6 Solving Queueing Models

6.1 Introduction
In this note we look at the solution of systems of queues, starting with simple isolated
queues. The benefits of using predefined, easily classified queues will become appar-
ent: many performance measures can be calculated directly from the parameters of the
model. Obviously the situation becomes more complicated when queues are connected to-
gether. However we see that even in this case deriving performance measures can be very
straightforward as we are able to consider each queue in isolation. Finally we summarise
the assumptions which we need to make in order to obtain these solutions.

6.2 Single Queues

We will assume that all the queues which we consider have a Markovian arrival process
and a Markovian service process and so the queue can be modelled as a Markov process.
The state transition diagram for a single-server queue with infinite capacity is shown in
Figure 7. Note that unlike the Markov processes which we have considered earlier in the
course this process has an infinite state space.

x0 x1 x2 x3 ...

Figure 7: The state transition diagram for a simple M/M/1 queue

If we write the global balance equations for this system we can soon recognise a regular
pattern emerging.
λπ0 = µπ1
(λ + µ)π1 = λπ0 + µπ2
(λ + µ)π2 = λπ1 + µπ3
..
.
Using simple algebra we can rewrite these as shown below:
λ
π1 = π0
µ
 2
λ λ
π2 = π1 = π0
µ µ
 3
λ λ
π3 = π2 = π0
µ µ
..
.
Performance Modelling LN-6

Remembering that traffic intensity, ρ is defined to be λ/µ (for the single server case,
λ/(c × µ) in the general case) we can see that for an arbitrary state xi

πi = ρi π0

and using the normalisation condition we see that

∞ ∞
X X 1
1 = πi = π0 ρi = π0 if ρ < 1
i=0 i=0
1−ρ

From this we can deduce that

π0 = 1 − ρ and πi = (1 − ρ)ρi for all i > 0.

Thus we have a symbolic evaluation of the steady state distribution for any M/M/1
queue, i.e. the steady state distribution expressed in terms of the parameters of the model,
λ and µ. This means that for any particular model we can find the steady state distribution
simply by evaluating the expressions above with our particular value of ρ = λ/µ. More
importantly, we can derive symbolic expressions for the important performance measures
we are likely to want to derive from a queue, and then evaluate those measures for a
particular model without having to carry out any numerical solution of global balance
equations.

Utilisation, U The queue is being utilised whenever it is non-empty; in other words

the utilisation, U , is 1−π0 .

Utilisation U =ρ

Mean number of customers in the queue, N This is the expectation of the number
of customers in the service facility as a whole, i.e.
∞ ∞
X X ρ
N = n × πn = n × (1 − ρ)ρn =
n=1 n=1
(1 − ρ)

ρ
No. in queue N=
1−ρ

Mean number of customers waiting, Nb This is the expectation of the number of

customers in the buffer i.e.
∞ ∞ ∞
X X
n
X ρ2
Nb = (n − 1) × πn = (n − 1) × (1 − ρ)ρ = ρ n × (1 − ρ)ρn =
n=1 n=1 n=1
(1 − ρ)

ρ2
No. in buffer Nb =
1−ρ

40
Performance Modelling LN-6

Mean response time, R Using Little’s Law we can calculate the mean response time
of the queue to be the mean number in the queue N , divided by the arrival rate λ, i.e.
ρ 1 1/µ 1
R = N/λ = × = = .
1−ρ λ 1−ρ µ(1 − ρ)

1
Response time R =
µ(1 − ρ)

Other common performance measures can be calculated in a similar way. Moreover we

can derive symbolic steady state distributions, and expressions for performance measures,
for the other standard queues. Thus, given almost any single queue model, in order to
derive a performance measure it is only necessary to select the appropriate formula from
a table and evaluate it using the parameters of your model. Some examples are shown in
Table 1. Most textbooks on performance models will contain these formulae.

Performance Measures M/M/1 M/M/c M/M/1/K

λ λ λ
Traffic Intensity ρ
µ c×µ µ

Utilisation U (1 − ρ)ρK
ρ ρ ρ(1 − )
(per server) 1 − ρK+1

c−1
!−1
c
(cρ) X (cρ)n 1−ρ
Prob. system is idle π0 1−ρ 1+ +
c!(1 − ρ) n=1 n! 1 − ρK+1

Prob. buffer (cρ)c

B ρ2 π0
non-empty c!(1 − ρ)

ρ ρ (K + 1)ρK+1
Mean no. in system N cρ + ρB/(1 − ρ) −
1−ρ 1−ρ 1 − ρK+1

ρ2 ρ 1 + KρK
Mean no. in buffer Nb ρB/(1 − ρ) −ρ
1−ρ 1−ρ 1 − ρK+1
 
1 1 B N
Mean response time R 1+ (1−ρ)ρK
µ(1 − ρ) µ c(1 − ρ) λ(1 − )
1−ρK+1

Table 1: Common performance measures

41
Performance Modelling LN-6

Example: Consider again the wireless access gateway discussed in the previous note.
Measurements have shown that packets arrive at a mean rate of 125 packets per second,
and are buffered until they can be transmitted. The gateway takes 2 milliseconds on
average to transmit a packet. The gateway currently has 13 places (including the packet
being transmitted) and packets that arrive when the buffer is full are lost. We wish to
find out if the buffer capacity is sufficient to ensure that less than one packet per million
gets lost.

We represent the gateway as a M/M/1/13 queue, with λ = 125 and µ = 1/0.002 = 500.
The utilisation of the gateway will be ρ = λ/µ = 0.25.
The loss rate is the arrival rate multiplied by the probability that the system is full, i.e.
λ × πK . The proportion of packets that are lost is simply πK (expected number lost per
time unit divided by the expected number arriving per time unit)
(1 − ρ)ρK
πK = = (0.75×0.2513 )/(1−0.2514 ) = (1.1176×10−8 )/0.99999999 = 1.12×10−8
1 − ρK+1
Thus the expected proportion of packets lost is 0.0112 every million packets, well within
the requirement.

Exercise: Using MatLab or otherwise, investigate how much the buffer size can be re-
duced before the loss rate becomes unacceptable. What happens if the arrival rate in-
creases to 175 packets per second?

6.3 Product Form Queueing Networks

Once we consider a network of queues rather than a single queue the possible forms of
the state spaces of the underlying Markov process become much more diverse. Therefore
we would not anticipate being able to derive symbolic steady state distributions of wide
applicability in the same way as we have done for single queues. However, for a large
class of queueing networks a straightforward and efficient means of solving models has
been found. These networks are known as product form networks.
The term product form comes from the fact that the steady state distribution of these
models can be derived as the product of the steady state distributions of each of the
constituent service centres.
Recall that the state of a single service centre can be characterised by the number
of customers currently in the system. In a queueing network the state of the system is
characterised by the number of customers waiting at each of the service centres. This
is usually represented as a tuple. For example, in a simple queueing network with two
service centres, such as the one shown in Figure 8, the state (n1 , n2 ) indicates that there
are n1 customers in service centre 1 (queueing or in service) and n2 customers in service
centre 2.
For a model, such as this one, to have a product form steady state distribution means
that the distribution can be expressed as a product of terms representing the steady states
of each of the service centres considered in isolation, e.g.
π(n1 , n2 ) = π1 (n1 ) × π2 (n2 )

42
Performance Modelling LN-6


λ - µ1 (1−p) -
6 p

q 
 (1−q) µ2 


Figure 8: Open Queueing Network

In terms of the system what this means is that the two service centres reach their equilib-
rium behaviours and then behave independently of each other. (Recall that a probability
of a coincidence of events can be expressed as the product of the probabilities of each
event if and only if those events are independent.)
The class of queueing networks in which service centres exhibit this form of independent
behaviour in equilibrium excludes some interesting and important system features as
was discussed in the previous lecture note. However when the necessary conditions are
satisfied, performance measures can be derived without resorting to the underlying Markov
process—just as in the case of a single queue, because in effect we only solve the network
one queue at a time.

6.3.1 Traffic equations

Using the already established formulae for individual service centres we only need to know
two parameters: the arrival rate at the queue, which we will denote by λi for the ith queue,
and the service rate at the queue, denoted µi . For a queueing network we will be given
µi for each queue in the network. However, we will only usually be given the arrival rate
of arrivals to the network from the environment (λ in the example shown in Figure 8).
Fortunately the following theorems, the decomposition principle of Poisson/exponential
distributions, and simple algebra, can help us to work out the arrival rate at each node
in the network.

Burke’s Theorem stated that a Poisson arrival process at a service centre with ex-
ponential service rates generates a Poisson process of departures. Moreover the rate of
departure is the same as the arrival rate. This result implies that each service centre
in a chain of simple exponential service centres with Poisson arrivals can be analysed
independently using results from simple queues.

Jackson’s Theorem generalised Burke’s Theorem to consider an arbitrary network of

exponential service centres each of which is driven by a Poisson arrival process. Although
the presence of feedback paths destroys the Poisson nature of the service centre arrivals,
Jackson showed that they still behave as if they were Poisson driven. Moreover the rela-
tionship between the arrival stream and the output stream is maintained. The networks
considered by Jackson were open.

43
Performance Modelling LN-6

Again, this implies that each service centre in the queueing network can be analysed
independently.

Gordon and Newell’s Theorem considered a modification of Jackson’s networks in

which the network is closed rather than open. Closed queueing networks are not driven by
any external Poisson arrival process. Instead, the number of customers in the network is
fixed, K. Whenever a customer departs one service centre in the network it immediately
requests service from another one. Hence customers never leave the network. Gordon and
Newell’s product form equation has the form:
n
1 Y
π(n1 , n2 , . . . , nk ) = πi (ni )
G(K) i=1

where G(K) is a normalisation constant chosen to ensure that the steady state probabil-
ities sum to 1.

Using these results we know that if the arrival rate at queue i is λi then the departure
rate will also be λi . Therefore, if all the departures from queue i go directly to queue i + 1
the arrival rate at queue i + 1 will also be λi , i.e. λi+1 = λi .
However, in general the Poisson departure stream from a queue may be split between
different queues and/or the environment, according to the routing probabilities. By the
decomposition principle we know that when the departure stream is split in this way each
of the resulting arrival streams will also be Poisson, with a rate obtained by multiplying
the departure rate by the probability of this branch of the routing probability.
If we analyse all the service centres of a queueing network in this way, expressing its
input stream as a sum of output streams from the environment or other service centres,
we obtain what are known as the traffic equations. For the simple network shown in
Figure 8, the traffic equations are as follows:

λ1 = λ + q × λ2
λ2 = p × λ1

We obtain one equation for each service centre, and we have one unknown for each service
centre, so we end up with n equations in n unknowns. Therefore we can solve the traffic
equations and find the arrival rate at each service centre. In the case above, assume that
λ = 10, p = 0.5 and q = 0.4. If we substitute λ2 = 0.5λ1 into the first equation we obtain

λ1 = 10 + 0.2λ1
0.8λ1 = 10

Therefore λ1 = 12.5 and substituting back we deduce that λ2 = 6.25.

The great advantage of solving traffic equations rather than the global balance equations
is that the number of equations we need to solve grows linearly with the number of service
centres, rather than exponentially, which is the case for global balance equations.

44
Performance Modelling LN-6

6.4 Assumptions
Queueing network modelling is inherently a top-down process. The underlying philoso-
phy is to begin by identifying the principal components of the system and the ways in
which they interact, then supply any details that prove to be necessary. This philosophy
inevitably leads to a large number of assumptions being introduced while a modelling
study is being conducted.
Many of these assumptions are the general ones which we have been making throughout
our study of Markov processes, although since we are avoiding the numerical solution of
the global balance equations we can now consider models which have an infinite number
of states. There are five assumptions which we must make about the behaviour of the
model to ensure that a product form solution exists. These are:

Each service centre is flow balanced; this means that the number of customers that
arrive at each centre is equal to the number of customers who depart.
The system exhibits one step behaviour; this means that no two customers in the
system “change state” at exactly the same time. Here changing state may mean a
customer finishing at one service centre and progressing to the next, or a customer
arriving at, or departing from the system.
The system has routing homogeneity. This means that the probability that a cus-
tomer completing service at service centre j now proceeds to service centre k is
independent of the current queue lengths at j or k, for all service centres in the
network.
The system has device homogeneity. The rate of completions of customers from a
service centre may vary with the number of jobs at that centre, but otherwise cannot
depend on the number or placement of customers within the network.
The system experiences homogeneous external arrivals; that is, the times at which
arrivals from outside the network occur may not depend on the number or placement
of customers within the network.

Jane Hillston [email protected]. January 25, 2017.

45