International Journal of Application or Innovation in Engineering& Management (IJAIEM)

Web Site: www.ijaiem.org Email: editor@ijaiem.org

Volume 3, Issue 4, April 2014 ISSN 2319 - 4847

Volume 3, Issue 4, April 2014 Page 406


Abstract
Queueing system considered in this paper suits for investigation of the two following situations appearing in telephone
networks. The first situation is typical for distributed telephone networks where the corresponding clients are not connected by the
direct physical channel and the preliminary establishing of a connection is required. During this phase of serving the client call,
reservation of some physical channels is performed and these channels become are unavailable for other calls. It may occur that
at the end of this phase the client who is destination of this call or its area are busy. In this case the work related to preliminary
establishing the connection is lost. The service provider does not get any profit from this call but he probably has a loss due to the
possible rejection of other calls due to the busyness of channels reserved for this call. The second variant concerns mobile
communication networks when provider promises that several first, seconds of conversation are free of charge for a client. Such a
promise attracts potential users of the network. But at the same time it creates problem to the provider. Some clients make
excessive use of possibility to have free conversation. Just before the free time expires, they stop conversation and make a new
call. So, they do not pay for their calls and probably create problems with access to the network of other clients who are ready to
pay for the calls but meet the server be busy.

Keywords: Mobile communication networks, Queueing system; Excessive use of possibility, Free conversation.

1. Introduction
The parallel systems total processing rate will fall below m when there are fewer than m jobs available. How close, then,
can we get to matching the performance of the fast processor with the corresponding set of slow processors, and how
should we schedule the parallel system to achieve its best performance? These issues are significant in the design and
operation of complex service systems, such as flexible manufacturing systems and computer communication networks
Burke (1956), Finch (1959), Papangelou (1972). In this paper we address such problems in the idealized setting of a
versatile service system model: a multiclass M/M/m queue with Bernoulli feedback. We shall thus consider the problem
of allocating dynamically m identical servers to customers in distributed M/M/m queueing network to minimize a
performance objective of expected linear holding costs, where E
u
[L
j
] represents the steady-state expected number of class j
customers in the system under policy u, and C
j
>0 their holding cost rate. Admissible scheduling policies make history-
dependent decisions, allow customer preemptions, and are nonidling (no server can lie idle when there are customers
waiting). Consider now the corresponding problem in which the m slow parallel servers are replaced by a pooled resource
consisting of one fast m-fold speed single server. While the parallel-server optimal scheduling problem is likely to be
computationally intractable, the solution for the pooled resource constitutes a classical result in the field of stochastic
scheduling: Klimov (1974, 1978) showed that the optimal policy is characterized by class-dependent priority-indices,
efficiently computed by an adaptive greedy algorithm, so it is optimal to give at each decision epoch higher service
priority to a customer with larger index. Clearly, Klimovs rule extends naturally into a simple heuristic for the parallel
server system: At each decision epoch, let servers select preemptively available customers with larger indices. The current
paper investigates the performance of this heuristic. In related work Weiss (1990, 1992, 1995) has analyzed the
performance of index-based heuristics in several models for the optimal scheduling of a batch of stochastic jobs on
parallel machines. He has argued that the index rules considered, which may be thought of as policies whose aim is to
drive down fastest the cost rate of waiting jobs, are suboptimal because of an end effect caused by the loss of processing
efficiency when the number of machines exceeds that of jobs present. He was able to bound the magnitude of this effect by
deriving and applying certain decomposition formulae for the systems total expected workload. He thus obtained
suboptimality bounds, independent of the batch size, for the index rules considered. Asymptotic optimality as the batch
size grows to infinity follows (Boxma, O. J.; 1989). Weiss further argued the importance of proceeding to analyze index
rules in more complex models incorporating job arrivals, such as queueing networks. This is the task we undertake in the
present paper.

In our analysis of the performance of Klimovs rule in the above multiclass M/M/m system we shall focus on the
following issues.
DISTRIBUTED QUEUEING SYSTEM WITH FREE
CONSERVATION: A MODEL FOR MOBILE
COMMUNICATION NETWORK

K. M. Sharma* and Yogesh Kumar Bhardwaj

B.S.A. College, Mathura (U.P.) India

International Journal of Application or Innovation in Engineering& Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 4, April 2014 ISSN 2319 - 4847

Volume 3, Issue 4, April 2014 Page 407

1. Approximate optimality. How far from the optimal cost can the expected cost under Klimovs rule be? How large can
the gap be between the expected cost achieved by Klimovs rule in the parallel and in the pooled systems? Can one obtain
simple bounds for the corresponding gaps?
2. Heavy traffic optimality. Does the relative suboptimality gap for Klimovs rule vanish in heavy traffic, as arrival rates
approach system capacity? Our findings support the claim that Klimovs rule is a good heuristic for the parallel-server
system:
We show that both its sub optimality gap and the gap between its expected cost in the parallel and in the pooled systems
are uniformly bounded with respect to (i) external arrival rates, as long as they stay within system capacity; and (ii) the
number of servers. The first such uniform boundness result implies its heavy-traffic optimality, in the following sense:
The relative sub optimality gap of Klimovs rule vanishes as external arrival rates approach system capacity. We note that
this notion of heavy-traffic optimality is not the standard one in the literature on queueing systems control (cf., Harrison
1998), where one typically considers the asymptotic behaviour of a sequence of systems appropriately scaled in time and
space. The form of heavy-traffic optimality established in this paper is technically simpler, yet we believe it has the
advantage of being intuitive.
In fact, we establish a stronger result, namely that the relative gap between the expected performance of Klimovs rule in
the parallel and in the pooled systems vanishes in heavy traffic, in the sense stated above. The fact that intelligent
dynamic scheduling of a queueing network may lead to an effective pooling of processing resources in heavy traffic has
been studied in a variety of models (see, e.g., the review paper by Kelly and Laws 1993). However, as pointed out by
Harrison (1998), studies of resource pooling have been largely heuristic to date. Harrison proves a resource pooling
result, and establishes a strong form of heavy-traffic optimality for a specific policy in the context of a model different
from the one discussed here.
The approach in this paper to a rigorous development of a resource pooling/heavy-traffic optimality result is radically
different and is based on an analysis of the systems region of achievable mean queue lengths. We believe that this
approach has the potential to be extended to more complex systems. Our mode of analysis is the so-called achievable
region approach to stochastic optimization. This approach was introduced in a seminal paper by Coffman and Mitrani
(1980) and has since been extended to ever more encompassing frameworks in Gelenbe and Mitrani (1980), Federgruen
and Groenevelt (1988), Ross and Yao (1989), Tsoucas (1991), Shanthikumar and Yao (1992), Bertsimas and Nirio-Mora
(1996), and Glazebrook, K.D. and Garbe, R. (1999). In all these analyses, the agenda outlined in Equations (1)(3) is
carried through in full. Bertsimas and Nirio-Mora (1996) use the achievable region approach to unify classical priority
index optimality results in a variety of problem domains, including deterministic machine scheduling (see Smith 1956),
multi-armed bandits (see Gittins and Jones 1974), and multiclass queueing networks (see Klimov 1974, 1978).
The paper proceeds as follows. The parallel-server system that is our prime object of study is described in 2. To assist the
reader we also give a brief account of an achievable region analysis of this system in the single-server (or pooled) case,
when Klimovs rule is optimal. In 35 we analyse the parallel server system according to the following three-step plan:

2. Mathematical Formulation of the Problem and Solution
The problem of enhancement of the system operation in these situations is very complicated. One of the possible ways for
its solving is separation of two phases of the service of a customer (channels reservation or free phase of the call is the
first phase and conversation, which is paid by a client, is the second phase) physically or virtually into separate stages
with the controlled access to the first phase depending on the situation at the second phase. To simplify capacity
planning, performance evaluation and parameters tuning for such a way of the system operation organization, the
following queueing model looks to be useful.
Tandem queueing system consists of two sequential multi-server systems. We suggest that the number of servers at the
first phase is infinite. Alternative assumption that it is finite, say V , simplifies investigation of the model. The number of
servers at the second phase is assumed to be equal to . 1 , > N N The servers at each phase are suggested to be identical.
Service time at a server at the k
th
phase is assumed to be exponentially distributed with the rate . 2 , 1 , = k
k

The customers arrive to the system according to a MAP ( ocess Arrival Markovian Pr ). The notion of the
MAPand its detailed description is given by Lucantoni (1991) were the currently standard in literature denotations for
the MAP are introduced. We denote the directing process of the MAP by . 0 , > t
t
The process , 0 , > t
t
is an
irreducible continuous time Markov chain with state space { }. ,..., 1 , 0 W The sojourn time of this chain in state is
exponentially distributed with the positive finite parameter .

When the sojourn time in the state expires, with

probability
) (
,
k
p
'
the process
t
jumps into the state ' and k customers arrive, W k , 0 , , 1 , 0 = ' = . The behavior of
the MAP is completely characterized by the matrices , 1 , 0 , = k D
k
defined by their entries
International Journal of Application or Innovation in Engineering& Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 4, April 2014 ISSN 2319 - 4847

Volume 3, Issue 4, April 2014 Page 408

1 , ) (
) 1 (
,
,
= =
'
'
k p D
k


and . , 0 , ) ( , , 0
, 0
W D k = = ' = =


The matrix
1 0
) 1 ( D D D + = represents the generator of the process 0 , > t
t
. The average arrival rate is defined by
e
1
D = where is the invariant vector of the stationary distribution of the Markov chain 0 , > t
t
. The vector is
the unique solution to the system . e 0 1 , ) 1 ( = = D Here and in the sequel e is the column-vector of appropriate size
consisting of 1s and 0 is the row-vector of appropriate size consisting of zeroes. The set of MAP s is dense in the set of
all point processes, hence, any ordinary arrival process can be approximated by a MAP . It is worth mentioning again
that the MAP in general is a correlated process. Hence, it is suitable to model the flows in modern communication
networks. The survey of research in the queues with the MAP can be found in Chakravarthy (2001).
Admission of customers to the system is controlled by means of a parameter . , 1 , N M M = If, at a customer arrival to
the first system of tandem, the number of customers at the second system does not exceed the threshold M then the
customer is admitted into the system and starts the service at the first phase. Otherwise, it leaves the system without
service (is lost). After the service at the first phase, the customer leaves the system with probability . 1 0 , < s q q With
the supplementary probability , 1 q the customers moves to the second system of tandem. If all servers of this system
are busy, the customer leaves the system (is lost). Otherwise, it is processed by a server at the second system and then
leaves the tandem.
We assume that the quality of the system operation is evaluated by means of the following cost criterion:
), 1 ( ) (
) 2 (
2
) 1 (
1
+ = M N d P c P c a M J
loss loss out
.(1)
where
out
is the rate of the flow of the customers which get successful service at the both systems of tandem, a is an
average profit obtained by the system from the service of one customer,
) (r
loss
P is probability of a loss of a customer who
wishes to get the service at the r th phase of the system,
r
c is the charge of the system when a customer is lost at the r th
phase of the system, , 2 , 1 = r d is a charge paid for reservation of one server at the second phase per time unit.
Our goal is to calculate the main performance measures of this tandem system and to show the effect of maximization of
cost criterion (1) via an appropriate choice of the threshold M .
To this end, we consider the three-dimensional continuous time Markov chain } { , 0 , , , > = t k i
t t t t
where
t
i is the
number of customers at the first phase,
t
k is the number of customers at the second phase at the moment t . Here
. , 0 , , 0 , 0 W N k i
t t t
= = > Denote ). 1 )( 1 ( , 1 + + = + = W N K W W

Theorem:
Generator A of the Markov chain , 0 , > t
t
has blocking structure . ) (
0 , , >
=
j i j i
A A Non-zero blocks
j i
A
,
of dimension
K K have the following form:
, 1 , )

) 1 ( ) ( (
1 1 1 1 1 1 ,
> + + =
+ +
i I E i q E I i E qi A
W N N N N i i

, 0 ,
1 1 ,
> =
+
i D E A
M i i

. 0 ,
~
) 1 ( ) (
1 2 1 1 0 ,
> + + =
+ +
i I E I i D E I D E A
W N K M N M i i

where
, , 0 ,
0
0
0
) 1 ( ) 1 (
) 1 (
) 1 (
N l
I
E
l N l N
l N l
l l N
l
l
=
|
|
.
|

\
|
=
+ +
+
+

1

+ N
E is the square matrix of dimension 1 + N which has entries
, , 0 , , )

(
, 1
N n n E
n n N
= '
' +

such that , 1 , 0 , 1 )

(
1 , 1
= =
+ +
N n E
n n N
, , 0 , , 1 , 0 )

(
, 1
N n n n n E
n n N
= ' + = ' =
' +

1
~
+ N
E is the square matrix of dimension 1 + N which has entries , , 0 , , )
~
(
, 1
N n n E
n n N
= '
' +
such that
, , 0 , )
~
(
, 1
N n n E
n n N
= =
+
, , 1 , )
~
(
1 , 1
N n n E
n n N
= =
+

, , 0 , , 1 , , 0 )
~
(
, 1
N n n n n n n E
n n N
= ' = ' = ' =
' +

International Journal of Application or Innovation in Engineering& Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 4, April 2014 ISSN 2319 - 4847

Volume 3, Issue 4, April 2014 Page 409

R
I is identity matrix of dimension R, is symbol of Kronecker product of matrices.
Let us denote W o N k i k k i i P k i p
t t t
t
, , , 0 , 0 }, , , { lim ) , , ( = = > = = = =

stationary probabilities of the
Markov chain . 0 , > t
t
It can be shown that, due to restricted access of customers into the system, the stationary
probabilities ) , , ( k i p exist for any choice of the system parameters. Let us enumerate probabilities ) , , ( k i p in the
lexicographic order and form row vectors
i
p of these probabilities corresponding to the value i of the first component of
the Markov chain.
It is well-known that the probability vectors
i
p satisfy the following system of linear algebraic equations:
1 ) , , , ) , , ,
2 1 2 1
0 0
= = e p p , (p 0 p p (p A .
This system is infinite, so the problem of its solving is non-trivial. Effective and numerically stable algorithm for
computing the vectors , 0 , > i
i
p of stationary probabilities for the Markov chain of the type, to which belongs the
Markov chain under consideration , 0 , > t
t
is described in Klimenok et al. (2005). So, the problem of computing these
vectors can be considered solved. Having these vectors been computed, we can calculate the main performance measures
of the tandem system including the measures involved in the cost criterion : ) (M J
- Probability
) 1 (
loss
P that an arbitrary customer will be lost at the first phase of the system is computed by

=
+
=
0
1 1
) 1 (
) (
1
i
M N i loss
D E I P e, p

- Average number
) 1 (
N of busy servers at the first phase of the system is computed by e. p
i
i
i N

=
=
0
) 1 (

- Intensity
) 1 (
out
of flow of customers, which get the service at the first phase of the system, is computed by
.
1 `
) 1 ( ) 1 (
N
out
=
- Average number
) 2 (
N of busy servers at the second phase of the system is computed by
. ) , 2 1
0
) 2 ( T
i
N , ( N e e, e, 0 p
i

=
=
- Intensity
out
of flow of customers, which get the service at the second phase of the system, is computed by
.
1 `
) 2 ( ) 2 (
N
out
=
- Probability
) 2 (
loss
P that an arbitrary customer will be lost at the second phase of the system (it leaves the system after the
service at the first phase due to the busyness of all servers at the second phase) is computed by
.
) 1 (
1
) 1 (
) 2 (
q
P
out
out
loss

4. Numerical Computation
Consider numerical example. Let the number N of servers at the second phase be equal to 50. Intensities of service are
given by . 1 . 0 , 2
2 1
= = Probability q is equal to 0.3. To illustrate the effect of maximization of the cost criterion
) (M J by means of appropriate choice of the threshold M and also to demonstrate importance of investigation of the
system with the MAP arrival process, we consider three different arrival processes. Arrival rate of all these three
processes is equal to 10. The first arrival process is assumed to be the stationary Poisson arrival process. It has correlation
between the successive inter-arrival times equal to 0 and variation equal to 1. The second process has correlation between
the successive inter-arrival times equal to 0.3868 and variation equal to 6.4788. The third process has correlation between
the successive inter-arrival times equal to 0.4627 and variation equal to 15.26. The curves 1-3 on the Fig.1 present the
dependence of cost criterion on the threshold M for these three arrival processes. Although arrival rate is the same, the
dependencies are quite different. One can see, that the optimal values of the threshold M for these curves are equal to
47, 45, and 38 correspondingly. Optimal values of the cost criterion are equal to 1045.02, 817.47, 245.24 while the values
of this criterion when no channel reservation is made at the second phase (i.e. 50 = = N M ) are equal to 714.45,
293.35, - 924.31 correspondingly. So, the optimal channel reservation gives essential profit.
International Journal of Application or Innovation in Engineering& Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 4, April 2014 ISSN 2319 - 4847

Volume 3, Issue 4, April 2014 Page 410

5. Discussion of Results and Conclusions
We have analyzed a simple heuristic index policy that extends Klimovs classical solution for the single-server case to the
general parallel-server model, presenting closed form sub optimality bounds that imply its asymptotic optimality in a
heavy-traffic limit. Ideas that emerge from our analysis include the following:
(1) Understanding of a simple single-server system has yielded useful insights into the performance of its more complex
parallel-server counterpart;
(2) Understanding the fundamental laws of a complex parallel-server model (flow balance and work decomposition) has
yielded a key to its analysis;
(3) Investigating strong linear programming relaxations of a complex stochastic optimization problem has yielded an
approximate and asymptotic analysis of a heuristic, which had resisted traditional approaches.

We believe these ideas, which guided our approach, should prove fruitful for addressing other complex stochastic
optimization problems. We refer the reader back to the discussion in the paragraph following Lemma 5 for indications of
further work to be done on the current model. In a companion paper (see Glazebrook and Nino-Mora 1999) we carry out a
corresponding analysis of priority index rules for scheduling Markovian multiclass queueing networks with multiple
service stations. Although such rules are known to perform poorly in general for the latter type of networks, in that paper
we present suboptimality bounds under appropriate light-traffic conditions. It should be remarked that while Klimovs
optimal solution for the single-server model applied to multiclass
M/G/1 networks, i.e., it was valid under general service time distributions, our analysis requires the latter to be
exponential. Extending our approach to a model with general service time distributions and nonpreemptive policies would
require the development and application of work decomposition laws for such a model. Carrying out this extension
remains a challenging problem.

REFERENCES
[1] Chakravarthy, S. (2001): The batch Markovian arrival process: a review and future work, Advances in Probability
Theory and Stochastic Processes, Notable Publications, New Jersey, pp. 2149.
[2] Klimenok, V., Kim, C.S., Orlovsky, D. and Dudin, A. (2005): Lack of invariant property of Erlang loss model in
case of the MAP input, Queueing Systems, Vol. 49, pp. 187213.
[3] Lucantoni, D. M. (1991): New results on the single server queue with a batch Markovian arrival process,
Communications in Statistics-Stochastic Models, Vol. 7, pp. 146.
[4] Bertsimas, D. and Nino-Mora, J. (1996): Conservation laws, extended polymatroids and multi-armed bandit
problems: a polyhedral approach to indexable systems, Math. Oper. Res., vol. 21, pp. 257306.
[5] Bertsimas, D. and Nino-Mora, J. (1999a): Optimization of multiclass queueing networks with changeover times via
the achievable region approach: Part I, the single-station case, Math. Oper. Res., vol. 24, pp. 306330.
[6] Bertsimas, D. and Nino-Mora, J. (1999b): Optimization of multiclass queueing networks with changeover times via
the achievable region approach: Part II, the multi-station case, Math. Oper. Res., vol. 24, pp. 331361.
[7] Boxma, O. J. (1989): Workloads and waiting times in single-server systems with multiple customer classes,
Queueing Syst., vol. 5, pp. 185214.
[8] Burke, P. J. (1956): The output of a queueing system, Oper. Res., vol. 4, pp. 699704.
[9] Coffman, E. and Mitrani, I. (1980): A characterization of waiting time performance realizable by single server
queues, Oper. Res., vol. 28, pp. 810821.
[10] Federgruen, A. and Groenevelt. H. (1988): Characterization and optimization of achievable performance in general
queueing systems, Oper. Res., vol. 36, pp. 733741.
[11] Finch, P. D. (1959): On the distribution of queue size in queueing problems, Acta Math. Hungar., vol. 10, pp. 327
336.
[12] Gelenbe, E. and Mitrani, I. (1980): Analysis and Synthesis of Computer Systems, Academic Press, London.
[13] Gittins, J.C. and Jones, D.M. (1974): A dynamic allocation index for the sequential design of experiments, Gani,
J., Sarkadi, K., and Vince, I., eds., Progress in Statistics: European Meeting of Statisticians, Budapest, 1972, North-
Holland, Amsterdam, pp. 241266.
[14] Glazebrook, K.D. and Garbe, R. (1999): Almost optimal policies for stochastic systems which almost satisfy
conservation laws, Ann. Oper. Res., vol. 92, pp. 1943.
[15] Harrison, J.M. (1998): Heavy traffic analysis of a system with parallel servers: asymptotic optimality of discrete-
review policies, Ann. Appl. Probab., vol. 8, pp. 822848.
[16] Kelly, F.P. and Laws, C.N. (1993): Dynamic routing in open queueing networks: Brownian models, cut constraints
and resource pooling, Queueing Syst. Theory Appl., vol. 13, pp. 4786.
[17] Klimov, G.P. (1974): Time sharing service systems I, Theory Probab. Appl., vol. 19, pp. 532551.
[18] Klimov, G.P. (1978): Time sharing service systems II. Theory Probab. Appl., vol. 23, pp. 314321.
International Journal of Application or Innovation in Engineering& Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 4, April 2014 ISSN 2319 - 4847

Volume 3, Issue 4, April 2014 Page 411

[19] Kumar, S. and Kumar, P.R. (1994): Performance bounds for queueing networks and scheduling policies, IEEE
Trans. Autom. Control, vol. 39, pp. 16001611.
[20] Papangelou, F. (1972): Integrability of expected increments of point processes and a related random change of time
scale, Trans. Amer. Math. Soc., vol. 165, pp. 483506.
[21] Ross, K.W. and Yao, D.D. (1989): Optimal dynamic scheduling in Jackson networks, IEEE Trans. Aut. Control,
vol. 34, pp. 4753.
[22] Shanthikumar, J.G. and Yao. D.D. (1992): Multi-class queueing systems: polyhedral structure and optimal
scheduling control, Oper. Res., vol. 40, pp. 293299.
[23] Smith, W.E. (1956): Various optimizers for single stage production, Naval Res. Logist. Quart., vol. 3, pp. 5966.
[24] Tsoucas, P. (1991): The region of achievable performance in a model of Klimov, Technical Report RC16543, IBM
T.J. Watson Research Center, Yorktown Heights, NY.
[25] Weiss, G. (1990): Approximation results in parallel machines stochastic scheduling Special Volume on Production
Planning and Scheduling M. Queyranne (ed.), Ann. Oper. Res., vol. 26, pp. 195242.
[26] Weiss, G. (1992): Turnpike optimality of Smiths rule in parallel machines stochastic scheduling, Math. Oper.
Res., vol. 17, pp. 255270.
[27] Weiss, G. (1995): On almost optimal priority rules for preemptive scheduling of stochastic jobs on parallel
machines, Adv. Appl. Probab., vol. 27, pp. 821839.




Figure 1: Dependence of cost criterion on the threshold M