Anda di halaman 1dari 17

This article was downloaded by: [Avinashilingham Inst for Hm Sci & Higher Ed for woman]

On: 25 February 2013, At: 01:01


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

Stochastic Analysis and Applications


Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/lsaa20

(X)
On an M /G/1 Retrial System with Two Types of
Search of Customers from the Orbit
a b c d
T. G. Deepak , A. N. Dudin , V. C. Joshua & A. Krishnamoorthy
a
Department of Mathematics, Indian Institute of Space Science and Technology,
Thiruvananthapuram, India
b
Department of Applied Mathematics and Computer Science, Belarusian State University,
Minsk, Belarus
c
Department of Mathematics, C.M.S College, Kottayam, India
d
Department of Mathematics, Cochin University of Science and Technology, Cochin, India
Version of record first published: 13 Dec 2012.

(X)
To cite this article: T. G. Deepak , A. N. Dudin , V. C. Joshua & A. Krishnamoorthy (2013): On an M /G/1 Retrial System with
Two Types of Search of Customers from the Orbit, Stochastic Analysis and Applications, 31:1, 92-107

To link to this article: http://dx.doi.org/10.1080/07362994.2013.741389

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to
anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,
proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.
Stochastic Analysis and Applications, 31: 92107, 2013
Copyright Taylor & Francis Group, LLC
ISSN 0736-2994 print/1532-9356 online
DOI: 10.1080/07362994.2013.741389

On an M X /G/1 Retrial System with Two Types


Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

of Search of Customers from the Orbit

T. G. DEEPAK,1 A. N. DUDIN,2 V. C. JOSHUA,3


AND A. KRISHNAMOORTHY4
1
Department of Mathematics, Indian Institute of Space Science and
Technology, Thiruvananthapuram, India
2
Department of Applied Mathematics and Computer Science, Belarusian
State University, Minsk, Belarus
3
Department of Mathematics, C.M.S College, Kottayam, India
4
Department of Mathematics, Cochin University of Science and
Technology, Cochin, India

Single server retrial queueing models in which customers arrive according to a


batch Poisson process are considered here. An arriving batch, nding the server
busy, enters an orbit. Otherwise, one customer from the arriving batch enters for
service immediately while the rest join the orbit. The customers from the orbit
(the orbital customers) try to reach the server subsequently with the inter-retrial
times exponentially distributed. Additionally, at each service completion epoch, two
different search mechanisms, that is, type I and type II search, to bring the orbital
customers by the system to service, are switched on. Thus, when the server is idle,
a competition takes place among primary customers, customers who come by retrial
and by two types of searches. The type I search selects a single customer whereas
the type II search considers a batch of customers from the orbit. Depending on the
maximum size of the batch being considered for service by a type II search, two
cases are addressed here. In the rst case, no restriction on batch size is assumed,
whereas in the second case, maximum size of the batch is restricted to a pre-assigned
value. We call the resulting models as model 1 and model 2 respectively. In all
service modes other than type II search, only a single customer is qualied for
service. Service times of the four types of customers, namely, primary, repeated,
and those who come by two types of searches are arbitrarily distributed (with
different distributions which are independent of each other). Steady state analysis is
performed and stability conditions are established. A control problem for model 2
is considered and numerical illustrations are provided.

Keywords Orbital customers; Queueing; Retrial queue; Search on the orbit.

Mathematics Subject Classication 60K25, 90B22.

Received June 3, 2012; Accepted September 5, 2012


Address correspondence to A. Krishnamoorthy, Department of Mathematics, Cochin
University of Science and Technology, Cochin 682 022, India; E-mail: achyuthacusat@
gmail.com

92
Retrial Queues with Orbital Search 93

1. Introduction
A retrial queueing system is characterized by the fact that a customer arriving
when all servers accessible for him are busy, leaves the service area and joins a
group of unsatised customers called orbit, but after a random amount of time, he
returns and repeats his demand for service. These arise frequently in the stochastic
modelling of telecommunications, computer systems, stock and ow etc. Review
of retrial queueing literature can be found in [1, 2, 4, 5, 7, 8, 14]. A number of
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

applications of retrial queues in science and engineering are given in [9]. In the retrial
set up, each service is preceded and followed by the server(s) idle time because of
the ignorance of the status of the server(s) and orbital customers by each other.
Some researchers have shown interest in designing retrial queueing models
that reduce the server(s)idle time. One way to achieve this is by the introduction
of search of orbital customers immediately after a service completion. Search for
orbital customers was introduced by Artalejo et al. [3] in which they considered an
M/G/1 retrial model where the server goes in search of customers immediately after
a service completion with probability depending on the orbit size at that moment.
They assumed that the time taken for search is negligible. Dudin et al. [6] generalizes
the result in [3] by introducing search time, two types of services to customers
(primary/orbital) and the external arrival process as a batch Markovian process.
Note that in [10], the authors examined a classical queue with search of customers
immediately on termination of a service.
This article generalizes the model discussed in [3] by introducing two types of
searches involving positive random duration, a batch Poisson arrival process and
different service time distributions to different class of customers. An arriving batch,
nding the server busy, enters an orbit. If the server is found to be idle by an arriving
batch, one customer from the arriving batch enters for service immediately while the
rest join the orbit. The orbital customers try to reach the server subsequently and the
time between successive retrials are exponentially distributed. Additionally, at each
service completion epoch, two different search mechanisms, say, type I and type II
search, to bring the orbital customers to the service station, are switched on. Thus,
when the server is idle, a competition takes place among primary customers, the
orbital customers by way of retrial and the two types of searches. The type I search
selects a single customer, whereas the type II search opts for a batch of customers
from the orbit. Depending on the maximum size of the batch being considered
for service by a type II search, two cases are addressed here. In the rst case, no
restriction on batch size is assumed, whereas in the second case, maximum size
of the batch is restricted to a preassigned value. We call the resulting models as
model 1 and model 2 respectively. In all service modes other than type II search,
only a single customer is qualied for service. Service times of the four types of
customers, namely, primary, repeated, and those who come by two types of searches
are arbitrarily distributed (with distinct distributions which are independent of each
other).
The present model is motivated by the situation in which when the number
of customers is below a pre-assigned level, they are provided service manually.
When this threshold is exceeded at a departure epoch, customers are served in
bulk using an expensive machine which can serve sufciently large number of
customers at a time. The xed cost of activating the machine is so high that the
minimum number of customers needed for machine service should be very large.
Another potential application of this type of service mechanism is in computer and
94 Deepak et al.

communication engineering where requests (messages or terminals or satellites) that


involve information of a general nature such as access to a common data base or a
common input-output device such as a laser printer or color plotter are handled in
groups.
The rest of the article is organized as follows. The mathematical model is
described in Section 2. Section 3 deals with the stationary distribution of the
embedded Markov chain in which Sections 3.1 and 3.2 are completely devoted to the
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

analysis of model 1 and model 2, respectively. Stationary distribution of system size


and orbit size at an arbitrary time point are given in Section 4. Some performance
measures are computed in Section 5 and numerical illustrations are provided in
Section 6.

2. The Mathematical Model


We consider a single server queueing system in which the arrivals occur according
to a batch Poisson process, that is, the customers arrive as a Poisson process with
parameter  in groups of random size C, where C has the distribution

PrC = n = cn  n 1

and the generating function



Cz = EzC  = cn zn  z 1
n=1

The probability that a total of n customers arrive in an interval of length t is thus


given by

n
et tk k
pn t = cn  n 0
k=0
k!

where cnk is the k-fold convolution of cn with itself (that is, the arrivals form a
compound Poisson process) and cn0 = 0n . The service mechanism of the system
is described in the following manner. The primary unit who meets an idle server
is served with service time having the distribution function B0 t, while the rest
join the orbit. Each unit in the arriving batch, nding the server busy, enters the
orbit and retry to access the server with an exponentially distributed time between
two successive retrials. Retrials to get the server occurs with a mean of 1/ i  i >
0, when the number of customers in the orbit is i. Additionally, at a service
completion epoch, type I and type II search mechanisms are switched on. If a retrial
customer reaches the idle server rst, the customer entering the service is served
with service duration having the distribution function B3 t while if the type I search
is turned out to be successful, the selected customer is served according to the
distribution function B1 t. In model 1, there is no restriction on the batch size of
customers taken into service by type II search, whereas in model 2, the maximum
size of the batch is restricted to a pre-assigned integer, say, K. In both models, the
batch service duration are assumed to have an arbitrary distribution function B2 t,
independent of each other. Denote by
r s, the Laplace-Stieltjes transform (LST)
r
and bi (assumed to be nite), the i th moment associated with the distribution
Retrial Queues with Orbital Search 95

function Br t r = 0 1 2 3. The duration of the type I (type II) search follows the
distribution function H1 tH2 t with LST h1 sh2 s.

3. Stationary Distribution of the Embedded Markov Chain


Denote by tn , the nth service completion epoch and in , the number of customers in
the orbit at tn + 0. Then in  n 1 is a Markov chain with state space as the set of
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

non-negative integers. Let pil be the one-step transition probability of the Markov
chain so that

pil = Prin+1 = lin = i

The following lemma, the proof of which follows immediately from the description,
gives expression for pil corresponding to model 1.

3.1. Model 1
Lemma 3.1. In model 1,


l+1
0
p0l = cj lj+1  l 0
j=1


li+1
0 1 1 2 2
pil =  i cj lij+1 + Fi li+1 + Fi l
j=1

3 3
+ Fi li+1  i 1 l i 1 (3.1)

and
2 2
pil = Fi l  i 1 0 l < i 1

where
    
l
et tj
r j
l = pl tdBr t = cl dBr t r = 0 1 2 3
0 0 j=0
j!

and for i > 0


 
i = e+ i t 1 H1 t1 H2 tdt
0
 
e+ i t 1 H2 tdH1 t
1
Fi =
0
 
e+ i t 1 H1 tdH2 t
2
Fi =
0
3
Fi = i i 
r
In the above lemma, note that Fi  r = 1 2 3 denote the probabilities that the idle
period of the server expires through an orbital unit(s) arrival, respectively, by type I
96 Deepak et al.

r
search or type II search or retrial. j denotes the probability that exactly j arrivals
r
occur during a service time of the rth type (according to Br , r = 0 1 2 3. j
occurs as coefcients in the probability generating function


  
eCzt dBr t r = 0 1 2 3
r
r z =
r  Cz = l zl =
l=0 0
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

Here onward, we restrict our analysis to the particular case that i = 1 0i  ,
r
which implies that i = and Fi = F r  r = 1 2 3 Now, we turn our attention to
the computation of the stationary probability vector

 = 0  1         

where

i = lim Prin = i


n

Before this, a natural question to be answered is: under what condition the
stationary probabilities exist? In the following theorem we investigate the ergodicity
of the Markov chain in model 1.

Theorem 3.2. In model 1, the Markov chain in  n 1 is always ergodic.

Proof. Here we use the criteria based on mean drift or, in other words, the theory
of Liapunov functions. The main result of this theory is the following Fosters
criterion (see [11]).
For an irreducible and aperiodic Markov chain with state space S, a sufcient
condition for positive recurrence is the existence of a non-negative function fs s S,
called Liapunovs test function, a positive number  and a nite A S such that the
mean drift

s = fs pss fs
s

is nite for all s A and s <  for all s A.


Take the Liapunovs test function fi = i. Then,


 l+1
0 0
0 = l cj lj+1 = EC1 + b1  1 < 
l=0 j=1

For i 1,

i = lpil i
l


 
li+1
0 

1
=  l cj lij+1 + F 1 l li+1
l=i1 j=1 l=i1



2 

3
+ F 2 l l + F 3 l li+1 i
l=0 l=i1
Retrial Queues with Orbital Search 97

0 1
=  EC1 + b1  1 + F 1 b1 EC 1
3 2
+ F 3 b1 EC 1 i 1F 2 b1 EC

since  + 3r=1 F r = 1.
Hence, i < ,  small, for large i.
Thus, in model 1, the chain is positive recurrent and hence it is ergodic.
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

3.1.1. Computation of Stationary Probabilities. Using the transition probabilities


given by Equation (3.1), we can compute the stationary probabilities l as shown
below.
In model 1,


l+1
0 

2
l = 0 cj lj+1 + i F 2 l
j=1 i=1
 

l+1 
li+1
0 1 3
+ i  cj lij+1 + F 1 li+1 + F 3 li+1  l 0 (3.2)
i=1 j=1

To solve these equations, we use the probability generating function z =



l=0 l z that converges in z < 1.
l

Multiplying each of the equations in 32 by the corresponding power of z,


summing up and rearranging the terms, we get

zz = 0 Cz 0 z + z1 0 F 2 2 z + z 0 Yz

where

Yz =  Cz 0 z + F 1 1 z + F 3 3 z (3.3)

Hence,

0 Cz 0 z zF 2 2 z Yz + zF 2 2 z


z =  (3.4)
z Yz

Let fz = z and gz = Yz.


It is a well-known result that on z = 1 +   small and positive,

Gz 1 + EG + o

for any nonnegative random variable G.


Using this result and (3.3), it can be seen that
0 1 3
gz 1 F 2 + EC  1 + b1  + F 1 b1 + F 3 b1  + o

Also,

fz = z = 1 + 
98 Deepak et al.

Hence, gz < fz on z = 1 + . So, by Rouches theorem, fz = z and fz +
gz = z Yz have the same number of zeroes inside z = 1 + . Clearly, fz has
exactly one zero inside z = 1 + . Therefore, z Yz also has exactly one zero
inside z = 1 +  and it is given by (see Lagranges theorem, [13, pp. 20, 237])


1 dn1
z = n1
Yzn z=0  (3.5)
n=1
n! dz
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

Since z is analytic in z < 1, the numerator of 34 must also be zero at z .
Hence,

z F 2 2 z 
0 =  (3.6)
z 1 + F 2 2 z  Cz  0 z 

The value of 0 together with 34 help us to nd various measures of effectiveness


of the system having practical signicance.
In Section 3.2, we explore the results related to model 2, which are analogous to
those discussed in Section 3.1 corresponding to model 1.

3.2. Model 2
Lemma 3.3. In model 2, the transition probabilities pil can be computed as


l+1
0
p0l = cj lj+1  l 0
j=1


li+1
0 1 1 2 2 3 3
pil =  i cj lij+1 + Fi li+1 + Fi l + Fi li+1  1 i K l i 1
j=1

2 2
pil = Fi l  1 i K 0 l < i 1 (3.7)

li+1
0 1 1 2 2 3 3
pil =  i cj lij+1 + Fi li+1 + Fi li+K + Fi li+1  i > K l i 1
j=1

and
2 2
pil = Fi li+K  i > K i K l < i 1

Theorem 3.4. In model 2, the Markov chain in  n 1 is ergodic if and only if
 
0 
3
r
EC 1 + b1  + F r b1 < 1 + K 1F 2  (3.8)
r=1

Proof. Assume that 38 is true. We have


0
0 = EC1 + b1  1 < 

and for 1 i K,
0 1
i =  EC1 + b1  1 + F 1 b1 EC 1
Retrial Queues with Orbital Search 99

3 2
+ F 3 b1 EC 1 i 1F 2 b1 EC < 

Now, for i > K,


 
0 
3
r r
i = EC 1 + b1  + F b1 1 + K 1F 2  < 
r=1
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

by (3.8). Hence, the chain in model 2 is ergodic if 38 is true.


To prove that (3.8) is necessary for ergodicity, we will use the following criterion
(see [12]):
An irreducible and aperiodic Markov chain with state space Z+ is not ergodic if the
mean drift

s = s pss s
s

is nite for all s Z+ , mean down drift



s = s spss Const
s <s

i.e., is bounded below, and there exists N such that s 0 for s N .


In model 2, the down drift is bounded below since in+1 in K, and if we
assume that 38 is not true, then s 0 for s > K so that by the above criterion,
the chain is not ergodic.
Hence, the theorem is proved.

Corollary 3.5. In Equation 38, if we let K  (by doing so, we are approximating
model 2 with model 1), then 38 is always true, which guarantees the ergodicity of the
chain in model 1 unconditionally, as we proved in Theorem 3.2.

Corollary 3.6. As a particular case, if we take K = 1, then the corresponding chain


will be of M/G/1 type and so by Theorem 1 in Dudin et al. [6], it can be proved that
the condition for existence of stationary distribution is
 
0 
3
r
EC 1 + b1  + F r b1 < 1
r=1

which agrees with 38 under the assumption K = 1.

Corollary 3.7. If the arrival process is Poisson, F 1 = F 2 = 0 and B0  = B3 , the
present model reduces to the classical M/G/1 retrial model with constant retrial rate.
3
Under this assumption, the Condition (3.8) reduces to b1 < F 3 , the result obtained
from [8, Theorem 1.5] by taking constant retrial rate instead of linear retrial rate.

Corollary 3.8. If the search time distributions Hi t expi  i = 1 2 and the retrial
r
rate is linear, that is, if i = i , then  i = + +

 Fi = + +
i
 r = 1 2 and
1 2 +i 1 2 +i
3
Fi = + + +i . In this case, it can be proved that the chain is positive recurrent if
i
1 2
3 2 3
EC b1 2 b1  < for model 1 and b1 EC < 1 for model 2.
100 Deepak et al.

In the above corollary, since the retrial rate is assumed to be linearly


proportional to the orbit size, as the orbit size grows, there is a great chance that
3
the server is always being occupied by an orbital customer so that Fi 1 whereas
1 2
Fi  Fi and  i 0. Again, in model 1, since the batch size has no restriction, the
throughput rate resulting from the batch service is also signicant in maintaining
stability whereas it is not signicant in model 2 because of the restriction on batch
size.
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

3.2.1. Computation of Stationary Probabilities. In model 2, z satises the relation


 
0 Cz 0 z Yz + zF 2 2 z K1
i=1 i + i=K i z
iK

z =  (3.9)
z Yz

Let


K1
K z =  i zi  (3.10)
i=0

Then,


K1 

1
i + i ziK = K 1 0 + z K z
i=1 i=K zK

Substituting in (3.9) and rearranging, we get

0 zK1 Cz 0 z Yz + F 2 2 z zK K 1 0  K z


z =  (3.11)
zK1 z Yz F 2 2 z

Now, 1 = 1 gives


K1
0 EC1 F 2 b1 + b1  Y  1 + K 1F 2  + F 2
2 0
K ii
i=1

= 1 Y  1 + F 2 K 1 b1 EC


2
(3.12)

Consider the denominator zK zK1 Yz F 2 2 z in 311.


Let

fz = zK

and

gz = zK1 Yz F 2 2 z

Then, it can be shown that on z = 1 + ,


2 0
gz 1 + K 11 F 2  + F 2 b1 EC + EC  1 + b1 
1 3
+ F 1 b1 + F 3 b1  + o
Retrial Queues with Orbital Search 101

and

fz = zK = 1 + K = 1 + K + o

Hence, gz < fz onz = 1 +  whenever

0 
3
r
EC 1 + b1  + F r b1  < 1 + K 1F 2 
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

r=1

which is the sufcient condition for positive recurrence given by (3.8) in Theorem 3.4.
Hence, under the condition for stability, by Rouches theorem, fzand fz + gz
have the same number of zeroes inside z = 1 + .
Clearly fz has K zeroes inside z = 1 + . Therefore, zK zK1 Yz
F 2 z also has K zeroes inside it. One of them is z = 1. The remaining K
2
2mi
1 zeroes are obtained by taking a = 0  = e K  m = 1 2     K 1 fz = z and
1
gz = zK1 Yz + F 2 2 z K in Lagranges theorem (see [13]) and are given by


 2mni
e Kdn1 K1
zm
n
= n1
z Yz + F 2 2 z K z=0  m = 1 2     K 1 (3.13)
n=1
n! dz

Since z is analytic in z < 1, the numerator of 311 must also be zero at each
zm  m = 1 2     K 1. Hence, we have


K1
0 zm K1 Czm  0 zm  zm K  + i zm K zm i 
i=1

F 2
2 zm  = 0 m = 1 2     K 1 (3.14)

Equations 312 and 314 (altogether K equations) can be solved for K unknowns
0  1      K1 . Once 0  1      K1 are known, the expression for z, given
by 311, can be used for computing various measures of effectiveness of practical
interest.

4. Stationary Distribution of System Size and Orbit Size at Arbitrary Time


Let pi r, i 0, r = 0 1 2 3 4 be the steady state probability that at an arbitrary
time point there are i customers in the system and the current service is in the rth
mode. Here, r = 4 corresponds to the case when the server is idle. The following
theorem gives the expression for the steady state probabilities.

Theorem 4.1. The steady state probabilities pi r, r = 0 1 2 3 4, which are same
for model 1 and model 2, are computed as follows:

pi 4 = 1 i  i 1
1
p0 4 = 1 0 


i
pi r = 1  r  i > 0 r = 1 2 3
l F r (4.15)
il
l=1
102 Deepak et al.

 

i 
i1 
il
pi 0 = 1 0  0 +  l
cj c  0
j ilj  i > 0

ij
j=1 l=1 j=1

where
  
l
et tj
 r =

j
cl 1 Br tdt r = 0 1 2 3
l
0 j!
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

j=0


  
 r z =
 r zl =
e1Czt 1 Br tdt
l
l=0 0

and the average inter-departure time


   
1 0 0 3
r
 = 0 + b1 + 1 0  1 + b1  + F r b1 
 r=1

Proof. Proof follows from the theory of Markov renewal process. Here,
 r denotes the probability that exactly l arrivals occur during a service
l
time of the rth type, which has not been completed so far. Introduce the

probability generating functions Pz r =  i=0 pi rz
i
 r = 0 1 2 3 4 and using

the normalizing condition that P1 = 1 where Pz = 4r=0 Pz r, we get the value
of .
In a similar manner, if we dene qi r i 0 r = 0 1 2 3 4 as the steady state
probability at an arbitrary time that there are i customers in the orbit and the
current service is in the rth mode, then

Theorem 4.2. In model 1,

qi 4 = 1 i  i 1
1
q0 4 = 1 0 


i+1
qi r = 1  r  i 0 r = 1 3
l F r il+1
l=1



qi 2 = 1  i 0
l F 2
2
(4.16)
i
l=1

and
 i+1 
 
i 
il+1
qi 0 = 1 0  0 +  l
cj c  0
j ilj+1  i 0

ij+1
j=1 l=1 j=1

In model 2, the only change is in the expression for qi 2 and it is given by
 

K 
i
qi 2 =  1  2
l F 2 +  2
K+l F 2  i 0
i il
l=1 l=1
Retrial Queues with Orbital Search 103

5. Some Performance Measures


Based on the stationary distributions for both the system size and orbit size
discussed above, we can compute some performance characteristics of the model as
follows:
1. Probability of the system being empty=p0 4
2. Probability that the server is idle=P1 4
in the rth mode=P1 r r = 0 1 2 3
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

3. Probability that the server is working


4. Probability that the orbit is empty= 4r=0 q0 r
5. Probability that an arriving unit is immediately taken into service= P14 
 EC
6. Average system size= 4r=0 P  1 r, where P  1 r = dz
d
Pz rz=1 
 
7. Average orbit size= 4r=0 Q 1 r where Qz r =  i=0 qi rz  r = 0 1 2 3 4
i

In order to discuss a control problem associated with model 2, we compute the


following measures that are important components of the cost function that we
intend to construct.
8. Mean number of units being served in a batch,


K1 

EB = ii + K i
i=1 i=K

= K 1 + K1 K 1


1
 EC1 F 2 b1 + b1  Y  1 F 2 
2 0
=
F 2 0
+ F 2 1 + b1 EC + Y  1 1
2

by using 312.
The expression for Yz given by 33 yields
   
1 0 0 3
r r
EB = 2 0 EC 1 + b1   b1  F b1 F
2 2
F r=1
 

0  r r


3
+ EC +  b1 + F b1 + F 2 1 
r=1

Now, we compute the mean number of service completions undergone in each


mode per unit of time.
To this end, consider the processin  rn  n 1 where in denotes the orbit size
at tn + 0 and rn , the service mode at tn 0 rn = 0 1 2 3. Let
r
 l = lim Prin = l rn = r l 0 r = 0 1 2 3
n

Then,

0 
l+1
0 
l 
li+1
0
 l = 0 cj lj+1 +  i cj lij+1
j=1 i=1 j=1

r 
l+1
r
 l = F r i li+1  r = 1 3 (5.17)
i=1
104 Deepak et al.

and
 
2 
K
2 
l
2
 l =F 2
i l + K+i li 
i=1 i=1

Consider the generating function



r z =
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

r
  l zl  r = 0 1 2 3
l=0

Multiplying each of the equations in 516 by zl and adding, we get

0 z = 1 Cz 0 z 0 +  z 0 



z
1
r z = z 0 F r r z r = 1 3

z
and
 

K 

2 z =
 i + i z iK
F 2 2 z
i=1 i=K+1

Then,
9. Mean number of service completions of rth type, r = 0 1 2 3 per unit of time



= 1 r 1
 l = 1 
r

l=0

Since each service of rth type, r = 0 1 3 includes exactly one customer, the rate
at which the customers leave the system after service completion of the said type
r 1.
is 1 
However, in the case of r = 2, since each batch contains on an average EB
customers, the rate at which the customers leave the system after being included
2 1EB.
in a batch is 1 
10. The rate at which the(individual)units leave the system after service completion,
that is the throughput rate,
 

3
Rout = 1 r 1 + 
 2 1EB
r=0r =2

1

= 0 + 1 0  + F 1 + F 3 + F 2 EB 

11. Mean number of customers in the orbit at a service completion epoch,

L =  1

12. The fraction of time the server is busy serving in mode r r = 0 1 2 3, is



Pr = pi r
i=1
Retrial Queues with Orbital Search 105

= 1 1 0 F r b1  r = 1 2 3
r

and

P0 = 1 0 b1 +  1 0 b1 
0 0

5.1. A Control Problem Related to Model 2


Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

The quality of the system in operation is evaluated by means of the unit time cost
function


3
TC = hL1 + dr Pr + cRout 
r=0

where h is the cost (penalty) per customer waiting in the orbit at a service
completion epoch; dr , the cost per unit time of serving in mode r r = 0 1 2 3 and
c is the service time cost per customer per unit time.
In the cost function, we assume that the penalty for waiting, charged at service
completion epochs, is proportional to the orbit size. The multiplier 1 is used to
normalize the penalty per unit of time.

6. Numerical Results
In order to illustrate the performance of the system, we present some numerical
results. As an illustration, we assume that the batch size of the arrivals follows
geometric distribution with mean size 2 (i.e, cn = 1/2n  n 1), search time
distribution function Hi t i = 1 2 is exponential with parameter i and service time
distribution function Bi t i = 0 1 2 3 is Erlang of order ni with parameter i so
that the mean service time is ni /i  i = 0 1 2 3.
Figure 1, associated with model 1, illustrate the dependence of the average
number L of customers in the system at a departure epoch and the probability
p0 4 of the system being empty, on the values of the parameter  of the arrival
process. From Figure 1, it can be seen that L increases and p0 4 decreases as 

Figure 1. Model 1 1 = 4, 2 = 3, = 5, n0 = 3, n1 = 5, n2 = 4, n3 = 3, 0 = 15, 1 = 20,


2 = 10, 3 = 14.
106 Deepak et al.

increases, as we expected. Similarly, if we consider the variation of the parameter


2 of the type II search time distribution or that of 2 of the batch service time
distribution, we can show that L decreases and p0 4 increases as 2 or 2 increases
as the case may be.
Figure 2, associated with model 2, illustrate the dependence of L and p0 4 on
the values of K, the maximum batch size allowed in a service, corresponding to three
sets of parameters. Here, we can see that as K increases, L decreases and p0 4
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

increases as we expected.
In Figure 3, we establish the dependence of the unit time cost function TC,
where we assumed that h = 3; d0 = d1 = d3 = 15; d2 = 50; c = 10, on the various
values of K. It can be seen that the graph of unit time cost function looks like that
of a convex function so that the optimal value for K corresponding to the selected
set of parameters can be identied from the graph. From Figure 3, we are getting
the optimal values of K as 4, 5, and 5 respectively.

Figure 2. Model 2  = 1, = 5, n0 = 3, n1 = 5, n2 = 4, n3 = 3, 0 = 15, 1 = 20, 2 = 10,


3 = 14; (i) 1 = 8, 2 = 5; (ii) 1 = 6, 2 = 4; (iii) 1 = 4, 2 = 3.

Figure 3. Model 2  = 1, = 5, n0 = 3, n1 = 5, n2 = 4, n3 = 3, 0 = 15, 1 = 20, 2 = 10,


3 = 14; (i) 1 = 8, 2 = 4; (ii) 1 = 7, 2 = 3; (iii) 1 = 4, 2 = 2.
Retrial Queues with Orbital Search 107

References
1. Artalejo, J.R. 1999. Accessible bibliography on retrial queues. Math. Comput. Model.
30:16.
2. Artalejo, J.R. 1999. A classied bibliography of research on retrial queues: Progress in
1990-99. TOP 7:187211.
3. Artalejo, J.R., Joshua, V.C., and Krishnamoorthy, A. 2002. An M/G/1 retrial queue
with orbital search by the server. In Advances in Stochastic Modelling (Artalejo, J.R. and
Downloaded by [Avinashilingham Inst for Hm Sci & Higher Ed for woman] at 01:01 25 February 2013

Krishnamoorthy. A. eds.). Notable Publications Inc., New Jersey, pp. 4154.


4. Artalejo, J.R., and Gomez-Corral, A. 2008. Retrial Queueing Systems-A Computational
Approach. Springer-Verlag, Berlin.
5. Artalejo, J.R. 2010. Accessible bibliography on retrial queues: Progress in 2000-2009.
Math. Comput. Model. 51:10711081.
6. Dudin, A.N., Krishnamoorthy, A., Joshua, V.C., and Tsarenkov, G.V . 2004. Analysis
of the BMAP/G/1 retrial system with search of customers from the orbit. Eur. J. Oper.
Res. 157:169179.
7. Falin, G.I. 1990. A survey of retrial queues. Queueing Syst. 7:127167.
8. Falin, G.I., and Templeton, J.G.C. 1997. Retrial Queues. Chapman and Hall, London.
9. Kulkarni, V.G., and Liang, H.M. 1997. Retrial queues revisited. In Frontiers in
Queueing: Models and Applications in Science and Engineering. (Dshalalow, J.H. ed.).
CRC Press Inc., Boca Raton FL, pp. 1934.
10. Neuts, M.F., and Ramalhoto, M.F. 1984. A service model in which the server is
required to search for customers. J. Appl. Probab. 21:157166.
11. Pakes, A.G. 1969. Some conditions for ergodicity and recurrence of Markov chains.
Oper. Res. 17:10581061.
12. Sennott, L.I., Humblet, P.A., and Tweedie, R.L. 1983. Mean drift and the non-
ergodicity of Markov chains. Oper. Res. 31(4):783788.
13. Takagi, H. 1991. Queueing Analysis-Vacation and Priority Systems- Part 1. Elsevier,
New York.
14. Yang, T., and Templeton, J.G.C. 1987. A survey on retrial queues. Queueing Syst.
2:201233.

Anda mungkin juga menyukai