Documente Academic
Documente Profesional
Documente Cultură
j
v=0
a
v
, j 0, with a
0
= 0 and u
0
= 1, by denition. Let
b
k
= u
k
/(u
k1
), k 1. The discrete GI process with nite suport can be represented as
an PH with representation (, T ), with = e
1
, where e
j
is the transpose of e
j
, and
T
i,i+1
= b
i
, 1 i n 1, and T
i,j
= 0, j = i + 1.
This we call the elapsed time representation of general discrete distribution. When
n = then we end up with the IPH, based on [52, theorem 2.3].
We point out that A can also be represented by another phase type distribution
based on remaining time with = a and
T
i,i1
= 1, 2 i n, and T
i,j
= 0, j = i 1.
This was used for the GI/G/1 system in [4].
Throughout this paper we shall use the term phase of arrival. We can practically
represent any discrete renewal type interarrival times using either of these phase type
systems.
2.2. Service times, vacation durations and operational times
By virtue of the fact that we can represent discrete independent and identically distrib-
uted (iid) service times, vacation durations and operational times the same way in which
we represent iid interarrival times we will not repeat the above information. We speak
10 A.S. ALFA
of the phase of arrival, the phase of service, the phase of vacation and the phase of oper-
ation from this point on, with n, m, r and s number of phases, respectively. We discuss
only the rules for going on vacation in this paper. We do not discuss the rule for return-
ing from vacation such as N-policy, T -policy, etc. The major difference between the
class of vacation models considered in this paper and the ones with rules for returning
from vacation (such as N-policy) is mainly at the boundary states. All the non-boundary
states aspects of both types of models are about the same.
3. Markov chains describing the systems
There are two types of Markov chains that we will encounter throughout this paper. Let
us call their respective transition matrices P
1
and P
2
. The matrices look as follows:
P
1
=
_
_
_
_
_
_
B C
E A
1
A
0
A
2
A
1
A
0
A
2
A
1
A
0
.
.
.
.
.
.
.
.
.
_
_
,
P
2
=
_
_
_
_
_
_
A
0,0
A
0,1
A
1,0
A
1,1
A
1,2
A
2,1
A
2,2
A
2,3
A
3,2
A
3,3
A
3,4
.
.
.
.
.
.
.
.
.
_
_
,
where B, C, E, A
0
, A
1
, A
2
, A
0,0
, A
0,1
and A
i,j
, i 1, i 1 j i + 1 may all be
block matrices. In both cases our interest is rst to obtain the probability distribution of
the steady state system, when it exists. We want to nd
x(c) = x(c)P
c
, x(c)1 = 1, c = 1, 2.
We let x(c) = [x
0
(c), x
1
(c), x
2
(c), . . .]. Later we will drop the index c in the brackets
when there is no possible confusion.
The transition matrix P
1
is a level-independent QBD and matrix P
2
is a level-
dependent QBD. There is extensive literature on P
1
and reasonable amount on P
2
. We
will discuss both briey for completeness.
3.1. Matrixgeometric solutions for P
1
3.1.1. The case when (n, m, r, s) <
In this case all the block matrices of P
1
are of nite order. From [30,45,46] it is known
that the Markov chain represented by P
1
is positive recurrent if
A
0
1 < A
2
1, where = A, 1 = 1, and A = A
0
+A
1
+A
2
.
For some special cases it is possible to obtain simpler expressions for the stability con-
dition.
VACATION MODELS IN DISCRETE TIME 11
Assuming that the stability condition holds we proceed to show how to obtain x.
There exists a matrix R which is the minimal non-negative solution to the matrix
quadratic equation
R = A
0
+RA
1
+R
2
A
2
.
For a stable system the spectral radius of R is less than 1. From the matrixgeometric
theorem we know that
x
i+1
= x
i
R = x
1
R
i
, i 1.
Suppose the vector [y
0
, y
1
] is obtained as the eigenvector corresponding to the eigen-
value of 1 for the matrix
_
B C
E A
1
+RA
2
_
.
Further let y = y
0
1 + y
1
(I R)
1
1, then x
0
= y
1
y
0
and x
1
= y
1
y
1
. We can now
compute x using the matrixgeometric theorem, after which we can obtain most of the
performance measures that we need.
In addition, we may talk about the matrix G which is the minimal non-negative
solution to the matrix quadratic equation
G = A
2
+A
1
G +A
0
G
2
.
The matrix G is stochastic if this Markov chain is positive recurrent. The matrix G is
used in studying the busy period of this system. A relationship between R and G is
R = A
0
(I A
1
A
0
G)
1
.
Nearly all ungated systems can be developed as Markov chains with this type of
QBD process.
3.1.2. The case when (n, m, r, s)
If at least one of the parameters n, m, r, s is equal to innity then the matrixgeometric
result has to be applied cautiously, when it is applicable. This type of situation happens
when we have the interarrival time, or the service time or the vacation duration, or the
operational time of the server having an IPH distribution. For example, if any of these is
a GI which is not of nite support then we have such a situation. In such cases we have
to appeal to other results to help us on a case by case basis.
The problem here is that all the block matrices B, C, E, A
0
, A
1
and A
2
are of
innite dimension. It was pointed out by Tweedie [59] that we still have the matrix
geometric solution in this case, provided the system is stable. The associated R matrix
is of innite dimension. However, the main difculty in this case is how to establish
stability. As pointed out by Neuts [47], in such cases we cannot establish stability easily
by using the mean drift result, A
0
1 < A
2
1. We have to resort to other approaches.
This is partly because the vector may not exist. There are no general results for such
cases. We have to study each case separately as a special case. Consider the case when
12 A.S. ALFA
the interarrival times, service times, vacation duration and also the operational times are
all of the general renewal types. In this case we may choose to represent all the four
distributions using the elapsed time or the remaining representations both which are
special cases of the IPH distribution. Let us use the remaining time representation. In
order to use the matrixgeometric result we need to establish the conditions for stability.
Consider the matrix A = A
0
+ A
1
+ A
2
. It will be of innite dimension. However, for
some of the cases to be discussed in this paper the matrix A will have a special structure.
For example, the model in section 4.1.1 will have an A matrix which will be closely
similar in structure to the matrix A in [4]. In this particular case, it is possible to extend
the results in [4] for establishing the stability conditions. After that has been done then
the matrixgeometric result can be applied. For each of the models in section 4 for which
we end up with a transition matrix of type P
1
and for which the associated matrix A is
of innite order, we have to establish stability conditions separately. It is not possible to
establish general stability conditions for all cases with innite-order matrix A.
The general results which we present in this paper apply only to the cases with
(n, m, r, s) < . Cases for which at least one of the parameters n, m, r or s is innity,
will have to be considered on a case by case basis. We will consider an example of such
a case in section 4.4.2.
3.2. Matrix-product solution for P
2
For this class of problems, results in the literature are only available for the cases with
(n, m, r, s) < . We therefore focus only on such results.
This system is a spatially inhomogeneous QBD process in discrete time, some-
times referred to as level-dependent QBD. Because of this spatial inhomogeneity we
cannot apply the standard matrixgeometric method as with P
1
. In order to calculate the
equilibrium distribution we rst have to evaluate the matrices R
k
, k 0, which are the
minimal non-negative solutions to
R
k
= A
k,k+1
+R
k
A
k,k
+R
k
[R
k+1
A
k,k1
], k 0,
where (R
k
)
ij
is the expected number of visits to state (k+1, j) before rst visit to level k
conditional on the process starting in state (k, i).
It is straightforward to show that the steady state vector x can be determined as
follows:
x
k
= x
k1
R
k1
= x
0
k1
j=0
R
j
, k 1,
where x
0
is obtained by solving the system of equations
x
0
[A
0,0
+ R
0
A
1,0
] = x
0
,
and then normalized by [x
0
k=0
[
k1
j=0
R
j
]]1 = 1, where R
1
= 1, by denition.
Note that provided the system is stable the solution to x
0
[A
0,0
+ R
0
A
1,0
] = x
0
satises
x
0
[
k=0
[
k1
j=0
R
j
]]1 < [8].
VACATION MODELS IN DISCRETE TIME 13
Ramaswami and Taylor [48] obtained explicit solution for R
i
, i = 0, 1, 2, 3, . . . .
Bright and Taylor [8] developed algorithms for this class of level-dependent QBD.
Note that the matrices G
k
are the minimal non-negative solutions to the matrix
quadratic equations
G
k
= A
k,k1
+A
k,k
G
k
+A
k,k+1
G
k+1
G
k
, k 0.
All gated systems are of this type. In some special cases we are able to exploit
the structure of the problem and obtain simpler results by converting them to level-
independent systems as an approximation.
4. Models
We dene the following indices as variables to be used in the sequel:
i = the number of items in the system.
i
= the phase of service interruption. This has m+1 phases including phase 0, when
there is no interrupted service.
l = the phase of vacation: vacation is phase type with representation (, L).
u = the time clock of a servers visit (or the number served so far by a server during
a visit) the use depends on the model. This could also represent the phase of the
operational time with representation (u, U).
We also dene the following parameters
N = the limited time of a single visit by a server for a time-limited service.
M = the limited number of customers to be served during a server visit for number-
limited service.
For some models where services may be interrupted we need to dene a matrix Q
which represents the phase at which an interrupted service begins when it is resumed,
given the interruption phase. The elements Q
j
1
,j
1
refer to the probability that a service in-
terrupted in phase j
1
resumes in phase j
1
when the service re-starts. For example, a pre-
emptive resume service has Q = I and a preemptive repeat service rule has Q = 1.
4.1. Standard vacation systems
In what we refer to as a standard type vacation system the server attends to the queue
until the system is empty (i.e. no waiting customers), in the case of exhaustive, or until
all the customers in the gate have been served, in the gated case, and then proceeds on a
vacation. We consider the two cases exhaustive and gated.
14 A.S. ALFA
4.1.1. Exhaustive
Consider the state space:
0
= (0, k, 0), k = 1, 2, . . . , n,
+
0
= (0, k, l), k = 1, 2, . . . , n, l = 1, 2, . . . , r,
V
i
= (i, k, l), i = 1, 2, . . . , k = 1, 2, . . . , n, l = 1, 2, . . . , r,
S
i
= (i, k, j), i = 1, 2, . . . , k = 1, 2, . . . , n, j = 1, 2, . . . , m.
For
0
= (0, k, 0), the rst 0 refers to an empty system, and the second 0 refers to
no vacation, i.e. the server is waiting at an empty queue for customers to arrive, after a
vacation. For the case
+
0
= (0, k, l), l refers to the phase of vacation. In both cases k
refers to the phase of arrival.
Let
0
=
+
0
0
and
i
=
V
i
S
i
, where
V
i
is associated with the vacation
states in which there are i customers in the system and
S
i
is associated with the service
states with i customers in the system. The state space for the single vacation system is
given by as
=
0
_
i=1
i
.
The state space for the multiple vacation system is given by as
=
+
0
_
i=1
i
.
It is immediately clear that when the system is in states
i
, i 1, the chain can
only make transitions to the states
i1
,
i+1
or remain in the states
i
. The transition
probabilities are not level-dependent, except for level i = 1. When the system is in
states
0
the chain can only have transitions to states
1
or remain in states
0
. This
Markov chain is thus a level-independent QBD. If we now label the states in lexico-
graphic order, and let the rst index, known as the level, be i, i 0, then the result-
ing transition matrix for this Markov chain can be represented as in P
1
. The matrix
geometric results apply.
We display the interior block matrices from P
1
for this system, i.e. the matrices
A
0
, A
1
and A
2
which respectively represent the transitions from
i
to
i+1
,
i
to
i
,
and
i
to
i1
, i 2. They are
A
0
=
_
_
T
0
_
L
_
T
0
_
L
0
_
0
_
T
0
_
S
_
, A
2
=
_
0 0
0 T
_
S
0
_
_
, and
A
1
=
_
T L T
_
L
0
_
0 T S +
_
T
0
_
S
0
_
_
.
We only explain the block A
0
, it is then easy to understand the rest. For block A
0
the
term (T
0
) L on the top left corner refers to a transition during a vacation with one
arrival; the term(T
0
)(L
0
) on the top right corner refers to a transition from vacation
VACATION MODELS IN DISCRETE TIME 15
to service commencement with one arrival; and the term (T
0
) S on the bottom right
corner refers to a transition during service with one arrival. The term on the bottom left
corner is zero because it refers to a transition from a service to a vacation with one arrival
which is not possible since if there is any customer in the system during a service the
server will not go on a vacation at all, since A
0
refers to a transition with at least one
customer in the system.
The boundary matrices are easy to obtain and can be found in [3]. They are omitted
here.
4.1.2. Gated
Consider the state space:
0
= (0, k, l), k = 1, 2, . . . , n, l = 1, 2, . . . , r,
V
i
= (i, k, l), i = 1, 2, . . . , k = 1, 2, . . . , n, l = 1, 2, . . . , r,
S
i
= (i, (i
, k, j)), i = 1, 2, . . . , k = 1, 2, . . . , n, i
= 1, 2, . . . , i 1, j =
1, 2, . . . , m,
t
i
= (i, ((i
= i), k, j)), i = 1, 2, . . . , k = 1, 2, . . . , n, j = 1, 2, . . . , m.
By the denition of the gated system we cannot consider a single vacation type.
Let
i
=
V
i
S
i
t
i
. The state space for the gated vacation system is given by
as
=
0
_
i=1
i
.
Once more, it is immediately clear that when the system is in states
i
, i 1 the
chain can only have transitions to states
i1
,
i+1
or remain in the states
i
. However,
because at level i we have a supplementary variable i
1
)
_
T
_
S
0
__
I(i 1) T
_
S
0
_
_
,
A
i,i+1
=
_
_
T
0
_
L e
1
_
T
0
_
L
0
_
0 I (i)
_
T
0
_
S
_
, and
A
i,i
=
_
T L e
1
T
_
L
0
_
0 I (i) T S +
I(i 1)
_
T
0
_
S
0
_
_
,
16 A.S. ALFA
where
I(v) =
_
0 0
I (v) 0
_
,
I (v) is an identity matrix of size v and every e
1
is of appropriate order as dened earlier.
In order to obtain x here we choose a level i = K, such that the probability of
arrivals during a vacation exceeding K is less than some very small value = 10
q
,
usually q 12. We then assume that the chances of the number in the gate ex-
ceeding K is negligible hence we approximate P
2
by P
2
(K), with A
K+v,K+v1
=
A
K+1
, A
K+v,K+v
= A
K
and A
K+v,K+v+1
= A
K1
, v 0. This now becomes a level-
independent system with a transition matrix similar to P
1
but with a large boundary.
4.2. Time-limited systems
For the time limited systems we introduce N, the time limit allowed at the queue during
each visit. We assume that N is deterministic. In a time-limited system the server visits
the queue to provide service until all customers are served (in the exhaustive case) or
N units of time has elapsed, whichever occurs rst. In this system a service could be
interrupted if the time clock N of the server visit is up during a service. The preemptive
rule is applied. Q
j
1
,j
1
is the probability that a service which was interrupted in phase j
1
is restarted in phase j
1
at the service resumption (Q
j
1
,j
1
was not considered in [1]). In the
gated case, service is provided until all customers within the gate are served or N units
of time has elapsed, whichever occurs rst.
4.2.1. Exhaustive
We consider only the preemptive case.
Consider the state space:
+
0
= (0, k, l), k = 1, 2, . . . , n, l = 1, 2, . . . , r,
V
i
= (i, k, l, j
), i = 1, 2, . . . , k = 1, 2, . . . , n, l = 1, 2, . . . , r, j
=
0, 1, 2, . . . , m,
S
i
= (i, u, k, j), i = 1, 2, . . . , u = 1, 2, . . . , N, k = 1, 2, . . . , n, j =
1, 2, . . . , m.
We assume that such time-limited vacations are multiple vacation types. Let
i
=
V
i
S
i
. The state space for this multiple vacation time-limited system is given by
as
=
+
0
_
i=1
i
.
This is a level-independent Markov chain which can be represented by the transi-
tion matrix P
1
. The block matrices are:
A
0
=
_
_
T
0
_
L I (m+ 1) e
1
_
T
0
_
L
0
Q
e
N
_
T
0
_
S
I(N 1)
_
T
0
_
S
_
,
VACATION MODELS IN DISCRETE TIME 17
A
1
=
_
T L I (m+ 1)
e
N
_
T S
+
_
T
0
_
_
S
0
e
1
__
e
1
T L
0
Q
I(N 1)
_
T S +
_
T
0
_
S
0
__
_
and
A
2
=
_
0 0
e
N
T
_
S
0
e
1
_
I(N 1) T
_
S
0
_
_
,
where e
v
are all of dimensions N 1,
S
= [0 S],
I(v) =
_
0 I (v)
0 0
_
, and Q
=
_
Q
_
,
4.2.2. Gated
We only consider the preemptive case. Consider the following state spaces:
0
= (0, k, l), k = 1, 2, . . . , n, l = 1, 2, . . . , r,
V
i
= (i, k, l), i = 1, 2, . . . , k = 1, 2, . . . , n, l = 1, 2, . . . , r,
S
i
= (i, (i
, u, k, j)), i = 2, . . . , k = 1, 2, . . . , n, i
= 1, 2, . . . , i 1, u =
1, 2, . . . , N, j = 1, 2, . . . , m,
t
i
= (i, ((i
= i), u, k, j)), i = 1, 2, . . . , u = 1, 2, . . . , N, k = 1, 2, . . . , n, j =
1, 2, . . . , m.
Let
i
=
V
i
S
i
t
i
. The state space for the gated time-limited vacation system
is given by as
=
0
_
i=1
i
.
The Markov chain representing this system is a level-dependent QBD of type P
2
. It is
fully described in [2]. We will therefore not present the block matrices here.
This level-dependent QBD has a special structure which makes it easier to study.
This feature is a property due to the discrete time analysis of the system. Each item
requires a minimum of at least one unit of time for service. Hence, after we have at least
N items in the gate then it implies that the server will have to stay for the whole N units
of time during a visit. Therefore, in such situations we simply say the number in the gate
is at least N and ignore information about the exact number in the gate as soon as that
number exceeds N. This results in a level-independent system with a very large set of
boundary states. The transition matrix then becomes:
18 A.S. ALFA
P =
_
_
_
_
_
_
_
_
_
_
_
_
_
A
0,0
A
0,1
A
1,0
A
1,1
A
1,2
A
2,1
A
2,2
A
2,3
.
.
.
.
.
.
.
.
.
A
N,N1
A
N,N
A
N,N+1
A
2
(N) A
1
(N) A
0
(N)
A
2
(N) A
1
(N) A
0
(N)
.
.
.
.
.
.
.
.
.
_
_
,
where A
0
(N), A
1
(N) and A
2
(N) are the respective non-level-dependent blocks.
We may now use the modied matrix-geometric approach for this problem.
4.3. Number limited systems
For the number-limited system the server visits the queue until it has served M customers
or the queue becomes empty (for the exhaustive case), whichever occurs rst. In the
gated case we only consider the number in the gate. The server attends to either M jobs
or the number of jobs in the gate, whichever is smaller.
4.3.1. Exhaustive
Consider the state space:
+
0
= (0, k, l), k = 1, 2, . . . , n, l = 1, 2, . . . , r,
V
i
= (i, k, l), i = 1, 2, . . . , k = 1, 2, . . . , n, l = 1, 2, . . . , r,
S
i
= (i, u, k, j), i = 1, 2, . . . , u = 0, 1, 2, . . . , M, k = 1, 2, . . . , n, j =
1, 2, . . . , m.
We assume that such number-limited vacations are multiple vacation types. Let
i
=
V
i
S
i
. The state space for this multiple vacation number-limited system is
given by as
=
+
0
_
i=1
i
.
This is a level-independent Markov chain which can be represented by the transi-
tion matrix P
1
. The block matrices are:
A
0
=
_
_
T
0
_
L e
1
_
T
0
_
L
0
_
0 I (M)
_
T
0
_
S
_
,
A
1
=
_
T L e
1
T
_
L
0
_
e
M
_
T
0
_
S
0
I (M) T S +
I(M 1)
_
T
0
_
S
0
_
_
,
A
2
=
_
0 0
e
M
T S
0
I(M 1) T
_
S
0
_
_
, where
I(M) =
_
0 I (M)
0 0
_
.
VACATION MODELS IN DISCRETE TIME 19
4.3.2. Gated
Consider the following state spaces:
0
= (0, k, l), k = 1, 2, . . . , n, l = 1, 2, . . . , r,
V
i
= (i, k, l), i = 1, 2, . . . , k = 1, 2, . . . , n, l = 1, 2, . . . , r,
S
i
= (i, (i
, u, k, j)), i = 2, . . . , k = 1, 2, . . . , n, i
= 1, 2, . . . , i 1, u =
1, 2, . . . , M, j = 1, 2, . . . , m,
w
i
= (i, ((i
= i), u, k, j)), i = 1, 2, . . . , u = 1, 2, . . . , M, k = 1, 2, . . . , n, j =
1, 2, . . . , m.
Let
i
=
V
i
S
i
w
i
. The state space for the gated number-limited vacation
system is given by as
=
0
_
i=1
i
.
The Markov chain representing this system is a level-dependent QBD of type P
2
.
However, similar to the case of gated time-limited system, we can study this as
level-independent system, by noting that once the number in the gate is at least M then
the server will attend to the queue until it has served exactly M customers. In this case
we do not need to keep track of the number in the gate whenever there is at least M
customers in the gate. Hence, after the number in the system is at least M the state space
becomes level-independent, if we do not need the information about the number in the
gate. The resulting transition matrix is level-independent with a large set of boundary
states. We do not give the details here in order to save on limited journal space. Details
are available in [3].
4.4. Random time limited
Here we consider a random time limited type of vacation in which the server (is oper-
ational and) attends to the queue for a random time period which we assume follows a
phase type distribution (u, U) of dimension s, with U
0
= 1 U1. We consider only the
preemptive case. Generally we will nd that the exhaustive time-limited case is a special
case of the random one. We present a more detailed analysis of this system by including
the boundary block matrices B, C and E and also by writing down the blocks of the
matrix R.
We will not consider the gated case since it is, to some extent, fairly similar in
structure to the gated time-limited case. The case of random number limited can also be
easily developed following the steps used in this section. Details of this can be found
in [3].
Consider the state space:
+
0
= (0, k, l), k = 1, 2, . . . , n, l = 1, 2, . . . , r,
V
i
= (i, k, l, j
), i = 1, 2, . . . , k = 1, 2, . . . , n, l = 1, 2, . . . , r, j
=
0, 1, 2, . . . , m,
20 A.S. ALFA
S
i
= (i, u, k, j), i = 1, 2, . . . , u = 1, 2, . . . , s, k = 1, 2, . . . , n, j = 1, 2, . . . , m.
We assume that such random visit systems are multiple vacation types. Let
i
=
V
i
S
i
. The state space for this multiple vacation random system is given by as
=
+
0
_
i=1
i
.
This is a level-independent Markov chain which can be represented by the transi-
tion matrix P
1
. The block matrices are:
B =
_
T L T
_
L
0
u
_
U
0
T U T
_
,
C =
_
_
T
0
_
L e
1
u
_
T
0
_
L
0
U
0
_
T
0
_
e
1
U
_
T
0
_
,
E =
_
0 0
U
0
T
_
S
0
_
U T S
0
_
,
A
0
=
_
_
T
0
_
L I (m+ 1) u
_
T
0
_
L
0
Q
U
0
_
T
0
_
S
U
_
T
0
_
S
_
,
A
1
=
_
T L I (m+ 1) T
_
L
0
u
_
Q
U
0
_
T S
+
_
T
0
_
_
S
0
e
1
__
U
_
T S +
_
T
0
_
S
0
__
_
and
A
2
=
_
0 0
U
0
T
_
S
0
e
1
_
U T
_
S
0
_
_
,
where all e
j
are all of dimensions s 1, and
S
= [0 S], and Q
=
_
Q
_
.
Let the R matrix associated with this Markov chain be written in smaller block
form as
R =
_
R
VV
R
VS
R
SV
R
SS
_
.
We can write the blocks of the matrix R, by applying the matrix quadratic equation
for R, as follows
R
VV
=
_
T
0
_
L I (m + 1) +R
VV
_
T L I (m+ 1)
_
+R
VS
_
U
0
_
T S
+
_
T
0
_
_
S
0
e
1
___
+(R
VV
R
VS
+R
VS
R
SS
)
_
U
0
T
_
S
0
e
1
__
,
R
VS
=u
_
T
0
_
L
0
Q
+R
VV
_
T
_
L
0
u
_
Q
_
VACATION MODELS IN DISCRETE TIME 21
+R
VS
_
U
_
T S +
_
T
0
_
S
0
___
+(R
VV
R
VS
+R
VS
R
SS
)
_
U T
_
S
0
__
,
R
SV
=U
0
_
T
0
_
S
+ R
SV
_
T L I (m+ 1)
_
+R
SS
_
U
0
_
T S
+
_
T
0
_
_
S
0
e
1
___
+
_
R
SV
R
VS
+R
2
SS
__
U
0
T
_
S
0
e
1
__
,
R
SS
=U
_
T
0
_
S + R
SV
_
T
_
L
0
u
_
Q
_
+R
SS
_
U
_
T S +
_
T
0
_
S
0
___
+
_
R
SV
R
VS
+R
2
SS
__
U T
_
S
0
__
.
These smaller blocks make computing R more efcient. They can still be partitioned
into much smaller blocks to take further advantage of the zero elements of the matrices.
We will not pursue this aspect in this paper.
After the R matrix has been obtained we can then follow the steps described in
section 3.1 to obtain the steady state distribution x.
4.4.1. Stability when all the dimensions of the parameters are nite
We consider the case when (n, m, r, s) < . For such cases we can use the mean drift
results of [45] to establish stability conditions. Consider the matrix A = A
0
+A
1
+A
2
,
which is given by
A =
_
_
T + T
0
_
L I (m + 1) u
_
T + T
0
_
L
0
Q
U
0
_
T + T
0
_
_
S
+ S
0
e
1
_
U
_
T + T
0
_
S + S
0
_
_
.
If we let the vector represent the invariant vector of A, then the condition for ergodic-
ity, A
0
1 < A
2
1 can be applied. For some special cases when Q = I we can obtain a
simplied version for this stability condition. When Q = 1, we can obtain a sufcient
condition for ergodicity.
4.4.2. Stability issues when at least one of the dimensions of the parameters is innite
Here we consider the case when (n, m, r, s) , with at least one of them equal to
innity, in which case all the block matrices of P
1
will be of innite-order. In this case
we cannot directly apply the mean drift result used in section 4.4.1. Nevertheless, under
certain conditions which we discuss here, the mean drift result may still be applicable.
The main issue is to show that the associated matrix A is irreducible. It was pointed out
by Tweedie [59] and Latouche and Ramaswami [30, p. 161] that if A is irreducible then
the mean drift result is a sufcient condition for the transition matrix P
1
to be positive
recurrent.
In order to make it apparent what the matrix A looks like, we perform a minor
rearrangement of the state space. We rearrange
S
i
as follows:
S
i
= (i, k, u, j), i = 1, 2, . . . , u = 1, 2, . . . , s, k = 1, 2, . . . , n, j = 1, 2, . . . , m.
22 A.S. ALFA
This simply allows us to swap the position of the arrival parameters with the operational
time parameter of the server. The resulting A matrix is denoted by
A and given as
A =
_
_
T + T
0
_
L I (m + 1)
_
T + T
0
_
u L
0
_
Q
_
T + T
0
_
U
0
_
S
+ S
0
e
1
_ _
T + T
0
_
U
_
S + S
0
_
_
.
We consider the case where all the distributions for the interarrival times, service times,
vacation durations, and operational times have general renewal distributions. If we rep-
resent these distributions using the remaining time format then they all have a special
IPH distribution. Each expression in the matrix
A has T + T
0
as the rst term, hence
the structure of
A is as follows
A =
_
1
2
3
4
_
,
where
v
=
_
_
_
_
_
_
_
_
v,1
v,2
v,3
v,4
v,5
. . .
v,0
v,0
v,0
v,0
.
.
.
_
_
, v = 1, 2, 3, 4.
Note that
w
=
w+1
v=w
v,0
, w = 1, 3, are stochastic matrices, and
w
=
w+1
v=w
z=1
v,z
, w = 1, 3, are also stochastic matrices.
By using the result in [4], which is based on [30,45,59], we can show that the
matrix
A is irreducible if
k
> 0,
j
> 0, u
u
> 0, and
l
> 0 (k, j, u, l) 1.
This condition only ensures that at least one element of
v,z
> 0, (v, z). The
cases when this type of condition does not apply were also discussed in [4]. This is not
a focus of the current paper and will not be pursued further here.
The example presented is one where it is possile to use the mean drift result of [45]
for establishing the stability conditions for cases that involve the special IPH. As pointed
out earlier, we have to consider them on a case by case basis.
5. Performance measures
First let us dene a column vector f
i
whose elements assume values of 0 or 1 only.
We let its transpose f
i
be a row vector which is of the same order as x
i
, and with each
element of x
i
having a relationship which corresponds to element in the same location
in f
i
. Let
i
represent the set of all the locations in f
i
for which there are ones, hence
all other locations have zeros. For example,
i
(i
= i
1
, i
2
) implies the value of ones in
VACATION MODELS IN DISCRETE TIME 23
all locations where the number in the gate is between i
1
and i
2
, inclusive. Hence we can
write the corresponding f
i
vector as f
i
(
i
) = f
i
(i
= i
1
, i
2
).
5.1. Queue length distributions
5.1.1. Level independent case
We start by assuming that, based on the results of section 3, the vector x has been
obtained. Our interest now is to obtain some key measures related to the queue length.
Let y
i
be the marginal probability of nding i customers in the system at an arbi-
trary time, then
y
i
= x
i
1.
Denote by
L
the mean number of customers in the system at an arbitrary time. We have
L
= x
1
(I R)
2
1.
Let v
0
be the probability that the server is on vacation, then
v
0
= x
0
1 +x
1
(I R)
1
f
1
_
l = 1, r
_
.
Note that we are assuming multiple vacation system here.
Dene y
v
i
as the conditional probability that there are i customers waiting, when
the server is on vacation. We have
y
v
i
=
x
i
f
i
(l = 1, r)
v
0
, i 1, and y
v
0
=
x
0
1
v
0
.
The conditional expectation of the number of customers waiting when the server is on
vacation,
v
, is given as
v
=
x
1
(I R)
2
f
1
(l = 1, r)
v
0
.
Denote by s
0
the probability that service is on-going at arbitrary time, i.e. that the server
is busy serving a customer, then we have s
0
= 1 v
0
.
Dene y
s
i
as the conditional probability that there are i customers waiting, when
the server is busy serving. We have
y
s
i
=
x
i
(1 f
i
(l = 1, r))
s
0
, i 1.
Let
s
be the conditional expectation of the number of customers waiting when the
server is busy serving then
s
= x
1
(I R)
2
1 f
1
(l = 1, r)
s
0
.
24 A.S. ALFA
5.1.2. Level dependent case
Once more we assume that we have obtained x as in section 3. Nowx can be partitioned
as follows: x = [x
0
, x
1
, x
2
, . . .], and further as x
i
= [x
i,1
, x
i,2
, . . . , x
i,i
1
, x
i,i
,
x
i,i
+1
, . . . , x
i,i
].
The marginal probability density, y
i
, of nding i customers in the system at an
arbitrary time is given as
y
i
= x
i
1.
Denote by
L
the mean number of customers waiting in the system at an arbitrary
time. We have
L
= x
0
_
=1
i=0
R
i
_
1.
The conditional probability y
v
i
, of nding i customers waiting in the system during
a vacation, is obtained as follows. Let v
0
be the probability that the server is on vacation
then
v
0
=x
0
1 +x
0
i=1
_
i1
=0
R
_
f
i
_
l = 1, r
_
, and
y
v
i
=
x
i
f
i
(l = 1, r)
v
0
, i 1,
y
v
0
=
x
0
1
v
0
.
v
, the conditional expectation of the number of customers waiting when the server is
on vacation, is given by
v
= x
0
=1
_
1
i=0
R
i
_
f
i
(l = 1, r)
v
0
.
The conditional probability, y
s
i
, of nding i customers waiting in the system during
service is obtained as follows. Let s
0
be the probability that service is on-going at an
arbitrary time. Then s
0
= 1 v
0
. We then have
y
s
i
=
x
i
[1 f
i
(l = 1, r)]
s
0
.
The conditional expectation of the number of customers waiting in the system when
the server is busy serving,
s
, is given by
s
= x
0
=1
_
1
i=0
R
i
_
1 f
i
(l = 1, r)
s
0
.
VACATION MODELS IN DISCRETE TIME 25
5.2. Distribution of work in the system at arbitrary times, during vacation and
during service
We need the following matrices in order to obtain information on work in the sys-
tem. Dene S
1
= S
0
, G
(1)
() = S
1
S
1
, 1, G
()
() = S
1
, 1, and
G
()
() = S
1
G
(1)
( 1) + SG
()
( 1), + 1, 2. The matrix G
()
()
is of dimension m m. Note that G
(0)
() = 0, 1, G
(v)
() = 0, < v, and
G
(0)
(0) = I, by denition. Also let D
0
= T and D
1
= T
0
.
Even though we can write the equations for the work in the system in a general
form for the systems considered in this paper, it would require too much complex no-
tations. We therefore opt to write the equations for the time-limited cases only for
both level-dependent and level-independent cases. For the level-independent case let
x
i
= [x
v
i
, x
s
i
], and for the level-dependent case let x
i,i
= [x
v
i,i
, x
s
i,i
], where the su-
perscript v represents vacation and s represents service, and x
i,i
implies x
i,i
=i
. Note
that for the level-independent case we have x
s
i
= [x
s
i,1
, x
s
i,2
, . . . , x
s
i,N
] and for the level-
dependent case we have x
s
i,i
= [x
s
i,i
,1
, x
s
i,i
,2
, . . . , x
s
i,i
,N
].
5.2.1. Level independent case
The probability V
p
a
that the amount of work in the system at the instant of return from a
vacation (i.e., at a polling instatnt) is a is given by
V
p
a
= c
1
_
x
v
0
_
D
1
L
0
_
G
(1)
a
1+
a
i=1
1
d=0
x
v
i
__
D
d
L
0
_
_
Q
G
(i+d)
(a)
_
1
_
_
, a 1,
where
c =
a=1
_
x
v
0
__
D
1
L
0
_
G
(1)
a
_
1 +
a
i=1
1
d=0
x
v
i
__
D
d
L
0
_
_
Q
G
(i+d)
(a)
__
1
_
.
The probability V
v
a
that the amount of work left behind just after the vacation starts
is a is given by
V
v
0
=b
1
N
u=1
x
s
1,u
_
D
0
S
0
_
1 and
V
v
a
=b
1
a
i=1
a
w=1
1
d=0
x
s
i,N
_
D
d
__
S
w1
S
0
_
G
(i+d1)
(a w + 1)
__
1, a 1,
where
b =
N
u=1
x
s
1,u
_
D
0
S
0
_
1
+
a=1
a
i=1
a
w=1
1
d=0
x
s
i,N
_
D
d
__
S
w1
S
0
_
G
(i
+d1)
(i w + 1)
__
1.
26 A.S. ALFA
Note that because of our rule for resolving conict this result would be different if we
consider work left behind just at the start of a vacation.
5.2.2. Level dependent case
The probability V
p
a
that the amount of work in the system at a polling instant (i.e., instant
of return from a vacation) is a is given by
V
p
a
= c
1
_
a
i=1
1
d=0
x
v
i,i
__
D
d
L
0
_
_
Q
G
(i+d)
(a)
__
1 +x
v
0
__
D
1
L
0
_
G
(1)
a
_
1
_
,
a 1,
where
c =
a=1
_
x
v
0
__
D
1
L
0
_
G
(1)
a
_
1 +
a
i=1
1
d=0
x
v
i,i
__
D
d
L
0
_
_
Q
G
(i+d)
(a)
__
1
_
.
The probability, V
v
a
, that the amount of work left behind just after the vacation
starts is a is given by
V
v
0
=b
1
N
u=1
x
s
1,1,u
_
D
0
S
0
_
1 and
V
v
a
=b
1
a
i=1
i
=1
a
w=1
1
d=0
x
s
i,i
,N
__
D
d
_
S
w1
S
0
__
G
(i
+d1)
(a w + 1)
_
1, a 1,
where
b =
N
u=1
x
s
1,1,u
_
D
0
S
0
_
1
+
a=1
a
i=1
i
=1
a
w=1
1
d=0
x
s
i,i
,N
__
D
d
_
S
w1
S
0
__
G
(i
+d1)
(a w + 1)
_
1.
Once more, note that because of our rule for resolving conict this result would be
different if we consider work left behind just at the start of a vacation.
5.3. Waiting time distribution
The waiting time distribution can be studied by using absorbing Markov chains in both
the level-dependent and level-independent cases. The absorbing states in each case are
the states when the target customer reaches the head of the line, which is the point we
assume the customers wait (in the queue) is over. The only additional wait may be, in
some cases, for the server to return from a vacation, but not for service completion of any
other customers before the target customer. We dene P
1
(w) and P
2
(w) as the transition
matrices for the absorbing Markov chains for the level-independent and level-dependent
VACATION MODELS IN DISCRETE TIME 27
cases, respectively. Both P
1
(w) and P
2
(w) are obtained by setting T := 1 and T
0
:= 0
in both P
1
and P
2
, respectively. We write this information as follows:
P
1
(w) = P
1
_
T := 1, T
0
:= 0
_
, P
2
(w) = P
2
_
T := 1, T
0
:= 0
_
.
Both matrices P
1
(w) and P
2
(w) are bidiagonal block matrices with diagonals and sub-
diagonals only. Matrix P
1
(w) has an identity matrix at the top left corner. Matrix P
2
(w)
has a stochastic block matrix at the top left corner the states in this rst block do not
communicate with the other remaining states in the other blocks.
Having obtained the transition matrices for the absorbing states we then have to
obtain the steady state distribution of the number of customers in the system as seen
by an arriving customer, which is then used in conjunction with the appropriate absorb-
ing Markov chain to obtain the waiting time. We only give an example with a level-
independent case the time-limited case, in particular. The other cases are available
in [4].
5.3.1. Waiting time distribution for the time-limited case
Keeping in mind that x
i
, i 0, has been partitioned as follows: x
i
= [x
v
i
x
s
i,1
. . . x
s
i,N
].
Dene the vectors z
i
, i 0, as the corresponding vectors to x
i
, where z
i
corresponds to
the steady state vector of the system as seen by an arriving customer. Then
z
v
0
=
1
x
v
0
__
T
0
_
L + L
0
__
+
1
N
u=1
x
s
1,u
__
T
0
_
S
0
__
;
z
v
i
=
1
_
x
v
i
__
T
0
_
L I (m+ 1)
_
+x
s
i,N
__
T
0
_
S
_
+x
s
i+1,N
__
T
0
_
S
0
__
, i 1;
z
s
i,1
=
1
_
x
v
i
__
T
0
_
L
0
Q
___
, i 1;
z
s
i,u
=
1
_
x
s
i,u1
__
T
0
_
S
_
+x
s
i+1,u1
__
T
0
_
S
0
___
, i 1, 2 u N.
Dene z
0
0
= z
v
0
(1 I), and for i 1 dene z
v,0
i,0
= z
v
i
(1 I), z
s,0
i,u
= z
s
i,u
(1 I). Let
z
0
i
= [ z
v,0
i,0
z
s,0
i,1
z
s,0
i,2
z
s,0
i,N
], i 1. Finally, let z
0
= [ z
0
0
, z
0
1
, z
0
2
, . . .]. Also let
z
n+1
= z
n
P, n 0.
We assume that z
n
is partitioned in the same way as z
0
. Let W
a
be the probability that a
customers waiting time is less than or equal to a units of time. Then we have:
Proposition 2.
W
a
= z
a
0
1, a 0.
6. Conclusion
We have shown that by using discrete time approach it is possible to formulate nearly all
vacation queues as QBD type of Markov chain and then use matrixgeometric method
28 A.S. ALFA
or matrix-product based method to analyze them. This approach is developed for al-
gorithmic tractability mainly. In applying matrixanalytic methods it is important that
the special structure associated with a problem is exploited. This is very critical, espe-
cially in the case of the vacation models which involve large order of matrices. It is even
more critical when the trafc intensity of such systems is close to 1, otherwise problems
related to computational feasibility ensue. In future works we study the algorithmic
aspects of specic systems.
The models covered in this paper are all single server systems with single arrivals
and non-batch service, including the GI/G/1 types but mainly for those GI and G which
can be represented by nite state Markov chains. The structure is the same as for the
SMP/MSP/1 types also. We also point out that the structure presented here can be easily
extended to batch arrivals and/or batch service, provided the batch size distribution has
a nite support. In that case we simply re-block the matrices and achieve QBDs.
Acknowledgements
The author acknowledges the anonymous referee and Associate Editor for their com-
ments which have denitely helped to improve the presentation of this paper. This re-
search is partially supported by grant No. OGP0006854 from the Natural Sciences and
Engineering Research Council of Canada.
References
[1] A.S. Alfa, A discrete MAP/PH/1 queue with vacations and exhaustive time-limited service, Oper.
Res. Lett. 18 (1995) 3140.
[2] A.S. Alfa, A discrete MAP/PH/1 vacation queue with gated time-limited service, Queueing Systems,
29 (1998) 3554.
[3] A.S. Alfa, Discrete time vacation models and matrixanalytic method, Internal Report, University of
Windsor (June, 2001).
[4] A.S. Alfa and W. Li, Matrixgeometric solution of the discrete time GI/G/1 system, Stochastic Mod-
els 17(4) (2001) 541554.
[5] J.R. Artalejo, Analysis of an M/G/1 queue with constant repeated attempts and server vacations,
Comput. Oper. Res. 24(6) (1997) 493504.
[6] B. Avi-Itzhak and P. Naor, Some queueing problems with the service station subject to server break-
downs, Oper. Res. 10 (1962) 303320.
[7] C. Blondia, Finite capacity vacation models with non-renewal input, J. Appl. Probab. 28 (1991) 174
197.
[8] L. Bright and P.G. Taylor, Calculating the equilibrium distribution in level dependent quasi-birth-and-
death processes, Stochastic Models 11(3) (1995) 497525.
[9] H. Bruneel, Analysis of discrete-time buffer with single server output, subject to interruption process,
in: Performance 84 (Elsevier, Amsterdam, 1984) pp. 103115.
[10] H. Bruneel, Analysis of an innite buffer system with random server interruption, Comput. Oper. Res.
11 (1994) 373386.
[11] X. Chao and Y.Q. Zhao, Analysis of multi-server queues with station and server vacation, European
J. Oper. Res. 110(2) (1998) 392406.
VACATION MODELS IN DISCRETE TIME 29
[12] B.D. Choi, Single server retrial queues with priority calls, Math. Computer Modelling 30(3/4) (1999)
732.
[13] S.C. Dafermos and M.F. Neuts, A single server queue in discrete time, Cahiers C.E.R.O. 13 (1971)
2340.
[14] B.T. Doshi, A note on stochastic decomposition in a GI/G/1 queue with vacations or set-up times,
J. Appl. Probab. 22 (1985) 419.
[15] B.T. Doshi, Queueing systems with vacations a survey, Queueing Systems 1 (1986) 2966.
[16] B.T. Doshi, Generalization of the stochastic decopmosition results for the single-server queues with
vacations, Stochastic Models 6 (1990) 307333.
[17] B.T. Doshi, Single server queues with vacations, in: Stochastic Analysis of Computer and Communi-
cations Systems, ed. H. Takagi (1990) pp. 217265.
[18] M. Eisenberg and K.K. Leung, A single-server queue with vacations and non-gated time-limited ser-
vice, Performance Evaluation 12 (1991) 115125.
[19] A. Federgruen and L. Green, Queueing systems with service interruptions, Oper. Res. 34(5) (1986)
752768.
[20] A. Federgruen and L. Green, Queueing systems with service interruptions II, Naval Res. Logistics 35
(1988) 345358.
[21] J.M. Ferrandiz, The BMAP/GI/1 queue with server set-up times and server vacations, Adv. in Appl.
Probab. 25 (1993) 235254.
[22] S.W. Fuhrman and R.B. Cooper, Stochastic decomposition in the M/G/1 queue with generalized
vacations, Oper. Res. 33 (1985) 11171129.
[23] D.P. Gaver, Jr., A waiting line with interrupted service including priorities, J. Roy. Statist. Soc. B 24
(1962) 7390.
[24] C.M. Harris and W.G. Marchal, State dependence in M/G/1 server vacation models, Oper. Res. 36
(1988) 560565.
[25] F. Ishizaki, T. Takine and T. Hasegawa, Analysis of a discrete-time queue with gated priority, Per-
formance Evaluation 23 (1995) 121143.
[26] F. Ishizaki, T. Takine, Y. Takahashi and T. Hasegawa, A generalized SBBP/G/1 queue and its appli-
cations, Performance Evaluation 21 (1994) 163181.
[27] J. Keilson, Queues subject to service interuptions, Ann. Math. Statist. 33 (1962) 13141322.
[28] J. Keilson and L.D. Servi, Oscillating random walk models for GI/G/1 vacation systems with
Bernoulli schedules, J. Appl. Probab. 23 (1986) 790802.
[29] B.K. Kumar and D. Arivudainambi, The M/G/1 retrial queue with Bernoulli schedules and general
retrial times, Comput. Math. Appl. 43(1/2) (2002) 1530.
[30] G. Latouche and V. Ramaswami, Introduction to Matrix Analytic Methods in Stochastic Modeling,
ASASIAM Series on Applied Probability (1999).
[31] H.W. Lee, B.Y. Ahn and N.I. Park, Decompositions of the queue length distributions in the MAP/G/1
queue under multiple and single vacations with N-policy, Stochastic Models 17(2) (2001) 157190.
[32] H.W. Lee, S.S. Lee, J.O. Park and K.C. Chae, Analysis of the M
x
/G/1 queue with N-policy and
multiple vacations, J. Appl. Probab. 31 (1994) 476496.
[33] T.T. Lee, M/G/1/N queue with vacation time and limited service discipline, Performance Evaluation
9 (1989) 181190.
[34] K.K. Leung and M. Eisenberg, A single-server queue with vacations and gated time-limited service,
IEEE Trans. Commun. 38(9) (1990) 14541462.
[35] Y. Levy and U. Yechiali, Utilization of the idle time in an M/G/1 queue, Managm. Sci. 22 (1975)
202211.
[36] Y. Levy and U. Yechiali, M/M/s queues with server vacations, INFOR 14 (1976) 153163.
[37] H. Li and T. Yang, A single-server retrial queue with server vacation and a nite number of input
sources, European J. Oper. Res. 85(1) (1995) 149160.
30 A.S. ALFA
[38] H. Li and Y. Zhu, M(n)/G/1/N queues with generalized vacations, Comput. Oper. Res. 24(4) (1997)
301316.
[39] D.M. Lucantoni, New results on the single server queue with a batch Markovian arrival process,
Stochastic Models 7 (1991) 146.
[40] D.M. Lucantoni, K.S. Meier-Hellstern and M.F. Neuts, A single -server queue with server vacations
and a class of non-renewal arrival processes, Adv. in Appl. Probab. 22 (1990) 676705.
[41] S.K. Matendo, A single-server queue with server vacations and a batch Markovian arrival process,
Cahiers C.E.R.O. 35 (1993) 87114.
[42] I.L. Mitrani and B. Avi-Itzhak, A many server queue with service interruptions, Oper. Res. 16 (1968)
628638.
[43] M. Miyazawa, Decomposition formulas for sngle-server queues with vacations: A unied approach
by rate conservation law, Stochastic Models 10 (1994) 389413.
[44] M.F. Neuts, A versatile Markovian point process, J. Appl. Probab. 16 (1979) 764779.
[45] M.F. Neuts, MatrixGeometric Solutions in Stochastic Models An Algorithmic Approach (Johns
Hopkins Univ. Press, Baltimore, MD, 1981).
[46] M.F. Neuts, Structured Stochastic Matrices of the M/G/1 Type and their Applications (Marcel Dekker,
New York, 1989).
[47] M.F. Neuts, Some promising directions in algorithmic probability, in: Advances in Matrix Analytic
Methods for Stochastic Models, eds. A.S. Alfa and S. Chakravarthy (Notable Publications, NJ, 1998)
pp. 429443.
[48] V. Ramaswami and P.G. Taylor, Some properties of the rate operators in level dependent quasi-birth-
and-death processes with a countable number of phases, Stochastic Models 12(1) (1996) 143164.
[49] H. Schellhaas, Single server queues with a batch Markovian arrival process and server vacations, OR
Spektrum 15 (1994) 189196.
[50] B. Sengupta, A queue with service interruption in an alternating random environment, Oper. Res. 38
(1990) 308318.
[51] D. Shi, J. Guo and L. Liu, SPH-distributions and the rectangle-iterative algorithm, in: MatrixAnalytic
Methods in Stochastic Models, eds. S. Chakravarthy and A.S. Alfa (Marcel Dekker, New York, 1996)
pp. 207224.
[52] D. Shi and D. Liu, Markovian models for non-negative random variables, in: Advances in Matrix
Analytic Methods for Stochastic Models, eds. A.S. Alfa and S. Chakravarthy (Notable Publications,
NJ, 1998) pp. 403427.
[53] M. Shomrony and U. Yechiali, Burst arrival queues with server vacations and random timers, Math.
Methods Oper. Res. 53 (2001) 117146.
[54] H. Takagi, Queueing Analysis: A Foundation of Performance Evaluation, Vol. 1: Vacation and Prior-
ity Systems, Part 1 (North-Holland, Amsterdam, 1991).
[55] H. Takagi, Time-dependent process of M/G/1 vacation models with exhaustive service, J. Appl.
Probab. 29 (1992) 418429.
[56] T. Takine and T. Hasegawa, A batch SPP/G/1 queue with multiple vacations and exhaustive service
disicpline, Telecommunication Systems 1 (1993) 195215.
[57] T. Takine and B. Sengupta, A single server queue with service interruptions, Queueing Systems 26
(1997) 285300.
[58] N. Tian and Z.G. Zhang, A discrete-time GI/Geo/1 queue with multiple vacations, Queueing Systems
40(3) (2002) 283294.
[59] R.L. Tweedie, Operator-geometric stationary distributions for Markov chains, with aplications to
queueing models, Adv. in Appl. Probab. 14(2) (1982) 368391.
[60] H.C. White and L.S. Christie, Queueing with preemptive priorities or with breakdown, Oper. Res. 6
(1958) 7995.
[61] Z.G. Zhang, R.G. Vickson and M.J.A. van Eenige, Optimal two-threshold policies in an M/G/1 queue
with two vacation types, Performance Evaluation 29(1) (1997) 6380.