Applied Mathematical Modelling 34 (2010) 931946

Contents lists available at ScienceDirect

Applied Mathematical Modelling

journal homepage: www.elsevier.com/locate/apm

The discrete-time MAP/PH/1 queue with multiple working vacations

Cosmika Goswami, N. Selvaraju *
Department of Mathematics, Indian Institute of Technology Guwahati, Guwahati 781 039, India

a r t i c l e i n f o a b s t r a c t

Article history: We consider a discrete-time single-server queueing model where arrivals are governed by
Received 25 June 2008 a discrete Markovian arrival process (DMAP), which captures both burstiness and correla-
Received in revised form 19 June 2009 tion in the interarrival times, and the service times and the vacation duration times are
Accepted 23 July 2009
assumed to have a general phase-type distributions. The vacation policy is that of a work-
Available online 8 August 2009
ing vacation policy where the server serves the customers at a lower rate during the vaca-
tion period as compared to the rate during the normal busy period. Various performance
Keywords:
measures of this queueing system like the stationary queue length distribution, waiting
Discrete-time queue
Working vacation
time distribution and the distribution of regular busy period are derived. Through numer-
Markovian arrival process ical experiments, certain insights are presented based on a comparison of the considered
Phase-type distribution model with an equivalent model with independent arrivals, and the effect of the parame-
Matrix-geometric method ters on the performance measures of this model are analyzed.
2009 Elsevier Inc. All rights reserved.

1. Introduction

Vacation queueing systems are a very important class of queues which are extensively used in the performance analysis
of communication systems, computer networks, production management and many other real life situations. In a queueing
context, a vacation is the period when the server is not attending to a particular targeted queue. The server may be under
repair, attending to other queues, or simply forced to stop serving. Many vacation policies have been presented and analyzed
in the literature. The monograph by Tian and Zhang [1] gives a comprehensive account of the various types of queues with
vacations and their applications.
Though the continuous-time vacation queueing models have received wide attention, many a time there is also a need for
discrete-time models. This is because in many communication and time-sharing systems, the time is discretized into time
slots which can be taken as one time unit. For example, in asynchronous transfer mode (ATM) networks, each cell has a xed
size and the service time of one cell can be used to discretize time and transition from one state to another is considered only
at the start or end of these epochs. Therefore, the discrete-time queueing models are taken to be more appropriate for such
computer and communication systems rather than the continuous ones and hence interest has grown for the discrete-time
models also. Takagi [2] gave a detailed analysis of Geo/G/1 queue with various vacation policies. Later, Zhang and Tian [3]
added to the model the multiple adaptive vacations. A discrete-time GI/Geo/1 queue with server vacation is presented by
Tian and Zhang [4].
Recently, Servi and Finn [5] studied an M/M/1 queueing model with a new type of vacation policy called a working vaca-
tion policy. In this policy, the server does not completely stop serving the customers during a vacation period but it serves at
a lower rate than in a normal busy period. Wu and Takagi [6] extended this work to M/G/1/WV model with generally dis-
tributed service times as well as vacation durations. Baba [7] considered the GI/M/1/WV system with the distribution of the

* Corresponding author. Tel.: +91 361 2582611; fax: +91 361 2582649.
E-mail address: nselvaraju@iitg.ernet.in (N. Selvaraju).

0307-904X/$ - see front matter 2009 Elsevier Inc. All rights reserved.
doi:10.1016/j.apm.2009.07.021
932 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946

vacation duration having an exponential distribution. Also, the nite buffer model GI/M/1/N/WV is presented by Banik et al.
[8] with multiple working vacation policy.
Similarly, in the discrete-time counterpart of the M/M/1/WV case, Tian et al. [9] analyzed the Geom/Geom/1/WV with
geometrically distributed vacation duration using quasi-birthdeath process and matrix-geometric solution method. Subse-
quently, Li et al. [10] presented the GI/Geo/1 queue with multiple working vacation, with geometrically distributed vacation
durations. Considering two variations of the arrival and departure schemes, they obtained the results on the stationary dis-
tributions and stochastic decomposition properties under both the schemes. We attempt to generalize these works with a
more general arrival process and phase-type distributions.
Now a days, for high-speed networks, burstiness in sources like voice, coded video and bulk data transfers are very impor-
tant factors in the analysis of system performance. They may even have a considerable correlation between interarrival times
of packets or cells. Thus a smooth Bernoulli process or an independent renewal arrival process is not an efcient assumption
for an arrival process of the network trafc in a discrete-time system since such a process may not capture these wide range
of ow characteristics. As a queueing systems performance is highly sensitive to correlation and bursty characteristics in the
arrival process, the Markov modulated Bernoulli process (MMBP) or more generally Markovian arrival process (MAP) is
widely used for a non-renewal type arrival processes in the performance evaluation of high-speed network trafc since it
can capture the properties of both burstiness and correlations. Keeping this aspect in mind, we extend the above mentioned
works to a more generalized model where the correlation can also be captured. The work that is related to ours in this direc-
tion is that of Alfa [11] who, using the matrix-geometric procedure, has analyzed the discrete-time MAP/PH/1 queue with
exhaustive and non-exhaustive service where the (usual) vacation follows a phase-type distribution. In this work, we study
the discrete-time MAP/PH/1 model with multiple working vacations whose durations are phase-type distributed random
variables.
The application of our model is mainly in telecommunication network. Servi and Finn, who introduced the working vaca-
tion policy, was motivated by a WDM optical access network. The wavelength division multiplexing (WDM) is a method of
transmitting data from different sources over the same bre optic link at the same time whereby each data channel is carried
in its own unique wavelength. The result is a link with an aggregate bandwidth that increases with the number of wave-
lengths employed. In this way, WDM technology can maximize the use of bre optic infrastructure that is available. So only
one bre link is required in place of two or more. Since many of the bandwidth-intensive applications require multiple ser-
vices for efciency purpose, WDM technique emerges as a promising solution to meet the rapidly growing demands on
bandwidth in present communication network. Servi and Finn showed the appropriateness of the working vacation model
for WDM optical access networks over the alternative of having no reconguration or reconguring all wavelengths. As for
the trafc model, the Markovian arrival process (MAP) is chosen due to the following reasons. The MAP is one of the most
powerful trafc model in terms of its ability to mimic complex statistical properties of a network trafc e.g. the self-similar
nature and the long-range dependent behavior. Also in the work of Servi and Finn, they have mentioned that the assumption
of uncorrelated arrivals may be problematic as the trafc sources are highly correlated.
The paper is organized as follows. In Section 2, we provide the description of the discrete-time DMAP/PH/1/WV(PH) mod-
el and the notations used. There we formulate the model as a quasi-birthdeath (QBD) process and derive the stability con-
dition. The stationary distribution and the steady state mean queue length are also obtained. Sections 3 and 4 give the
waiting time distribution and the distribution of regular busy period respectively. Finally in Section 5, we have given numer-
ical examples where we provide self explanatory tables and graphs to show the difference of this model to that which as-
sumes independent arrivals. Also we have presented the effect of model parameters on the performance measures.

2. The quasi-birthdeath (QBD) process model

We consider here a discrete-time single-server queue where the customers arrive according to a Markovian arrival pro-
cess in discrete-time (DMAP). The DMAP process is a semi-markov process described by two n dimensional substochastic
matrices D0 and D1 and is such that I  D0 is nonsingular and is denoted by DMAPD0 ; D1 [12,13]. Here D0 ij represents
the transition probability from phase i to phase j with no arrival and D1 ij represents the transition probability from phase
i to phase j with only one arrival. The matrix D D0 D1 is an irreducible stochastic matrix. If p is the invariant probability
vector of the Markov chain described by D satisfying pD p and pe 1, then the customer arrival rate can be given by
k pD1 e, where e is the column vector of ones of appropriate dimension.
The service times and the duration of working vacations are assumed to follow discrete phase-type (PH) distributions. A
discrete PH-distribution denotes the distribution of time until absorption in a nite Markov chain whose transition proba-
bility matrix is of type
" #
S S0
PS ;
0 1

and is denoted by PHb; S, where b is the initial probability vector satisfying be 1 and S0 e  Se. During the normal busy
period the server serves the customers according to a PHb; S of dimension m and rate lb with lb nS0  be, where n is the
stationary probability vector of S S0 b satisfying nS S0 b n and ne 1. And, the service during a vacation is also of a
phase-type distribution, PHb; S, of dimension m, having the same initial probability vector b but with different transition
 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946 933

probabilities. During the working vacation period the server serves at a rate lv such that lv < lb . The rate lv  nS0  be,
where  nS S0 b 
n and 
ne 1.
The duration of a working vacation is also phase-type, PHd; L, of dimension r with duration rate h and the transition
probability matrix P L similar to the one given above with L0 e  Le and de 1. The mean duration of vacation is given
by v dI  L1 e. From a normal busy period the server goes to working vacation as soon as the system becomes empty.
If a customer arrives during a working vacation period then the system serves the customer at the rate lv . When the vacation
ends, if at least one customer is there in the system at the end of the working vacation, the server switches its service rate
from lv to lb and a normal busy period starts. Otherwise, the system goes to another working vacation. At the end of a vaca-
tion if there is a customer whose service is not yet completed, then the ongoing service is continued until completion but
with the higher service rate. So we need to keep track of the phase of arrival process, phase of service time and the phase
of vacation duration. The customers are served according to the arrival order i.e., FCFS. The interarrival times, service times
and vacation times are assumed to be mutually independent.
In discrete-time queueing systems, the arrivals, the departures and the end of vacations may occur at the same time. It is
assumed that time is slotted in intervals of queue length with length of slot being unity. Without loss of generality, we can
assume that if there is an arrival at time t, then its precise occurrence lies somewhere inside the interval t; t , where t is
the moment immediately after t. On the other hand, if there is a departure at time t, then its precise occurrence lies some-
where inside the interval t ; t, where t is the moment immediately before t. This model is referred as the early arrival sys-
tem (EAS). The beginning and ending of vacations also occurs at the instant t .
Given the just dened sign convention, the state of the system under consideration can then be given by a discrete-time
Markov chain

D fNt ; Q t ; Jt ; V t ; K vt [ Q t ; Jt ; K bt ; t 0; 1; 2; . . .g;
where N t is the number of customers in the system at time t , Q t represents the state of the system at time t i.e., Q t 0, if
the system is on vacation and Q t 1, if the system is on normal busy period, J t is the phase of arrival, V t is the phase of the
vacation duration, K vt gives the phase of service during working vacation and K bt gives the phase of the service in normal busy
period. This Markov chain is a discrete-time quasi-birthdeath (QBD) process with the state space of D given by
E fff0g  f0; i; jgg [ fN  ff0; i; j; kg [ f1; i; kggg; 1 6 i 6 n; 1 6 j 6 r; 1 6 k 6 mg
as, when there are no customers in the system there is no need to keep track of the phase of service.
Using the lexicographical order of the states i.e., 0; 0; 1; 1; . . . ; 0; 0; 1; r; 0; 0; 2; 1; . . . ; 0; 0; n; r; 1; 0; 1;
1; 1; . . . ; 1; 0; n; r; m; 1; 1; 1; 1 . . . ; 1; 1; n; m; 2; 0; 1; 1; 1 . . . ; we get the transition probability matrix P of
the QBD process is given by
0 1
B00 B01
B C
B B10 A1 A0 C
B C
PB
B A2 A1 A0 C;
C
B C
@ A2 A1 A0 A
: : :
" #  
where B00 D0  L L0 d, B01 D1  L  b D1  L0 b, B10 D0  L L0 d  S0 , Ai
Ai3 Ai2
, i 0; 1; 2,
D0  S0 d 0 Ai0
A13 D0  L  S D1  L  S0 b, A12 D0  L0  Se  n D1  L0  S0 b, A10 D0  S D1  S0 b, A03 D1  L  S, A02 D1 
L  Se  n, A00 D1  S, A23 D0  L  S b, A22 D0  L0  S0 b, A20 D0  S0 b.
0  0

The vector n gives the initial probability of switched service from working vacation to normal busy period. A0 ; A1 ; and A2
are square matrices of dimension nmr 1, B00 is a square matrix of dimension nr, B01 is of dimension nr  nmr 1 and B10
is of dimension nmr 1  nr.

2.1. Stability condition

Theorem 1. The QBD process D is positive recurrent if and only if k < lb :

Proof. We assume that the QBD is irreducible. Since all Ai , i 0; 1; 2 are upper triangular so the matrix A A0 A1 A2 is
reducible and stochastic matrix. The matrix A is
2  3
D  L  S S0 b D  L0  Sem  n S0 b
A4 5:
0 D  S S0 b
Now, Theorem 7.3.1 in [14] gives the condition for positive recurrence of the QBD when A is reducible. The matrix A can
be written after permutation of its rows and columns as
934 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946

" #
C 1 0
A ;
C 2 C 0

where C 1 is irreducible and stochastic and C 0 is sub-stochastic. Since the matrices A0 ; A1 and A2 are non-negative, they are
similarly structured and we have for i 0; 1; 2
" 1
#
Ci 0
Ai 2 0
;
Ci Ci
1 2 0
with C i Ai0 ; C i Ai2 and C i Ai3 . The theorem states that the QBD is positive recurrent if and only if

^ C 1
p 2 e >p ^ C 1
0 e; 1
1 1 0 1
where p
^ is the stationary probability vector of C . Here C D  S S b and p
^ p  n such that p
^C p
^ and p
^ e 1.
From (1) we have

p  nD0  S0 be > p  nD1  Se:

Adding term (pD1  nS0 be) both sides of the above inequality gives

pD0 D1 e  nS0 be > pD1 e  nS S0 be;

which nally reduces to k < lb . h

2.2. The rate matrix R

Since the Markov chain is a QBD process, it can be analyzed by the matrix-geometric method [12]. To calculate the invari-
ant vector of the process, it is necessary to solve for the minimum non-negative solution to the matrix equation
R A0 RA1 R2 A2 and this solution is called the rate matrix, denoted by R. The matrix R is a square matrix of dimension
nmr 1. The matrix R can be computed using the well-known algorithms [14]. From the previous result, we get that for
k < lb the Markov chain D is positive recurrent and the matrix R has all its eigenvalues in the unit disk [12].
The matrices Ak ; k 0; 1; 2 are all upper triangular, and so R will also have the same structure, say
 
R00 R01
R :
0 R11

If the dimensions of the matrices Ai s i 0; 1; 2, are higher, so will be the matrix R. To have computational efciency we
can break the equation R A0 RA1 R2 A2 into three matrix equations as

R00 A03 R00 A13 R200 A23

R01 A02 R00 A12 R01 A10 R200 A22 R00 R01 R01 R11 A20 2
R11 A00 R11 A10 R211 A20

from which R00 ; R01 ; R11 can be easily computable than computing the matrix R. A simple approach for computing Ri;j is as
follows. Let Rni;j be the value of Ri;j at the nth iteration, where R0i;j 0. We insert Rni;j into the RHS of the equations to obtain Rn1 i;j
on the LHS. This iteration procedure is continued until jRn1 n n
i;j u; v  Ri;j u; v j < ; 8i; j; 8u; v , where Ri;j u; v is the u; v th
element of the matrix Rni;j and  is a very small positive number, usually about 1012 .

2.3. Stationary distribution

Let x be the stationary probability vector associated with the transition matrix P such that

xP x; xe 1:
Let us partition the vector x into x x0 x1 x2 . . ., where x0 is a row nr-vector, xk k P 1 is a row nmr 1-vector. The
partition is based on the number of customers in the system. Each of the vector xk can further be partitioned, for computa-
tional purposes, into xk0 xk1 , for k P 1, where xk0 is a nmr-vector and xk1 is a nm-vector and this further partition is by the
state of the server (in vacation or not). Then, we have from the matrix-geometric method that

xk1 xk R; for k P 1: 3

We calculate x0  x1  from

x0 x1  x0 x1  BR; 4
 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946 935

where
2 3
  D0  L L0 d D1  L  b D1  L 0 b
B00 B01 6 7
BR 6 0
4 D0  L L d  S
0
A13 R0 A23 A12 R0 A22 R10 A20 7
5:
B10 A1 RA2
D 0  S0 d 0 A10 R1 A20

Now, since I  A1  RA2 is nonsingular, it is easy to see that x1 x0 B01 I  A1  RA2 1 . The vector x0 is then obtained from
the normalization condition

x0 e x0 B01 I  A1  RA2 1 I  R1 e 1: 5

Calculating the inverses of I  A1  RA2 and I  R can be carried out in a simpler manner because they are both
upper triangular blocks, and this fact can be utilized to compute the inverses block by block.
Matrix-geometric method is a signicant modelling tool because it gives the ability to construct and analyze, in a
unied way and in an algorithmically tractable manner, a wide class of stochastic models [15]. One of the practical
advantage of this method is that these elementary matrix operations can easily be programmed for a high-speed com-
puter. This approach of solving makes the expressions in matrix form simpler than the corresponding expressions gi-
ven in terms of eigenvalues. Also the fundamental matrices have direct probabilistic interpretation, while the
eigenvalues do not. The geometric relation between the vectors of the stationary probability distribution allows for
simple closed form formulas for the computation of measures of interest such as the system queue length and busy
cycle.

2.4. The queue length

The vector xi is the joint probability of the number of customer, arrival phase, working vacation phase, service phase dur-
ing busy period or service phase during working vacation period, with i customers in the system.
Let yi be the marginal probability density of nding i customers in the system at an arbitrary time t, then

yi xi e;

and the mean number of customers in the system is

X
1
EN iyi x1 I  R2 e: 6
i0

3. The waiting time

The distribution of waiting time can be studied by using absorbing Markov chain. The absorbing states are the states
when the target customer reaches the head of the queue, which is the point we assume the customers wait in queue is over.
With the help of the stationary distribution of the number of customer in the system for this Markov chain we can nally get
the waiting time distribution of the customers [16].
First let us nd the state of the system at the arrival of an arbitrary customer. Let zi be the stationary probability vector of
nding i customers in the system by an arriving customer and partitioning zi for i P 1, we get

zi zi0 zi1  and z0 z00 with ze 1;

where zij represents the probability vector that an arriving customer nds i customers in the system with system is in state
jj 0 for the system on working vacation period and j 1 for the system on normal busy period). This can happen as a re-
sult of three mutually exclusive cases. When the system is in vacation, it may be empty (j 0; i 0) or it may have more
than one customer (j 0; i P 1). But when the system is busy there will be at least one customer (j 1; i P 1) in the system
because if the system becomes empty it always goes for vacation. So we dont have the state j 1; i 0. According to these
cases we get the following probabilities

Case 1: The probability that the arriving customer nds the system on working vacation with no customer i.e., j 0; i 0
is,
1h i
z0 x0 D1  L L0 d x1;0 D1  L L0 d  S0 x1;1 D1  S0 d :
k
The rst term is the probability that in the time slot of the arriving customer (D1 ) there is no customer in the system
x0 and system is in working vacation (L) or goes to another vacation after completing one (L0 d). The second term
says, if there is a customer in the system x1;0 then there must be a service completion within the slot S0 so that
the arriving customer (D1 ) nds the system empty. The system is in working vacation (L) or after completing both
936 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946

the service and the vacation, it goes to another vacation (L0 d). The third term means that if the system was in normal
busy period with one customer x1;1 and the service of that customer completes (S0 d), then the system goes to vaca-
tion and the arriving customer (D1 ) nds the empty system.
Case 2: The probability that the arriving customer nds the system on working vacation period with more than one cus-
tomer i.e., j 0; i P 1 will be,

1h i
zi0 xi0 D1  L  S xi1;0 D1  L  S0 b :
k
The rst term is the probability that in the time slot of the arriving customer (D1 ), the system is in vacation with i
customers xi0 and there are no vacation termination (L) and no service completion S during the slot. The second
term gives that if there are i 1 customers in vacation xi1;0 then, there must be a service completion within the
slot S0 b without vacation termination (L) so that the arriving customer (D1 ) nds i customers and nds the system
in vacation.
Case 3: The probability that the arriving customer nds the system on normal busy period with more than one customer i.e.,
j 1; i P 1 is,

1h   i
zi1 xi1 D1  S xi1;1 D1  S0 b xi0 D1  L0  Se  n xi1;0 D1  L0  S0 b :
k
The rst term gives that there are i customers in the system xi1 in normal busy period and the ongoing ser-
vice of the customer is not completed yet S, so the arriving customer (D1 ) nds i customers in busy system.
The second term gives that if there are i 1 customers in the busy period xi1;1 then there must be a ser-
vice completion S0 b within the slot, so that the arriving customer (D1 ) nds i customer in the busy system.
The third term says if there are i customers in the system with the system in vacation xi0 , then with the
termination of the vacation L0 and switching the service rate from lower to the higher one Se  n within
the slot, the system comes to the busy period with i customers. The last term is the probability that when
there are i 1 customers in the in vacation xi1;0 and the vacation terminates (L0 ) with the completion
of ongoing service S0 b within the slot, then the arriving customer (D1 ) nds the system in busy period with
i customers.

We can easily show that

X
1 X
1
z0 e zi0 e zi1 e 1:
i1 i1

e to be the transition probability matrix

In order to determine the waiting time distribution of the customers, we dene P
for the absorbing Markov chain where the state of absorption is the one when the customer reaches the head of the queue.
e as
Putting D0 1 and D1 0 in P, we get P
0 1
e 00
B 0
Be e1 C
B B 10 A 0 C
B C
eB
P e2
A e1
A 0 C;
B C
B e2 e1 C
@ A A 0 A
: : :
where
" # " # " #
e 00 L L0 d; e 10 L L0 d  S0 e1 L  S L0  Se  n e2 L  S0 b L 0  S0 b
B B ; A and A :
S0 d 0 S 0 S0 b
Let us dene ~z0 ~z00 ; ~z01 ; ~z02 ; . . . with ~z0i ~z0i;0 ~z0i;1  for i P 1 and

~z00 z0 e  Ir ;
~z0i;0 zi;0 e  Irm i P 1;
~z0i;1 zi;1 e  Im i P 1;

where Ir is the identity matrix of dimension r, then we get

e n P 0:
~zn1 ~zn P;

If W is the random variable giving the waiting time of a customer in queue and sj is the probability that a customers wait-
ing time is less than or equal to j unit of time then we have
 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946 937

j
sj ~z0 e; j P 0:
~zj
0 can be calculated recursively to an appropriate truncation level form as follows
h i
~zn1 ~z0 L L0 d ~z1;0 L L0 d  S0 ~z1;1 S0 d;
n n n
0

~zn1 ~zi;0 L  S ~zi1;0 L  S0 b;

n n
i;0 i P 1:
h i
~zi;1 ~zi;0 L  Se  n ~zi;1 S ~zi1;0 L0  S0 b ~zn
n1 n 0  n n 0
i1;1 S b; i P 1:

We can obtain the mean waiting time using Littles law

EW EN=k:

4. The regular busy period and the busy cycle

In a vacation queue, the regular busy period is the duration in which the server works at the normal rate lb , let it be de-
noted by Dv . Between two consecutive regular busy period the total time duration of the consecutive working vacations is
the total working vacation period V g . According to our vacation policy the system goes to regular busy period from the work-
ing vacation when the system nds at least one customer waiting at the end of vacation, otherwise it goes to another vaca-
tion. Therefore the busy cycle, Bc , the time period between two consecutive regular busy period ending instant can be given
as the sum

Bc Dv V g :
Let Gj i be a matrix of dimension m  m with its entries Gj
u;v i representing the probability that the service time of j
th
customers last i units of time given that the service of the rst customer starts in phase u and that of the j customer ends
in phase v. Then

G1 i Si1 S0 b; i P 1;
Gi i S0 bi ; i P 1;
Gj i S0 b Gj1 i  1 S Gj i  1; i P j 1; j P 2:
Let di;j be the probability that j customers arrive in i units of time, then
 
i
di;j Dij j
0 D1 :
j
Also we dene the matrix Bj i of dimension mn  mn such that its entries Bj
u;v i represent the probability that a busy
period initiated by j customers lasts i units of time given that the rst customers service and arrival are in phase
u; 1 6 u 6 mn, and that the service of the last customer and arrival are in phase v ; 1 6 v 6 mn, then
!
X
ij X
l
j
B i G i  j
Di0 j
G i  l  dil;k Bk l:
li k1

Let bi j be the probability that a busy period initiated by j customers lasts i units of time. As we are dealing with discrete-
time systems the service time of j customers must be at least equal to j, thus

bi j 0; for i < j;
wBj ie; otherwise;
where w be the stationary probability vector that arrival phase is i and service phase is j for i 1; . . . ; n; j 1; . . . ; m. The
probability that a busy period initiated by j customers lasts i units of time is given by [17]
  ij X
X l    
il ilk k
bi j w Gj i  Di0 e w Gj i  l  D0 D1 e bl k;
li k1
k

where the rst term on the right hand side (RHS) is the probability that the service time of j customers lasts i units of time
and during that period no new customer join the queue. The second term on the RHS is due to the probability that the service
time of j customers lasts i  l units of time. During the rst i  l units of time, k new customers join the queue. These k cus-
tomers initiate a busy period that lasts l units of time.
Let v i be the probability that total vacation duration lasts i units of time which is same as the probability that there is no
customer in the system for i  k units of time and the system does not return from vacation when customer arrives during
last k units of time. Then,
938 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946

X
i h ik
vi x0 eik x1 I  R1 e1  e ;
k0

where e1 is the column vector of zeros and 1 in the 1st position.

The mean duration of regular busy period and mean of total vacation duration will be
X
1
EDv ibi j;
ij
X
1
EV g iv i :
i0

and therefore we can have the mean busy cycle as

EBc EDv EV g :

5. Numerical examples

Our model DMAP/PH/1/WV(PH) captures both burstiness and correlation of the arrival process which are highly essential
features to be considered to measure the performances. To show the impact of burstiness and correlation in mean queue
length and the mean waiting time, we compare our model with an equivalent model that does not capture the correlation
property. For simplicity, we have taken a special case of DMAP, the 2-state Markov modulated Bernoulli process (MMBP) and
presented some numerical examples. In the rst and the second parts, we have given the DMAP representations of the MMBP
arrival process and the PH renewal process respectively. The third part gives the comparison of the performance measures,
like mean queue length and mean waiting time, of MMBP/Geo/1/WV(Geo) model with that of a PH/Geo/1/WV(Geo). By this
comparison we can see the impact of autocorrelation on the performance measures of a working vacation model. And in the
last part we will observe only the MMBP model and see how the performances changes with the change of autocorrelation.
Also we have provided some self explanatory tables to show the differences in distributions.

5.1. MMBP process and autocorrelation

b and the arrival rate

For a two state MMBP arrival process which is characterized by the transition probability matrix P
matrix K dened as
   
b p 1p k1 0
P and K ;
1q q 0 k2
and the total arrival rate k is,
k1 1  q k2 1  p
k : 7
2pq
The squared coefcient of variance, c2 , of the interarrival times of the MMBP process [18],
2k1 1  q k2 1  p 2k1 1  p k2 1  qk1 1  q k2 1  pp q  1
c2
k1 1  q k2 1  p k1 k2 p q  1 2  p  q2 k1 1  q k2 1  p k1 k2 p q  1
k1 1  q k2 1  p
 1 8
2pq
The autocorrelation function of the interarrival time with step 1 is
k1 k2 k1  k2 2 1  p1  qp q  12
w1 9
c2 2  p  q2 k1 1  q k2 1  p k1 k2 p q  12
The squared coefcient of variance, c2 , of the interarrival times is an important measure of the degree of trafc burstiness
which is dened as the ratio of the variance to the square of mean of interarrival times. The autocorrelation coefcient, which
is the ratio of covariance of interarrival times to the variance, is used to measure the amount of correlation in the process.
The DMAP representation of a MMBP process is D0 PI b  K, D1 P b K and this gives
   
p1  k1 1  p1  k2 pk1 1  pk2
D0 and D1 :
1  q1  k1 q1  k2 1  qk1 qk2
Here we have four variables, p; q; k1 ; k2 and two equations, one of total arrival rate k and the other of squared coefcient of
variation, c2 . For given values of k and c2 , we can obtain the corresponding MMBP parameters pk1 and qk1 as a function of
k1 from the formulas by xing the other arrival rate k2 . Therefore by xing two variables, k1 and k2 we can get a MMBP pro-
cess for the given k and c2 .
 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946 939

5.2. PH renewal process

A PH renewal process, meansthat the

 interarrivals
  have PH distribution and they are independent of each other. For PH
a b g
renewal process, PHa; T if T ; T0 and a a1 ; a2  then it can be represented as a DMAP process with
c d h
 
a1 g a2 g
D0 T and D1 T 0 a :
a1 h a2 h
In this process, we have eight variables and got ve equations from the conditions Te T 0 e; ae 1 and the equations
of mean and c2 variance=mean2 . So we need to x three variables.
If we want to have a PH process which is equivalent to this MMBP process and we can relate the three variables of PH
process to that of MMBP. Different ways of choosing these variables gives different PH arrival processes. Since we are free
to choose three variables, let us pick them in such a way that the transition matrix D0 T of PH process becomes equivalent
to D0 of the MMBP process i.e., any three variables of T is taken to be equal to their respective variables in D0 , e.g.

a p1  k1 ; c 1  q1  k1 and b 1  p1  k2 : 10
Another way of choosing the variables of PH process is to make equivalent the matrix D1 T 0 a to the D1 of the MMBP
process for given k and c2 .

5.3. Comparison of MMBP/Geo/1/WV(Geo) and PH/Geo/1/WV(Geo)

In the examples below we show how the performances of our model differs from general PH renewal process which has
independent arrivals, by keeping the service time distribution and the working vacation duration as geometric for all the
cases to see the inuence of the arrival process.
For given arrival rate k and c2 we can independently choose two variables for MMBP arrival process and three variables
for PH arrival process. For k 0:3 and c2 2, xing the rates k1 0:5 and k2 0:1, we get the MMBP process as
   
b 0:9815 0:0185 and K 0:5 0
P for which
0:0185 0:9815 0 0:1
   
0:4907 0:0167 0:4907 0:0019
D0 and D1 :
0:0093 0:8833 0:0093 0:0981
The autocorrelation (w) of this MMBP process is 0.1321. In Fig. 1 we can see the change of autocorrelation of this MMBP
process with the change in k1 .
For the PH process PHa; T having the parameters same with the MMBP process k 0:3; c2 2, and choosing the three
variables using relation (10), we get
   
0:4907 0:0167 0:4926
T ; T0 ;
0:0093 0:8833 0:1074
and a 0:8333; 0:1667 .
The service process and vacation durations for both the models are taken to be geometric. The parameter values are taken
as lb 0:7 and h 0:1. Generating the matrix R from the Naoumovs algorithm and using (6), we have calculated the mean
queue lengths for both the models with varying parameter values. We can see in Fig. 2 that the queue lengths of both the
models have a small difference in mean queue lengths but this difference increases 25% as we reduce the value of k from 0.3
to 0.2. Fig. 3 shows the difference between the mean queue lengths of these MMBP and PH models if the vacation duration
rate h is increased from 0.1 to 0.8.
In Fig. 4, the change in queue length with h is shown for different values of k. This also shows that the queue lengths of a
working vacation model with correlation cannot be assumed to be same with that having independent arrivals. The increase
in the value of c2 also shows quite a difference in queue lengths for MMBP and PH models. The queue lengths differs more
and more as the value of c2 increases as in Fig. 5. Finally, if we change the model by changing the value of k1 we can see how
signicantly the queue lengths differs with the change in c2 in Fig. 6.
From these observations, we can conclude that there is a signicant difference in queue lengths when we consider the
correlation in the arrival process (MMBP model) to that of the case where the arrivals are assumed to be independent
(PH model). Therefore, assuming the interarrivals to be independent when it is not might lead to inaccurate prediction of
performance measures. Similar behaviors are also observed for the mean waiting times.
We have seen that for c2 1 the mean queue lengths of MMBP model and PH model comes very close, in Fig. 5. Here the
amount of correlation in the MMBP process is 0.0281. For such a situation we present in Table 1, the distributions of mean
queue lengths for both the models and show that though the mean queue lengths are close enough, the distributions are not
same. For lb 0:5 if k 0 the values of yk are same for both of them. Then for k 1 to k 4 the values of yk for MMBP model
(yM P P M
k ) is smaller than that of PH model (yk ) and after that (yk ) is smaller than (yk ). We have tabled the change for four different
values of lb . So same mean does not give the same distribution of those two models.
940 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946

0.14
p,q=0.95
p,q=0.9815
0.12

Autocorrelation function of interarrival times 0.1

0.08

0.06

0.04

0.02

0
0 0.5 1 1.5 2
1

Fig. 1. Autocorrelation vs. k1 , with k2 0:1.

We can also have situation like in Fig. 7 where the queue lengths of MMBP model becomes less than that of the PH model.
The mean waiting times also show similar behaviors in this case.
So, whatever way we choose the parameters of the models, because of the correlation of the MMBP process, there are
considerable difference in mean queue lengths and also in mean waiting times which cannot be neglected if we wish to have
a good estimation of the performance measures. Thus the correlation in the arrival process in quite an important feature to

5
MMBP,=0.2
PH, =0.2
4.5
MMBP,=0.3
PH, =0.3
4

3.5
Mean Queue Length

2.5

1.5

0.5

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7

v

Fig. 2. Mean queue length vs. lv , with lb 0:7, h 0:1 and c2 2.

 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946 941

6
MMBP,=0.1
PH, =0.1
MMBP,=0.3
5 PH, =0.3
MMBP,=0.8
PH, =0.8

4
Mean Queue Length

0
0 0.1 0.2 0.3 0.4 0.5 0.6
v

Fig. 3. Mean queue length vs. lv , with lb 0:7; k 0:2and c2 2.

6
MMBP,=0.2
PH, =0.2
MMBP,=0.3
5 PH, =0.3

4
Mean Queue Length

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

Fig. 4. Mean queue length vs. h, with lb 0:7; lv 0:5 and c2 2.

942 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946

2
MMBP, c 2=1
PH, c 2=1
MMBP, c 2=2
PH, c 2=2

1.5
Mean Queue Length

0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

Fig. 5. Mean queue length vs. h, with lb 0:7; lv 0:5 and k 0:3.

1.15
MMBP, 1=0.5
PH, =0.5
1.1 1
MMBP,1=0.7
PH, =0.7
1.05 1
MMBP, =0.9
1
PH, =0.9
1
1
Mean Queue Length

0.95

0.9

0.85

0.8

0.75
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
2
c

Fig. 6. Mean queue length vs. c2 , with lb 0:7; lv 0:3; h 0:4 and k 0:3.
 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946 943

Table 1
Comparing the distribution of number of customers in the MMBP/Geo/1/WV(Geo) model and PH/Geo/1/WV(Geo) model for
k1 0:7; k2 0:1 k 0:4; h 0:7; lv 0:4 and c2 1.

k lb 0:48 lb 0:5 lb 0:52 lb 0:54

yM
k yPk yM
k yPk yM
k yPk yM
k yPk

0 0.1634 0.1634 0.1953 0.1953 0.2246 0.2245 0.2514 0.2514

1 0.1761 0.1783 0.2047 0.2070 0.2293 0.2315 0.2505 0.2525
2 0.1375 0.1420 0.1514 0.1561 0.1607 0.1653 0.1664 0.1708
3 0.1088 0.1116 0.1133 0.1156 0.1135 0.1154 0.1110 0.1124
4 0.0862 0.0875 0.0847 0.0853 0.0801 0.0803 0.0739 0.0736
5 0.0683 0.0686 0.0633 0.0630 0.0565 0.0558 0.0491 0.0482
6 0.0541 0.0537 0.0473 0.0465 0.0399 0.0388 0.0327 0.0315
7 0.0428 0.0421 0.0354 0.0343 0.0281 0.0270 0.0217 0.0206
8 0.0339 0.0330 0.0264 0.0254 0.0198 0.0187 0.0145 0.0135
9 0.0268 0.0259 0.0198 0.0187 0.0140 0.0130 0.0096 0.0088
10 0.0213 0.0203 0.0148 0.0138 0.0099 0.0091 0.0064 0.0058

0.88
MMBP,1=0.7
PH, 1=0.7
MMBP,1=0.9
0.86
PH, 1=0.9

0.84
Mean Queue Length

0.82

0.8

0.78

0.76
0 0.1 0.2 0.3 0.4 0.5 0.6
v

Fig. 7. Mean queue length vs. lv , with lb 0:7; h 0:8; k 0:4 and c2 0:5.

consider which indicates that this DMAP/PH/1/WV model can give a better estimation of the performance measures when-
ever interarrivals are correlated.

5.4. Autocorrelation and MAP/Geo/1/WV(Geo) model

The autocorrelation coefcient for the interarrival times is the measure the degree of autocorrelation. This correlation
parameter is a function of the arrival rate and the transition probabilities as given in Eq. (9) for the MMBP arrival process.
Figs. 8 and 9 show the impact of correlation on the mean queue length of MMBP/Geo/1/WV(Geo) model under different
utilization or system load (q) and burstiness c2 . These results are obtained by xing the values k2 0:9; lb 0:5;
lv 0:4; and h 0:7. The graph shows that for low utilization, correlation has stronger impact on mean queue length for
higher bursty trafc i.e., for low utilization, higher bursty trafc and higher correlation can cause a signicant rise in the
queue length of the working vacation model. This impact gradually increases as the utilization becomes higher and the cor-
relation approaches one.
944 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946

16
c2=1
2
c =2
14 2
c =6
2
c =10

12
Mean Queue Length

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1

Fig. 8. Mean queue length vs. w1 , with lb 0:5; lv 0:4; h 0:7; k2 0:9 and q 0:65.

200
c 2 =1
c 2 =2
180 c 2 =6
2
c =10

160

140
Mean Queue Length

120

100

80

60

40

20

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

1

Fig. 9. Mean queue length vs. w1 , with lb 0:5; lv 0:4; h 0:7; k2 0:9 and q 0:9.

In Table 2, we have presented the queue length distribution for different values of q of an MAP/Geo/1/WV (Geo) model.
The matrices of the MAP arrival process as given in [19] are
 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946 945

Table 2
Distribution of number of customers in the MAP/Geo/1/WV(Geo) system with parameters: n 7, k 0:0925, lv 0:03 and h 0:08.
k q 0:9250 q 0:8409 q 0:7708 q 0:7115 q 0:6607 q 0:6167 q 0:5781
yk yk yk yk yk yk yk
0 0.0549 0.1142 0.1627 0.2030 0.2371 0.2662 0.2915
1 0.0347 0.0669 0.0891 0.1046 0.1155 0.1233 0.1289
2 0.0545 0.1031 0.1354 0.1573 0.1725 0.1832 0.1907
3 0.0441 0.0767 0.0932 0.1008 0.1036 0.1036 0.1021
4 0.0482 0.0804 0.0945 0.0996 0.1002 0.0986 0.0960
5 0.0433 0.0666 0.0726 0.0714 0.0674 0.0625 0.0576
6 0.0426 0.0619 0.0642 0.0606 0.0553 0.0499 0.0450
7 0.0396 0.0532 0.0514 0.0454 0.0390 0.0332 0.0284
8 0.0377 0.0474 0.0432 0.0363 0.0299 0.0245 0.0204
9 0.0354 0.0412 0.0351 0.0277 0.0214 0.0167 0.0131
10 0.0334 0.0363 0.0290 0.0216 0.0159 0.0119 0.0090
11 0.0314 0.0317 0.0236 0.0165 0.0115 0.0081 0.0059
12 0.0296 0.0277 0.0194 0.0128 0.0084 0.0057 0.0039
13 0.0278 0.0242 0.0158 0.0098 0.0061 0.0039 0.0026
14 0.0262 0.0212 0.0129 0.0075 0.0044 0.0027 0.0017
15 0.0246 0.0185 0.0106 0.0058 0.0032 0.0018 0.0011

2 3 2 3
0 1 0 0 00 0 0 0 0 0 0 0 0
60
6 0 1 0 00 770 6 0
6 0 0 0 0 077 0
6 7 6 7
60
6 0 0 0 0 0 0 77
6 0
6 0 0 1 0 0 077
6 7 6 7
D0 6 0 0 0 0:95 0 0 0 7; D1 6 0 0 0 0 0:05 0 0 7;
6 7 6 7
60
6 0 0 0 0 0:05 0 77
6 0
6 0 0 0:95 0 0 077
6 7 6 7
40 0 0 0 0 0:95 0:05 5 4 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0:8 0:2 0 0 0 0 0
and p 0:0018 0:0023 0:0023 0:9011 0:0451 0:0451 0:0023:
Here the values of the parameters are k 0:0925, lv 0:03 and h 0:08. The values of service rate is taken to be
lb 0:1; 0:11; 0:12; 0:13; 0:14; 0:15 and 0:16. The table shows that if the trafc intensity q decreases we get higher proba-
bility of system remaining empty. From k 0 to k 1 the probability decreases but again increase for k 2. After k 2
the probabilities keeps on decreasing. Which means that the probability to have a long queue is quite small. This distribution
of number of customers depict the fact that even in the presence of correlation we can reduce the queue lengths for higher
trafc intensities with the working vacation policy.

6. Conclusions

We have considered the discrete-time MAP/PH/1 model with multiple working vacations. Such models have practical
implications as majority of the trafc sources that high speed networks support are highly bursty and, more importantly,
correlated in nature. Our study highlights the impact of the correlation factor on the mean queue lengths and the mean wait-
ing times. This work can be extended to the cases with batch Markovian arrival process (BMAP), with nite servers, and con-
sidering vacation interruptions and server breakdowns.

Acknowledgement

The authors wish to record their gratitude to the anonymous referees for their helpful comments.

References

[1] N. Tian, Z.G. Zhang, Vacation Queueing Models: Theory and Applications, Springer, New York, 2006.
[2] H. Takagi, Queueing Analysis, Discrete Time Systems, vol. 3, Elsevier Science Publishers, Amsterdam, 1993.
[3] Z.G. Zhang, N. Tian, Geo/G/1 queue with multiple adaptive vacations, Queueing Syst. 38 (2001) 419429.
[4] N. Tian, Z.G. Zhang, The discrete time GI/Geo/1 queue with multiple vacations, Queueing Syst. 40 (2002) 283294.
[5] L.D. Servi, S.G. Finn, M/M/1 queue with working vacations (M/M/1/WV), Perform. Eval. 50 (2002) 4152.
[6] D. Wu, H. Takagi, M/G/1 queue with multiple working vacations, Perform. Eval. 63 (7) (2006) 654681.
[7] Y. Baba, Analysis of a GI/M/1 queue with multiple working vacations, Oper. Res. Lett. 33 (2005) 201209.
[8] A.D. Banik, U.C. Gupta, S.S. Pathak, On the GI/M/1/N queue with multiple working vacations analytic analysis and computation, Appl. Math. Modell.
31 (9) (2007) 17011710.
[9] N. Tian, Z. Ma, M. Liu, The discrete time Geom/Geom/1 queue with multiple working vacations, Appl. Math. Modell. 32 (12) (2007) 29412953.
[10] J.H. Li, N.S. Tian, W.Y. Liu, Discrete-time GI/Geom/1 queue with multiple working vacations, Queueing Syst. 56 (2007) 5363.
[11] A.S. Alfa, A discrete MAP/PH/1 queue with vacations and exhaustive time-limited service, Oper. Res. Lett. 18 (1995) 3140.
[12] M.F. Neuts, Matrix-Geometric Solutions in Stochastic Models, John Hopkins University Press, Baltimore, 1981.
946 C. Goswami, N. Selvaraju / Applied Mathematical Modelling 34 (2010) 931946

[13] L. Breuer, D. Baum, An Introduction to Queueing Theory and Matrix-Analytic Methods, Springer, Netherlands, 2006.
[14] G. Latouche, V. Ramaswami, Introduction to Matrix Analytic Methods in Stochastic Modeling, ASA-SIAM Series on Applied Probability, 1999.
[15] M.F. Neuts, Some promising directions in algorithmic probability, in: A.S. Alfa, S. Chakravarthy (Eds.), Advances in Matrix-analytic Methods for
Stochastic Models, Notable Publications, 1998, pp. 429443.
[16] A.S. Alfa, Matrix-geometric solution of discrete time MAP/PH/1 priority queue, Naval Res. Logist. 45 (1998) 2350.
[17] I. Frigui, A.S. Alfa, Analysis of a time limited polling system, Comput. Commun. 21 (1998) 558571.
[18] J. Liao, L. Lemin, S. Hairong, The inuence of burstiness and correlation of trafc on an ATM multiplexer, in: International Conference on
Communication Technology (ICCT96), Beijing, China, 1996.
[19] U.C. Gupta, S.K. Samanta, R.K. Sharma, M.L. Chaudhry, Discrete-time single-server nite-buffer queues under discrete Markovian arrival process with
vacations, Perform. Eval. 64 (2007) 119.