

• W = (w1 , . . . ,w|T | ) is an array whose entry wi is a function of

[C(ti ) 7→ R+ ] such that ∀c ∈ C(ti ) : wi (c) ∈ R+

– is a (possibly marking dependent) rate of a negative exponential dis-

tribution specifying the firing delay with respect to colour c, if ti ∈ T1
or
9.3 Further Remarks on GSPNs 159

colour 2

t1 p1
t2

p2

colour 1
p3 t3

Figure 9.18 2 class -/M/1-”preemptive-resume priority” queue modelled by a GSPN;

w2 = w3

– is a (possibly marking dependent) firing weight with respect to colour

c, if ti ∈ T2 .

A more severe disadvantage of GSPNs is the difficulties which arise if one mod-
els queues, which we discussed in Ch. 3. Modelling scheduling strategies is par-
ticularly difficult.
Fig 9.18 illustrates this point. The GSPN models a queue which serves two
colours of tokens according to the pre-empt resume priority strategy (cf. Sec. 3.8).
A newly arriving token with higher priority (colour 1) pre-empts the service of
the token being served at that moment. The pre-empted token will continue his
task, at the point of interruption, after completion of service of all tokens with
higher priority. Transition t1 is enabled if pre-emption is to occur. Firing of t1
pre-empts tokens of colour 2. The markings of p1 and p2 keep track of this event.
This is the reason why the service of a token of colour 1 is modelled by two
timed transitions, t2 and t3 , with identical parameters. Note that this model of a
pre-emptive priority scheduling strategy with resumption is only correct if expo-
nentially distributed service times are assumed. Other service time distributions
require more complex models.
Also simple scheduling strategies, like FCFS, might not be easy to represent
using plain GSPN elements. Consider a queue where 2 colours of tokens arrive
according to exponential distributed interarrival times and whose service times
are exponentially distributed. If the scheduling strategy is FCFS, i.e. our system
is an M/M/1-FCFS queue, we have to encode the colour of the token in each
position of the queue. Assume that an upper bound for the number of tokens in
the queue is given, e.g. 3, then the GSPN in Fig. 9.19 would model the queue
accurately.
Transitions t1 and t2 model the arrival of a token of either colour and transitions
t3 and t4 model the service of the token in front of the queue. The places pi1
and pi2 represent position i of the queue and the place pi0 ensures that this posi-
tion is occupied by at most one token. Since entering a position is modelled by
160 9 Generalized Stochastic Petri Nets

p31 p21 p11

t1 t3 colour a

p30 p20 p10

t2 t4 colour b
p32 p22 p12

Figure 9.19 GSPN model of a FCFS queue

immediate transitions, a token entering position 3 of the queue will immediately

advance to position 2 and 1, if they are free.
This way of modelling a FCFS queue with GSPN elements works fine if an up-
per bound for the number of tokens is known a priori. Suppose, however, that
we want to perform several experiments with different initial markings. This will
necessitate a modification of the GSPN model of the queue for each experiment.
If there is no upper bound known beforehand it is even more difficult. Addition-
ally, if the service time of a queue is specified by, e.g. a Coxian distribution, it
becomes just about impossible to model such a queue with a GSPN.
In an attempt to resolve these difficulties, GSPNs were extended to Queueing
Petri nets (QPNs).

Exercise 9.1

1. Consider a GSPN with a finite reachability set. Convince yourself that for
such a GSPN the existence of home states is necessary and sufficient for
the existence of the steady state distribution, provided there are no timeless
traps.

2. Determine the steady state probability of the Producer-Consumer exam-

ple 9.1 for w1 = 12.3,w2 = 0.56,w3 = 2.0,w4 = w5 = 32.1,w6 = 2.0.

Exercise 9.2 Give the formal counterparts of Def. 5.4 and 5.5 for GSPNs.

Exercise 9.3

• Apply the algorithm given in Fig. 9.8 to the GSPN of Fig. 9.6.

• Prove that the GSPN given in Fig. 9.7 does not have a timeless trap.

• Design an algorithm for the construction of the reduced embedded MC

where vanishing states are eliminated on the fly.

Exercise 9.4

1. Prove that the GSPN of Fig. 9.13 has no live transitions.

9.3 Further Remarks on GSPNs 161

p1
t1
t3
t2 p2

Figure 9.20 An unbounded GSPN model

2. Prove the statements in the text concerning liveness of the GSPNs in Fig. 9.10
and 9.12.

Exercise 9.5 Show that the GSPN of Fig. 9.14 has no home states and that for
the underlying Place-Transition net home states do exist.

Exercise 9.6 Prove or disprove each conjecture:

Given a GSPN satisfying condition EQUAL-Conflict. Then

1. PN live =⇒ GSPN live.

2. GSPN live =⇒ PN live.

3. PN has home states =⇒ GSPN has home states.

Exercise 9.7

1. Design a GSPN model for a queue which serves 3 colours of tokens accord-
ing to a 2-stage Coxian distribution and the scheduling strategy depicted in
Fig. 9.18.

2. Design a GSPN representation of a M/C/1-LCFSPR queue with a 2-stage

Coxian distribution and 2 colours of tokens, where at most 4 tokens are
present. LCFSPR stands for Last Come First Served Preemptive Resume,
which means that a newly arriving token preempts the token in service. After
completing service of the new arrived token, the preempted token continues
at the point of interruption.

Exercise 9.8 Consider the unbounded GSPN of Fig. 9.20 with

w1 = λ,w2 = µ,w3 = 1. Determine the corresponding steady state distribution
and conditions for its existence.
Hint: Have a look at Fig. 3.5 and Eqs. 3.5 and 3.6 on page 62.
162

10 Queueing Petri Nets

Queueing Petri Nets (QPNs; F. Bause, H. Beilner, P. Kemper [18, 22, 29]) try
to eliminate the problem of representing scheduling strategies by integrating the
concept of queues (cf. Ch. 3) into a coloured version of GSPNs (CGSPNs). This
is done by partioning the set of places into two subsets: queueing places and
ordinary places.
A queueing place (cf. Fig. 10.1) consists of two components: the queue and a
depository for tokens which have completed their service at this queue. Tokens,
when fired onto a queueing place by any of its input transitions, are inserted
into the queue according to the scheduling strategy of the queue. Tokens in a
queue are not available for the transitions. After completion of its service, the
token is placed onto the depository. Tokens on this depository are available to
all output transitions of the queueing place. An enabled timed transition will fire
after an exponentially distributed time delay and an immediate transition fires
immediately as in GSPNs.
Since we are now no longer forced to describe scheduling strategies with only
GSPN elements, the description of systems is simplified. Since the arrival pro-
cess is determined by the firings of the input transitions of a queueing place, we
will omit its specification in the Kendall notation of a queue (cf. Sec. 3).
Example 10.1 Consider the FCFS queue with 2 colours of tokens as described
in Sec. 9.3 (see also Fig. 9.19). Using queueing places we have no problem to
describe a model even for an unlimited number of tokens. Fig. 10.2 shows the
corresponding QPN employing the shorthand notation for queueing places as
given in Fig. 10.1.
The definition of a QPN is as follows:
Definition 10.1 A Queueing Petri net (QPN) is a triple
QP N = (CGSP N,P1 ,P2 ) where

queue depository

Figure 10.1 A queueing place and its shorthand notation

 163

t1 a
-/M/1-FCFS

t2 b

Figure 10.2 QPN model of a FCFS queue with 2 colours of tokens

x x
a
p1 -/C/1-FCFS
x x x
a -/C/1-PS x
p2
x x x x x x x
-/C/1-IS x
-/C/1-FCFS

Figure 10.3 QPN model of a central server system with memory constraints

• CGSP N is the underlying Coloured GSP N ,

• P1 ⊆ P is the set of queueing places and

• P2 ⊆ P is the set of ordinary places, P1 ∩ P2 = ∅,P = P1 ∪ P2 .

Note that if P1 = ∅ the QPN describes a CGSPN so that the following relation-
ship holds: CGSPNs ⊂ QPNs.

Example 10.2 Fig. 10.3 shows an example of a central server system with mem-
ory constraints. Place p2 represents several terminals, where users start jobs
after a certain thinking time. These jobs request service at the CPU (represented
by the -/C/1-PS queue where C stands for a Coxian distribution) and two I/O
systems (represented by the -/C/1-FCFS queues). To enter the subsystem (CPU
+ I/O systems) each job has to allocate a certain amount of memory. For sim-
plicity the memory size needed by each job is assumed to be the same, which is
represented by a token of colour a on place p1 .
164 10 Queueing Petri Nets

10.1 Quantitative Analysis of QPNs

Since we have introduced queueing places, which consist of queues and depos-
itories, a marking M of a QPN consists of two parts M = (n,m) where n
specifies the state of all queues and m is the marking of the underlying CGSPN.
For a queueing place p ∈ P1 , m(p) denotes the marking of the depository. The
initial marking M0 of the QPN is given by M0 = (O,m0 ), where O is the state
describing that all queues are empty and m0 is the initial marking of the CGSPN.
Similar to GSPNs the firing of immediate transitions has priority over the firing
of timed transitions and the service of tokens in queues. Thus, as in GSPNs, the
reachability set of a QPN comprises vanishing and tangible states. If at least one
immediate transition is enabled, M is a vanishing state and tangible otherwise.
Thus

ENT (M ) := {(t,c)| m[(t,c) > and

if ∃t0 ∈ T2 ,c0 ∈ C(t0 ) : m[(t0 ,c0 ) > then t ∈ T2 }

where m[(t,c) > denotes the usual enabling in CPNs (cf. Def. 6.5). ENT (M ) is
the set of enabled transitions at a marking M with respect to a colour c ∈ C(t),
keeping the priority of immediate transitions in mind.
QPNs can be analysed like GSPNs by calculating the steady state distribution of
the reduced embedded Markov chain. For the remaining discussion we consider
only QPNs with finite reachability sets. To simplify notation let us assume that all
queues of the QPN are of type -/M/1-PS and that the first |P1 | places are queueing
places. The service rate of a colour-c token in the queue of a queueing place p
is defined as µ(p,c). A possible state descriptor for such a queue is a function
n(p) : C(p) 7→ N0 where n(p)(c) denotes the number of colour-c tokens in the
queue.
State transitions occur due to the arrival or the service of tokens. With respect to
a service of a colour c ∈ C(p) token the state n(p) changes to

n(p)(c̃) − 1

0 if c̃ = c
n (p)(c̃) =
n(p)(c̃) otherwise
The corresponding rate is given by
n(p)(c)
r(n(p)(c)) = P µ(p,c)
c0 ∈C(p) n(p)(c0 )
1
Recall that each token is served with P n(p)(c0 )
of the server’s capacity
c0 ∈C(p)
and that n(p)(c) colour-c tokens are in the queue. If the queue is empty, i.e.
P 0
c0 ∈C(p) n(p)(c ) = 0, we define r(n(p)(c)) = 0.
The possible state transitions of the QPN can be described in the following way.
Let M = (n,m) be a marking of the QPN. A directly reachable marking M 0 =
(n0 ,m0 ) is defined as follows. If n(p)(c) > 0 for some p ∈ P1 ,c ∈ C(p) then M 0
with respect to a service in place p according to colour c is given by
10.1 Quantitative Analysis of QPNs 165

if p̃ 6= p or c̃ 6= c

0 n(p̃)(c̃)
n (p̃)(c̃) = (10.1)
n(p̃)(c̃) − 1 otherwise
if p̃ 6= p or c̃ 6= c

m(p̃)(c̃)
m0 (p̃)(c̃) = (10.2)
m(p̃)(c̃) + 1 otherwise

and we denote this case by M [(p,c) >QP N M 0 . Eq. (10.1) says that a token of
colour c has left the queue in p and Eq. (10.2) tells us that this token is now on
the depository of p. If (t,c0 ) ∈ ENT (M ), then M 0 is given by

n0 (p)(c) = n(p)(c) + I + (p,t)(c0 )(c), ∀p ∈ P1 ,c ∈ C(p) (10.3)

− 0

m(p)(c) − I (p,t)(c )(c) if p ∈ P 1
m0 (p)(c) = (10.4)
m(p)(c) + I + (p,t)(c0 )(c) − I − (p,t)(c0 )(c) otherwise

This means, tokens fired onto a queueing place are inserted into the queue (cf.
Eq. (10.3)); the tokens at the depository are removed and ordinary places are
marked according to the usual firing rule of CPNs (cf. Eq. (10.4)). We will denote
this case by M [(t,c0 ) >QP N M 0 .
To specify the transition probabilities of the QPN’s embedded Markov chain1
we now have to take also the service rates into consideration. For a marking
M = (n,m) define
X X
Z(M ) = r (n(p)(c))
p∈P1 c∈C(p)

which is the sum of all rates out of M due to services in all queueing places of
P1 . Furthermore define

T (M ) = {t ∈ T |∃c ∈ C(t) : (t,c) ∈ ENT (M )}

With that the transition probability between marking Mk and Mr are defined as
P [Mk → Mr ] =
1
Note that this technique of using the embedded Markov chain is only applicable if the transition
probabilities out of a marking (state) are independent of the time spend in that state (semi-
Markov process). This imposes a restriction on the allowed queues of the QPN.
166 10 Queueing Petri Nets
 X


 wi (c)
(ti ,c)∈ENT (M ):Mk [(ti ,c)>QP N Mr


if T (M ) ⊆ T2


wj (c0 )

 X






 (tj ,c0 )∈ENT (Mk )




 X
wi (c)





 (ti ,c)∈ENXT (M ):Mk [(ti ,c)>QP N Mr


if T (M ) ⊆ T1

wj (c0 ) + Z(Mk )
(tj ,c0 )∈ENT (Mk )










 if T (M ) ∩ T2 = ∅
X r(n(p)(c))


and


wj (c0 ) + Z(Mk )





 Mk [(p,c) >QP N Mr


 (tj ,c0 )∈ENT (Mk )





0 otherwise

(10.5)
The first expression in Eq. (10.5) is the probability that an enabled immediate
transition fires. The second and the third expressions are, respectively, the prob-
abilities that a timed transition fires and that a service completion in a queueing
place occurs. The probabilities in the latter cases (cf. page 131) are given by the
rate causing the state transition divided by the sum of all rates leading out of that
marking.
Having determined the probabilities P [Mk → Mr ] we can proceed as in the
GSPN case by defining the matrices C,D,E and F (cf. Eq. (9.2) on page 142)
and solving the global balance equations of the reduced embedded Markov chain

π̃ = π̃ × P 0 ;
X
π̃s = 1.
Ms ∈T̂

where P 0 is given by Eq. (9.6) and T̂ is the set of tangible states (cf. page 142).
Again we see that a steady state distribution of the QPN exists iff there is no
timeless trap and the reachability set of the QPN has home states.
Given the steady state distribution π̃ of the reduced embedded Markov chain,
the mean number of visits to marking Ms between two consecutive visits of
marking Mj is vsj = π̃π̃sj . Thus the mean cycle time tc (Mj ), i.e. the mean time
between two consecutive visits of Mj , is given by

1 X
wk (c) + Z(Ms ))−1
X
tc (Mj ) = π̃s (
π̃j (t
Ms ∈T̂ k ,c)∈ENT (Ms )

and since the mean sojourn time ts in marking Mj ∈ T̂ is

wk (c) + Z(Mj ))−1

X
ts (Mj ) = (
(tk ,c)∈ENT (Mj )
10.1 Quantitative Analysis of QPNs 167

we obtain the steady state distribution π of the QPN by

( t (M )
s j
tc (Mj ) if Mj ∈ T̂
πj = (10.6)
0 if Mj ∈ V̂
as in the case of the quantitative analysis of GSPNs.
Example 10.3 Consider the QPN of Fig. 10.4 with
µ(p1 )(a) = 3,µ(p1
)(b) = 2,w1 (a) = 3,w1 (b) = 2,w2 (a) = w2 (b) = 1.
1 if c̃ = c
Define δc (c̃) =
0 otherwise

t1

x x

p1 a
p2
b
-/M/1-PS
x x

t2

Figure 10.4 C(p1 ) = C(p2 ) = C(t1 ) = C(t2 ) = {a,b}

The reachability set comprises the following markings:

M0 = ((0),(0,δa + δb )), M1 = ((δa ),(0,δb )), M2 = ((δb ),(0,δa )),
M3 = ((0),(δb ,δa )), M4 = ((0),(δa ,δb )), M5 = ((δa + δb ),(0,0)),
M6 = ((δa ),(δb ,0)), M7 = ((δb ),(δa ,0)), M8 = ((0),(δa + δb ,0))
and the corresponding reachability graph is shown in Fig. 10.5. For a marking,
e.g. M5 = ((δa + δb ),(0,0)) the first component denotes n(p1 ) which is here
the sum of two functions δa and δb . The second component of M5 , that is (0,0),
denotes the marking of the underlying CGSPN.
The markings M0 ,M1 ,M2 ,M3 ,M4 are vanishing and the markings M5 ,M6 ,M7 ,M8
are tangible. The matrix P of the embedded Markov chain is
1 1
 
0 2 2 0 0 0 0 0 0

 0 0 0 0 0 1 0 0 0 

0 0 0 0 0 1 0 0 0
 
 
0 0 0 0 0 0 1 0 0
 
   
C D  
P = = 0 0 0 0 0 0 0 1 0 
E F 
2 3


 0 0 0 0 0 0 5 5 0 

2 3

 0 5 0 0 0 0 0 0 5


3 2
0 0 0 0 0 0 0
 
5 5
 
3 2
0 0 0 5 5 0 0 0 0
168 10 Queueing Petri Nets

M0

(t2 ,a) (t2 ,b)

M1 M2

(t2 ,b) (t2 ,a)

(t1 ,b) (t1 ,a)
M5
(p1 ,b) (p1 ,a)

M6 M7

(p1 ,a) (p1 ,b)

(t2 ,a) M8 (t2 ,b)

(t1 ,a) (t1 ,b)

M3 M4

Figure 10.5 Reachability graph

Since matrix
1 1
 
1 2 2 0 0

 0 1 0 0 0 

−1
(I − C) = 0 0 1 0 0
 

0 0 0 1 0
 
 
0 0 0 0 1

we get
2 3
 
0 5 5 0
2 3

0 0 
P 0 = F + E × (I − C)−1 × D =  5 5
 
3 2
0 0

 5 5 
3 2
0 5 5 0
Solving
8
0
X
π̃P = π̃ and π̃i = 1
i=5

yields π̃5 = π̃6 = π̃7 = π̃8 = 41 . Thus vkj = 1,∀j,k ∈ {5,6,7,8}. The sojourn
time for tangible markings is given by ts (M5 ) = ( 21 × 3 + 12 × 2)−1 = 52
and ts (Mj ) = (3 + 2)−1 = 51 ,∀j ∈ {6,7,8}. With that the cycle time is given
by tc (Mj ) = 1,∀j ∈ {5,6,7,8}, which yields the steady state distribution for
tangible states: π5 = 25 and πj = 51 ,∀j ∈ {6,7,8}.
10.2 Qualitative Analysis of QPNs 169

As we have seen, QPNs allow to describe queues in a convenient way without

having to specify queues by pure Petri net elements. An appropriate state de-
scriptor for a −/M/1 − F CF S queue serving several colours of tokens would
be a string in (C(p))∗ , where p is the queueing place. A state va ∈ (C(p))∗ ,a ∈
C(p), would be changed to bva if a token of colour b ∈ C(p) arrives or alterna-
tively to v if the token of colour a completes its service.
However the complexity of the quantitative analysis, determined by the size and
structure of the reachability set, is still the same as that obtained by modelling
the queue with CGSPN elements.

10.2 Qualitative Analysis of QPNs

As we have seen, qualitative properties of a QPN are essential for a successful

quantitative analysis of the net. Because QPNs are based on CPNs, we are again
interested in employing the analysis procedure shown in Fig. 9.5. Since CGSPNs
⊆ QPNs, all negative examples given in Sec. 9.2 hold for QPNs as well.
First let us investigate timeless traps. If we neglect the specific values of the rates
and consider only the priority relation imposed on the firing of transitions, the
enabling rule in QPNs is given by

Definition 10.2 A transition t ∈ T of a QPN is enabled at a marking M with

respect to a colour c ∈ C(t), denoted by M [(t,c) >QP N , iff (t,c) ∈ ENT (M ).

According to this enabling rule all definitions for Place-Transition nets can be
defined similarly for QPNs. A timeless trap is now defined as follows:

Definition 10.3 A QPN has a timeless trap, iff ∃M ∈ R(QP N,M0 ) :

∀k ∈ N : ∃σ ∈ TC∗ : |σ| ≥ k, such that M [σ >QP N and
∀σ̃ ∈ TC∗ : M [σ̃ >QP N =⇒ σ̃ ∈ (T2 )∗C ,
where YC := ∪y∈Y ∪c∈C(y) (y,c) for Y ⊆ P ∪ T .

As in GSPNs timeless traps cannot occur if condition NoTT (cf. Def. 9.5) holds
for QPNs. For QPNs we are able to establish a generalised version of this con-
dition: If an immediate transition has at least one queueing place as an input
place, this transition cannot contribute to a timeless trap, since tokens fired onto
a queueing place are only available to the output transitions after being served.
To serve a token in a queueing place the QPN has to be in a tangible state. Since
we are now dealing with Coloured Petri nets all colours of a transition have to be
considered, because they characterise the different “firing modes” of a transition.

Definition 10.4 (Condition NoTT for QPNs)

Condition NoTT :⇐⇒ ∀T̃ ⊆ (T2 )C \ (P1 )C • : •T̃ 6= T̃ •
170 10 Queueing Petri Nets

For the •-notation of these sets, see Def. 6.4 on page 118. Condition NoTT en-
sures the absence of timeless traps for QPNs with finite reachability sets and the
proof is similar to the one given for Th. 9.2.

Theorem 10.1 Let QPN have a finite reachability set.

Condition NoTT (for QPNs) =⇒ QPN has no timeless trap.

Obviously Th. 9.3 also holds in this context and condition NoTT can be verified
by a similar algorithm as given in Fig. 9.8. We start initially with TH := (T2 )C \
(P1 )C • and use the •-notation according to Def. 6.4.
Assuming that tokens are not created in queues, which holds for all queues we
have learned about so far, we get the following result directly:

Theorem 10.2 CPN bounded =⇒ QPN bounded

The converse is, as in the case of GSPNs, not true, which is shown by Fig. 9.9.
For the derivation of further results, especially on liveness and existence of
home states, we restrict ourselves to QPNs which have an EFC-net structure.
Remember that we have established several useful results for this class of PNs in
Sec. 5.4.3.

10.2.1 Qualitative Analysis of EFC-QPNs

The structure of a QPN is determined by a Coloured Petri net. Unfolding even

a strongly connected CPN might yield several isolated Place-Transition nets.
Fig. 10.6 shows a simple example of this. We know that a timed transition of
a CGSPN “unfolds” into a set of timed transitions of the GSPN according to
Def. 9.8. Each of the isolated Place-Transition nets has at least one transition,
which is derived by unfolding a timed transition, or there is at least one place,
which is derived by unfolding a queueing place. We will denote this property by
(∗). With that we can exploit the results for EFC-GSPNs.
Taking the colours of the underlying CPN into consideration, condition EQUAL-
Conflict can be defined for QPNs as well.

(t,a)
(p,a)
t
a x
p
b
(t,b)
(p,b)
x

Figure 10.6 Unfolding a simple CPN

10.2 Qualitative Analysis of QPNs 171

Definition 10.5 A QPN satisfies condition EQUAL-Conflict, iff

∀t,t0 ∈ T : •({t}C ) ∩ •({t0 }C ) 6= ∅ =⇒ {t,t0 } ⊆ T1 or {t,t0 } ⊆ T2 .

If you think of the unfolded Petri net of the CPN underlying the QPN, this con-
dition describes the same condition as that in Def 9.7 and it is again of particular
importance for EFC-nets.

Definition 10.6 A QPN is an EFC-QPN iff its underlying unfolded CPN consists
of EFC-nets.

Remember that unfolding a CPN might yield several isolated Place-Transition

nets. For an EFC-QPN these Place-Transition nets must all be EFC-nets. Ac-
cording to property (∗) all these EFC-nets have at least one transition, derived
by unfolding a timed transition, or at least one place derived by unfolding a
queueing place. If we assume that the CPN is bounded and live (cf. Def. 6.6
on page 118) this means that all EFC-nets are strongly connected as implied
by Th. 5.1. With that we easily see that condition EQUAL-Conflict implies the
absence of timeless traps.

Theorem 10.3 Given an EFC-QPN whose underlying CPN is live and bounded,
then
Condition EQUAL-Conflict =⇒ QPN has no timeless traps.

Proof. Although the proof is similar to the proof of Th. 9.5 we will go more into
the details; just to show how the unfolding of a net can be used for this purpose.
Assume the QPN has a timeless trap, i.e. ∃M ∈ R(QP N,M0 ) :
∀k ∈ N : ∃σ ∈ TC∗ : |σ| ≥ k, such that M [σ >QP N and
∀σ̃ ∈ TC∗ : M [σ̃ >QP N =⇒ σ̃ ∈ (T2 )∗C .
Since T is a finite set, there is at least one element (t̂,ĉ) whose number of occur-
rences in σ is not bounded.
Since the CPN is bounded, Th. 10.2 ensures boundedness of the QPN implying
the finiteness of R(QP N,M0 ). Thus there exists a terminal strongly connected
subset CM ⊆ R(QP N,M ). Define
T̃ := {(t,c)|∃M 0 ∈ CM : M 0 [(t,c) >QP N }. The assumption implies T̃ ⊆
(T2 )C and since CPN is live and bounded, the corresponding unfolded net con-
sists of strongly connected Place-Transition nets. Let us consider the correspond-
ing Place-Transition net which comprises (t̂,ĉ). Because of (∗) there is at least
one transition in the net, which does not contribute to the timeless trap and thus
is not a member of T̃ . Define P̃ := •T̃ , then one of the following two cases holds
with respect to the Place-Transition net being considered:

1. ∃(pr ,cr ) ∈ P̃ :
(pr ,cr )• ⊇ {(ti ,ci ),(tj ,cj )} and (ti ,ci ) ∈ T̃ ,(tj ,cj ) ∈ TC \ T̃ .
The assumption of a timeless trap yields ti ∈ T2 . Thus if M 0 [(ti ,ci ) >QP N
for some marking M 0 ∈ CM , the EFC-net structure and condition EQUAL-
Conflict imply M 0 [(tj ,cj ) >QP N contradicting (tj ,cj ) 6∈ T̃ .
172 10 Queueing Petri Nets

t1

x x

p1 a
p2
b

-/M/1-LCFSPR x
x

t2

Figure 10.7 Non live QPN

2. ∃(tr ,cr ) ∈ T̃ :
(tr ,cr )• ⊇ {(pk ,ck ),(pl ,cl )} and (pk ,ck ) ∈ P̃ ,(pl ,cl ) ∈ PC \ P̃ .
Since (tr ,cr ) ∈ T̃ ,tr fires infinitely often with respect to cr ∈ C(tr ). Since
(pl ,cl ) 6∈ P̃ = •T̃ , the place pl is not bounded with respect to colour
cl ∈ C(pl ), contradicting the boundedness of the QPN. t
u
Considering the theorems we have established for EFC-GSPNs one would also
expect that condition EQUAL-Conflict is sufficient for liveness in the QPN case.
Unfortunately this is not the case, since it might now be possible that the schedul-
ing strategy prevents tokens from being served. Fig. 10.7 shows an example of
this. A token being served at the queueing place p1 immediately returns to this
place and preempts the token in service. Thus the token of colour a or the token
of colour b will never complete its service. If at least one token of each colour
of the queueing place has the opportunity of being served in every state of the
queue, such situations cannot occur. This gives rise to the following definition.
Definition 10.7 A QPN satisfies condition EQUAL-Service iff ∀p ∈ P1 : the ser-
vice time distribution of the queue in p is Coxian and if |C(p)| > 1 then the
scheduling strategy is PS or IS.
If |C(p)| = 1 we can choose an arbitrary scheduling strategy. We only have to
demand that the server stays never idle while there are customers present, which
is one of the main features of so-called work conserving scheduling strategies
[102]. Note that all scheduling strategies we have looked at in Ch. 3 satisfy this
requirement.
Condition EQUAL-Service states that the queue can be analysed by means of
Markovian techniques, since the service time is exponentially distributed in each
stage of service of a token and that for queues serving several colours of tokens
all tokens are served simultaneously.2 With that we can exploit Th. 9.5 to estab-
lish the following:
2
[18] gives a general definition of condition EQUAL-Service also comprising further scheduling
strategies.
10.2 Qualitative Analysis of QPNs 173

(t̃1 ,c) (tˆ1 ,c) (t̃2 ,c) (tˆ2 ,c)

···

···
(t1 ,c) (t2 ,c)

Figure 10.8 GSPN representation of a Coxian distributions

Theorem 10.4 Given an EFC-QPN satisfying Condition EQUAL-Service and

whose underlying CPN is live and bounded, then

1. Condition EQUAL-Conflict ⇐⇒ QPN is live

2. Condition EQUAL-Conflict =⇒ QPN has home state.

Proof. Since for each queueing place p the service distribution is Coxian and
in case of |C(p)| > 1 the scheduling strategy is PS or IS, each queue can be
unfolded and the service of each token colour can be represented by the GSPN
subnet shown in Fig. 10.8. (t̃i ,c) represents the exponential service time distri-
bution for a token in stage i of its service, (tˆi ,c) represents entering stage (i + 1)
of service and (ti ,c) the completion of service. Note that the rates and the firing
weights of the transitions (t̃i ,c),(tˆi ,c),(ti ,c) are now possibly marking depen-
dent. The resultant GSPN satisfies condition EQUAL-Conflict and the statement
of this theorem follows from Th. 9.5. t
u

If condition EQUAL-Service does not hold, we have already seen that live-
ness does not extend from the CPN to the QPN, even for simple net structures.
Fig. 10.7 also shows that home states might not exist. But even if the QPN is
live, condition EQUAL-Service is essential for preserving the existence of home
states. Fig 10.9 shows a further example of a QPN with no home states. The
queues in the queueing places p2 and p5 serve arriving tokens according to the
FCFS scheduling strategy and the service times are exponentially distributed. If
transition t1 (t3 ) is enabled in the QPN, the token on the input place p1 (p4 ) is
“transferred” to place p2 (p5 ) and a token of the complementary colour is created
on place p3 (p6 ). Transitions t2 and t4 “merge” tokens of the same colour. Serv-
ing tokens according to the FCFS rule yields two strongly connected subsets in
the reachability set of the QPN, such that home states do not exist. These subsets
are characterised by the sequence of tokens in both queues: either “a” ahead of
“b” or “b” ahead of “a”.
174 10 Queueing Petri Nets

-/M/1-FCFS

x x
t1 t2
p2

x x x
if (x = a)
then b p3
else a

a p1
b p4

-/M/1-FCFS

x x
x x
p5

t4 t3
x if (x = a)
then b
p6 else a

Figure 10.9 QPN with no home states

10.3 Some Remarks on Quantitative Analysis

As we have seen, introducing time into Place-Transition nets might change the
behaviour and thus the properties of the net significantly. Most of the results pre-
sented in Sec. 9.2 and 10.2 only hold for nets with a restricted structure. Most
nets in practice unveil a more complex net structure and thus only some proper-
ties, like boundedness or P-invariants of the Place-Transition net, extend to the
time-augmented net. Nevertheless, a qualitative analysis of the Place-Transition
net is often useful to gain insight into the behaviour of the net. Since it is often
desirable that the “functioning” of the net does not depend on specific timing
constraints, like e.g. in mutual exclusion problems, such an analysis should be
performed first. Unfortunately, determining essential properties for a quantita-
tive analysis, must often only be done by inspecting the reachability set of the
time-augmented net.
Even if the time-augmented Petri net satisfies all positive qualitative properties
for a quantitative analysis, determining a solution of the global balance equations
becomes intractable for large nets. Hence several authors (cf. [13, 25, 26, 43,
65, 89, 111, 179]) have investigated time-augmented Petri nets with product-
10.3 Some Remarks on Quantitative Analysis 175

form solutions, similar to product-form queueing networks [102] and developed

algorithms based on the product-form property [164, 165].
Another approach is to structure the model hierarchically, e.g. by place refine-
ment, and to exploit the hierarchical structure also for the quantitative analysis
such that models with millions of states can be analysed, see e.g. [27]. The gen-
eral idea is similar to the one used for the analysis of superposed GSPNs [63],
where the net is composed of several components interacting via a set of synchro-
nised transitions ST . Considering these components in isolation gives reachabil-
ity sets RS i whose individual size is usually much smaller than the size of the
reachability set of the whole net. The infinitesimal generator Q of the whole
reachability set RS can then be expressed by a Kronecker based representation
of the following form (cf. [105])
M X O
Q= Qil + Qit
t∈ST

where the Qi∗ are (|RS i | × |RS i |)-matrices.

The advantage of this representation is a shift from space to time concerning
the complexity of iterative solution techniques for solving the global balance
equations. In many cases the sizes of the reachability sets of the components
are much smaller than the corresponding state-transition matrix Q of the whole
GSPN. Thus the information on the state space structure is stored in a set of
smaller matrices and an element of Q can be determined by some additional cal-
culations (cf. [52]). This approach has its greatest benefits when all matrices Qi∗
fit into main memory whereas Q would not. Then the additional operations spec-
ified by the Kronecker operators can be neglected in contrast to time-consuming
swapping operations of the computing equipment. A similar Kronecker based
representation can be defined for nets with specific hierarchical structures, thus
profiting from the same effect.

Exercise 10.1 Calculate the steady state distribution of the central server model
of Fig. 10.3. Assume that all rates and firing weights are 2.0.

Exercise 10.2 Determine the reachability sets of the QPN of Fig. 10.9 and its
underlying CPN.
176

11 Further Reading

Introductory papers on Stochastic Petri nets are, for example, those by Ajmone-
Marsan [2] and Pagnoni [129].
In this introduction we have only discussed some of the most important time-
augmented Petri net models, which can be analysed using Markovian analy-
sis techniques. There are several other models integrating timing aspects into
Petri nets. We will briefly describe some of them using the classification of time-
augmented Petri nets given on page 129:

Timed Places Petri Nets:

Timed Petri Nets: In [166] J. Sifakis attaches real values zi ∈ R+ to each

place pi of the Petri net. The constant zi determines the dwelling time
of a token on that place. After this time has expired the token becomes
available to all output transitions of pi . Enabled transitions have to fire
immediately. In [167] it is shown that this model is equivalent to the
TTPN model of C. Ramchandani. In [166, 167] a relationship between
the initial marking, the dwelling times zi and the firing frequencies of
transitions is established.
A modified version of the TPPN model by Sifakis is investigated in
[55].
Stochastic Petri Nets: C.Y. Wong, T.S. Dillon and K.E. Forward [181] de-
fine a stochastic version of the TPPN model by Sifakis, where the
dwelling times of tokens are now characterised by exponential distri-
butions such that Markovian analysis techniques are applicable.

Timed Transitions Petri Nets:

Preselection models:
Timed Petri Nets:
• One of the first models considering Petri nets and “time”
was developed by C. Ramchandani [144]. The constant fir-
ing time, i.e. the time delay between enabling and firing,
of a transition ti ∈ T is characterised by a real variable
zi ∈ R+ . If transitions are in conflict the one to reserve to-
kens first is chosen randomly. It might be possible that after
a transition has reserved tokens it becomes enabled again.
Thus a transition might be enabled simultaneously to itself.
In [143] marked graph structures of this TTPN model are
investigated and, similar to [144], the minimum cycle time
 177

of transitions, i.e. the minimum time between two consecu-

tive firings of a transition, is calculated (see also [125]).
• R.R Razouk and C.V. Phelps [147] define a model similar
to that of Merlin (see Merlin’s TTPNs). The behaviour of a
transition is determined by a pair of real values (zmin ,zmax ),zmin ∈
R+ ,
zmax ∈ R+ ,zmin ≤ zmax ; if a transition is enabled at time
τ it must not reserve tokens before time τ + zmin . After
reservation it fires exactly at time τ + zmax .
Stochastic Petri Nets: W.M. Zuberek [185, 186] associates a rate,
specifying an exponentially distributed firing time, and a proba-
bility c(t) with each transition. In case of a conflict c(t) deter-
mines the probability with which transition t reserves tokens on
the input places. Zuberek also extends the underlying Petri net
by introducing inhibitor arcs.
Race models:
Timed Petri Nets: P. Merlin [121] attaches to each transition a pair
of real values (zmin ,zmax ),zmin ∈ R+ ,
zmax ∈ R+ ,zmin ≤ zmax and defines the behaviour as follows:
If a transition is enabled at time τ then it must not fire before
time τ + zmin . Afterwards the transition may fire at any point
in time in the time interval [τ + zmin ,τ + zmax ] provided it is
still enabled. At latest at time τ + zmax the transition has to fire.
Merlin shows that his model is adequate for modelling time-outs
of communication protocols, which is also investigated in [61].
In [33] the analysis of this TTPN model is considered.
Stochastic Petri Nets:
• M. Ajmone-Marsan and G. Chiola [8, 9] extend the GSPN
model by introducing transitions with a deterministic firing
time (Deterministic Stochastic Petri nets; DSPNs). If at all
markings amongst timed transitions (i.e. transitions with an
exponentially distributed firing time) only one determinis-
tic transition is enabled, the Petri net can be mapped onto a
Semi-Markov process. The analysis of DSPNs is also con-
sidered in [10, 81, 84, 113]. The analysis of DSPNs and
also more general Non-Markovian Stochastic Petri Nets is
considered in [82].
• J.B. Dugan [66] defines Extended Stochastic Petri nets (ES-
PNs) by allowing firing times of transitions to be specified
by an arbitrary continuous time distribution. As in SPNs
each transition is a timed transition. Amongst inhibitor arcs
the underlying Petri net is extended by further arcs: counter-
alternate and probabilistic arcs.
178 11 Further Reading

Since 1985 an international workshop on Timed and Stochastic Petri nets is held
biennial. The title of the first workshop in 1985 was “Timed Petri Nets” [134],
which afterwards changed to “Petri Nets and Performance Models” [135, 136,
137, 138, 139, 140, 141, 142]. Some publications on Timed and Stochastic Petri
Nets can also be found in the annual international conference “Application and
Theory of Petri Nets” and in the “Advances in Petri Nets”.
Further information on QPNs can be accessed via WWW using URL http:
//ls4-www.informatik.uni-dortmund.de/QPN/
Tool support is available for several Stochastic Petri net formalisms (e.g. GSPNs
[50, 51], DSPNs [83, 85, 184, 183, 112, 114, 113], QPNs [27, 30]), see also
[73, 74]. A list on Petri net tools can be accessed via WWW using URL http:
//www.informatik.uni-hamburg.de/TGI/PetriNets/tools.
A proposal has been published defining a common interface description also for
Stochastic Petri nets [28, 47], thus exhibiting the possibility to exchange model
descriptions between different sets of tools.
 179

12 Application Examples

In this chapter we present two examples of modelling systems using Stochas-

tic Petri nets. We will employ the QPN formalism, since it comprises SPNs and
GSPNs, and also offers convenient description facilities for queueing situations.
In Sect. 12.1 we address the well-known problem of sharing resources by differ-
ent processes. In Sect. 12.2 a node of a DQDB network is modelled by a QPN
subnet. In this section we will use the more general definition of QPNs including
so-called immediate queueing places as given in [18].

12.1 Resource Sharing

This example considers the general problem of modeling sharing of

processors/machines by different processes.
Imagine that each process is given by a process schemata specified by an ordinary
Place-Transition net. Since scheduling in QPN models is described by queueing
in places, obviously activities sharing resources should be attached to the places
of the Place-Transition net. Assume, e.g., the following description of process
schemata.
A user specifies a series of n simulation runs, which might be executed in parallel
(see Fig. 12.1). After termination of all simulation runs a special program is in-
voked, collecting the simulation results and preparing them for printing. Finally
the results are printed. The second process schemata specifies a user who pre-
pares a LATEX-document for printing by starting two sequential calls of this type-
setting program (just to ’get cross-references right’). For both process schemata
we have to specify additionally the service time distribution of each task, e.g. the
parameters of a Coxian distribution and the weights of immediate transitions.
For simplicity we will neglect these parameters here.
After definition of the load, we have to determine the machines executing these
tasks. In this example we consider a small network comprising three worksta-
tions and one printer. Each workstation works off all tasks in a processor sharing
manner and the printer services jobs in order of their arrival.
Mapping of these processes (tasks) to processors (machines) is defined statically
by the users. The user initiating the simulation runs, e.g., decides to dedicate
ni runs to workstation i (n = n1 + n2 + n3 ) and starting the collector task
at workstation 3. The user of LATEX on the other hand independently decides
to use workstation 3 for all his tasks. Attaching distinct colours to the places
and transitions of both process schemata now yields the QPN model given in
Fig. 12.2. Activities represented by places of the P/T-net are ’mapped’ to the
180 12 Application Examples

s
j

?
J
j j simulation
 J
^
... s
j
runs
? ? ?
@
j j
R
@ ?
LATEX1
n
? ?

j collect j
? ?
LATEX2

? ?

j print j
? ?
print

Figure 12.1 Two process schemata

corresponding timed queueing places of the QPN. Places Pi ,i = 1,2,3, model the
three workstations and place P5 the printer. The most interesting queueing place
is P 3, where n3 simulation runs and afterwards a collector job compete with a
LATEX job (LATEX1 or LATEX2) for service at the CPU. Note that this QPN models
an absorbing Markov chain (cf. Sect. 2.1.2) and the mean time until absorption
reflects the mean time until both results have been printed.

12.2 Node of a DQDB network

The Distributed Queue Dual Bus (DQDB) protocol has been proposed as the
subnetwork for the IEEE 802.6 MAN for the interconnection of Local Area Net-
works, servers, workstations and other devices. The basis of a DQDB network
is a pair of unidirectional buses as illustrated in Fig. 12.3. Nodes are connected
to both buses by a logical OR-writing tap and a reading tap. The Head Station
(frame generator) continuously generates frames. Each frame is subdivided into
slots of equal size and each slot in turn has several fields. The End Station (slave
frame generator) terminates the forward bus and removes all incoming frames
and generates slots at the same transmission rate and of the same sort on the
other bus.
Data to be transmitted, is split up into segments of fixed size fitting on a slot’s
information field. Each slot can be allocated either to isochronous traffic (called
12.2 Node of a DQDB network 181

start sim runs start text proc

1‘sim S 1‘LATEX1 L

x x

n3 ‘x
n1 ‘x x
n2 ‘x

CLS CLS CLS LATEX1

P1 P2 P3
-/Cox/1-PS LATEX2
-/Cox/1-PS -/Cox/1-PS

x x x x

[x=sim] [x=sim] [x=sim]

[x=LATEX2 or
collect x=collect]
x
P4
n‘x P5 CL
[x=sim]
-/Cox/1-FCFS

colour C = with collect;

colour L = with LATEX1 | LATEX2;
colour S = with sim;
colour CS = with C | S;
colour CL = with C | L;
colour LS = with L | S;
colour CLS = with C | L | S;

Figure 12.2 QPN model of two processes sharing resources

pre-arbitrated slots or PA-slots) or non-isochronous traffic (called queued arbi-

trated slots or QA-slots). Access to QA-slots is controlled by the following pro-
tocol, which operates on both buses symmetrically, so it is sufficient to consider
only one direction.
In the following we assume data segments to be sent in the direction of the D-bus
and requests in the direction of the R-bus (cf. Fig. 12.3). Segments arriving at a
node are placed into a so called ’local queue’ which is served in FCFS manner.
The first segments (only) at all non-empty local queues for one bus constitute a
distributed queue whose service discipline is FCFS. This distributed queue does
not represent any existing physical queue, but is only established virtually by
means of the following mechanism. Each local queue can be in two states: idle
182 12 Application Examples

R-bus
-
6 6 6
? ? ?
Head/End Node Node Node Head/End
... ...
Node 1 i N Node
6 6 6
 ? ? ?
D-bus

Figure 12.3 DQDB architecture

or countdown. If no segment is waiting for transmission the corresponding local

queue is in the idle state. If a segment arrives at a node for transmission on the
D-bus, an empty local queue changes to its countdown state. For each segment
to be send, a request is transmitted on the opposite bus, thus informing upstream
nodes (with regard to the direction of the D-bus) to let through an empty slot
on the D-bus for its own waiting segment. A request can only be transmitted if
the corresponding bit in the slot on the R-bus is not set, otherwise this indicates
the request of a downstream node. Each node keeps track of its own pending
requests by incrementing a request queue counter (RQC). When a local queue
is idle, it keeps count of the number of outstanding requests from downstream
nodes via a so-called request counter. For each empty slot passing on the D-bus
this counter is decremented, accounting in this way for a served segment of a
downstream node on the D-bus. If a local queue changes to its countdown state
due to the arrival of a segment, the current value of its request counter is copied
to a so-called countdown counter. This special counter is decremented whenever
an empty slot passes on the D-bus until it reaches 0. Following this instant, the
node transmits its own segment in the first empty slot on the D-bus.
For this example we employ the more general definition of QPNs [18]. Pure
scheduling aspects can be described by immediate queueing places. In contrast to
timed queueing places, tokens on those places can be viewed as being “served”
immediately. The service in immediate queueing places has priority over the
scheduling/service in timed queueing places and firing of timed transitions, sim-
ilar to the priority of the firing of immediate over timed transitions.
A simple example of an immediate queueing place is given in Fig. 12.4. Here
the definition of the scheduling strategy is as follows: serve tokens immediately
when present according to FCFS, but only if the depository is empty. If there is
a token at the depository all tokens are blocked in their current position.1 Since
the depository is emptied by transition t in Fig. 12.4, this transition determines
the “real” service rate. We will now use this type of immediate queueing place,
which will be called “FCFS with blocking” for the modelling of a node in a
1
Similar to immediate transition also probabilities have to be specified for the “service” in im-
mediate queueing places. For the special strategy chosen here the specific values of these prob-
abilities do not matter. A formal definition of this strategy is given in [18].
12.2 Node of a DQDB network 183

FCFS with
blocking

Figure 12.4 An immediate queueing place

FCFS with
R bus depart blocking T4 P3
noreq R D bus arrive
T
req req seg seg tok F -/Cox/1-FCFS
NR req S
RQC P1 1‘tok
Arrivals

if y = seg x x
seg
then 1‘tok
else empty
x
if x = req
FCFS with [x=free]
y T3
then 1‘req blocking
else empty T2
RS
P2
T1
if y = req
then 1‘tok
else empty x
x case y of
if x = req seg => 1‘busy
then 1‘req
req => 1‘free
else empty T if x = free
then k‘tok
-/Cox/1-FCFS else empty
P4 F
NR k‘tok
R bus arrive D bus depart

colour F = with free | busy;

colour T = with tok;
colour RS = with req | seg;
colour NR = with noreq | req;
colour S = with seg;
colour R = with req;

Figure 12.5 QPN model of a DQDB node

DQDB network.
Fig. 12.5 presents a QPN model for an arbitrary node of a DQDB network. The
definition of the colour sets and the initial marking is located near the corre-
sponding places. E.g. place P 4 has colour set T and will initially contain k to-
kens of type tok. The guard of a transition is given in brackets. E.g. transition
T 2 is enabled if the free variable x is bound to free ([x = free]).
The R- and D-bus are represented by immediate transitions T 1 and T 2,T 3 resp.
Since DQDB mainly deals with two queues: the local queue at each node and the
distributed queue consisting of all first segments at all non-empty local queues,
these queues are modelled explicitly. Part of the distributed queue is modeled by
the immediate queueing place P 2. The state of this queue reveals the information
that node has about the actual distributed queue due to arriving requests on the
184 12 Application Examples

R-bus. Note that e.g. ’Node 1’ does not have any information on the state of
the distributed queue, because it does not know of segments to be transmitted
at other nodes. The basic idea of the QPN model is to represent the value of
the request counter by inserting each request into the FCFS queue at place P 2.
This queue might also contain one segment (the first of the local queue) at most,
which is ensured by place P 3. All other segments have to wait in the immediate
queueing place P 1, whose queue also obeys an FCFS scheduling strategy with
blocking. Place P 4 is complementary to P 2 thus ensuring that the number of
tokens in P 2 will never exceed k = ’maximum value of request counter’ + 1
(first segment of the local queue). Service of the queue is performed by firing of
immediate transition T 2, if an empty slot on the D-bus arrives at this node. In
case of serving a segment at P 2 the formerly empty slot is marked busy. If P 2 is
empty (implying P 4 contains k tokens of type tok) or a busy slot (token of type
busy) arrives, T 3 fires immediately.
Several models of such nodes can be interconnected by immediate transitions to
constitute the R- and D-bus yielding a model of the whole DQDB network.
 185

13 Solutions to Selected Exercises

Solution to Ex. 1.5

kp(1 − p)k−1

Solution to Ex. 1.1

• Probability of the same birthday for at least two persons is
365!
p=1−
365n (365 − n)!

• p ≤ 0.5 for n ≤ 22
Solution to Ex. 1.2 In this case there are two mutually exclusive hypotheses:
G1 “The student knows the correct answer”, and
G2 “The student is guessing”
with probabilities P {G1 } = p and P {G2 } = 1 − p. We observed, event E,
that the student had the answer wrong. The conditional probabilities of E,
given the hypotheses, are
1
P {E|G1 } = 0 P {E|G2 } = 1 −
m
According to the Theorem of Total Probability,
1
P {E} = (1 − )(1 − p).
m
Using Bayes Theorem we get the answer

P {E|G2 }P {G2 }
P {G2 |E} =
P {E|G1 }P {G1 } + P {E|G2 }P {G2 }
= 1

which is to be expected.
Solution to Ex. 2.1
• The stochastic process changes state every day and since we know
that two sunny days never follow one another the transition probabil-
ities are not strictly state independent. Assuming that the process is
Markovian is thus an approximation only.
186 13 Solutions to Selected Exercises

• If we let S,R and D represent a sunny, rainy and dull day respectively,
then we can define a new stochastic process by considering the states
SS,SR,SD etc. which would represent the weather on two consecu-
tive days. We need to determine new transition probabilities as well.
A Markovian process will approximate this new process more closely
than in the previous case.

Solution to Ex. 2.5 Draw the process as a discrete time Markov chain with 4
states of which state 0 is one absorbing state (the other is clearly C). Then
 
0 p
Q=
q 0

and
 
q 0
R=
0 p

From these we can compute


1 p 

N=  1−pq 1−pq 
q 1
1−pq 1−pq

and the steady state probability matrix NR. From the latter the probability
of starting in state 1 and ending in state 0 is given by
q
n10 = .
1 − pq

The expected duration of the game is given from Th. 2.5 by

1+p
1 − pq

Solution to Ex. 2.8 A car spends 64 percent of its time in Town 2.

Solution to Ex. 2.9 There are 1.62 computers broken on the average and they
8 6 3
break down at a mean rate of 2.76 (= 2 × 21 + 4 × 21 + 6 × 21 ) per day.

Solution to Ex. 3.3 For random variables X1 , . . . ,Xn we have

X X
E[ Xi ] = E[Xi ]
i i

so that the mean arrival rate to the central computer is 10

15 . Since the av-
erage number in the central computer is 5, the average time in the central
computer is thus 7.5 seconds.
 187

Solution to Ex. 3.4

πk = (k + 1)ρk π0
= (k + 1)ρk (1 − ρ)2
2ρ
N =
1−ρ

Solution to Ex. 3.11 Introducing a speed vector ci , which describes the speed of
the server when i customers are present, allows one to write the solution to
both parts (a) and (b) using one formula:
K
n! Y ρni i
P [(n1 , . . . ,nK )] = P [(0, . . . ,0)] Qn ( ) (13.1)
i=1 ci i=1 ni !

where
λi
ρi =
µi
K
X
ρ = ρi (< 1, ergodicity condition)
i=1
XK
n = ni
i=1
∞
ρj
)−1
X
P [(0, . . . ,0)] = ( Qj
j=0 i=1 ci

For PS we have ci = 1,∀i, giving:

K
Y ρni i
P [(n1 , . . . ,nK )] = (1 − ρ)n!
i=1
ni !

and for the IS case ci = i,∀i, we have

K
ρni
P [(n1 , . . . ,nK )] = e−ρ
Y
i

i=1
ni !

The mean number of customers in the latter case is

∞
X ∞ X
X K
N= ... ( nj )P [(n1 , . . . ,nK )] = ρ
n1 =0 nK =0 j=1

Eq. 13.1 is also valid for an M/Cox/m queue by setting ci = min(i,m).

The reader should note that the steady state probability distribution of each
of these queues, is independent of the particular Coxian distribution of the
corresponding service time and only depends on the mean value.
188 13 Solutions to Selected Exercises

t2

p5

p1 p2
t1 t4 p4

p3

t3

Figure 13.1 Counterexample for “Boundedness ⇒ Pos. P-Invariant covering”

Solution to Ex. 5.11 The coverability tree is not sufficient to determine liveness.
Fig. 5.16 displays two Petri nets which have the same coverability tree. The
Petri net displayed on the left side is not live (e.g. M0 [t1 t2 > M 0 yields a
marking M 0 in which no transition is enabled) whereas the Petri net on the
right side is live.

Solution to Ex. 5.14 The safe Place-Transition net of Fig. 13.1 is not covered by
positive P-invariant, since there is no positive P-invariant covering place p5 .

Solution to Ex. 5.17

1. This statement follows directly from theorem 5.16, since a trap is

marked in every marking M covering the given marking M0 . This
property of an EFC-net is also known as the liveness monotonicity
property of EFC-nets.
2. Fig. 9.12 (cf. [150]) shows a counterexample. If the token at the place
in •t2 is removed the Petri net is live. With the displayed initial mark-
ing, which covers the former marking, the net is not live.

Solution to Ex. 5.19

1. Fig. 13.2 (cf. [39]) shows a counterexample.

2. Since the given proof of theorem 5.19 does not use the fact that the
Petri net is an EFC-net, this statement is valid as well. The EFC prop-
erty is needed for the converse implication.

Solution to Ex. 5.21 The first and the third algorithm are correct implementa-
tions. The fourth algorithm is an incorrect solution for the famous dining
 189

Figure 13.2 EFC-net whose initial marking is not a home state

philosopher problem. The system might reach a deadlock state, e.g. if all in-
volved processes perform their “P(sem[i])”-operation simultaneously. The
second algorithm is also incorrect, since both processes might enter the crit-
ical region simultaneously. One can see that as follows:
Statements of process P0 Statements of process P1
flag[1] := true;
turn = 0 =⇒ enter while loop
flag[0] = false
=⇒ skip inner while loop
flag[0] := true;
turn = 0
=⇒ enter critical region
turn := 1;
turn = 1 =⇒ skip while loop
enter critical region

Solution to Ex. 9.6 All conjectures do not hold, since we have the following
counterexamples:

1. see Fig. 9.10,

2. see Fig. 9.12,
3. see Fig. 9.14.

Solution to Ex. 9.8 This simple unbounded GSPN is quite interesting, since the
corresponding Markov process has no steady state distribution irrespective
of the values of λ and µ.
190 13 Solutions to Selected Exercises

λ λ λ λ λ λ λ λ λ λ

... -n -n+1 ... n ...

-1 0 1 ... n-1

µ µ µ µ µ µ µ µ µ µ

Figure 13.3 Markov chain for unbounded GSPN

The most important step proving this is in finding an appropriate state rep-
resentation for the infinite Markov chain. Since transition t3 is immedi-
ate, we have that whenever M (p1 ) > 0, M (p2 ) = 0 holds and similarly
M (p2 ) > 0 implies M (p1 ) = 0. Thus M (p1 ) − M (p2 ) is a sufficient state
descriptor and the potentially two dimensional Markov chain of the GSPN
can be represented in one dimension as shown in Fig. 13.3.
The global balance equations are given by

πi (λ + µ) = λπi−1 + µπi+1 i = −∞,...,∞ (13.2)

which rewrites as

λ(πi − πi−1 ) = µ(πi+1 − πi ) i = −∞,...,∞ (13.3)

If system (13.3) could be solved subject to 0 ≤ πi ≤ 1,∀i and ∞

P
i=−∞ πi =
1, then a stationary distribution would in fact exist and would be given by
this solution. However, it is not difficult to verify that such a solution cannot
exist.
Define τi = πi+1 − πi . With that Eq. (13.3) rewrites to
λ
τi = τi−1
µ
giving
λ
τi = ( )i τ0
µ
implying
λ
πi+1 − πi = ( )i (π1 − π0 ) i = −∞,...,∞
µ
We can now distinguish the following cases:
P∞
1) π1 = π0 implying πi+1 = πi ,∀i showing that i=−∞ πi = 1 can not be
satisfied.
2) π1 6= π0 . Consider
n n
X X λ
πn+1 − π−n = (πi+1 − πi ) = ( )i (π1 − π0 )
i=−n i=−n
µ
 191

If π1 > π0 we can conclude πn+1 > n(π1 −π0 ) contradicting πn+1 ≤

1 and assuming π1 < π0 implies πn+1 < n(π1 − π0 ) + 1 leading to a
further contradiction for arbitrarily large n, namely πn+1 < 0.1

The reader should note that all conclusions hold (and thus no steady state
distribution exists) irrespective of the values of λ and µ!
The unbounded GSPN of Fig. 9.20 can, e.g., be used to describe the core
transshipping activities in logistic nodes (cf.[24]). Vehicles arrive at a rate
of λ delivering a unit to the store (p1 ) of the logistic node, whereas vehicles
wanting to load a unit are arriving at a rate of µ and have to wait (p2 ) when
their demands cannot be satisfied because of an empty storage. The time for
loading and unloading etc. is neglected in the model.
In [23] this effect is described in more detail and it is also shown that a
simulation of such a system runs the immanent danger of leaving this effect
uncovered and assuming a stationary process.

1
Pn
Note that in i=−n ( µλ )i . . . we have n terms with a factor of α1 := µλ and similarly n terms
with the reciprocal factor α2 := µλ . Thus α1 ≥ 1 or α2 ≥ 1 holds irrespective of the values of
λ and µ.
192 13 Solutions to Selected Exercises
 193

Bibliography
[1] G.A. Agha, F. De Cindio, and G. Rozenberg, editors. Concurrent Object-
Oriented Programming and Petri Nets, Advances in Petri Nets 2001, vol-
ume 2001. Lecture Notes in Computer Science, Springer-Verlag, 2001.
[2] M. Ajmone-Marsan. Stochastic Petri nets: An elementary introduction. In
Advances in Petri Nets, pages 1–29. Lecture Notes in Computer Science,
Vol. 424, Springer-Verlag, 1989.
[3] M. Ajmone-Marsan, editor. Proceedings of the 14th International Con-
ference on Application and Theory of Petri Nets, Chicago (USA), volume
691. Lecture Notes in Computer Science, Springer-Verlag, June 1993.
[4] M. Ajmone-Marsan, G. Balbo, A. Bobbio, G. Chiola, G. Conte, and
A. Cumani. On Petri nets with stochastic timing. In Proceedings of the
International Workshop on Timed Petri Nets, Torino, pages 80–87, 1985.
[5] M. Ajmone-Marsan, G. Balbo, A. Bobbio, G. Chiola, and A. Cumani. The
effect of execution policies in the semantics and analysis of Stochastic
Petri nets. IEEE Transactions on Software Engineering, 15(7):832–846,
1989.
[6] M. Ajmone-Marsan, G. Balbo, and G. Conte. Performance Models of Mul-
tiprocessor Systems. MIT Press Series in Computer Science, 1986.
[7] M. Ajmone-Marsan, G. Balbo, G. Conte, S. Donatelli, and G. Franceschi-
nis. Modelling with Generalized Stochastic Petri Nets. Wiley Series in
Parallel Computing. John Wiley and Sons, 1995.
[8] M. Ajmone-Marsan and G. Chiola. On Petri nets with deterministic and
exponential transition firing times. In Proceedings of the 7th European
Workshop on Application and Theory of Petri Nets, Oxford, pages 151–
165, 1986.
[9] M. Ajmone-Marsan and G. Chiola. On Petri nets with deterministic and
exponentially distributed firing times. In G. Rozenberg, editor, Concur-
rency and Nets. Advances in Petri Nets, pages 132–145. Springer-Verlag,
1987.
[10] M. Ajmone-Marsan and G. Chiola. Improving the efficiency of the anal-
ysis of DSPN models. In Proceedings of the 9th European Workshop on
Application and Theory of Petri Nets, Venice, pages 183–201, 1988.
[11] M. Ajmone-Marsan, G. Conte, and G. Balbo. A class of Generalised
Stochastic Petri Nets for the performance evaluation of multiprocessor sys-
tems. ACM Transactions on Computer Systems, 2:93–122, 1984.
194 13 Bibliography

[12] P. Azema, editor. Proceedings of the 18th International Conference on

Application and Theory of Petri Nets, Toulouse (France), volume 1248.
Lecture Notes in Computer Science, Springer-Verlag, June 1997.

[13] G. Balbo, S.C. Bruell, and M. Sereno. Arrival theorems for product-form
stochastic Petri nets. In SIGMETRICS, pages 87–97, May 1994.

[14] S. Balsamo and V. de Nitto-Persone. A survey of product-form queue-

ing networks with blocking and their equivalences. Annals of Operation
Research, 48:31–61, 1994.

[15] F. Baskett, K. M. Chandy, R. R. Muntz, and F. G. Palacios. Open, closed

and mixed networks of queues with different classes of customers. In
Journal of the ACM 22, pages 248–260, 1975.

[16] B. Baumgarten. Petri-Netze. Grundlagen und Anwendungen. BI Wis-

senschaftsverlag, Mannheim, 1990.

[17] F. Bause. Funktionale Analyse zeitbehafteter Petri-Netze. DeutscherUni-

versitätsVerlag, 1992.

[18] F. Bause. Queueing Petri nets: a formalism for the combined qualita-
tive and quantitative analysis of systems. In Proceedings of the 5th In-
ternational Workshop on Petri Nets and Performance Models, Toulouse
(France), pages 14–23. IEEE, October 1993.

[19] F. Bause. Petri nets and priorities. Technical Report 564, Fachbereich
Informatik der Universität Dortmund (Germany), 1995.

[20] F. Bause. On the Analysis of Petri Nets with Static Priorities. Acta Infor-
matica, 33(7):669–685, 1996.

[21] F. Bause. Analysis of Petri Nets with a dynamic priority method. In

P. Azema and G. Balbo, editors, Proc. of 18th International Conference on
Application and Theory of Petri Nets, Toulouse, France, June 1997, vol-
ume LNCS 1248, pages 215–234, Berlin, Germany, June 1997. Springer-
Verlag.

[22] F. Bause and H. Beilner. Eine Modellwelt zur Integration von

Warteschlangen- und Petri-Netz-Modellen. In Proceedings of the 5th
GI/ITG-Fachtagung, Messung, Modellierung und Bewertung von Rechen-
systemen und Netzen, pages 190–204. Braunschweig, Gesellschaft für In-
formatik (GI), Germany, September 1989.

[23] F. Bause and H. Beilner. Intrinsic problems in simulation of logistic net-

works. In 11th European Simulation Symposium and Exhibition (ESS99),
Simulation in Industry, Erlangen (Germany), pages 193–198, October
1999.
 195

[24] F. Bause and H. Beilner. A short note on synchronisation in open systems.

In Petri Net Newsletters, No. 57, pages 9–12. Special Interest Group On
Petri Nets and Related System Models, Gesellschaft für Informatik (GI),
Germany, 1999.

[25] F. Bause and P. Buchholz. Aggregation and disaggregation in product

form Queueing Petri nets. In Proceedings of the Seventh International
Workshop on Petri Nets and Performance Models, June 3-6, 1997, Saint
Malo, France, pages 16–25. IEEE Computer Society, 1997.

[26] F. Bause and P. Buchholz. Queueing Petri nets with product form solution.
Performance Evaluation, 32(4):265–299, 1998.

[27] F. Bause, P. Buchholz, and P. Kemper. QPN-Tool for the specification and
analysis of hierarchically combined Queueing Petri Nets. In H. Beilner
and F. Bause, editors, Quantitative Evaluation of Computing and Com-
munication Systems, volume LNCS 977, pages 224–238. Springer-Verlag,
Berlin, 1995.

[28] F. Bause and H. Beilner (eds.). Performance tools - model interchange for-
mats. Technical Report 581, Fachbereich Informatik der Universität Dort-
mund (Germany); also presented at the Joint International Conference on
“Modelling Techniques and Tools for Computer Performance Evaluation”
and “Measuring, Modelling and Evaluating Computing and Communica-
tion Systems”, Sept. 1995, Heidelberg (Germany), 1995.

[29] F. Bause and P. Kemper. Queueing Petri nets. In Proceedings of the 3rd
Fachtagung Entwurf komplexer Automatisierungssysteme, Braunschweig.
Technische Universität Braunschweig, May 1993.

[30] F. Bause and P. Kemper. QPN-Tool for the qualitative and quantitative
analysis of Queueing Petri Nets. In G. Haring and G. Kotsis, editors, Proc.
of the 7th International Conference on Computer Performance Evaluation,
Modelling Techniques and Tools, Vienna (Austria), LNCS 794, pages 321–
334. Springer-Verlag, Berlin, 1994.

[31] G. Berthelot. Checking properties of nets using transformations. In Ad-

vances in Petri Nets, pages 19–40. Springer-Verlag, 1985.

[32] G. Berthelot. Transformations and decompositions of nets. In Petri Nets:

Central Models and Their Properties, Advances in Petri Nets, Part I, pages
359–376. Springer-Verlag, 1986.

[33] B. Berthomieu and M. Menasche. A state enumerative approach for an-

alyzing time Petri nets. In W. Brauer, editor, Applications and Theory of
Petri Nets. Selected Papers from the 3rd European Workshop on Applica-
tions and Theory of Petri Nets, Varenna. Informatik-Fachberichte. No. 66,
1982.
196 13 Bibliography

[34] E. Best. Structure theory of Petri nets: The free choice hiatus. In Advances
in Petri Nets, Part I, pages 168–205. Springer-Verlag, 1986.

[35] E. Best and J. Desel. Partial order behaviour and structure of Petri nets.
Technical Report 373, Arbeitspapiere der GMD, Gesellschaft für Mathe-
matik und Datenverarbeitung, Sankt Augustin, Germany, 1989.

[36] E. Best, J. Desel, and J. Esparza. Traps characterize home states in free
choice systems. Theoretical Computer Science, 101:161–176, 1992.

[37] E. Best and M.W. Shields. Some equivalence results for free choice nets
and simple nets on the periodicity of live free choice nets. In Proceedings
of the 8th Colloquium on Trees and Algebra and Programming, pages 141–
154. Lecture Notes in Computer Science, 159, 1983.

[38] E. Best and P.S. Thiagarajan. Some classes of live and safe Petri nets. In
Advances in Petri Nets, pages 71–94. Springer-Verlag, 1987.

[39] E. Best and K. Voss. Free choice systems have home states. Acta Infor-
matica, 21:89–100, 1984.

[40] J. Billington, editor. Proceedings of the 17th International Conference

on Application and Theory of Petri Nets, Osaka (Japan), volume 1091.
Lecture Notes in Computer Science, Springer-Verlag, June 1996.

[41] J. Billington, editor. Application of petri nets to communication networks,

Advances in Petri Nets 1999, volume 1605. Lecture Notes in Computer
Science, Springer-Verlag, 1999.

[42] G. Bolch, S. Greiner, H. de Meer, and K.S. Trivedi. Queueing Networks

and Markov Chains. John Wiley and Sons, 1998.

[43] R.J. Boucherie. A characterisation of independence for competing markov

chains with applications to stochastic Petri nets. In in [138], pages 117–
126, 1993.

[44] W. Brauer, W. Reisig, and G. Rozenberg, editors. Petri Nets: Applications

and Relationships to Other Models of Concurrency. Advances in Petri Nets
1986, Part II, volume 255. Lecture Notes in Computer Science, Springer-
Verlag, 1987.

[45] W. Brauer, W. Reisig, and G. Rozenberg, editors. Petri Nets: Central Mod-
els and Their Properties. Advances in Petri Nets 1986, Part I, volume 254.
Lecture Notes in Computer Science, Springer-Verlag, 1987.

[46] S.L. Brumelle. A generalization of L = λW to moments of queue length

and waiting times. Operations Research, 20:1127–1136, 1972.
 197

[47] P. Buchholz and P. Kemper. APNNed – a net editor and debugger within
the APNN toolbox. In J. Desel, P. Kemper, E. Kindler, and A. Oberweis,
editors, Forschungsbericht: 5. Workshop Algorithmen und Werkzeuge für
Petrinetze. Universität Dortmund, Fachbereich Informatik, 1998.

[48] P.J. Burke. The output of a queueing system. Operations Research, 4:699–
704, 1956.

[49] M.L. Chaudhry and J.G.C. Templeton. A First Course in Bulk Queues.
John Wiley and Sons, 1983.

[50] G. Chiola. A graphical Petri net tool for performance analysis. In Proceed-
ings of the 8th European Workshop on Application and Theory of Petri
Nets, Zaragoza (Spain), pages 317–331, 1987.

[51] G. Chiola. GreatSPN 1.5 software architecture. Computer Performance

Evaluation, pages 121–136, 1992.

[52] G. Ciardo and A. S. Miner. A data structure for the efficient kronecker so-
lution of GSPNs. In Proc. 8th Int. Workshop on Petri Net and Performance
Models (PNPM’99), 8-10 October 1999, Zaragoza (Spain), pages 22–31,
1999.

[53] E. Cinlar. Introduction to Stochastic Processes. Prentice Hall, 1975.

[54] J.-M. Colom and M. Koutny, editors. Proceedings of the 22nd Inter-
national Conference on Application and Theory of Petri Nets, Newcas-
tle upon Tyne (UK), volume 2075. Lecture Notes in Computer Science,
Springer-Verlag, June 2001.

[55] J.E. Coolahan and N. Roussopoulos. A timed Petri net methodology for
specifying real-time system requirements. In Proceedings of the Interna-
tional Workshop on Timed Petri Nets, Turino, pages 24–31, 1985.

[56] P.J. Courtois and P. Semal. Bounds for transient characteristics of large
or infinite Markov chains. In W.J. Stewart, editor, Numerical Solutions of
Markov Chains, pages 413–434. Marcel Dekker, Inc., 1990.

[57] D.R. Cox. A use of complex probabilities in the theory of complex pro-
cesses. Proc. Cambridge Phil. Soc., 51:313–319, 1955.

[58] G. de Michelis and M. Diaz, editors. Proceedings of the 16th International

Conference on Application and Theory of Petri Nets, Turin (Italy), volume
935. Lecture Notes in Computer Science, Springer-Verlag, June 1995.

[59] J. Desel, editor. Proceedings of the 19th International Conference on Ap-

plication and Theory of Petri Nets, Lisboa (Portugal), volume 1420. Lec-
ture Notes in Computer Science, Springer-Verlag, June 1998.
198 13 Bibliography

[60] J. Desel and J. Esparza. Free Choice Petri Nets. Cambridge University
Press, 1995.

[61] M. Diaz. Petri net based models in the specification and verification of pro-
tocols. In G. Rozenberg, editor, Petri Nets: Applications and Relationship
to Other Models of Concurrency. Advances in Petri Nets. Proceedings of
an Advanced Course, Bad Honnef, pages 135–170. Lecture Notes in Com-
puter Science, Vol. 255, Springer-Verlag, 1986.

[62] Stephen R. Donaldson. Algorithms and complexity of Petri net transfor-

mations. Master’s thesis, Department of Computer Science, University of
Cape Town, 1993.

[63] S. Donatelli. Superposed generalized stochastic petri nets: definition and

efficient solution. In R. Valette, editor, Lecture Notes in Computer Sci-
ence; Application and Theory of Petri Nets 1994, Proceedings 15th In-
ternational Conference, Zaragoza, Spain, volume 815, pages 258–277.
Springer-Verlag, 1994.

[64] S. Donatelli, editor. Proceedings of the 20th International Conference

on Application and Theory of Petri Nets, Williamsburg, Virginia, (USA),
volume 1639. Lecture Notes in Computer Science, Springer-Verlag, June
1999.

[65] S. Donatelli and M. Sereno. On the product form solution for stochastic
Petri nets. In Proceedings of the 13th International Conference on Appli-
cation and Theory of Petri Nets, Sheffield (UK), pages 154–172, 1992.

[66] J.B. Dugan. Extended Stochastic Petri Nets: Applications and Analysis.
PhD thesis, Department of Electrical Engineering, Duke University, 1984.

[67] H. Ehrig, G. Juhas, J. Padberg, and G. Rozenberg, editors. Unifying Petri

Nets, Advances in Petri Nets 2001, volume 2128. Lecture Notes in Com-
puter Science, Springer-Verlag, 2001.

[68] S. Eilon. A simpler proof of L = λW . Operations Research, 17:915–916,

1969.

[69] J. Esparza. Synthesis rules for Petri nets, and how they lead to new re-
sults. In J.C.M. Baeten and J.W. Klop, editors, Proceedings of CON-
CUR’90: Theories of Concurrency: Unification and Extension, pages 183–
198. Springer-Verlag, Berlin, 1990.

[70] J. Esparza. Reduction and synthesis of live and bounded free choice nets.
Technical report, Hildesheimer Informatikberichte, Institut für Informatik,
Universität Hildesheim, 1991.
 199

[71] J. Esparza and C. Lakos, editors. Proceedings of the 23nd International

Conference on Application and Theory of Petri Nets, Adelaide (Australia),
volume 2360. Lecture Notes in Computer Science, Springer-Verlag, June
2002.

[72] J. Esparza and M. Silva. Top-down synthesis of live and bounded free
choice nets. In Proceedings of the 1tth International Conference on the
Application and Theory of Petri Nets, Paris, pages 63–83, 1990.

[73] F. Feldbrugge. Petri Net Tool Overview 1989, chapter in [157], pages 151–
178. Lecture Notes in Computer Science, Springer-Verlag, 1990.

[74] F. Feldbrugge. Petri Net Tool Overview 1992, chapter in [161], pages 169–
209. Lecture Notes in Computer Science, Springer-Verlag, 1993.

[75] W. Feller. An Introduction to Probability Theory and its Applications,

volume 1. Academic Press, 1968.

[76] G. Balbo G. Chiola, M. Ajmone-Marsan and G. Conte. Generalized

stochastic Petri nets: A definition at the net level and its implications. IEEE
Transactions on Software Engineering, 19(2):89–107, February 1993.

[77] E. Gelenbe and G. Pujolle. Introduction to Queueing Networks. John

Wiley & Sons, 1987.

[78] H.J. Genrich and K. Lautenbach. Synchronisationsgraphen. Acta Infor-

matica, 2:143–161, 1973.

[79] H.J. Genrich and K. Lautenbach. System modelling with high level Petri
nets. Theoretical Computer Science, pages 109–136, 1981.

[80] H.J. Genrich and P.S. Thiagarajan. A theory of bipolar synchronization

schemes. Theoretical Computer Science, 30:241–318, 1984.

[81] R. German. New results for the analysis of deterministic and stochastic
Petri nets. In Proc. IEEE International Performance and Dependability
Symposium, Erlangen (Germany), pages 114–123, 1995.

[82] R. German. Performance Analysis of Communication Systems: Modeling

with Non-Markovian Stochastic Petri Nets. John Wiley and Sons, 2000.

[83] R. German, C. Kelling, A. Zimmermann, and G. Hommel. TimeNET: a

toolkit for evaluating non-markovian stochastic Petri nets. Performance
Evaluation, 1995.

[84] R. German and C. Lindemann. Analysis of stochastic Petri nets by the

method of supplementary variables. Performance Evaluation, 20:317–
335, 1994.
200 13 Bibliography

[85] R. German and J. Mitzlaff. Transient analysis of deterministic and stochas-

tic Petri nets with TimeNET. In H. Beilner and F. Bause, editors, Quantita-
tive Evaluation of Computing and Communication Systems, volume LNCS
977, pages 209–223. Springer-Verlag, Berlin, 1995.
[86] R. Goodman. Introduction to Stochastic Models. Benjamin/Cummings
Publishing Company, Inc, 1988.
[87] W.J. Gordon and G.F. Newell. Closed queueing networks with exponential
servers. Operations Research, 15(2):254–265, 1967.
[88] S. Haddad. Generalization of reduction theory to coloured nets. In Pro-
ceedings of the 9th European Workshop on Application and Theory of Petri
Nets, Venice, 1988.
[89] W. Henderson and P.G. Taylor. Embedded processes in stochastic Petri
nets. IEEE Transactions on Software Engineering, 17:108–116, 1991.
[90] R.A. Howard. Dynamic Probabilistic Systems. Volume I: Markov Models.
Volume II: Semi-Markov and Decision Processes. John Wiley & Sons,
Inc., 1971.
[91] J.R. Jackson. Networks of waiting lines. Oper. Res., 5:518–521, 1957.
[92] N.K. Jaiswal. Priority queues. Academic Press, 1968.
[93] K. Jensen. Coloured Petri nets and the invariant method. Mathematical
Foundations on Computer Science, Lecture Notes in Computer Science,
118:327–338, 1981.
[94] K. Jensen. Coloured Petri nets. In G. Rozenberg, editor, Petri Nets: Cen-
tral Models and Their Properties. Advances in Petri Nets. Proceedings of
an Advanced Course Lecture Notes in Computer Science, volume 254, No.
1, pages 248–299. Springer-Verlag, Berlin, 1986.
[95] K. Jensen, editor. Proceedings of the 13th International Conference on
Application and Theory of Petri Nets, Sheffield (UK), volume 616. Lecture
Notes in Computer Science, Springer-Verlag, June 1992.
[96] K. Jensen. Coloured Petri Nets; Basic Concepts, Analysis Methods and
Practical Use. EATCS Monographs on Theoretical Computer Science,
Vol. 1 Basic Concepts, 1997.
[97] K. Jensen. Coloured Petri Nets; Basic Concepts, Analysis Methods and
Practical Use. EATCS Monographs on Theoretical Computer Science,
Vol. 2 Analysis Methods, 1997.
[98] K. Jensen. Coloured Petri Nets; Basic Concepts, Analysis Methods and
Practical Use. EATCS Monographs on Theoretical Computer Science,
Vol. 3 Practical Use, 1997.
 201

[99] K. Jensen and G. Rozenberg, editors. High-level Petri Nets. Theory and
Application. Springer-Verlag, 1991.
[100] J.W. Jewell. A simple proof of L = λW . Operations Research, 15:1109–
1116, 1967.
[101] R. Johnsonbaugh and T. Murata. Additional methods for reduction and
expansion of marked graphs. IEEE Transactions on Circuit and Systems,
28(10):1009–1014, 1981.
[102] K. Kant. Introduction to Computer System Performance Evaluation.
McGraw-Hill, Inc., 1992.
[103] J.G. Kemeny and J.L. Snell. Finite Markov Chains. Springer-Verlag, New
York, 1960.
[104] P. Kemper. Linear time algorithm to find a minimal deadlock in a strongly
connected free-choice net. In M. Ajmone-Marsan, editor, Proceedings
of the 14th International Conference on Application and Theory of Petri
Nets. Springer-Verlag, Berlin, 1993.
[105] P. Kemper. Numerical analysis of superposed GSPNs. IEEE Transactions
on Software Engineering, 22(9):615–628, September 1996.
[106] P. Kemper. Reachability analysis based on structured representations. In
Lecture Notes in Computer Science; Proc. 17th International Conference
in Application and Theory of Petri Nets (ICATPN’96), Osaka, Japan, vol-
ume 1091, pages 269–288. Springer-Verlag, June 1996.
[107] P. Kemper and F. Bause. An efficient polynomial-time algorithm to decide
liveness and boundedness of free choice nets. In K. Jensen, editor, Pro-
ceedings of the 13th International Conference on Application and Theory
of Petri Nets, volume LNCS 616, pages 263–278. Springer-Verlag, Berlin,
1992.
[108] L. Kleinrock. Queueing Systems. Volume 1: Theory. John Wiley and Sons,
1975.
[109] R. Kosaraju. Decidability of reachability in vector addition systems. In
Proceedings of the 14th Annual ACM Symposium on Theory of Computing,
pages 267–280, 1982.
[110] R. Lal and U.N. Bhat. Reduced systems in Markov chains and their appli-
cations in queueing theory. Queueing Systems, 2:147–172, 1987.
[111] A.A. Lazar and T.G. Robertazzi. Markovian Petri net protocols with prod-
uct form solution. Performance Evaluation, 12:67–77, 1991.
[112] C. Lindemann. DSPNexpress: a software package for the efficient solution
deterministic and stochastic Petri nets. Performance Evaluation, 1995.
202 13 Bibliography

[113] C. Lindemann. Performance Modelling with Deterministic and Stochastic

Petri Nets. John Wiley & Sons, 1998.

[114] C. Lindemann, A. Thümmler, A. Klemm, M. Lohmann, and O.P. Wald-

horst. Quantitative system evaluation with dspnexpress 2000. In Proc.
2nd Int. Workshop on Software and Performance, Ottawa (Canada), pages
12–17, 2000.

[115] J.D.C. Little. A simple proof of L = λW . Operations Research, 9:383–

387, 1961.

[116] K. Marshall and R. Wolff. Customer average and time average - queue
lengths and waiting times. J. of Appl. Probability, 8:535–542, 1971.

[117] J. Martinez and M. Silva. A simple and fast algorithm to obtain all in-
variants of a generalized Petri net. In Application and Theory of Petri
Nets, Selected Papers from the First and Second European Workshop on
Application and Theory of Petri Nets. Informatik-Fachberichte, No. 52,
Springer-Verlag, 1981.

[118] E.W. Mayr. An algorithm for the general Petri net reachability problem.
In Proc. of the 13th Annual ACM Symp. on Theory of Computing, pages
238–246, 1981.

[119] E.W. Mayr. An algorithm for the general Petri net reachability problem.
SIAM, Journal Comput., 13(3):441–460, August 1984.

[120] J. McKenna. A generalization of littles’s law to moments of queue lengths

and waiting times in closed, product-form queueing networks. J. Appl.
Prob., 26:121–133, 1989.

[121] P. Merlin. A Study of the Recoverability of Computing Systems. PhD

thesis, Department of Information and Computer Science, University of
California, Irvine (USA), 1974.

[122] R. Milner, R. Harper, and M. Tofte. The definition of standard ML. MIT
Press, 1990.

[123] A.S. Miner and G. Ciardo. Efficient reachability set generation and storage
using decision diagrams. In Proc. of the 20th Int. Conf. on Application
and Theory of Petri Nets 1999, Williamsburg, VA (USA). Lecture Notes in
Computer Science Vol. 1639, pages 6–25, 1999.

[124] M.K. Molloy. On the Intergration of Delay and Throughput Measures

in Distributed Processing Models. PhD thesis, University of California,
1981.

[125] T. Murata. Petri nets: Properties, analysis and applications. Proceedings

of IEEE, 77(4):541–580, April 1989.
 203

[126] T. Murata and J.Y. Koh. Reduction and expansion of live and safe marked
graphs. IEEE Transactions on Circuit and Systems, 27(1):68–71, 1980.

[127] S. Natkin. Les Reseaux de Petri Stochastiques et leur Application a

l’Evaluation des Systemes Informatiques. PhD thesis, CNAM, Paris, 1980.

[128] M. Nielsen, editor. Proceedings of the 21st International Conference on

Application and Theory of Petri Nets, Aarhus (Denmark), volume 1825.
Lecture Notes in Computer Science, Springer-Verlag, June 2000.

[129] A. Pagnoni. Stochastic nets and performance evaluation. In Petri Nets:

Central Models and Their Properties. Advances in Petri Nets, pages 460–
478. Lecture Notes in Computer Science, No. 254, 1986.

[130] J.L. Peterson. Petri nets. ACM Computing Surveys, 9:242–252, September
1977.

[131] J.L. Peterson. A note on coloured Petri nets. Information Processing

Letters, 11:40–43, Aug. 1980.

[132] J.L. Peterson. Petri Nets and the Modelling of Systems. MIT Press Series
in Computer Science, 1981.

[133] C.A. Petri. Kommunikation mit Automaten. PhD thesis, Universität Bonn,
1962.

[134] Proceedings of the International Workshop on Timed Petri Nets, Turino

(Italy). IEEE Computer Society Press, July 1985.

[135] Proceedings of the 2nd International Workshop on Petri Nets and Perfor-
mance Models, Madison (USA). IEEE Computer Society Press, 1987.

[136] Proceedings of the 3rd International Workshop on Petri Nets and Perfor-
mance Models, Kyoto (Japan). IEEE Computer Society Press, 1989.

[137] Proceedings of the 4th International Workshop on Petri Nets and Per-
formance Models, Melbourne (Australia). IEEE Computer Society Press,
1991.

[138] Proceedings of the 5th International Workshop on Petri Nets and Perfor-
mance Models, Toulouse (France). IEEE Computer Society Press, 1993.

[139] Proceedings of the 6th International Workshop on Petri Nets and Perfor-
mance Models, Durham (USA). IEEE Computer Society Press, 1995.

[140] Proceedings of the 7th International Workshop on Petri Nets and Perfor-
mance Models, Saint Malo (France). IEEE Computer Society Press, 1997.

[141] Proceedings of the 8th International Workshop on Petri Nets and Perfor-
mance Models, Zaragoza (Spain). IEEE Computer Society Press, 1999.
204 13 Bibliography

[142] Proceedings of the 9th International Workshop on Petri Nets and Perfor-
mance Models, Aachen (Germany). IEEE Computer Society Press, 2001.

[143] C.V. Ramamoorthy and G.S. Ho. Performance evaluation of asynchronous

concurrent systems using Petri nets. IEEE Transactions on Software En-
gineering, 6(5):440–449, 1980.

[144] C. Ramchandani. Analysis of Asynchronous Concurrent Systems by Timed

Petri Nets. PhD thesis, Department of Electrical Engineering, MIT (USA),
July 1973.

[145] M. Rauhamaa. A comparative study of methods for efficient reachability

analysis. Technical Report Research Report, Helsinki University of Tech-
nology, September 1990.

[146] M. Raynal. Algorithms for mutual exclusion. MIT Press, 1986.

[147] R.R. Razouk and C.V. Phelps. Performance analysis using Petri nets. In
Proceedings of the IFIP WG 6.1 4th International Workshop on Proto-
col, Specification, Testing, and Verification, pages 561–576, Skytop Lodge
(USA), 1984.

[148] W. Reisig. Petri nets with individual tokens. In Applications and Theory of
Petri Nets. Selected Papers from the 3rd European Workshop on Applica-
tions and Theory of Petri Nets, Varenna (Italy). Informatik-Fachberichte,
No. 66, Springer-Verlag, 1982.

[149] W. Reisig. Petri Nets. An Introduction, volume 4. EATCS Monographs on

Theoretical Computer Science, Springer-Verlag, 1985.

[150] W. Reisig. Petrinetze. Eine Einführung. Springer Verlag, 1986.

[151] T.G. Robertazzi. Computer Networks and Systems: Queueing Theory and
Performance Evaluation. Springer, 1990.

[152] S.M. Ross. Introduction to Probability Models. Academic Press, fourth

edition, 1989.

[153] G. Rozenberg, editor. Advances in Petri Nets 1984, volume 188. Lecture
Notes in Computer Science, Springer-Verlag, 1985.

[154] G. Rozenberg, editor. Advances in Petri Nets 1985, volume 222. Lecture
Notes in Computer Science, Springer-Verlag, 1986.

[155] G. Rozenberg, editor. Advances in Petri Nets 1987, volume 266. Lecture
Notes in Computer Science, Springer-Verlag, 1987.

[156] G. Rozenberg, editor. Advances in Petri Nets 1988, volume 340. Lecture
Notes in Computer Science, Springer-Verlag, 1988.
 205

[157] G. Rozenberg, editor. Advances in Petri Nets 1989, volume 424. Lecture
Notes in Computer Science, Springer-Verlag, 1990.

[158] G. Rozenberg, editor. Advances in Petri Nets 1990, volume 483. Lecture
Notes in Computer Science, Springer-Verlag, 1991.

[159] G. Rozenberg, editor. Advances in Petri Nets 1991, volume 524. Lecture
Notes in Computer Science, Springer-Verlag, 1991.

[160] G. Rozenberg, editor. Advances in Petri Nets 1992, volume 609. Lecture
Notes in Computer Science, Springer-Verlag, 1992.

[161] G. Rozenberg, editor. Advances in Petri Nets 1993, volume 674. Lecture
Notes in Computer Science, Springer-Verlag, 1993.

[162] G. Rozenberg, editor. Advances in Petri Nets 1998, volume 1499. Lecture
Notes in Computer Science, Springer-Verlag, 1998.

[163] V.M. Savi and X. Xie. Liveness and boundedness analysis for Petri nets
with event graph modules. In Proceedings of the 13th International Con-
ference on Application and Theory of Petri Nets, pages 328–347. Springer-
Verlag, 1992.

[164] M. Sereno and G. Balbo. Computational algorithms for product form solu-
tion stochastic Petri nets. In Proceedings of the 5th International Workshop
on Petri Nets and Performance Models, Toulouse (France), pages 98–107.
IEEE Computer Society Press, Oct. 1993.

[165] M. Sereno and G. Balbo. Mean value analysis of stochastic petri nets.
Performance Evaluation, 29(1):35–62, 1997.

[166] J. Sifakis. Use of Petri nets for performance evaluation. In Measuring,

Modeling and Evaluation of Computer Systems, Proceedings of the 3rd In-
ternational Workshop on Modeling and Performance Evaluation of Com-
puter Systems, Amsterdam, pages 75–93, 1977.

[167] J. Sifakis. Performance evaluation of systems using Petri nets. In

W. Brauer, editor, Application and Theory of Petri Nets, Selected Papers
from the First and Second European Workshop on Application and Theory
of Petri Nets, pages 307–319. Informatik-Fachberichte, No. 52, 1980 and
1981.

[168] Y. Souissi and G. Memmi. Compositions of nets via a communication

medium. In Proceedings of the 10th Workshop on Application and Theory
of Petri Nets, Bonn (Germany), pages 292–311, 1989.

[169] W.J. Stewart. Introduction to the numerical solution of Markov chains.

Princeton University Press, 1994.
206 13 Bibliography

[170] S. Stidham. A last word on L = λW . Operations Research, 22:417–421,

1974.

[171] I. Suzuki and T. Murata. Stepwise refinement of transitions and places. In

W. Brauer, editor, Application and Theory of Petri Nets. Selected Papers of
the First and Second Workshop on Application and Theory of Petri Nets,
pages 136–141. Informatik-Fachberichte, No. 52, 1980.

[172] I. Suzuki and T. Murata. A method of stepwise refinement and abstraction

of Petri nets. Journal on Computing and System Sciences, 27(1), 1983.

[173] E. Teruel and M. Silva. Liveness and home states in equal conflict systems.
In Proceedings of the 14th International Conference on Application and
Theory of Petri Nets, Chicago (USA), pages 415–432, 1993.

[174] E. Teruel and M. Silva. Well-formedness of equal conflict systems. In

Proceedings of the 15th European Workshop on Application and Theory
of Petri Nets, Zaragoza (Spain), pages 491–510, 1994.

[175] P.S. Thiagarajan and K. Voss. In praise of free choice nets. In Advances in
Petri Nets, pages 438–454. Springer-Verlag, 1984.

[176] R. Valette. Analysis of Petri nets by stepwise refinement. Journal of Com-

puter and Systems Sciences, 18:35–46, 1979.

[177] R. Valette, editor. Proceedings of the 15th International Conference on Ap-

plication and Theory of Petri Nets, Zaragoza (Spain), volume 815. Lecture
Notes in Computer Science, Springer-Verlag, July 1994.

[178] W. Vogler. Live and bounded free choice nets have home states. In
Petri Net Newsletters, pages 18–21. Special Interest Group On Petri Nets
and Related System Models, Gesellschaft für Informatik (GI), Germany, 4
1989.

[179] I.Y. Wang and T.G. Robertazzi. Service stage Petri net models with prod-
uct form. Queueing Systems, 7:355–374, 1990.

[180] W. Witt. A review of L = λW and extensions. Queueing Systems, 9:235–

268, 1991. A correction note for this paper was published 1992 in Vol. 12,
pp. 431-432 of the same journal.

[181] C.Y. Wong, T.S. Dillon, and K.E. Forward. Timed place Petri nets with
stochastic representation of place time. In Proceedings of the International
Workshop on Timed Petri Nets, Turino, pages 96–103, 1985.

[182] S.S. Yau and C.R. Chou. Control flow analysis of distributed comput-
ing system software using structured Petri net models. In Proceedings
of the Workshop on Future Trends of Distributed Systems in the 1990s,
HongKong, pages 174–183, 9 1988.
 207

[183] A. Zimmermann, J. Freiheit, R. German, and G. Hommel. Petri net mod-

elling and performability evaluation with timenet 3.0. In 11th Int. Conf.
on Modelling Techniques and Tools for Computer Performance Evaluation
(TOOLS’2000), volume LNCS 1786, pages 188–202. Springer-Verlag,
2000.

[184] A. Zimmermann, R. German, J. Freiheit, and G. Hommel. Timenet 3.0

tool description. In Int. Conf. on Petri Nets and Performance Models
(PNPM’99), Zaragoza (Spain), 1999.

[185] W.M. Zuberek. Timed Petri nets and preliminary performance evaluation.
In Proceedings IEEE 7th Annual Symposium on Computer Architecture,
pages 89–96, 1980.

[186] W.M. Zuberek. Performance evaluation using extended Petri nets. In Pro-
ceedings of the International Workshop on Timed Petri Nets, Turino, pages
272–278, 1985.
208 13 Bibliography
Index 209

Index
ENT (M ), 139, 164 — MVA, 73
PC , 169 BCMP-type queues, 70
SM S , 117 BCMP-type queues, 72
TC , 169 Bernoulli distribution, 24
#, 112 Bernoulli random variable, 16
ω, 91 Binomial random variable, 16
nth central moment, 19 birth-death process, 61
nth moment, 19 boundedness, 87, 89
N, 23, 97 — in EFC-nets, 107
N0 , 23 — in GSPNs, 151
R, 117 — in marked graphs, 105
R+ , 140 — in QPNs, 170
Z, 97 — in state machines, 102
bulk arrival, 72
absorbing bulk service, 72
— chain, 35
— state, 29 central server model, 163
addition of multi-sets, 117 CGSPN, 158, 163
Advances in Petri Nets, 126 chain, 21
age memory, 129 Chapman-Kolmogorov equation, 26, 49, 50
algorithm for condition NoTT, 152 closure
analysis procedure for time-augmented Petri — reflexive and transitive, 83
nets, 149 coefficient of variation, 20
aperiodic state, 30 colour
Application and Theory of Petri Nets, 126 — multi-set, 116
application examples, 179 — of a token, 116
arc, 77 — of a transition, 116
— counter-alternate, 177 colour function, 118
— probabilistic, 177 Coloured GSPN, 158, 163
arrival rate, 57 Coloured Petri net, 116, 118
arrival theorem, 73 — colour function, 118
asymmetric choice net, 125 — expression representation, 121
atomic net, 112 — graphical representation, 121
atomicity of statements, 113 — meta language ML, 121
average value, 19 — Turing machine power of, 125
— with infinite number of colours, 125
backward incidence function, 82 combined qualitative and quantitative anal-
backward incidence matrix, 95 ysis, 148
bag, 117 complexity
— multi-set, 117 — of reachability problem, 113
batch arrival, 72 condition
batch service, 72 — EQUAL-Conflict, 156
Bayes’s theorem, 15 — EQUAL-Service, 172
BCMP network, 72 — NoTT, 150
— arrival theorem, 73 — NoTT for QPNs, 169
— convolution algorithm, 73 — of ergodicity, 63
210 Index

Condition-Event nets, 87 — memoryless property, 65

conditional probability, 14 — minimum of exponential distributions,
conference 69, 132
— Application and Theory of Petri Nets, — Poisson, 64
126 DQDB, 180
— Petri Nets and Performance Models, 178 DSPNs, 177
conflict, 80 DTMC, 25
confusion, 140
contact, 87 EC-net, 125
continuous random variable, 18 EFC-GSPN, 155
continuous time Markov process, 46 EFC-net, 101
convolution algorithm, 73 — liveness monotonicity property, 188
counter-alternate arc, 177 EFC-QPN, 171
coverability tree, 91 eliminating vanishing states on the fly, 151,
— inner node of the tree, 91 160
— leaves of the tree, 91 embedded Markov chain of a GSPN, 141
— state space explosion, 94 embedded Markov process, 41, 53
— strongly connected component, 93 empty firing sequence, 85
covered by empty set of places, 84
— cycles, 105 enabled transition, 78, 84
— P-components, 107 — in a GSPN, 149
— positive P-invariants, 99 — in a QPN, 169
— positive T-invariants, 100 enabling memory, 129
covering of markings, 91 equal conflict net, 125
Coxian distribution, 48, 172 ergodic Markov chain, 32, 52
CPN, 116, 118 ergodicity, 63
critical section, 113 ESPL-net, 101
CTMC, 46 ESPNs, 177
cumulative distribution function, 17 — counter-alternate arc, 177
customer classes, 70 — probabilistic arc, 177
customers, 55 exhaustive set of events, 13
cycle, 104 exponential function, 18
expression representation in CPNs, 121
deadlock, 106 extended conflict set, 140
deadlock/trap-property, 106 extended free choice net, 101
— dt-property, 106 extended simple net, 101
decidability Extended Stochastic Petri nets, 177
— of liveness, 113
— of reachability problem, 113, 125 fairness, 112
depository, 162 FC-net, 101
Deterministic Stochastic Petri nets, 177 FCFS, 159, 162, 169
discrete time Markov chain, 25 FCFS with blocking, 182
Distributed Queue Dual Bus, 180 finite MC, 33
distribution firing delay, 140
— Bernoulli, 24 firing policy, 129
— Coxian, 48, 172 — age memory, 129
— GSPN representation of a Coxian dis- — enabling memory, 129
tribution, 173 — resampling, 129
— Laplace transform, 48 firing sequence, 85
Index 211

— empty, 85 — throughput at an immediate transition,

firing transition, 78, 84 147
firing vector, 96 — timed transition, 139
firing weight, 141 — timeless trap, 145, 149
flow in = flow out, 62 — Turing machine power, 158
flow relation, 83 — vanishing marking, 141
folding of a Petri net, 121 — visit ratio, 142
forward incidence function, 82 — with a timeless trap, 150
forward incidence matrix, 95 — with no home states, 155
free choice net, 101 — with no live transition, 154
functional analysis, 129 geometric random variable, 16, 44
global balance equations, 52, 62
Generalized Stochastic Petri net graphical representation of a CPN, 121
— with no steady state distribution, 190 GSPN, 139, 140
Generalized Stochastic Petri net, 139, 140 guard of a transition, 121
— algorithm for condition NoTT, 152
— analysis procedure, 149 hierarchical analysis of QPNs, 175
— boundedness, 151 holding time, 41
— coloured, 158, 163 home states, 87
— combined qualitative and quantitative — in EFC-GSPNs, 156
analysis, 148 — in EFC-nets, 108
— condition EQUAL-Conflict, 156 — in EFC-QPNs, 173
— condition NoTT, 150 — in marked graphs, 105
— confusion, 140 — in state machines, 102
— describing scheduling strategies, 159 homogeneous Markov process, 23
— embedded Markov chain, 141
— enabled transition, 149 iff, 83
— extended conflict set, 140 immediate queueing place, 182
— extended free choice net, 155 immediate transition, 139, 141
— — liveness, 156 incidence functions of a Petri net, 82
— firing delay, 140 incidence matrix, 95
— firing weight, 141 incomplete reachability analysis, 125
— home states, 156 independent random variables, 14
— immediate transition, 139 infinite number of colours in CPNs, 125
— Kronecker based representation of gen- infinite server, 69
erator matrix, 175 infinitesimal generator, 52
— liveness, 153 — Kronecker based representation, 175
— modified analysis procedure, 156 infinitesimal rates, 49
— qualitative analysis, 148, 155 inhibitor arcs, 125
— quantitative analysis, 141 initial marking, 82
— random switch, 140 inner node of a tree, 91
— reduced embedded MC, 144 input elements, 83
— representation of a Coxian distribution, input places, 83
173 input transitions, 83
— semi-Markov process, 141 interarrival time, 56
— superposed, 175 interface description, 178
— switching distribution, 140 irreducible Markov chain, 29
— tangible marking, 141 IS, 69
isolated place, 77
212 Index

isolated transition, 77 — absorbing state, 29

isomorphic graphs, 135 — aperiodic state, 30
— birth-death process, 61
joint probability density function, 20 — Chapman-Kolmogorov equation, 26, 49,
50
Kendall notation, 57 — continuous time, 46
Kronecker based representation, 175 — discrete time, 25
— embedded, 41, 53
language of a Place-Transition net, 113 — embedded chain of a GSPN, 141
Laplace transform, 48 — ergodic, 32, 52, 63
last come first served pre-emptive resume, — finite, 33
161 — global balance equations, 52, 62
LCFSPR, 161 — holding time, 41
leaves of a tree, 91 — homogeneous, 23
Little’s law, 58, 66 — infinitesimal generator, 52
— moments, 60 — — Kronecker based representation, 175
liveness, 87, 89 — infinitesimal rates, 49
— decidability, 113 — interval transition probabilities, 42
— in EFC-GSPNs, 156 — irreducible, 29
— in EFC-nets, 106 — local balance equations, 52
— in EFC-QPNs, 173 — mean recurrence time, 30
— in GSPNs, 153 — numerical analysis, 136
— in marked graphs, 105 — of a SPN, 132
— in state machines, 101 — periodic state, 30, 32
— monotonicity property of EFC-nets, 188 — power method, 136
load/machine aspect, 179 — recurrent nonnull, 30, 31
local balance equations, 52 — recurrent null, 30, 31
— recurrent state, 30
M/M/1 queue, 65 — reduced embedded chain of a GSPN,
M/M/1-FCFS queue, 159 144
M/M/m queue, 68 — reduced embedded chain of a QPN, 164
mailing group, 126 — reducible, 33
mapping of processes onto processors, 179 — semi, 40
marked graph, 101 — — embedded Markov process, 44
— Synchronisationsgraph, 125 — — GSPN, 141
— T-graph, 125 — — steady state, 44
marked set of places, 84 — sojourn time, 32, 41, 47, 132
marking, 84 — state space explosion, 135
— covering, 91 — state transition diagramme, 28
— initial, 82 — stationary, 23
— of a QPN, 164 — stationary distribution, 31
— reachable, 86 — steady state distribution, 31, 51
— tangible, 141 — stochastic process, 21
— vanishing, 141 — time before absorption
Markov process — — mean, 37–39
— with no steady state distribution, 190 — — variance, 38, 39
Markov chain, 21, 23 — transient, 35
Markov process, 21, 23 — transient state, 30
— absorbing, 35, 36 — transition probability, 25
Index 213

— visit ratio, 32, 142, 166 — ω, 91

Markov property, 23 — Advances in Petri Nets, 126
mean, 19 — arc, 77
mean cycle time, 142, 166 — asymmetric choice net, 125
mean number of customers, 66 — atomic net, 112
mean number of tokens, 134 — backward incidence function, 82
mean number of visits, 142, 166 — backward incidence matrix, 95
mean recurrence time, 30 — boundedness, 87, 89
mean sojourn time, 142, 166 — colour function, 118
mean time before absorption, 37, 39 — coloured, 116, 118
mean value analysis, 73 — Condition-Event nets, 87
memoryless property, 65 — conference, 126
Merlin’s TTPNs, 177 — conflict, 80
meta language ML, 121 — confusion, 140
MG-net, 101 — contact, 87
— Synchronisationsgraph, 125 — coverability tree, 91
— T-graph, 125 — covered by cycles, 105
minimum cycle time, 176 — covered by P-components, 107
minimum of exponential distributions, 69, — covered by positive P-invariants, 99
132 — covered by positive T-invariants, 100
modified analysis procedure for time-augmented
— covering of markings, 91
Petri nets, 156 — cycle, 104
multi-set, 116 — deadlock, 106
— SM S , 117 — deadlock/trap-property, 106
— addition of, 117 — empty set of places, 84
— bag, 117 — enabled transition, 78, 84
— multiplication with ordinary number, — equal conflict net, 125
117 — extended free choice net, 101, 188
mutual exclusion problem, 113 — extended simple net, 101
mutually exclusive events, 14 — fairness, 112
mutually exclusive exhaustive events, 14 — firing sequence, 85
MVA, 73 — firing transition, 78, 84
— firing vector, 96
Non-Markovian Stochastic Petri Nets, 177 — flow relation, 83
numerical analysis of Markov processes, — folding of, 121
136 — forward incidence function, 82
— forward incidence matrix, 95
occurrence sequence, 85 — free choice net, 101
ordinary Petri nets, 77 — home state, 87
output elements, 83 — incidence function, 82
output places, 83 — incidence matrix, 95
output transitions, 83 — incomplete reachability analysis, 125
— inhibitor arcs, 125
P-component, 107 — initial marking, 82
P-invariant, 97 — input elements, 83
performance analysis, 129 — isolated elements, 77
periodic state, 30, 32 — language of, 113
persistence, 113 — liveness, 87, 89
Petri net, 77 — mailing group, 126
214 Index

— marked graph, 101 — transformation, 110

— marked set of places, 84 — transition, 77, 81
— marking, 84 — trap, 106
— occurrence sequence, 85 — tube, 106
— output elements, 83 — unfolding of a coloured, 120, 170
— P-component, 107 — vector addition system, 96
— P-invariant, 97 — weakly connected, 83
— persistence, 113 — WWW, 126
— place, 77, 81 Petri net standard, 126
— Place-Transition net, 81 Petri Nets and Performance Models (work-
— postset, 83 shop), 178
— preset, 83 place, 77, 81
— priorities, 125 — input, 83
— reachability graph, 90 — isolated, 77
— reachability problem, 113 — output, 83
— reachability set, 86 — queueing, 162
— reachability tree, 90 — redundant, 110
— reduction analysis, 110 Place-Transition net, 81
— reduction rule, 110 Poisson
— redundant place, 110 — distribution, 64
— redundant transition, 142 — process, 63
— reversibility, 113 postset, 83
— S-graph, 125 power method, 136
— S-invariant, 99 pre-empt resume priority, 159
— safe, 87 preemptive repeat, 70
— self-loop, 85 preemptive resume, 70
— simple cycle, 104 preselection models, 129
— simple net, 101 preset, 83
— siphon, 106 priorities between transitions, 125
— Special Interest Group, 126 priority service, 70
— standard, 126 probabilistic arc, 177
— state machine, 101 probability density function, 18
— Stelle, 99 probability mass function, 16
— strongly connected, 84, 89 probability of transition firings, 134
— stubborn set method, 125 processor sharing, 69
— subnet, 107 Product-form Queueing Network, 70
— subnet generated by X, 107 Product-form Queueing Network, 72
— symbolic model checking, 125 PS, 69
— symmetry method, 125
— synchronic distance, 113 QPN, 162
— Synchronisationsgraph, 125 qualitative analysis, 129
— synthesis analysis, 110 — of EFC-GSPNs, 155
— synthesis of, 111 — of EFC-QPNs, 170
— system, 82 — of GSPNs, 148
— T-graph, 125 — of QPNs, 169
— T-invariant, 100 quantitative analysis, 129
— terminal strongly connected component, — of GSPNs, 141
93 — of QPNs, 164
— token, 77 queue, 55, 162
Index 215

Queueing Petri net, 162 — unfinished work, 55

— boundedness, 170 — utilisation, 66
— condition EQUAL-Service, 172 — waiting line, 55
— condition NoTT, 169 — waiting time, 57
— depository, 162 — work-conserving, 72
— enabled transition, 169
— extended free choice, 171 race models, 129
— — home states, 173 Ramchandanis TTPNs, 176
— — liveness, 173 — minimum cycle time, 176
— hierarchical analysis, 175 random switch, 140
— marking, 164 random variable
— non-live, 172 — nth central moment, 19
— qualitative analysis, 169, 170 — nth moment, 19
— quantitative analysis, 164 — average value, 19
— queue, 162 — Bayes’s theorem, 15
— queueing place, 162 — Bernoulli, 16
— reduced embedded MC, 164 — Binomial, 16
— timeless trap, 169, 171 — coefficient of variation, 20
— visit ratio, 166 — conditional probability, 14
— with no home states, 174 — continuous, 18
— WWW, 178 — cumulative distribution function, 17
queueing place, 162 — exhaustive set of events, 13
Queueing system, 55 — exponential, 18
— arrival rate, 57 — geometric, 16, 44
— batch arrival, 72 — independence, 14
— batch service, 72 — joint probability density function, 20
— bulk arrival, 72 — mean, 19
— bulk service, 72 — mutually exclusive events, 14
— central server model, 163 — mutually exclusive exhaustive events,
— customer, 55 14
— customer classes, 70 — probability density function, 18
— FCFS, 159 — probability mass function, 16
— global balance equations, 62 — standard deviation, 20
— infinite server, 69 — statistical independence, 14
— interarrival time, 56 — total probability, 14
— Kendall notation, 57 — variance, 19
— Little’s law, 58 Razouk and Phelp’s TTPNs, 177
— — moments, 60 reachability graph, 90
— M/M/1 queue, 65 reachability problem, 113
— M/M/m queue, 68 — complexity of, 113
— mean population, 66 — decidability of, 113, 125
— preemptive repeat, 70 reachability set, 86
— preemptive resume, 70 reachability tree, 90
— priority service, 70 reachable marking, 86
— processor sharing, 69 Readers/Writers-Problem, 98
— queue, 55, 162 recurrent nonnull state, 30
— scheduling strategy, 55 recurrent null state, 30
— server, 55 recurrent state, 30
— service rate, 57
216 Index

reduced embedded Markov chain of a GSPN, stationary distribution, 31

144 stationary Markov process, 23
reduced embedded Markov chain of a QPN, statistical independence, 14
164 steady state distribution, 31, 51
reduction analysis, 110 Stelle, 99
reduction rule, 110 Stochastic Petri net
redundant place, 110 — workshop, 178
redundant transition, 142 Stochastic Petri net, 131
reflexive and transitive closure, 83 — interface description, 178
resampling, 129 — isomorphic graphs, 135
resource sharing, 179 — Markov process, 132
reversibility, 113 — mean number of tokens, 134
— probability of transition firings, 134
S-graph, 125 — qualitative analysis, 129
S-invariant, 99 — quantitative analysis, 129
safe Petri net, 87 — throughput at a timed transition, 134
scheduling strategy, 55, 162 — timed transition, 131
— FCFS, 159, 162 — tool list, 178
— infinite server, 69 — tool support, 178
— LCFSPR, 161 — unstable iteration process, 137
— pre-empt resume priority, 159 Stochastic Petri Nets, 129
— preemptive repeat, 70 stochastic process, 21
— preemptive resume, 70 strongly connected component, 93
— priority service, 70 strongly connected Petri net, 84, 89
— processor sharing, 69 stubborn set method, 125
— work conserving, 172 subnet, 107
self-loop in a Petri net, 85 subnet generated by X, 107
semi-Markov process, 40, 141 superposed GSPNs, 175
server, 55 switching distribution, 140
service rate, 57 symbolic model checking, 125
Sifakis TPPNs, 176 symmetry method, 125
simple cycle, 104 synchronic distance of two transitions, 113
simple net, 101 Synchronisationsgraph, 125
siphon, 106 synthesis analysis, 110
SM-net, 101 synthesis of Petri nets, 111
— S-graph, 125 system, 82
sojourn time, 32, 41, 47, 132
Special Interest Group on Petri Nets and T-graph, 125
Related Models, 126 T-invariant, 100
special symbol ω, 91 tangible marking, 141
speed vector, 187 tangible state, 141
SPL-net, 101 terminal strongly connected component, 93
SPNs, 131 throughput
standard deviation, 20 — at a timed transition, 134
state machine, 101 — at an immediate transition, 147
— S-graph, 125 time before absorption
state space, 21 — mean, 38
state space explosion, 94, 135 — variance, 38
state transition diagramme, 28 Timed Petri Nets, 129
Index 217

Timed Petri Nets (workshop), 178 — timeless, 145, 149

Timed Places Petri net, 129 — timeless in a QPN, 169
timed queueing place, 182 — timeless in an EFC-GSPN, 156
timed transition, 131, 139, 140 — timeless in an EFC-QPN, 171
Timed Transitions Petri net, 129 TTPN, 129
timeless trap, 145, 149 — DSPN, 177
— in EFC-GSPNs, 156 — ESPN, 177
— in EFC-QPNs, 171 — Merlin, 177
— in QPNs, 169 — Non-Markovian, 177
token, 77 — preselection, 129
— colour of, 116 — race, 129
— mean number of, 134 — Ramchandani, 176
— probability density function, 134 — Razouk and Phelp, 177
tool — Zuberek, 177
— interface description, 178 tube, 106
— list, 178 Turing machine power
— support, 178 — of CPNs, 125
total probability, 14 — of GSPNs, 158
TPPN, 129 — Petri nets with inhibitor arcs, 125
— Sifakis, 176
transformation of a Petri net, 110 unfinished work, 55
transient Markov chain, 35 unfolding of a CPN, 120, 170
transient state, 30 unmarked set of places, 84
transition, 77, 81 unstable iteration process in SPNs, 137
— colour of, 116 utilisation, 66
— enabled, 78, 84
— enabled in a GSPN, 149 vanishing marking, 141
— enabled in a QPN, 169 — elimination on the fly, 151, 160
— firing, 78, 84 vanishing state, 141
— firing delay, 140 variance, 19
— firing sequence, 85 vector addition system, 96
— firing time, 131 visit ratio, 32, 142, 166
— firing weight, 141
— guard, 121 waiting line, 55
— immediate, 139, 141 waiting time, 57
— input, 83 weakly connected Petri net, 83
— isolated, 77 work conserving scheduling strategy, 172
— live, 87 work-conserving, 72
— output, 83 workshop
— probability of firing, 134 — Petri Nets and Performance Models, 178
— rate, 131 — Timed Petri Nets, 178
— redundant, 142 World Wide Web
— throughput at a timed, 134 — Computer Science Bibliography, 73
— throughput at an immediate, 147 — further search facilities, 73
— timed, 131, 139, 140 — PN information, 126
transition probability, 25 — QN information, 73
trap — QPN information, 178
— of a Petri net, 106
— of matrix C, 145 Zuberek’s TTPNs, 177

View publication stats