Discrete-Time Multiserver Queues with Geometric Service Times

Peixia Gao∗, Sabine Wittevrongel, Herwig Bruneel

SMACS† Research Group,

Department of Telecommunications and Information Processing,

Ghent University, Sint-Pietersnieuwstraat 41, B-9000 Gent, Belgium.

Scope and purpose

Discrete-time queueing models are widely used to study the behavior of various types
of telecommunication and computer systems. Most of the existing studies of discrete-time
multiserver queueing models assume that the service times of the customers are constant.
Recently, however, there has been an increased interest in discrete-time models with non-
deterministic service times, due to the more and more complicated and irregular service
mechanisms in nowadays telecommunication networks. In view of this, this paper focuses on
a discrete-time queueing system with multiple servers and geometrically distributed service
times. We show that the queueing model can be analyzed by means of a generating-functions
approach. The results of the analysis are important to evaluate performance measures such
as packet loss and delay in communication and computer networks.

∗
Corresponding author. Tel.: +32-9-264-89-01; fax: +32-9-264-42-95.
E-mail address: [email protected] (P. Gao)
†
SMACS: Stochastic Modeling and Analysis of Communication Systems

1
Abstract
In this paper, a discrete-time multiserver queueing system with infinite buffer size and
general independent arrivals is considered. The service times of a packet served by one of the
servers are assumed to be independent and identically distributed according to a geometric
distribution. Each packet gets service from only one server. In the paper, the behavior of
the queueing system is studied analytically by means of a generating-functions approach.
This results in closed-form expressions for the mean values, the variances and the tail dis-
tributions of the system contents and the packet delay. Some numerical examples are given
to illustrate the analysis.

Keywords: Discrete-time queueing, multiserver system, geometric service times, system con-
tents, delay

2
1 Introduction
Discrete-time queueing theory has received considerable attention during the last decades
because of its direct applicability in the study of many computer and communication systems
in which time is divided into fixed-length time intervals (“slots”). The reader is referred to
the books [1]-[6] and the references therein for an extensive treatment of a wide variety of
discrete-time queueing models. There are a number of analysis techniques for discrete-time
models, ranging from computer simulation to the numerical solution of the associated set
of linear balance equations and various types of analytical or semi-analytical methods. In
case of computer simulation, a program is built to simulate the behavior of the system in
software; a typical simulation technique is discrete event simulation ([7]). With respect to the
mathematical solution techniques, a distinction can be made between algorithms that allow
to efficiently calculate the performance measures numerically (e.g. matrix-analytic methods
([8], [9])) and analytical methods that lead to (exact or approximate) closed-form expressions
(e.g. probability generating functions based techniques ([4], [10]) or methods based on the
maximum entropy principle ([11])).
In this paper we focus our attention on the specific class of discrete-time queues with
multiple servers (or output channels). To the best of our knowledge, in most of the existing
literature on discrete-time multiserver queueing models, the service (or transmission) times
of customers (or packets) are assumed to be constant, equal to one slot (see [12]-[15]) or
multiple slots ([16]). These assumptions are appropriate when modeling for instance buffers
in ATM-based integrated-services digital networks, in view of the fixed packet (cell) length
used in these networks ([17]). Only a few authors have considered more general discrete-time
multiserver models with non-deterministic (variable) service times, although there are both
fundamentally theoretic as well as more practical motives to do this, given the more and
more complicated and irregular service mechanisms in the nowadays internet. We mention
[18], where the service times are assumed to be geometrically distributed; somewhat related
models with randomly interrupted servers are studied in [19] and [20]. Geometric service
times are also considered in [21] and [22], while [23]-[25] deal with general service times, but
only for the case of one single server.
In line with the above, we are concerned in this paper with generalizing the service-
time distribution. More specifically, as a first step, we consider a discrete-time queueing

3
system with geometric service times and a general uncorrelated arrival process, and present
an analytical technique based on generating functions for the performance analysis of the
system. The model under study has also been treated in [18]. However, as opposed to [18],
we obtain explicit expressions for the probability generating functions (pgf’s) of not only
the system contents, but also the total delay experienced by customers. Furthermore, these
pgf’s are used to derive various additional performance measures. Specifically, expressions
are derived for the mean values, the variances and the tail probabilities of system contents
and delay.
The outline of the paper is as follows. In section 2, the queueing model under study is
described. In section 3, we focus on the system contents and derive expressions for the pgf,
the mean value, the variance and the tail probabilities of the system contents. The packet
delay is analyzed in section 4, and again the pgf, mean value, variance and tail probabilities
are obtained. In section 5, some special cases are considered in order to check the obtained
results. In section 6, some numerical examples are given and discussed. Finally, the paper
is concluded in section 7.

2 Queueing Model
In this paper, we consider a discrete-time queueing system with multiple output channels
and an infinite storage capacity for packets. The number of output channels (servers) will
be denoted by c (c > 0). Packets arrive at the input of the system in a stochastic manner,
wait in a buffer for some time and are then transmitted via one of the c output channels,
after which they finally leave the system. The time axis is assumed to be divided into fixed-
length time intervals, referred to as slots and chronologically labeled. New packets enter
the buffer according to a general uncorrelated arrival process, i.e., the numbers of packets
arriving in the buffer during the consecutive slots are modeled as independent and identically
distributed (i.i.d.) random variables, with common pgf A(z).
The service process is modeled as follows. The service (or transmission) of a packet can
start or end at slot boundaries only. Hence, the service time of a packet always consists of
an integer number of full slots. This also implies that the service of a packet which arrives
in an empty system cannot start before the end of the packet’s arrival slot. One packet can
only get service from one of the servers. The service times of the packets are assumed to be

4
i.i.d. random variables and geometrically distributed with parameter 1 − µ (0 < µ ≤ 1), i.e.,
with common probability mass function (pmf)

g(n) = Prob[service of a packet takes n slots]

(1)
= µ(1 − µ)n−1 , n ≥ 1,

and corresponding pgf

∞
X µz
G(z) = g(n)z n = . (2)
n=1
1 − (1 − µ) z

Moreover, the service and arrival processes are assumed to be mutually independent. Packets
are served according to a first-come-first-served (FCFS) queueing discipline.
It is well known (see e.g. [2]) that a queueing system can reach a steady state if and only
if the average amount of “work” (expressed in slots of service time) entering the system per
slot is strictly less than what the system can handle per slot. In view of the above model
description, this equilibrium condition can be expressed as

0
A (1) < cµ, (3)

where we have used primes to indicate derivatives. In the analysis that follows, we assume
that the condition (3) is satisfied.

3 System Contents
This section deals with the analysis of the system contents. First, we calculate the pgf of
the system contents and next, we use this pgf to derive the mean, the variance and the tail
probabilities of the system contents.

3.1 The pgf of the system contents

Let us denote by vk the system contents (i.e., the total number of packets in the queueing
system, including the ones being transmitted, if any) at the beginning of slot k, and by ak
the number of packet arrivals during slot k. Then, in view of the assumptions of the model

5
in section 2, we have

vk+1 = vk − tk + ak , (4)

where

(vk ,c)−
X
tk = tk,j , (5)
j=1

4
with (., .)− = min(., .). The random variable tk denotes the total number of departures at
the end of slot k, i.e., the total number of packets whose service is completed at the end of
slot k.
Note that tk,j corresponds to the number of departures at the end of slot k via the j-th
output channel when there is a packet receiving service from this output channel. In view of
the geometric distribution of the packet service times, tk,j is a Bernoulli distributed random
variable with parameter µ, i.e.,


1 with probability µ,
tk,j =

0 with probability 1 − µ.

We define the conditional pgf Ti (z) as

4
Ti (z) = E[z tk |(vk , c)− = i], i = 0, 1, ... , c,

where E[·] denotes the expected value of the expression between square brackets. Then,
from (5) we have

Ti (z) = (1 − µ + µz)i . (6)

Now let Vk (z) denote the pgf of vk . Using standard z-transform techniques, we can trans-
late the system equation (4) into the z-domain, as follows:

6
 £ ¤
Vk+1 (z) =A(z) E z vk −tk
½ c−1
X · ¸¾ (7)
1 i 1 1
=A(z) Vk (z) Tc ( ) + Prob[vk = i]z Ti ( ) − Tc ( ) .
z i=0
z z

When the steady state is reached, the pgf’s Vk (z) and Vk+1 (z) converge to their equilib-
rium version V (z), which is the pgf of the system contents v at the beginning of an arbitrary
slot in the steady state. Taking limits for k → ∞ in (7) and solving the resulting equation
for V (z), we then obtain the following expression:

Pc−1
A(z) [ Ti ( z1 ) − Tc ( z1 ) ]v(i)z i
i=0
V (z) =
1 − Tc ( z1 )A(z)
Pc−1 (8)
A(z)z c H(z)c i=0 [ H(z)
i−c
− z i−c ] v(i)
= ,
z c − H(z)c A(z)

4
where H(z) = µ + (1 − µ)z, and the constants v(i) are probabilities, defined as v(i) =
Prob[v = i], i = 0, 1, ..., c − 1.
In order to determine V (z) completely, we need to find the c unknown constants v(i),
i = 0, 1, ..., c − 1. These can be obtained by invoking the analyticity of the pgf V (z) inside
the unit disk (z : |z| < 1) of the complex z-plane and the normalization condition V (1) = 1.
Specifically, by means of Rouché’s theorem ([26]), it can be shown that the denominator of
the right-hand side of (8) has exactly c − 1 roots inside the unit disk. We denote these roots
by zj , j = 1, 2, ..., c − 1. Since V (z) is analytic for |z| < 1, the numerator of the right-hand
side of (8) must also be zero at these points. Hence, we have

c−1
X
[ H(zj )i−c − zji−c ] v(i) = 0, j = 1, 2, ..., c − 1. (9)
i=0

From the normalization condition V (1) = 1 and equation (8), we find that

c−1
X 0
µ (c − i) v(i) = cµ − A (1). (10)
i=0

Note that whenever the condition (3) for the existence of a steady state is fulfilled, the
right-hand side of (10) is strictly positive. From equations (9) and (10), the constants v(i)

7
(i = 0, 1, ..., c − 1) can be calculated. Once V (z) is determined, some important performance
measures of the queueing system can be obtained, for instance the mean value, the variance
and the tail distribution of the system contents.

3.2 Mean value and variance of the system contents

The mean system contents can be obtained by taking the first-order derivative of equation
(8) with respect to z in z = 1. By applying the rule of de l’Hospital twice to (8), we get

0
E[v] = V (1)
c−1
X
µ(2 − µ)
= (c2 − i2 ) v(i) (11)
2 [ cµ − A0 (1) ] i=0
0 0
A∗ + 2cA (1) − (2 − µ) [ c2 µ − A (1) ]
+ ,
2 [ cµ − A0 (1) ]

00 0 00
where A∗ = A (1)−2A (1)2 and A (1) is the second-order derivative of A(z) with respect to z
at z = 1. With equation (11), the calculation of the mean system contents is straightforward,
once the unknown probabilities v(i) (i = 0, 1, ..., c − 1) have been obtained. Higher-order
moments of the system contents can be derived in a similar way. For instance, taking the
second-order derivative of equation (8) with respect to z at z = 1 and using (10) and (11),
we find

c−1
X
00 1 µ µ[3]
V (1) = − 0 (c3 − i3 ) v(i)
3 cµ − A (1) i=0
0 0 0 0
(µ[3] + c) [ 2A (1)V (1) + A∗ ] − c(c + 1) µ[2] [ µV (1) − A (1) ] (12)
−
(2 − µ)[ cµ − A0 (1) ]
0 0 0 0 000
µ[3] [ c3 µ − A (1) ] − 3 A∗ [ V (1) − 2A (1) ] + 6A (1)3 − A (1)
+ ,
3 [ cµ − A0 (1) ]

where µ[ζ] = µ2 − ζµ + ζ. Expression (12) can be used to calculate the variance of the system
contents through the relation

00 0 0
V ar[v] = V (1) + V (1) − V (1)2 . (13)

8
3.3 Tail probabilities of the system contents

Another important performance characteristic for a queueing system is the tail distribution
of the system contents. This can be used to estimate the packet loss probability that would
be observed in case of a buffer with finite storage capacity, namely by approximating the
packet loss probability for a system of size N by the probability that the system contents in
the corresponding infinite system exceeds N (see e.g. [27]). We will use here the technique
presented in [15] and [28] to derive the tail distribution of the system contents. Specifically,
we have that for sufficiently large values of N , the tail distribution of the system contents
can be approximated as

zv−N −1
Prob[v > N ] ≈ −Cv . (14)
zv − 1

In the above expression, zv is the real positive pole of V (z) with the smallest modulus out-
side the unit disk, i.e., the dominant pole of V (z), and Cv is the residue of V (z) at z = zv .
From equation (8), it follows that zv is a real positive zero of the denominator of V (z). The
residue Cv can be calculated from (8) as

c−1 · ¸
zv H(zv )A(zv )2 X i zvi
Cv = − H(zv ) − v(i), (15)
Y i=0
A(zv )

0
where Y = zv H(zv ) A (zv ) − cµ A(zv ).

4 Packet Delay
In this section, we derive the pgf of the delay experienced by packets in the queueing system.
The delay of a packet is defined as the total number of slots between the end of the slot
during which the packet arrived in the system and the end of the slot where the packet’s
transmission finishes and the packet leaves the system. From the pgf of the delay, we will be
able to calculate various performance measures, such as the mean value, the variance and
the tail distribution of the packet delay.

9
4.1 The pgf of the delay

Let us consider an arbitrary packet, denoted by P (called the tagged packet). Upon arrival,
it will find a number of other packets in the queueing system. Just after the end of the
arrival slot of P, all packets that arrived in the previous slots, but have not been taken into
service yet, and all packets that arrived in the same slot as the tagged packet, but before this
packet, are waiting in the queue in front of the tagged packet (due to the FCFS queueing
discipline). Whenever there are output channels available for transmission at the beginning
of a slot, the packet at the head of the queue is selected for service, until eventually the
tagged packet itself is served and in the end leaves the system. We now assume that the
packet P arrives in the system during slot k. Let us denote by τk its arrival instant, by vk the
number of packets present in the system at the beginning of slot k, and by qk the number of
packets staying in the system at τk , except the tagged packet itself and the ones that leave
the system at the end of slot k, if such packets exist. Then qk can be expressed as follows:

qk = vk − tk + fk , (16)

where fk denotes the number of packets arriving during slot k before P. It can be shown (see
[2]) that the pgf F (z) of fk is given by

A(z) − 1
F (z) = 0 . (17)
A (1)(z − 1)

Owing to the independent nature of the arrival process, we have (i) that vk has the same
probability distribution as v, which denotes the system contents at the beginning of a ran-
dom slot in the steady state, and (ii) that vk and fk are statistically independent of each
other. From equations (16) and (4), we can derive the pgf Q(z) of qk when the steady state
is reached as

F (z)V (z)
Q(z) = . (18)
A(z)

We now first consider the waiting time experienced by the tagged packet P. The waiting

10
time of a packet is defined as the number of slots between the end of the packet’s arrival
slot and the beginning of the slot when the packet’s transmission starts. We define wk as
the waiting time that P experiences if it arrives during slot k. In order to derive the pgf of
wk , let us briefly discuss a close relationship between the waiting time wk and the random
variable qk (i.e., the number of packets present in the system right after the arrival slot of
the tagged packet P, that have been or will be selected for service before P).
Let us observe the slots following the k-th slot. During the (k + 1)-th slot, there will be
P
tk+1 (= cj=1 tk+1,j ) packets leaving the system at the end of slot k + 1, if qk ≥ c (i.e., just
e
after the end of that slot or at the beginning of slot k + 2, there will be qk − e
tk+1 packets in
the system with priority over the tagged packet P to be taken into service), or the tagged
packet will get into service, if qk < c (i.e., the tagged packet stops waiting). If the tagged
packet didn’t get into service, then a similar remark can be made for the (k + 2)-th slot:
either there will be qk − e
tk+1 − e
tk+2 packets with service priority over the tagged packet at
the end of slot k + 2, if qk − e
tk+1 ≥ c, or the packet P will get into service, if qk − e
tk+1 < c.
We observe that the tagged packet will still be waiting for service during the (k + i + 1)-th
slot only if qk − e
tk+1 − e
tk+2 − ... − e
tk+i ≥ c. This leads to the following relationship between
wk and qk :

w k > i ⇔ qk − e
tk+1 − e
tk+2 − ... − e
tk+i ≥ c.

If we introduce the following random variables :

4 4
sk,0 = c; sk,i = e
tk+1 + e
tk+2 + ... + e
tk+i + c, i ≥ 1,

then the above equation for the waiting time can be rewritten as

wk > i ⇔ qk ≥ sk,i . (19)

Note that since the variables e

tk are i.i.d. random variables, the pgf of sk,i can be expressed as

4
Sk,i (z) = E[z sk,i ] = z c Tc (z)i . (20)

11
 We now begin with the derivation of the pgf Wk (z) of the waiting time wk . From equation
(19) and the easily proven identity

∞
Wk (z) − 1 X i
= z Prob[wk > i],
z−1 i=0

we can write

∞
Wk (z) − 1 X i
= z Prob[qk ≥ sk,i ]
z−1 i=0
∞ X
X ∞
= z i Prob[qk = j] Prob[qk ≥ sk,i |qk = j].
i=0 j=0

Since qk and the sk,i ’s are statistically independent, this can be rewritten as

∞ j ∞
Wk (z) − 1 X X X
= Prob[qk = j] Prob[sk,i = n]z i . (21)
z−1 j=0 n=c i=0

Next, by using (20) and the probability generating property of pgf’s, i.e.,

1 dn ¯ 1 dn £ c ¤ ¯¯
¯
Prob[sk,i = n] = n
S k,i (x)¯ = n
x Tc (x)i ¯ ,
n! dx x=0 n! dx x=0

and working out the sum over i, we find

Wk (z) − 1 X
∞ Xj
X∞
1 ∂n ³ c ´¯
i ¯
= Prob[qk = j] n
x [ T c (x)z] ¯
z−1 j=0 n=c i=0
n! ∂x x=0

∞ j µ ¶¯ (22)
X X 1 ∂n xc ¯
= Prob[qk = j] n
¯ .
j=0 n=c
n! ∂x 1 − zTc (x) x=0

Note that working out the sum over i requires that |zTc (x)| < 1 in the neighborhood of
x = 0. This condition will always be fulfilled for |z| ≤ 1, since |Tc (x)| < 1 for |x| < 1. In
order to derive the partial derivatives in (22), we replace the argument by its partial fraction
expansion. Considering z to be a constant and x the variable of interest, we find

12
 c−1
X
xc 1 −xcp
=− c + , (23)
1 − zTc (x) µ z p=0 zTc0 (xp ) (x − xp )

where the xp ’s are the solutions of the equation

1 − zTc (x) = 0. (24)

Clearly, the xp ’s are functions of z. However, to ease the notation, we write xp instead
of xp (z). Since Tc (x) is a polynomial of degree c, equation (24) has c solutions, which we
assume to be distinct. Notice that (24) has a solution at x = 1, when z = 1. We denote this
0
solution as x0 . Since Tc (1) > 0, this solution is unique. Substituting (23) in equation (22),
we find after some manipulations that

c−1 ∞
Wk (z) − 1 X −xcp X
= 0 Prob[qk = j](x−c −1−j
p − xp ).
z−1 p=0
zT c (x p ) (1 − x p ) j=c

When the steady state is reached, Wk (z) becomes independent of k and converges to a
limit function W (z). By using the definition of the pgf Q(z), we then get

c−1 · ¸
W (z) − 1 X −xcp −c 1 1
= x − Q( )
z−1 p=0
zTc0 (xp ) (1 − xp ) p xp xp
c−1
X c−1
X 1 − xpc−j−1
+ Prob[q = j] .
j=0 p=0
zTc0 (xp ) (1 − xp )

Using the partial fraction expansions

X c−1
1 −1
= 0 ;
1 − zTc (x) p=0
zTc (xp ) (x − xp )
c−1
1 − xc−j−1 X −(1 − xc−j−1 p )
= 0
1 − zTc (x) p=0
zTc (xp ) (x − xp )

at x = 1, we finally find the following expression for the pgf of the packet waiting time:

13
 c−1
X xpc−1 1
W (z) = (z − 1) 0 Q( ). (25)
p=0
zTc (xp ) (1 − xp ) xp

Next, we derive the expression for the pgf of the packet delay. Let the random variable d
indicate the delay experienced by the tagged packet P (when the system has reached equilib-
rium), and let D(z) denote the corresponding pgf. Note that the packet delay is equal to the
sum of the packet waiting time and the packet transmission time. From (2) and (25), we have

D(z) = W (z) · G(z)

µ(z − 1) X
c−1
xc−1 1 (26)
p
= 0 Q( ).
1 − (1 − µ)z p=0 Tc (xp )(1 − xp ) xp

We now proceed with calculating some performance measures related to the delay.

4.2 Mean value and variance of the delay

In a similar way as for the mean system contents, the mean delay experienced by packets
can be found by calculating the first derivative of D(z) at z = 1. From equation (26), we find

0 0 0
E[d] = D (1) = W (1) + G (1). (27)

0
Clearly, the derivation of W (1) is a major contribution to get the mean delay. In order to
0
calculate W (1), we take the first-order derivative of W (z) in (25) with respect to z at z = 1,
using the rule of de l’Hospital twice because for the term where p = 0, x0 = 1 if z = 1. With
the following derivative, found by differentiating (24) with respect to z:

dxp (z) 1
=− 2 0 ,
dz z Tc (xp (z))

we find

· ¸ X
1 ¯¯
c−1
0 1 (c − 1)(µ − 2) 0 xc−1
p
W (1) = + Q (1) + 0 Q( )¯ , (28)
cµ 2 p=1
T c (x p ) (1 − x p ) x p z=1

14
 0
where Q (1) is the first-order derivative of Q(z) evaluated at z = 1. Combining (8), (17) and
(18), the second part of equation (28) can be expressed as:

1 ¯¯ ] ¯¯
c−1
X c−1 c−1 c−i
xc−1
p 1 X X xp [ 1 − (1 − µ + µxp )i
Q( ) ¯ = v(i) ¯ . (29)
p=1
Tc0 (xp ) (1 − xp ) xp z=1 A0 (1) i=0 p=1
Tc0 (xp ) (1 − xp )2 z=1

Using partial fraction expansion again, we have:

c−1
X i ¯
xc−i [ 1 − (1 − µ + µx)i ] i−c
−xc−i
p [ 1 − (1 − µ + µxp ) ] ¯
=µ + 0 ¯ . (30)
1 − Tc (x) p=0
T c (x p ) (x − x p ) z=1

Taking the derivative on both sides of (30) with respect to x, moving the term with p = 0
to the left-hand side and then taking the limit for x → 1, we find (by using the rule of de
l’Hospital several times) that

c−1 c−i
X xp [ 1 − (1 − µ + µxp )i ] ¯¯ 2−µ
¯ = i(c − i). (31)
p=1
Tc0 (xp ) (1 − xp )2 z=1 2c

Combining (31), (29), (28), (11) and (10), we finally have

0 E[v] 1
W (1) = 0 − .
A (1) µ

It is clear from the above equation and (27) that the mean packet delay and the mean system
contents are related through Little’s theorem ([26]), which is a check for the correctness of
our analysis.
Calculation of the second-order derivative of D(z) at z = 1 yields the variance of the
delay through the relation

00 0 0
V ar[d] = D (1) + D (1) − D (1)2 , (32)

which is also called the delay jitter. From (26), we have

00 00 0 0 00
D (1) = W (1) + 2 W (1) G (1) + G (1). (33)

15
Taking the second-order derivative of (25) and applying the rule of de l’Hospital to the term
for p = 0, we have

c−1
( "
00 2 X T1 (xp (1)) S1 1 xp (1)
W (1) = 0 −T1 (xp (1))xp (1)c−3 + 1− +
A (1) p=1 3 3
c µ Xp1 2 µ A( xp1(1) ) − 1
# )
3
T (x
1 p (1)) x p (1) T (x
1 p (1)) S 2
+(1 − c)xp (1)2 + +
cµXp1 Xp1 T1 (xp (1)) c2 µ2 Xp21 (34)
½ 0 0 000
1 00 2ξA (1) + A∗ 0 ξ − A (1) ∗ A (1)
+ 2 2 V (1) + V (1) + A +
cµ A0 (1) A0 (1) 3A0 (1)
00
h (c + 1)µ2 i¾
−A (1) − (c − 1) +ξ−µ ,
6

where
c−1 ½
X h T (x (1)) ii
1 p
S1 = [(c + i)(µ − 1) − cµxp (1)]
i=0
xp (1)
h 1 ii ¾
+ [(c + i)(µ − 1) − iµ xp (1)] v(i),
xp (1)
c−1 ½h ¾
X T1 (xp (1)) ii h 1 ii
S2 = − v(i),
i=0
xp (1) xp (1)

Xp1 = xp (1) − 1,

ξ = 2 − µ − c.

From the analysis above, an explicit expression for the packet delay jitter or the variance of
the packet delay can be obtained.

4.3 Tail probabilities of the delay

The tail distribution of the packet delay, for a sufficiently large value of T , can be expressed as

zd−T −1
Prob[d > T ] ≈ −Cd , (35)
zd − 1

when D(z) has a singular dominant pole zd , and where Cd is the residue of D(z) at z = zd , or

· µ ¶ ¸
1 C2d zd−T −1
Prob[d > T ] ≈ − C1d − 2 + T + , (36)
zd − 1 zd zd − 1

16
when D(z) has a dominant pole zd with multiplicity 2, and where C1d and C2d are determined
by

d £ ¤
C1d = lim (z − zd )2 D(z) , C2d = lim (z − zd )2 D(z). (37)
z→zd dz z→zd

In order to find the dominant pole zd , note that from (2), (25) and (26), it follows that
poles of D(z) are to be found among the following z’s:
• z so as to make 1 − (1 − µ)z = 0. This requires z = 1/(1 − µ).
• z so as to make 1/xp (z) (for some p) a pole of Q(z). This requires z = 1/Tc ( z1∗ ), where
z∗ is a pole of Q(z).
• z so as to make xp (z) − 1 = 0 (for some p). This case, however, does not yield a pole
since it requires z = 1 (see (24)).
As in [20], it can be shown that if z∗ is a pole of Q(z) with z∗ 6= zv , then |1/Tc ( z1∗ )| >
|1/Tc ( z1v )|, and hence a pole of D(z) given by z = 1/Tc ( z1∗ ) cannot be the dominant pole.
Thus, in order to find zd , we need to compare the modulus of 1/(1 − µ) and 1/Tc ( z1v ). As a
result, we have to distinguish the following three cases:
Case 1. When 1/Tc ( z1v ) < 1/(1−µ), there is a dominant pole zd with multiplicity 1, given by

1
zd = = A(zv ),
Tc (1/zv )

and the tail distribution of the delay is given by the geometric form (35), where

G(A(zv )) [A(zv ) − 1]2

Cd = Cv .
A0 (1) zvc (zv − 1)2

We notice that this is always valid for c = 1 because of the monotonically increasing
character of Tc (z) along the positive real axis, i.e. 1/T1 ( z1v ) < 1/T1 (0) = 1/(1 − µ).
Case 2. When 1/Tc ( z1v ) > 1/(1 − µ), the dominant pole zd also has a multiplicity 1 and is
equal to

17
 1
zd = . (38)
1−µ

Hence, in this case, we also have a tail distribution of the form (35), where

µ 1
Cd = − 2
W( ).
(1 − µ) 1−µ

Case 3. When 1/Tc ( z1v ) = 1/(1 − µ), we get a dominant pole zd of multiplicity 2, given by

1 1
zd = = A(zv ) = 1/Tc ( ), (39)
1−µ zv

and the tail distribution of the delay has the form (36), where C1d and C2d are
obtained from (26) and (37) as
( "
zv H(zv ) X h
c−1
µ [1 − A(zv )] Cv [A(zv ) − 1]
C1d = − [ cµ − iH(zv ) ] zvi
1−µ c2 µ2 A0 (1)zvc (zv − 1)2 cµCv i=0
i zv H(zv ) h 0
+ (1 − µ)zv A(zv ) i H(zv )i v(i) + − [ µ + cµ − 2H(zv ) ]A (zv )
2Y
i Y h (3c − 1)z − (1 + 2c + cz )H(z )
00 v v v
+ zv H(zv )A (zv ) −
cµA(zv ) zv − 1
# )
cµ + zv H(zv )A (zv ) i X xc−1
0
p Q(1/x p )
+ + 0 ,
A(zv ) − 1 p
T c (x p )(1 − x p )

µ A(zv ) [A(zv ) − 1]2

C2d = −Cv .
(1 − µ)A0 (1)zvc (zv − 1)2

In the last term of the expression for C1d , xp is decided by Tc (xp ) = 1/A(zv ), except
the term for 1/xp = zv .

5 Special Cases
In this section, we consider some special cases (c = 1 and µ = 1) in order to check the results
obtained above.

18
 When c = 1, i.e., for a discrete-time queueing system with one output channel and geo-
metric service times, the pgf of the system contents can be derived from (8) as

0
[ µ − A (1) ] (z − 1)A(z)
V (z) = . (40)
z − H(z)A(z)

From (26) and considering the fact that x0 is the root of the equation 1 − zT1 (x) = 0, the
pgf of the packet delay can be expressed as

D(z) =G(z) Q (G(z))

0
[µ − A (1)] z [ A(G(z)) − 1] (41)
= .
A0 (1) [ z − A(G(z)) ]

Equations (40) and (41) correspond exactly to the results found in [2].
When µ = 1, i.e., for a queueing system with c output channels and deterministic service
times of one slot, the pgf of the system contents can be derived from (8) as

Pc−1 c
A(z) i=0 (z − z i ) v(i)
V (z) = . (42)
z c − A(z)

Considering the fact that the xp ’s are the roots of the equation 1 − zxc = 0, the pgf of the
packet delay can be simplified from (25) and (26) as

c−1
c 1 X zc − 1
D(z ) = Q(αj z), (43)
c j=0 1 − (αj z)−1

where α = exp(2Πι/c) (ι is the imaginary unit). Equations (42) and (43) correspond exactly
to the results found in [14].

6 Numerical Examples
In order to illustrate the results obtained above, let us consider a number of numerical ex-
amples. Throughout this section, we assume that the number of output channels c equals
1, 2, 4 or 16. The number of packets that arrive during a slot has a geometric distribution, i.e.,

1
A(z) = .
1 + λ − λz

19
 In Figure 1, the mean system contents (divided by c) is plotted versus the average carried
load ρ (= λ/cµ) for the four different numbers of output channels. All curves have the same
µ = 0.75. Note that the scaling of the mean system contents by a factor c is done to allow
a fair comparison between systems with dedicated or shared buffers. We observe that for a
given ρ, when one shared buffer with multiple output channels is used to transport a given
mean number of packets per slot, the equivalent mean system contents (per output channel)
is less than in case several dedicated buffers with a single output channel are used. This
also confirms the conclusion that the shared resource policy has an advantage over the non-
shared one. For a given number of output channels, the mean system contents increases with
increasing values of ρ. A similar figure can be plotted for the mean packet delay in terms of
ρ, in view of Little’s law.
In Figure 2, the variance of the packet delay is shown (for µ = 0.75 and c = 1, 2, 4, 16)
versus ρ. Clearly, for a given value of ρ, the delay jitter for the system with more output
channels is less than the jitter for systems with less output channels.
In Figures 3 and 4, we have plotted the tail distributions of the system contents and the
packet delay, for µ = 0.75 and ρ = 0.25. To fairly compare systems with dedicated or shared
buffers, the tail distribution of the system contents Prob[v > N ] is shown in Figure 3 as a
function of N/c. This can be used to approximate the equivalent packet loss probability in a
system with c output channels and a finite capacity N/c per output channel. Once again, we
observe that for a system with more output channels, its equivalent packet loss probability
is much less than for systems with less output channels. In Figure 4, the probability that
the delay exceeds some given threshold T is plotted versus T for the four different numbers
of output channels. Clearly, for a given T , the tail distribution of the delay decreases as the
number of the output channels increases.
In Figure 5, we have plotted the quantiles of the system contents, i.e., the amount of
buffer space required to yield a loss probability of 10−6 and 10−12 , versus the value of ρ, for
µ = 0.75 and c = 4. For high values of ρ (near the limiting value ρ = 1.0), a large storage
capacity will be needed to keep the loss probability at an acceptable level.

20
7 Concluding Remarks
In this paper, we have developed an analytical technique for the analysis of a discrete-time
multiserver queueing system with geometric service times and a general uncorrelated arrival
process. Our approach is based on the use of probability generating functions and leads
to closed-form expressions for the mean values, the variances and the tail distributions of
the system contents and the packet delay. The obtained expressions are easy to evaluate
numerically.
Several further generalizations of the model could be investigated. For instance, more
general non-geometric service-time distributions could be allowed or some degree of corre-
lation between the numbers of packet arrivals in different slots could be introduced in the
model. A first result in this direction is reported in [29], where our analytical technique has
been adapted to deal with the case of a correlated two-state Markovian arrival process.

Acknowledgements
The authors wish to thank the anonymous referees for their valuable comments and sugges-
tions.

References
[1] Hunter JJ. Mathematical techniques of applied probability, Volume 2, Discrete time
models: techniques and applications. New York: Academic Press, 1983.

[2] Bruneel H, Kim BG. Discrete-time models for communication systems including ATM.
Boston: Kluwer Academic Publishers, 1993.

[3] Woodward ME. Communication and computer networks: modelling with discrete-time
queues. London: Pentech Press, 1993.

[4] Takagi H. Queueing analysis, A foundation of performance evaluation, Volume 3:

discrete-time systems. Amsterdam: North-Holland, 1993.

[5] Robertazzi TG. Computer networks and systems: queueing theory and performance
evaluation. New York: Springer-Verlag, 2000.

21
[6] Daduna H. Queueing networks with discrete time scale: explicit expressions for the
steady state behavior of discrete time stochastic networks. New York: Springer-Verlag,
2001.

[7] Kurose JF, Mouftah HT. Computer-aided modeling, analysis, and design of communi-
cation networks. IEEE Journal on Selected Areas in Communications 1988;6(1):130-45.

[8] Neuts MF. Matrix-geometric solutions in stochastic models: an algorithmic approach.

Baltimore: Johns Hopkins University Press, 1981.

[9] Blondia C, Casals O. Statistical multiplexing of VBR sources: a matrix analytic ap-
proach. Performance Evaluation 1992;16:5-20.

[10] Wittevrongel S, Bruneel H. Discrete-time ATM queues with independent and correlated
arrival streams. In: Kouvatsos D (Ed.). Performance evaluation and applications of
ATM networks. Boston: Kluwer Academic Publishers, 2000. p.387-412.

[11] Kouvatsos DD, Tabet-Aouel NM, Denazis SG. ME-based approximations for general
discrete-time queueing models. Performance Evaluation 1994;21(1-2):81-109.

[12] Chu WW. Buffer behavior for Poisson arrivals and multiple synchronous constant out-
puts. IEEE Transactions on Computers 1970;C-19:530-4.

[13] Li S-Q. A general solution technique for discrete queueing analysis of multimedia traffic
on ATM. IEEE Transactions on Communications 1991;39:1115-32.

[14] Bruneel H, Steyaert B, Desmet E, Petit G. An analytical technique for the derivation of
the delay performance of ATM switches with multiserver output queues. International
Journal of Digital and Analog Communication Systems 1992;5:193-201.

[15] Bruneel H, Steyaert B, Desmet E, Petit G. Analytic derivation of tail probabilities

for queue lengths and waiting times in ATM multiserver queues. European Journal of
Operational Research 1994;76:563-72.

[16] Bruneel H, Wuyts I. Analysis of discrete-time multiserver queueing models with constant
service times. Operations Research Letters 1994;15:231-6.

[17] De Prycker M. Asynchronous transfer mode: solution for broadband ISDN. New York:
Ellis Horwood, 1991.

22
[18] Rubin I, Zhang Z. Message delay and queue-size analysis for circuit-switched TDMA
systems. IEEE Transactions on Communications 1991;39:905-14.

[19] Georganas ND. Buffer behavior with Poisson arrivals and bulk geometric service. IEEE
Transactions on Communications 1976;24:938-40.

[20] Laevens K, Bruneel H. Delay analysis for discrete-time queueing systems with multiple
randomly interrupted servers. European Journal of Operational Research 1995;85:161-
77.

[21] Hsu J. Buffer behavior with Poisson arrival and geometric output processes. IEEE Trans-
actions on Communications 1974;22:1940-1.

[22] Gao P, Wittevrongel S, Bruneel H, Zhang S. A discrete-time queueing system with

correlated arrivals and geometric service times. Belgian Journal of Operations Research,
Statistics and Computer Science (JORBEL) 2002;41(3).

[23] Briem U, Theimer TH, Kröner H. A general discrete-time queueing model: analysis
and applications. Proceedings of the 13th International Teletraffic Congress, ITC 13,
Copenhagen, 1991. p. 13-9.

[24] Bruneel H. Performance of discrete-time queueing systems. Computers & Operations

Research 1993;20(3):303-20.

[25] Herrmann C. The complete analysis of the discrete time finite DBMAP/G/1/N queue.
Performance Evaluation 2001;43(2-3):95-121.

[26] Kleinrock L. Queueing systems, Volume I: theory. New York: Wiley, 1975.

[27] Bisdikian C, Lew JS, Tantawi AN. On the tail approximation of the blocking probability
of single server queues with finite buffer capacity. Proceedings of the Second Interna-
tional Conference on Queueing Networks with Finite Capacity, Research Triangle Park,
1993. p. 267-80.

[28] Woodside CM, Ho EDS. Engineering calculation of overflow probabilities in buffers with
Markov-interrupted service. IEEE Transactions on Communications 1987;35:1272-7.

[29] Gao P, Wittevrongel S, Bruneel H. Discrete-time multiserver buffer systems with cor-
related arrivals and geometric service times. Submitted for publication.

23
 Peixia Gao has been working as a Ph.D. student at the SMACS Research Group,
Department TELIN (Telecommunication and Information Processing), Faculty of Applied
Sciences, Ghent University, Belgium since 1999. Her main research interests include the
stochastic modeling of communication networks and the analysis of discrete-time queueing
models.
Sabine Wittevrongel was born in Gent, Belgium, in 1969. She received the M.S.
degree in Electrical Engineering and the Ph.D. degree in Applied Sciences from Ghent Uni-
versity, Belgium, in 1992 and 1998, respectively. Since September 1992, she has been with
the SMACS Research Group, Department TELIN, Faculty of Applied Sciences, Ghent Uni-
versity, first in the framework of various projects, from October 1994 to September 2001, as
a researcher of the Fund for Scientific Research - Flanders (Belgium) (F.W.O.), and since
October 2001 as a full time Professor. Her main research interests include discrete-time
queueing theory, the performance evaluation of communication networks, the study of traffic
control mechanisms, and the analysis of ARQ protocols.
Herwig Bruneel was born in Zottegem, Belgium, in 1954. He received the M.S. de-
gree in Electrical Engineering, the degree of Licentiate in Computer Science, and the Ph.D.
degree in Computer Science in 1978, 1979 and 1984 respectively, all from Ghent University,
Belgium. He is full time Professor in the Faculty of Applied Sciences and head of the Depart-
ment TELIN at the same university. He also leads the SMACS Research Group within this
department. His main personal research interests include stochastic modeling and analysis
of communication systems, discrete-time queueing theory, and the study of ARQ proto-
cols. He has published more than 170 papers on these subjects and is coauthor of the book
H. Bruneel and B.G. Kim, “Discrete-Time Models for Communication Systems Including
ATM” (Kluwer Academic Publishers, Boston, 1993). Since October 2001, he serves as the
Academic Director for Research Affairs at Ghent University.

24
 Figure captions

Figure 1. Mean system contents versus average carried load ρ.

Figure 2. Variance of the packet delay versus average carried load ρ.
Figure 3. Tail distribution of the system contents, Prob[v > N ], versus N/c.
Figure 4. Tail distribution of the packet delay, Prob[d > T ], versus T .
Figure 5. Buffer space N required to yield an overflow probability of 10−6 and 10−12 , versus
the average carried load ρ.
 mean system contents divided by c
6

4
c=1

3
2

4
2

16

0 0.2 0.4 0.6 0.8 1

rho

Figure 1
 variance of packet delay
60

50

40

c=1
30

20

10 4

16
0 0.2 0.4 0.6 0.8 1
rho

Figure 2
 Prob[system contents>N]

.1

.1e–1

.1e–2

.1e–3

1e–05

1e–06

1e–07 c=16 4 2 1

1e–08

1e–09

1e–10 0 2 4 6 8 10 12 14 16 18 20
N/c

Figure 3
 Prob[packet delay>T]
.1e–1

.1e–2
c=1

.1e–3

1e–05
2

1e–06 4

16

1e–07

1e–08 4 6 8 10 12 14
T

Figure 4
 buffer space
400

300

200 1e–12

100 1e–6

0 0.2 0.4 0.6 0.8 1

rho

Figure 5