CHAPTER 13 WAITING LINE MODELS

13.1 Basic Queuing Models

Queuing systems consist of customers receiving service
performed at service stations. Customers are in a
calling population. They will approach the service
system and line up, be served, after which they leave
the service system and rejoin the calling population.
The structure of this process can be visualized in the
figure below.

Each service system has only a single waiting line and

typically consists of a number of c parallel service
stations, each performing the same type of service.
Parallel service stations are usually referred to as
channels.

Kendall (1918-2007) devised a taxonomy in 1953

consisting of three descriptors, and later extended to
six. We introduce the complete six-component system
first, but then use only the more compact 3-component
system. The notation is
A/B/C K/N/D,
where each letter is a placeholder for one component
of waiting lines.
A describes the arrival process. We use M, if arrivals
are random and follow a so-called Poisson process
(where the M stands for Markovian). The Poisson
distribution describes such a process. Other processes
include D, deterministic arrivals, or G, where
arrivals follow some general distribution.
B describes the service process. Again, the letter M
denotes Markovian, i.e., exponential service time,
while D and G (or, more precisely, GI)
symbolize deterministic and general (general
independent) service times, respectively.
C indicates the number of parallel service stations.

K describes the number of customers that can be

accommodated in the system. If not used, K is assumed
to be infinite.
N denotes the size of the calling population. If not
used, N is assumed to be infinite.
D denotes the queuing discipline. Typical disciplines
are FCFS or FIFO (first-come, first-served or first-in,
first-out), LCFS or LIFO (last-come, first-served or
last-in, first-out), or SIRO (service in random order).
There are also priority queues, in which some
customers receive preferential treatment, as in health
care, where serious cases will be treated first. If no
queuing discipline is specified, FCFS applies.
It is useful to think of waiting lines as buffers between
arrival and service (similar to inventories that can be
thought of as buffers between the inflow & outflow of
materials). A good image would be one of a water tank.
While a waiting line grows with the inflow, it shrinks
whenever service is provided and customers leave the
system.
In queuing theory, we distinguish between steady states
occurring if the system has been running for a very
long time, and transient states which depend on the
initial state of the system. An example is the opening of

a new checkout counter in a store. Initially, no one is in

the system and waiting times are short, but as the
counter has operated for some time, the system
becomes more congested and no longer dependent on
the opening conditions.
In queuing one computes measures of interest from
key input factors or queue characteristics. Such
measures include average waiting times, the
probability that a newly arriving customer will have to
wait, the average length of a queue, and others. We use
the following notation:
is the mean arrival rate, the average number of
customers who actually arrive at the system per time
unit.
is the mean service rate, the average number of
customers who can be served by a single service
station per time unit.
Note that while expresses an actual observable fact,
indicates a capability. The actual number of
customers served does not only depend on the
service stations capabilities, but also on the number
of customers requiring service (which depends on
).

The inverse values of and also have important

interpretations. If a service station can deal with =
12 customers per hour, then 1/ = 1/12
[hours/customer] = 5 [minutes/customer] is the
average service time. Similarly, if the average arrival
rate is = 10 [customers/hour], then 1/ = 1/10
[hours/customer] = 6 [minutes/customer], so that on
average, six minutes elapse between two successive
arrivals. This is referred to as the (average)
interarrival time.
For a single service station, the arrival rate cannot
exceed the service rate, since then the waiting line
would grow towards infinity. This gives rise to a
feasibility condition. We define the traffic intensity
(sometimes also referred to as utilization rate) = /
and require < c, where c denotes the number of
parallel service stations. For a single-server system
with = 24 customers per hour, feasibility requires >
24, i.e., the average service time cannot exceed 150
seconds.
On the output side are also measures of interest:
Pn: the probability of n customers in the system
(with P0, being the probability of an idle system),

Ws: the average waiting time per customer,

Wq: the average time a customer is in the queue,
Ls: the average number of customers in the system,
Lq: the average number of customers in the queue.
Some general relations hold in queuing, regardless of
the specific system under consideration. The first is
Ws = Wq + 1/.

(1)

Another useful relation, known as Littles formula

(based on work published by J.D.C. Little in 1961) is
L W ,
where the stand for either the subscript s or the
subscript q.
The simplest queuing model is the M/M/1, for which
key formulas are:

1
W

P0 = 1
s

2
n
Wq

Pn = P0 Lq

1
Ls

Example 1: Customers arrive at a bank at a rate of 30

per hour. Arrivals are random and service time is
exponential, so that the M/M/1 model applies. The
clerks service time is 90 seconds. Therefore = 30 and
= 45, and as = / = 30/45 = < 1, the system has
a steady state. The probability of an idle bank teller is
P0 = 1 = . The probability that at least two
customers are waiting is P3 + P4 + P5 + = 1 P0 P1
P2 = 1 ()1 ()2 278 .2963 or slightly
less than one third. On average, there are Lq = 1.3333
customers waiting in line and the average time for a
customer in the system is Ws = 1/15 {hour] = 4 minutes.
Suppose that in an M/M/1 system with = 20, the
decision maker specifies that the probability of three
or more customers in the system should not exceed
95%. This probability is 1 P0 P1 P2 = 1 (1)
(1) 2(1) = 1 1 + + 2 2 + 3 = 3,
and since it should not exceed 95%, we obtain 3 =
3
3

.95
or

> 8,000/.95 or > 20.3449.

3

Suppose now that the service rate follows some general

distribution with mean service time 1/ and variance
2. This is an M/G/1 model, for which there are some
elegant formulas. As in all single channel systems in a

steady state, P0 = 1 . Furthermore, the PollaczekKhintchine formula is

2 2 2
Lq
.
2(1 )
The values of Ls, Wq, and Ws can then be computed
based on the general relations (1) and (2). (Note that
the M/M/1 model is a special case of the M/G/1 model
with 2 = 1/ 2 which in the above expression results in
2
Lq
from the M/M/1 model. Similarly, with the
1
M/D/1 model, the variance 2 = 0, and the Pollaczek2
Khintchine formula reduces to Lq
, so that the
2(1 )
number of customers waiting in the deterministic
model is only half compared to the standard model
with exponential service time).
Example 2: Arrivals of customers at a single service
desk follow a Poisson distribution averaging = 15
arrivals per hour, while the service time follows a
general distribution with an average of 3 minutes and
a standard deviation of 6 minutes. Then 1/ = 1/20
[hours] and 2 = 1/(10)2 = 1/100, and on average Lq =
5.625 customers are waiting in line, with an average

waiting time Wq = .375 hours = 22.5 minutes. If the

service time were exponential, the M/M/1 model
applies and Lq = 2.25 and Wq = 9 minutes, while the
deterministic model has Lq = 1.125 and Wq = 4.5
minutes.
Consider now multiple service stations. We distinguish
between a multi-station service center (a bank
system) and several parallel single-service centers (a
supermarket system). In a bank system, there is a
single queue and thus a single multi-server system,
whereas a supermarket has multiple single-server
systems.
First consider multiple single-server systems as simple
extensions of the previous models. Assume no balking
(arriving customers join the queue regardless of its
length) and no jockeying (customers do not change
queues even if it might result in shorter waiting times).
For example, assume Poisson arrivals with = 45
customers per hour, and exponentially distributed
service time with mean 1/ = 2 minutes (i.e., = 30
customers per hour). Feasibility requires that = /
= 45/30 = 1 does not exceed the number of service
stations c, so that two or more service stations are
needed. Suppose that c = 3, so that we now deal with
three single-service systems, or 3 M/M/1. While on

average = 45 customers are waiting, each system

receives only a third of this number, as customers
randomly choose the system to be served by. For each
of the three systems, we have an effective arrival rate
of ' = 15 which will then replace in all formulas.
Feasibility is achieved with ' = '/ = 15/30 = < 1
for each system, and the M/M/1 formulas in Table 1
1
apply. For instance, Ws =
= 1/(3015) = 1/15
'
2
hours = 4 minutes, and Lq =
= ()2/(1) = , so
1
that on average half a customer is waiting in line in
each of the three subsystems. In all of the three service
systems combined, 1.5 customers will be waiting for
service.
Now consider a single M/M/3 system such as the one in
a bank with three tellers. Some formulas are in the
Table below, with W computed by Littles formula (2).
P0

n
n!

P0 ,if 0 n c

Pn n

P0 ,if n c
c !(1 )

c !c nc

i 0
c
c 1
Lq
P
Ls = Lq +
2 0
(c 1)!(c )

c 1

i
i!

In our example, first find the probability of no

customers in the system. Here,
P0

1
3
(3
/
2)
[1 32 98 ]
3!(1 3/2
3 )

4
19 .

Then compute Lq = 9/38 .2368, Ls = 9/38 + 3/2

1.7368 and with Littles formula, we obtain Wq = Lq/
= .3158 minutes and Ws = Ls/ = 2.3158 minutes.
The differences between the two models are that in the
former model with separate queues, it is possible that
a customer is waiting, while a service station is idle.
This is possible since we assumed jockeying was
prohibited.
An extension is a model with self-service, derived from
the standard M/M/c model by letting c tend to infinity.
Such a model will have no waiting time, so that Lq =
Wq = 0, Ws = 1/ and Ls = . One can show that in this
e n
case, Pn
for n 0.
n!

For more complex queuing models, it is useful to have

tables or graphs giving the decision maker a quick
idea of what to expect, without complex computations.
Entire books with queuing graphs and queuing tables
exist. The figure below shows an example of such
graphs. This one depicts a c M/M/1 graph, where the
abscissa shows the utilization rate ' and the ordinate
the total number of customers in all systems,
. The
steep solid line shows the case of a single facility (i.e., c
= 1); the broken, broken-and-dotted, broken-andtwice-dotted, and dotted lines are for c = 2, 3, 5, and 10
service stations.

Finally, we look at multiple-server problems from a

different angle. Suppose that the decision maker
specifies a certain service level and asks how many
service stations are needed for a given level of service.
Questions like this arise in the public service, e.g., how
many police officers are needed to protect an area.
Example: On average twenty customers per hour
arrive at a hospital. There are c doctors available.
Given that each doctor can deal with on average ten
patients per hour, hospital administrators want to
ensure that the average waiting time for any patient
does not exceed 10 minutes. What is the smallest
number of doctors that allows this?
Solution: The system is a set of c parallel M/M/1
systems, and the effective arrival rate at each is then '
= /c. From Table 12.1, the average waiting time is
'

Wq

'

/ c , can be
, which, with

c
/c
2/c
W

rewritten as q
=
. The average
/c
10 20 / c
waiting time must be no higher than 1/6 [hour], which
can be written as

2/c
1/6.
10 20 / c
2
1
, or c
10c 20 6
3.2. At least four doctors are needed for the desired
service.
This inequality can be written as

We also note that with three doctors, average waiting

time is 12 minutes, while four doctors result in an
average waiting time of only 6 minutes.

13.2 Optimization in Queuing

Queuing models can also be used for optimization.
Example: At a tool crib expensive tools are kept and
signed out by clerks at a service desk to workers, when
needed. Demand for the tools occurs randomly,
averaging 100 times/hour. There are costs of the clerks
as well as costs for lost time of workers who are
waiting to sign out tools. With c clerks, and $c and $w
being the hourly wages of a clerk and a worker,
respectively, total costs are
C = (clerks) + (workers lost time) = c$c + $wLs.

How many clerks would minimize the overall costs?

Assume that the signing out procedure follows c
M/M/1 systems and that the hourly clerk wage is $ c =
10, while a workers lost hour costs $w = 25.
Solution: With = 100 and = 20 (so that = 5), the
feasibility condition c indicates that at least six
clerks are needed. The costs for the clerks increases
linearly with the number of clerks, whereas the cost
for workers lost time decreases hyperbolically with an
increasing number of clerks, as shown in the Figure
below. (We display the costs for all real values of c to
show the shapes of the curves.)

For very few clerks (c = 6), waiting time for workers

are long, resulting in high costs. With more clerks,
waiting times decrease and also the costs. Eventually,
the benefit of an additional clerk is outweighed by his
cost. Total costs are computed for c = 6, 7, 8, clerks
until optimum has been found:
c = 6: = 100/6 = 16.6667, = 16.6667/20 = .8333,
Ls = 5 in each system, cost = 6(10) + 5(6)(25) = 810.
c = 7: = 100/7 = 14.2857, = .7143, Ls = 2.5 in each
system, cost = 7(10) + 2.5(7)(25) = 507.5.
c = 8: = 100/8 = 12.5, = .6250, Ls = 1.6667 in each
system, cost = 8(10) + 1.6667(8)(25) = 413.34.
c = 9: = 100/9 = 11.1111, = .5556, Ls = 1.25 in each
system, cost = 9(10) + 1.25(9)(25) = 371.25.
c = 10: = 100/10 = 10, = .5, Ls = 1 in each system,
cost = 10(10) + 1(10)(25) = 350.
c = 11: = 100/11 = 9.0909, = .4545, Ls = .8333 in each
system, cost = 11(10) + (.8333)7(25) = 339.16.
c = 12: = 100/12 = 8.3333, =.4167, Ls = .7143 in each
system, cost = 12(10) + (.7143)12(25) = 334.29.

c = 13: = 100/13 = 7.6923, = .3846, Ls = .625 in each

system, cost = 13(10) + .625(13)(25) = 333.13.
c = 14: = 100/14 = 7.1429, = .3571, Ls = .5556 in
each system, cost = 14(10) + .5556(14)25 = 334.43.
It is optimal to have c = 13 parallel service stations.
Consider also retraining of clerks. The retraining time
includes the actual retraining and costs for time the
clerk is absent during training.
Example: Customers arrive at a rate of = 30 per
hour. Keeping them waiting costs $20 per hour. The
service rate is = 40, but with training it can be
increased up to 60. Training costs 2( 40)2. Assume
that an M/M/1 system is used. To what service rate
should clerks be trained?
Solution: The average waiting costs in an M/M/1

2
system are Lq
=
, and the cost function
( )
is then
C = (retraining) + (waiting) =

900
= 2( 40) + 20
.
( 30)
2

The Figure below shows the total costs. The function

reaches a minimum at = 41.12358 with costs C =
41.8742.

Total costs for other service rates are:

The queuing costs for trained clerks are:

40
45

41
42
50
41.9113 43.7143 218

60
810

Leaving the clerk essentially un(re-)trained costs $45,

less than 10% off optimum, while training the clerk up
to capacity will cost 18 times more.