You are on page 1of 26

AN M/G/1 RETRIAL QUEUE WITH UNRELIABLE SERVER FOR

STREAMING MULTIMEDIA APPLICATIONS

Nathan P. Sherman
Department of Operational Sciences
Air Force Institute of Technology
2950 Hobson Way (AFIT/ENS)
Wright Patterson AFB, OH 45433-7765, USA
Email: Nathan.Sherman@pentagon.af.mil

Jeffrey P. Kharoufeh1
Department of Mechanical & Industrial Engineering
Northeastern University
334 Snell Engineering Center
Boston, MA 02115-5000, USA
Email: jkharoufeh@coe.neu.edu

Mark A. Abramson
Department of Mathematics and Statistics
Air Force Institute of Technology
2950 Hobson Way (AFIT/ENC)
Wright Patterson AFB, OH 45433-7765, USA
Email: Mark.Abramson@afit.edu

Final version, June 2007

Abstract
As a model for streaming multimedia applications, we study an unreliable retrial
queue with infinite-capacity orbit and normal queue for which the retrial rate and
the server repair rate are controllable. Customers join the retrial orbit if and only if
their service is interrupted by a server failure. Interrupted customers do not rejoin the
normal queue but repeatedly attempt to access the server at i.i.d. intervals until it is
found functioning and idle. We provide stability conditions, queue length distributions,
stochastic decomposition results, and performance measures. The joint optimization
of the retrial and server repair rates is also studied.

Keywords: Retrial queue, unreliable, streaming multimedia.

1
Corresponding author. Phone: (937) 479-9095.

1 Introduction

In this paper, we analyze an M/G/1 retrial queue with an unreliable server whose
orbit and normal queue have infinite storage capacity and whose retrial and server repair

rates are controllable. Customers in service join the retrial orbit if and only if they are
interrupted by a server breakdown and do not rejoin the normal queue, but rather attempt

to access the server directly at random intervals independently of arrivals or other retrial

customers. However, these interrupted customers can regain access to the server only when it

is operational and idle and repeat service until they have been completed. Arriving customers

who find a failed server join the normal queue. We allow for both active breakdowns which

occur during a service cycle, and idle breakdowns which occur while the server is not failed
but idle. The server may not breakdown while under repair. The times between customer

arrivals, breakdowns, retrials, and repairs are assumed to be exponentially distributed while
the service times are general.

Over the past two decades, advances in telecommunications and computer network-
ing technologies have reinvigorated the study of queueing systems in general and retrial

queueing systems in particular. The model we present here is well-suited to model computer

network streaming multimedia applications. The primary (or normal) queue is similar to a
1-persistent carrier-sense multiple-access (CSMA) system. When the oldest packet in the

normal queue detects that the transmission medium (or server) is free, transmission begins

immediately. If the communication medium fails during transmission, the packet is sent to

a retrial queue which is analogous to a non-persistent CSMA system. If the medium is un-

available (i.e., busy or failed), then the retrial packet waits a random amount of time before

checking the status of the medium again. This process repeats until the retrial packet finds

the transmission medium operational and idle. In this sense, the model is a priority queue

wherein the packets that are not interrupted by a transmission failure have non-preemptive
priority over those awaiting availability of the transmission medium. An important applica-

tion is that of streaming voice or video wherein transmitted packets are used for playback

1

[20]. if they are not received within a given time threshold. [15]. By developing analytical expressions for congestion and delay mea- sures in stable systems. an interesting stability result emerges.upon reception and also stored for future use. In addition to its practical relevance. the authors did not consider an unreliable server. For retrial models with no waiting room and server breakdowns. we illustrate the means by which to simultaneously select the optimal retrial rate and repair capacity to minimize a long-run average cost criterion. [7]. Packets that are interrupted during transmission can still be used for later playback from the stored copy of the stream. it is possible to simultaneously select a packet retrial rate and server repair rate that minimize the long-run average cost of holding customers in either queue. but their transmission time is no longer important. [27]) force these customers into the orbit while others ([4]. [3]. Although [10] considered an M/G/1 retrial queue with infinite- capacity orbit and normal queue. The seminal papers in this area are [1] and [15]. [6]. they are effectively useless. The pres- ence of an infinite waiting space for primary customers introduces an interaction between the two infinite queues. [23]. We will show that the steady state orbit size and the overall sys- tem size both possess a stochastic decomposition property. These packets correspond to the priority customers. Some models (cf. customers arriving to find the server unavailable (busy or failed) join the orbit. This dynamic does not exist in the vast majority of retrial queueing models which include only an infinite retrial orbit and do not consider an infinite waiting space for primary arrivals. All models considering retrial queues with server breakdowns assume an M/G/1/1 loss system with the exception of [6] and [21]. The literature addressing retrial queues with unreliable servers is relatively sparse but growing at a rapid pace. [16]. namely that the normal queue may remain stable even if the condition for system stability is violated. The packets used for immediate playback are time-sensitive in that. [2]. These packets constitute customers in the orbit. Moreover. the model we present also exhibits extremely interesting mathematical properties that warrant investigation in their own right. [11]. 2 . [5]. [25]. Using the steady state distributions and corresponding queueing performance measures.

[22]) and selection of the optimal retrial rate ([8]. researchers such as [17]. [15]). an optimal T -policy and he computes the optimal retrial rate that minimizes costs using an N -policy. [27]). both active and idle breakdowns ([2]. With the exception of two cases ([3] and [25]). For general queueing systems (non-retrial queues). In particular. [5]. Our model differs from these in that arriving customers who find the server busy or failed join the normal queue whereas interrupted customers always join the orbit and attempt to re-access the server at random intervals. The current literature addressing the optimal design or control of unreliable retrial queues is very sparse.[26]) provide the option of joining the orbit or departing the system. For retrial systems with no breakdowns and zero capacity in the normal queue. An informal. and [24] have considered optimal N -policies wherein the server remains idle until exactly N (N ≥ 1) customers are present in the queue. In that work. [10]. [16]. [19]. these models also either force. [25]. in- service customers interrupted by a server failure to join the orbit. [11]. we consider the problem of simultaneously selecting an optimal retrial rate and optimal repair rate with the objective of minimizing the long-run average operating 3 . graphical approach to the optimal control and design of a retrial queue with vacations was presented in [20] wherein the authors examined the impact of the retrial rate. this paper is concerned with the analysis and control of an M/G/1 retrial queue with an infinite-capacity orbit and normal queue. the arrival rate. [23]). [3]. and [26]) consider orbits as FCFS queues. As a model for streaming multimedia applications. and like our model. [13]. the number of input sources. vacations ([7]. [14]). or provide the option for. active breakdowns ([6]. It appears that only [7] formally addressed these issues for a retrial queue with vacations. however some models (cf. Most orbits are assumed to behave as infinite-server queues with identical exponential service times. the most common optimal control strategies include the optimal routing of arriving customers ([9]. [20]. the author presents an optimal N -policy. [4]. and the service rate on the mean waiting time and throughput. [7]. A variety of failure types are considered in the literature including starting failures ([16]. [26]).

Service times form an independent and identically distributed (i. or busy (operational and occupied). using the method of supplementary variables and a classical generating function approach.i. we show that the steady state length of the retrial queue and the system size can be stochastically decomposed.f. derive the queue length distributions. 1 − B(x) For s ≥ 0. section 5 presents and illustrates a nonlinear optimization problem for the optimal selection of the retrial and server repair rates.cost which is a function of the key queueing performance measures.f. by means of generating functions. normal queue and in service). Using these results. Finally. To this end. probability density function (p.) B. 2 Model Description Customers arrive to the system according to a homogeneous Poisson process with rate λ > 0.) b. independent of the server’s status. We provide a necessary and sufficient condition for stability of the orbit and system as well as a (distinct) condition for stability of the normal queue. key queueing perfor- mance measures. x ≥ 0. The remainder of the paper is organized as follows. we obtain the joint generating function of the orbit size and normal queue size as well as the generating function for the overall system size (the total number of customers in orbit. Moreover. as well as the steady state distribution of the server’s status. we derive the steady state joint distribution of the orbit size and normal queue size when the server is idle (operational and not occupied). let Z ∞ ∗ b (s) = e−sx b(x)dx 0 4 . Section 2 provides the model description and mathematical notation. In section 3 we establish stability conditions and. In section 4 we present stochastic decomposability results for queue length distributions.d.) sequence of random variables with absolutely continuous distribution function (d.d. failed (non-operational and being repaired). and service completion rate b(x) µ(x) = .

j. The arrival. Ut = 0. Ut = 0. 0.. service.1. k ≥ 0 t→∞ πk. Let Nt denote the total number of customers in the system at time t (i. Denote by Qt the number of customers in the normal queue at time t.0 = lim P (Qt = k.0. or busy state. and let Rt denote the number of customers in the orbit at time t. failed. St = 1).1 (x) = lim P (Qt = k. Now define π0. j. An in-service customer interrupted by a server failure enters the orbit and spends an exponential amount of time there with rate θ > 0. Xt < x). if the server is not occupied at time t while St describes the operational status of the server at time t defined by ½ 1. St = 1. Qt ⇒ Q. if the server is not failed at time t St = . repair. St = 0). excluding any customer that might be in service. after which it either enters service (if possible) or remains in the orbit for an additional exponentially distributed time with rate θ. and in service).0.denote the Laplace transform of b. The repair time is exponentially distributed with rate parameter α > 0. j. St ⇒ S and Nt ⇒ N where “⇒” denotes weak convergence. if the server is occupied at time t Ut = 0. Rt = j. Rt ⇒ R. k ≥ 0 t→∞ as the limiting probabilities that the system is in an idle. Rt = j. normal queue. failure. and retrial processes are assumed to be mutually independent.j. Server failures occur according to a Poisson process with rate ξ > 0 when the server is not being repaired.j. Rt = j. Ut = 1. Ut . Assume that as t → ∞. Xt ) : t ≥ 0} describes the state of the system. Rt .1 = lim P (Qt = 0. The random variable Ut is the occupation status of the server given by ½ 1. in orbit. if the server is failed at time t Let Xt denote the elapsed service time of the customer in service at time t so that the continuous-time stochastic process. j ≥ 0 t→∞ πk. St . {(Qt . With the transform variables z1 and z2 corresponding to the orbit size and normal queue 5 .e. respectively.

1 (z1 .size. 0 Let p denote the joint probability mass function (p.0. z1 ) with respect to the normal queue size. j=0 These are.1.f. z1 .0 (z1 ) and φk.) of R and Q while q denotes the p. respectively. define ∞ X φ0. z2 ) + ψ1.0 (z1 . z1 ) = z1 πk. of N .1 (z1 ) = z1j π0.j.1 (x.1 (z1 ) + ψ0.1 (x. In a similar manner. and πk.j.1 (x. πk.1 (x. z). The joint p. respectively.1.1 . z2 ) = z2k φk.0 (z. ∞ X ψ0. z2 ).1.1 (z.0.0. the generating functions for π0. of the orbit and normal queue size when the server is not failed and busy. z2 )dx.m.1 . j=0 X∞ φk.0.1.1 (x) with respect to the orbit size. Further define. z2 ) = p(j. we denote by ∞ X H(z) = q(j)z j = φ0.0 (z1 ).0.1 (z) + ψ0. In the next section.0 (z1 ) = z1j πk. j=0 the generating function for the overall system size. z) + zψ1. k)z1j z2k = φ0. j=0 ∞ X j φk.j.1 (x.0 (z1 .j.j.0 . is given by Z ∞ ψ1.0.f. we denote by ∞ X X ∞ G(z1 .m.0. k=0 j=0 the joint generating function for the orbit and normal queue size. z2 ) = ψ1.f. z2 ) = z2k φk.g. we provide stability conditions and formally derive the generating functions defined in this section. z1 ).1.1 (x). Subsequently. By summing over the three distinct and exhaustive server states.0.j. k=0 X∞ ψ1.1 (z1 . z1 .0. 6 . k=0 the generating functions for φk. we use these to characterize queue length distributions and performance measures.0.0 .

z2 ) = b∗ (ξ + λ(1 − z2 )) + . Let Nr and Nq respectively denote the number of customers entering the orbit and normal queue during (0. T ]. j ≥ 0. independent of the server’s status. (α + λ(1 − z2 ))(ξ + λ(1 − z2 )) Now let us define the quantity d ¯ λ(1 − b∗ (ξ))(α + ξ) ¯ ρ1 = − Q(1.d. As in Aissani and Artalejo [4]. and the distribution of the system size. define the fundamental server period as the time from which a service cycle begins until the next time at which the server is able to initiate a new service cycle. |z2 | ≤ 1. we obtain the joint distribution of the orbit size and normal queue size.f. j) = P (Nr = i. we first provide a lemma that is needed to char- acterize the stability conditions and steady state distributions. 1 − ε)¯ = . Lemma 1 For either |z1 | < 1 or |z1 | ≤ 1 and ρ1 > 1 the relation z2 − Q(z1 . Subsequently. respectively. and let a(i. dε ε=0 αξ where b∗ (ξ) is the Laplace transform of the service time p. Before proceeding to the main result. Using these definitions we have the following important result which is needed to obtain our main results. Denote this random duration by T . i=0 j=0 Then one can verify (see [4]) that αz1 ξ(1 − b∗ (ξ + λ(1 − z2 )) Q(z1 . z2 ) = a(i. Define the generating function ∞ X X ∞ Q(z1 . failed. we provide a necessary and sufficient condition for stability of the overall queueing system and derive the steady state joint distribution of the orbit and normal queue size when the server is idle. Nq = j). evaluated at ξ. we obtain standard queueing performance measures as well as the limiting distribution of the server’s status. or busy. i. j)z1i z2j . z2 ) 7 .3 Stability Analysis and Steady State Equations In this section. |z1 | ≤ 1. Additionally.

1) = 1. Theorem 1 states that the long-run proportion of time the server is available for serving customers must exceed the long-run proportion of time the server is busy if the system is to remain stable. inside the region |z2 | < 1. consider the quantity ρ1 and the case when z1 = 1. z2 ). z2 ) are respectively given by ( Z α ) αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ) 1 1 λ(1 − g(u)) + ξ(1 − α+λ(1−g(u)) ) φ0. if z1 = 1 and ρ1 > 1.1 (z1 . Thus. ψ0. for each z2 inside the unit disk |z2 | < 1. g(1) is the smallest positive real zero with g(1) < 1 if ρ1 > 1. Proof. as a function of z2 . applying Rouche’s theorem to the function z2 and the generating function Q(z1 . p. and g(1) = 1 if ρ1 ≤ 1. On the other hand. the generating functions φ0. In case z1 = 1.1 (z1 . if ρ1 ≤ 1. z2 ). g(z1 ). (1) αb∗ (ξ)ξ In such case. Now for the first part. ξb∗ (ξ)(α + ξ) θ z1 g(u) − u (2)  n o ˆ 2 ))[α + λ(1 − z2 )] − z1 (1 − B(z  (g(z1 ) − z1 )[α + λ(1 − g(z1 ))] (z2 − B(z ˆ 2 ))[ξ + λ(1 − z2 )]) ψ0. g(1) = 1 is the unique zero. g(z1 ). The function Q(z1 .1 (z1 ). and ψ1. z2 ).0. 0) < 1 and Q(1. 351-352]. as well as the generating functions φ0. we conclude that there is one and only one zero. and this completes the proof.0. z2 ) is monotonically increasing in z2 for z2 ∈ [0.0 (z1 .0. z2 ) =  (z2 − B(zˆ 2 )) [α + λ(1 − z2 )] [ξ + λ(1 − z2 )] − αξ(1 − B(z ˆ 2 ))z1 ) ˆ 2 ))(z2 − z1 )(1 − g(z1 ))[α + ξ + λ(1 − g(z1 ))] λz1 (1 − B(z ξφ0.has. Theorem 1 The queueing system is stable if and only if ρ < 1 where λ(1 − b∗ (ξ))(α + ξ) ρ= . on |z1 | < 1. 1] such that 0 < Q(1. z2 ). The proof is similar to that of Theorem 3 in [18. one and only one zero.1 (z1 ) = exp − du . positive real zero with g(1) < 1.1 (z1 ). then g(1) is the minimal.0 (z1 . Using Lemma 1. and ψ1. For the second part. we now characterize the stability condition for the overall system size (and orbit size). ψ0.0.1 (z1 ) + ˆ ˆ (z2 − B(z2 )) [α + λ(1 − z2 )] [ξ + λ(1 − z2 )] − αξ(1 − B(z2 ))z1 (g(z 1 ) − z1 )[α + λ(1 − g(z1 ))] (3) 8 .0 (z1 .

(5) through (11) by z1j and summing over all j. z1 )dx (13) Z ∞ 0 d (λ + ξ)φ0.1 + ξ π0.j.j.and ( (z2 − z1 )(1 − g(z1 ))[α + ξ + λ(1 − g(z1 ))][α + λ(1 − z2 )] ψ1.1 (x)dx (5) 0 Z ∞ (α + λ)πk.0.j. we obtain.j.1.1 (x.0 = ξπ0.j.0.1 (x.0.1 = απ0.j−1.j.0 + λπ0.1 (x.1 (x) (8) dx d πk.j.1 (x.1 (z1 ) + θz1 φ0.1.1 (x) (9) dx with boundary conditions Z ∞ π0.1 (x. the following equations: Z ∞ (α + λ)φ0.j.1.0 = λπk−1. z1 ) + λφk−1.j.1 (10) 0 Z ∞ πk.0 (z1 ) + µ(x)φ0.1 (x)dx. z2 ) = ˆ 2 )) [α + λ(1 − z2 )] [ξ + λ(1 − z2 )] − αξ(1 − B(z (z2 − B(z ˆ 2 ))z1 ) (1 − z2 )(g(z1 ) − z1 )[α + λ(1 − g(z1 ))][α + ξ + λ(1 − z2 )] ˆ 2 ))φ0.0.1 (x) = −[µ(x) + λ + ξ]π0.1 (z1 ) λ(1 − B(z − (4) ˆ ˆ (z2 − B(z2 )) [α + λ(1 − z2 )] [ξ + λ(1 − z2 )] − αξ(1 − B(z2 ))z1 (g(z1 ) − z1 )[α + λ(1 − g(z1 ))] where ˆ 2 ) = b∗ (ξ + λ(1 − z2 )) B(z and g(z1 ) verifies αξz1 [1 − b∗ (ξ + λ(1 − g(z1 )))] g(z1 ) = b∗ (ξ + λ(1 − g(z1 ))) + .1. z1 ) = −[µ(x) + λ + ξ]φ0.j.1.0.1.0 (z1 ) + ξz1 φk.0 + ξ πk.0.1 (x)dx (7) 0 d π0.0.1. z1 ) (16) ∂x 9 .0.0.1 (x)dx + (j + 1)θπ0.0.1.j+1.j.1. the balance equations are Z ∞ (α + λ)π0.l (x) = 0. For j ≥ 0 and k ≥ 1 with πk.j.1 (0) = απ1.1.1. z1 )dx (14) dz1 0 ∂ φ0.1 + µ(x)π1.j.0 + µ(x)π0.1 (x.0.j.0.j.0.i.1.1.j.1 (z1 ) = αφ0. (11) 0 Multiplying both sides of Eqs.0 (z1 ) = ξφ0.1.1 (z1 ) + ξz1 φ0.0. z1 ) (15) ∂x ∂ φk.1.j.0.0.1. respectively.1 (0) = απk+1.j−1. [α + λ(1 − g(z1 ))][ξ + λ(1 − g(z1 ))] Proof.0.−1.1.0 + µ(x)πk+1.1 (x)dx (6) 0 Z ∞ (λ + ξ + jθ)π0.1. z1 ) = −[µ(x) + λ + ξ]φk.0.0 (z1 ) = λφk−1.1 (x) + λπk−1.1.1 (x. z1 )dx (12) 0 Z ∞ (α + λ)φk.1.j.j.1 (x) = −[µ(x) + λ + ξ]πk.1 (x.1 (z1 .

with boundary conditions Z ∞ φ0.1 (x. (19) 0 Performing a similar operation on Eqs. we have from (23) · µ ¶¸ d α θ(z1 − g(z1 )) φ0.1 (x.0 (z1 )] dz1 z Z 2 1 ∞ + µ(x)[ψ1.1 (z1 ) = 0.1 (x. and (22) in (21) we obtain θ(z2 − z1 ) dzd1 φ0. (18) 0 Multiplying both sides of Eq.1 (z1 ) (17) 0 Z ∞ φk.1 (z1 ) + ξz1 ψ1. ∂ ψ1. we obtain Z ∞ [α + λ(1 − z2 )]ψ0.1 (z1 ) + θ φ0.1 (0.1 (z1 ) + [ψ0.1 (0.1 (x.0. z2 )e−[ξ+λ(1−z2 )]x (1 − B(x)). z1 .0. (20). for any z1 in the unit disk.0 (z1 ) + µ(x)φk+1. z1 ) = αφ1.0. (14). we obtain ψ1.0 (z1 . z1 )dx + θφ0.1 (x.0 (z1 . z1 .1. (12) by z20 and Eq.1 (z1 ) ψ1. z2 ) = h ∗ i . z2 ). (21) z2 0 Solving Eq.1 (x.1 (x. z1 .1 (0.0. z2 ) = λφ0. z2 ) = −[µ(x) + ξ + λ(1 − z2 )]ψ1.0. (23) αz1 ξ(1−b (ξ+λ(1−z2 )) z2 − b∗ (ξ + λ(1 − z2 )) + (α+λ(1−z 2 ))(ξ+λ(1−z2 )) Now by Lemma 1.0. z2 ) − φ0. z1 . (20) ∂x and d α ψ1.0. z1 )dx. z1 .1 (z1 ) + λ(1 − g(z1 )) + ξ 1 − φ0. (23) has. z1 .1 (0. (22) Using Eqs.1.0 (z1 ) + λφ0.1 (0.1 (z1 ) − [λ(1 − z2 ) + ξ(1 − α+λ(1−z α 2) )]φ0.0. (19).1 (x. we obtain. z2 )dx.0.1.1 (z1 ) + µ(x)φ1. z2 ) = ψ1. (15) and (16) as well as (17) and (18). This must also be a zero for the numerator. (24) dz1 α + λ(1 − g(z1 )) 10 .0. z2 ) − φ0.0. z2 ) = ξφ0. z1 ) = αφk+1. a zero in the region |z2 | < 1. z1 . z1 . therefore. the denominator of Eq.1.0. z1 )]dx.1. respectively. (13) by z2k and summing over all k ≥ 0.

we apply the normalization condition. In particular.0.1 (z1 ) in Eq. g(z1 )) has one and only 11 . assume that ρ > 1 and the system is stable. |z2 | ≤ 1. Note that ¯ ¯ 1 d ¯ d ¯ αξ(1 − b∗ (ξ)) b∗ (ξ) −1 ¯ g(z1 )¯ = Q(z1 .0. λξ(1 − b∗ (ξ))(α + ξ) ξρ lim h(z1 ) = − ∗ ∗ =− < ∞.1 (1) + ψ0.0. φ0. which yields αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ) K= = φ0. ξb∗ (ξ)(α + ξ) the steady state probability that the server is not failed and idle. for ρ ≤ 1. 1 dz1 z1 =1 dz1 z1 =1 αξ − λ(1 − b∗ (ξ))(α + ξ) b∗ (ξ) −ρ where ρ is defined in Eq.1 (1. z1 − g(z1 ) and note that. g(z1 ))¯¯ = = .0. We then observe that. 1) = 1. for ρ < 1. (25) makes all generating functions analytic in |z1 | ≤ 1. z1 →1 αξb (ξ) − λ(1 − b (ξ))(α + ξ) 1−ρ Thus. (2). and h(z1 ) can be defined at the point z1 = 1. call it β. we conclude that h(z1 ) is analytic on the open disk |z1 | < 1. φ0. and (4) are obtained up to the multiplicative constant K. when ρ > 1. To obtain this constant.0. while for ρ > 1 this quantity has one and only one zero. To see that ρ < 1 is also necessary for system stability. For ρ < 1.1 (z1 ) = 0. (1). By Lemma 1. (3). 1) + ψ1. such that β ∈ (0. dz1 is verified by the function  ³ ´   1 Z 1 λ(1 − g(u)) + ξ 1 − α α+λ(1−g(u))  φ0.In order to solve the differential equation (24).1 (z1 ) = K exp − du (25)  θ z1 g(u) − u  where K is a constant of integration. define the function ³ ´ α λ(1 − g(z1 )) + ξ 1 − α+λ(1−g(z1 )) h(z1 ) = . d φ0. the quantity k(z1 ) never becomes zero in |z1 | < 1.1 (1). the differential equation. 1) (see for example [12] or [18]). (24). on the closed disk |z1 | ≤ 1. g(z1 )). Now rearranging Eq.0 (1.1 (z1 ) + h(z1 )φ0. Therefore. we first examine the function k(z1 ) = z1 − g(z1 ) = z1 − Q(z1 . the function z1 − Q(z1 .0.

0.1 (β) = 0 since the coefficient of φ0.0.1 (1) = 0 since dzd1 g(z1 )¯z1 =1 = 1 when ρ = 1.0. j ≥ 0. however. independent of server status.1 (1) = 0. g(z1 )) never becomes zero in the unit disk |z1 | < 1.1 . Therefore.one zero. suppose that ρ = 1 and the system is stable.0. Corollary 1 For ρ < 1. z1 →1 1−ρ Differentiating Eq. we obtain the relation µ ¶ ¯ α+ξ d ¯ −λ g(z1 )¯¯ φ0. to satisfy the above relation. z2 ) = n o  [g(z ) − z ][α + λ(1 − g(z ))] (z − B(z ˆ 2 ))[α + λ(1 − z2 )][ξ + λ(1 − z2 )] − αξz1 (1 − B(z ˆ 2 )) 1 1 1 2  ˆ λ(1 − B(z2 ))[α + ξz1 + λ(1 − z2 )](1 − z2 )[α + ξ + λ(1 − z2 )] [α + ξ + λ(1 − z2 )]  − n o+ ˆ ))[α + λ(1 − z )][ξ + λ(1 − z )] − αξz (1 − B(z [α + λ(1 − z )] (z − B(z ˆ )) [α + λ(1 − z2 )]  2 2 2 2 2 1 2 × φ0. When ρ = 1.0.0. Finally. α dz1 z1 =1 ¯ implying that φ0.1 (β) = 0.j.1 (β) does not equal zero. We must therefore conclude that ∞ X φ0. α + λ(1 − β) which implies that φ0. it is not possible to find positive values π0.1 = 0. as well as the distribution of the overall system size. Using Theorem 1.1 (z1 ) (26) 12 .0. a contradiction.0.j. (24) yields µ ¶ α λ(1 − β)ξ 1 − φ0. in this case ξρ lim h(z1 ) = − = −∞. we next obtain the joint distribution of the orbit and normal queue size. the function z1 − Q(z1 . denoted by β. j=0 However. so the system is not stable.1 (β) = β j π0. the function h(z1 ) is not analytic in |z1 | < 1. z2 ) and H(z) are given by   ˆ 2 ))[α + ξz1 + λ(1 − z2 )](z2 − z1 )(1 − g(z1 ))[α + ξ + λ(1 − g(z1 ))] λ(1 − B(z G(z1 . Substituting z1 = β in Eq. the probability generating functions G(z1 . this contradicts the hypothesis that the system is stable. we conclude that the system cannot be stable unless ρ < 1.0. Thus. with 0 < β < 1. However. thus. (24) and evaluating at the point z1 = 1.

and N . denote by WR . λ2 E(S 2 ) E(Q) = . as well as their probability distributions. Furthermore. Corollary 2 The steady state mean orbit size.1 (z).d. normal 13 . and mean number in system are respectively given by " # ρ α ∗ ∗ ∗ ˆ 0 ξb (ξ)[ξ − λ(1 − b (ξ))] + (α + ξ)[λ(1 − b (ξ)) − ξ B ] ξ E(R) = · + (28) 1−ρ α+ξ b∗ (ξ)[αξ − λ(1 − b∗ (ξ))(α + ξ)] θ ξ 3 b∗ (ξ) − (α + ξ)2 [ξ Bˆ 0 − λ(1 − b∗ (ξ))] E(Q) = λ .and n o ˆ [α + ξ + λ(1 − z)] αξ B(z) ˆ − λB(z)(1 − z)[α + ξ + λ(1 − z)] H(z) = n o φ0. respectively. mean normal queue size. (26) and (27) can be used to obtain the mth moment (m ≥ 1) of R. (28) to (30). the time spent in the orbit. B. Q. (29) ξb∗ (ξ)(α + ξ) [αξ − λ(1 − b∗ (ξ))(α + ξ)] and ˆ0 λb∗ (ξ) {ξ 3 + (1 − b∗ (ξ))[αξ(α + 2ξ) + λ(α + ξ)2 ]} − λξ(α + ξ)2 B ξρ E(N ) = ∗ ∗ ∗ + . WQ and W . (30) ξb (ξ)(α + ξ)[αξb (ξ) − λ(1 − b (ξ))(α + ξ)] θ(1 − ρ) where ¯ Z ∞ d ˆ ¯ ˆ 0 B = ¯ B(z2 )¯ =λ xe−ξx b(x)dx. The first moments are provided in the following corollary. it can be shown that E(R) = 0. (27) ˆ ˆ [α + λ(1 − z)] αξ B(z) − λ(z − B(z))[α + ξ + λ(1 − z)] Using standard methods. Eqs. 2(1 − λE(S)) These expressions are consistent with results for the standard M/G/1 queue with no failures and no retrials.f. dz2 z2 =1 0 Let S denote the duration of an arbitrary service time with c.0. 2(1 − λE(S)) and λ2 E(S 2 ) E(N ) = λE(S) + . Then as ξ → 0 in Eqs.

we characterize the steady state distribution of the server’s status by directly applying the results of Theorem 1. and thus. The expected values of these random variables are obtained by applying Little’s Law to Eqs. normal queue customers experience a greater effective service rate than do retrial customers. Hence.1 (z1 ) = − . we demonstrate that both the orbit and system size exhibit a stochastic decomposition property which has been observed for the system size distribution of many 14 . and E(W ) = λ−1 E(N ). Let pI . z2 ) = . z1 →1 α+ξ z2 →1 and λ(1 − b∗ (ξ)) pB = lim ψ1.queue and system by an arbitrary customer in the long-run. and pB respectively denote the limiting probability that the server is idle.. pF . the steady state distribution of the server’s status is given by α λ(1 − b∗ (ξ)) pI = lim φ0. i. E(WR ) = λ−1 E(R). z2 ) = .0. Owing to the nature of the orbit dynamics. retrial customers are subordinate to normal customers and may be served only when the server is idle and operational. failed. E(WQ ) = λ−1 E(Q). 4 Stochastic Decomposition In this section. That is. (28) to (30). the normal queue can be stable even if the orbit stability condition is violated. z1 →1 α+ξ ξb∗ (ξ) ξ pF = lim ψ0. or busy. Finally. Corollary 3 For ρ < 1. and by (29) we see that ρ1 < 1 is necessary for the stability of Q. respectively. Corollary 2 also confirms that ρ < 1 is necessary for the stability of R (and N ). it is possible that the orbit may continue to grow while the normal queue remains stable. z1 →1 ξb∗ (ξ) z2 →1 In the following section we show that the orbit and system size can be stochastically de- composed before considering the optimal selection of the retrial and repair rates in section 5.0 (z1 .e.1 (z1 .

[4.e.. Therefore. (31) θ z1 g(u) − u The following two propositions describe the decomposability of the orbit and system size distributions. ˆ + V. (3). retrial queues.M/G/1 models including those with vacations. let V be a random variable whose generating function is given by ( Z 1 λ(1 − g(u)) + ξ(1 − α ) 1 α+λ(1−g(u)) ) E(z V ) = exp − du . using L’Hospital’s rule.e. R=R (32) Proof. and (27) depend on the retrial rate θ only through the generating function E(z1V ). including the decomposability of the vector of server status and the orbit size. Let R the steady state orbit size in the instantaneous feedback model. i. 1) = AG (z1 ) θ→∞ where. 11. it can be shown that · ¸ λ(α + ξz1 )(1 − g(z1 ))[α + ξ + λ(1 − g(z1 ))] AG (z1 ) = (1 − ρ) 1 + .. in the standard M/G/1 retrial queue (i. ξ(α + ξ)(g(z1 ) − z1 )[α + λ(1 − g(z1 ))] 15 . 27]).. 16. Note that Eqs. one of which is the steady state orbit size in the instantaneous feedback model and the other is V . Let N function of R ˆ denote the steady state total number of customers in the system in the instantaneous feedback model and denote its generating function by ˆ E(z N ) for |z| ≤ 1. Falin and Templeton [12] provide several stochastic decomposition results. and denote the generating ˆ by E(z Rˆ ) for |z| ≤ 1. (4). one with no infinite waiting space and no server breakdowns). (2). (26). Finally. an instantaneous feedback model). Allowing θ → ∞ in our model yields a model in which retrial customers instanta- ˆ denote neously attempt to re-access the server (i. we may write the generating function for ˆ as R ˆ E(z1R ) = lim G(z1 . |z| ≤ 1. Proposition 1 The random variable R may be expressed as the sum of two independent random variables. and breakdowns (cf.e.

. one of which is the steady state system size in the instantaneous feedback model and the other is V . The proof is analogous to that of Proposition 1. ( Z 1 λ(1 − g(u)) + ξ(1 − α ) 1 α+λ(1−g(u)) ) E(z N ) = AH (z) exp − du θ z1 g(u) − u ˆ = E(z N )E(z V ) ˆ = E(z N +V ).e. Note that by setting z1 = z2 = z in Eqs. 16 . we may write ( Z 1 λ(1 − g(u)) + ξ(1 − α ) 1 α+λ(1−g(u)) ) E(z1R ) = AG (z1 ) exp − du θ z1 g(u) − u ˆ = E(z1R )E(z1V ) ˆ = E(z1R+V ).0 (z. ˆ (α + ξ)[α + λ(1 − z)] αξ B(z) ˆ − λ(z − B(z))[α + ξ + λ(1 − z)] The generating function for N is the product of these two. 1)..0. Proposition 2 The random variable N may be expressed as the sum of two independent random variables. (3).1 (z.e. Similar behavior may be observed for the steady state system size as noted in Proposition 2. and (4) we obtain E(z N ) ≡ H(z) = φ0. (2). ˆ is given by The generating function of N ˆ E(z N ) = lim H(z) = AH (z) θ→∞ where n o ˆ α(1 − ρ)[α + ξ + λ(1 − z)] αξ B(z) ˆ − λB(z)(1 − z)[α + ξ + λ(1 − z)] AH (z) = n o .1 (z) + ψ0. ˆ + V. N =N (33) Proof. i. z). i. z) + zψ1.Now since E(z1R ) ≡ G(z1 .

we solve the optimization problem λ(1 − b∗ (ξ))2 Minimize C(θ. Required for the optimization are the queueing performance measures E(R). α) = cS + cR E(R)E(WR ) + cQ E(Q)E(WQ ) ξ 2 b∗ (ξ)2 Subject to λ(1 − b∗ (ξ))(α + ξ) − αξb∗ (ξ) < 0 (34) cθ θ + cα α ≤ D (35) θ. λ(1 − b∗ (ξ))/ξb∗ (ξ). α > 0 (36) where D is a fixed budget (D < ∞). and E(WQ ). We also provide two illustrative examples using distinct service time distributions. E(Q). Constraint (34) enforces the stability condition (ρ < 1) and (35) is a budget constraint that limits the attainable repair capacity and the rate at which interrupted customers may attempt to re-access the server. and the cost of holding customers in the orbit. subject to a budget constraint. 5 Optimal Retrial and Repair Rates We now consider the simultaneous optimal selection of the retrial and repair rates that minimize the long-run average operating costs. as well as the expected number of customers in service. The coefficient cθ is the cost of one “unit” of retrial rate while cα is the cost of one “unit” of repair rate. (28) and (29). the cost of holding customers in the normal queue. and the expected time to complete service. The cost function includes the cost of service. 17 . (1 − b∗ (ξ))/ξb∗ (ξ). In the next section we formulate an optimization problem for the selection of the optimal retrial and repair rates that minimize the long-run average cost of operating the queueing system. Using Eqs. The cost per unit time per customer in service is cS while the holding costs per unit time per customer in the orbit and normal queue are respectively denoted by cR and cQ . E(WR ).

the following functions are defined: ˆ 0] ξb∗ (ξ)[ξ − λ(1 − b∗ (ξ))] + (α + ξ)[λ(1 − b∗ (ξ)) − ξ B fR1 (α) = α . Strict convexity of the objective function. does not affect the convexity of C. (α + ξ) [αξ − λ(1 − b∗ (ξ))(α + ξ)] The next two lemmas are needed to prove the strict convexity of C(θ.1 Convexity Analysis The uniqueness of a global solution to the above optimization problem can be estab- lished by showing it is a convex program. The feasible region. α). α) on X.) of Eq. θ[αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)] λ(1 − b∗ (ξ)) fR (θ. (37) depends on neither θ nor α. 18 . which will now be proved.5. α) = fR1 (α) + fR2 (θ.s. The existence of a solution will then be shown directly. will complete the uniqueness proof. defined by X = {(θ. α) : λ(1 − b∗ (ξ))(α + ξ) − αξb∗ (ξ) < 0. and hence. b∗ (ξ) and ˆ 0 − λ(1 − b∗ (ξ))] ξ 3 b∗ (ξ) − (α + ξ)2 [ξ B fQ (α) = . α) gives à !2 λ(1 − b∗ (ξ))2 ˆ 0 − λ(1 − b∗ (ξ))] ξ 3 b∗ (ξ) − (α + ξ)2 [ξ B C(θ. (37) θ[αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)] The first term on the right-hand side (r. is a convex set since it is defined by a finite set of linear constraints. Expanding the terms of C(θ. α) = cS + cQ λ ξ 2 b∗ (ξ)2 ξb∗ (ξ)(α + ξ) [αξ − λ(1 − b∗ (ξ))(α + ξ)] à ξb∗ (ξ)[ξ − λ(1 − b∗ (ξ))] + (α + ξ)[λ(1 − b∗ (ξ)) − ξ B ˆ 0] + cR λ αλ(1 − b∗ (ξ)) ∗ b (ξ)[αξ − λ(1 − b∗ (ξ))(α + ξ)][αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)] ¶2 λξ(α + ξ)(1 − b∗ (ξ)) + . α > 0}. θ.h. α) = . [αξ − λ(1 − b∗ (ξ))(α + ξ)][αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)] λξ(α + ξ)(1 − b∗ (ξ)) fR2 (θ. To prove the convexity of the other two terms. cθ θ + cα α ≤ D.

∂α2 and thus. λ(1 − b∗ (ξ)) − ξ Bˆ 0 is strictly increasing for ξ ∈ [0. 0 Since its derivative with respect to ξ. α) is strictly convex on X. and assuming stability. Proof. fR1 (α) is strictly convex for all α > 0. shows that ∂ 2 fR2 (θ. We first establish the positivity of the quantity µ Z ∞ ¶ ∗ ˆ 0 λ(1 − b (ξ)) − ξ B = λ 1 − −ξx (ξx + 1)b(x)e dx . respectively. Now the second derivative of fR1 with respect to α is given by ½ ∂ 2 fR1 (α) αb∗ (ξ)[ξ − λ(1 − b∗ (ξ))]2 [ξb∗ (ξ) − λ(1 − b∗ (ξ))] = 2ξ ∂α2 [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]2 [αξ − λ(1 − b∗ (ξ))(α + ξ)]2 ˆ 0 ] {α[ξ − λ(1 − b∗ (ξ))][ξb∗ (ξ) − λ(1 − b∗ (ξ))] + λ2 ξ(1 − b∗ (ξ))2 } [λ(1 − b∗ (ξ)) − ξ B + [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]2 [αξ − λ(1 − b∗ (ξ))(α + ξ)]2 ˆ 0 ][αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]2 λ(1 − b∗ (ξ))(α + ξ)[ξ − λ(1 − b∗ (ξ))][λ(1 − b∗ (ξ)) − ξ B + [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]3 [αξ − λ(1 − b∗ (ξ))(α + ξ)]3 λξb∗ (ξ)(1 − b∗ (ξ))[ξ − λ(1 − b∗ (ξ))]2 [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]2 + [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]3 [αξ − λ(1 − b∗ (ξ))(α + ξ)]3 λ(1 − b∗ (ξ))(α + ξ)[ξb∗ (ξ) − λ(1 − b∗ (ξ))][λ(1 − b∗ (ξ)) − ξ Bˆ 0 ][αξ − λ(1 − b∗ (ξ))(α + ξ)]2 + [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]3 [αξ − λ(1 − b∗ (ξ))(α + ξ)]3 ¾ λξb∗ (ξ)(1 − b∗ (ξ))[ξ − λ(1 − b∗ (ξ))][ξb∗ (ξ) − λ(1 − b∗ (ξ))][αξ − λ(1 − b∗ (ξ))(α + ξ)]2 + . ∞) and thus attains its minimum value of zero at the left endpoint ξ = 0.Lemma 2 The function fR2 (θ. ∂ 2 fR1 (α) > 0. [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]3 [αξ − λ(1 − b∗ (ξ))(α + ξ)]3 It follows from the stability condition that [αξ − λ(1 − b∗ (ξ))(α + ξ)] > 0. [ξb∗ (ξ) − λ(1 − b∗ (ξ))] > 0. Z ∞ λξ x2 b(x)e−ξx dx. 0 is strictly positive. Taking second partial derivatives of fR2 with respect to θ and α. Since λ(1 − b∗ (ξ)) − ξ Bˆ 0 is nonnegative. and [ξ − λ(1 − b∗ (ξ))] > 0. α) 2λξ(1 − b∗ (ξ))(α + ξ) = >0 ∂θ2 θ3 [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)] 19 .

Consequently. and thus. Theorem 2. Then it can be shown that λ2 ξ 4 b∗ (ξ)(1 − b∗ (ξ))2 {3ξ 2 b∗ (ξ) + 4[αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]} det(H(θ. Hence. dα2 which establishes strict convexity of fQ for α > 0. Lemma 3 The function fQ2 (α) is strictly convex for all α > 0. α)) = > 0. α). and Lemma 2 ensures that λ(1 − b∗ (ξ)) − ξ Bˆ 0 is nonnegative. θ4 [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]4 Hence. (α + ξ)3 [αξ − λ(1 − b∗ (ξ))(α + ξ)]3 It follows from the stability condition that [αξ−λ(1−b∗ (ξ))(α+ξ)] > 0 and [ξ−λ(1−b∗ (ξ))] > 0. fR2 (θ. α) + fQ2 (α). Theorem 2 The cost function C(θ. fR2 is strictly convex on X.and ∂ 2 fR2 (θ. α). A CP guarantees that any stationary point (Karush-Kuhn-Tucker point) is 20 . The strict convexity of fQ2 (α) follows directly. The second derivative of fQ with respect to α is given by ( d2 fQ (α) ˆ 0 − λ(1 − b∗ (ξ))]} (α + ξ)[ξ − λ(1 − b∗ (ξ))]{ξ 3 b∗ (ξ) − (α + ξ)2 [ξ B = 2ξ 2 dα2 (α + ξ)3 [αξ − λ(1 − b∗ (ξ))(α + ξ)]3 ¾ ξb∗ (ξ)[αξ − λ(1 − b∗ (ξ))(α + ξ)]{2ξ 2 + 3[αξ − λ(1 − b∗ (ξ))(α + ξ)]} + . α) is strictly convex on X. α) is strictly convex on X. In particular. d2 fQ (α) > 0. along with the convexity of X. The proof follows directly from Lemmas 2 and 3. and thus C(θ. show that the optimization problem is a convex program (CP). is strictly convex on X. α) denote the Hessian matrix of fR2 (θ. the strict convexity of fQ2 (α) for all α > 0 ensures that fR2 (θ. Proof. fR = fR1 + fR2 is strictly convex on X. ∂α2 θ[αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]3 Let H(θ. The following theorem is the main result of this section. Proof. α) 2λξ 3 b∗ (ξ)(1 − b∗ (ξ))[ξb∗ (ξ) − λ(1 − b∗ (ξ))] = > 0.

Although X is bounded. α) −2cR λ2 ξ(α + ξ)(1 − b∗ (ξ)) = 2 fR (θ. D − cα α ∗ θ∗ = . it is not closed. and we may substitute D − cα α θ= cθ into C(θ. the equation. ∂C(θ. the optimal repair rate. but to ensure existence of a solution. Thus. we note that. for all values of θ and α. (θ∗ . α). α)  b∗ (ξ)[αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)]2 [αξ − λ(1 − b∗ (ξ))(α + ξ)] n o  2 ∗ α ξ[λ(1 − b ∗ (ξ)) − ξ ˆ 0 ] + b∗ (ξ)[αξ − λ(1 − b∗ (ξ))(α + ξ)]2  B ξ b (ξ) + + θ[αξb (ξ) − λ(1 − b∗ (ξ))(α + ξ)] b∗ (ξ)[αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)][αξ − λ(1 − b∗ (ξ))(α + ξ)]2  ∗ ( ) ˆ 0 − λ(1 − b∗ (ξ))] + 2ξb∗ (ξ)[αξ − λ(1 − b∗ (ξ))(α + ξ)] ξ 3 b∗ (ξ) − (α + ξ)2 [ξ B −2λcQ fQ (α) < 0. 21 . the budget constraint (35) is always binding. cθ may be solved to obtain the optimal retrial rate θ∗ . α) is monotonically decreasing in both θ and α and is bounded below by λ(1 − b∗ (ξ))2 ˆ 0 ]2 λ3 (1 − b∗ (ξ))2 [λ(1 − b∗ (ξ)) − ξ B ˆ 0 − (1 − b∗ (ξ))]2 λ[ξ B cS +cR +cQ . α) < 0 ∂θ θ [αξb∗ (ξ) − λ(1 − b∗ (ξ))(α + ξ)] and ∂C(θ.a global minimizer. the cost function C(θ. b∗ (ξ)2 (α + ξ)2 [αξ − λ(1 − b∗ (ξ))(α + ξ)]2 Hence. is the unique root that satisfies Eq. the global minimizer. (34). denoted by α∗ . ξ 2 b∗ (ξ)2 b∗ (ξ)2 [ξb∗ (ξ) − λ(1 − b∗ (ξ))]2 [ξ − λ(1 − b∗ (ξ))]2 ξ 2 b∗ (ξ)2 [ξ − λ(1 − b∗ (ξ))]2 Therefore. Differentiating the objective function with respect to α and setting it equal to zero. the feasible region must be closed and bounded. and hence. To circumvent this complication. α∗ ) is a stationary point of C on X. α) = ∂α  n o ˆ 0 ] + ξb∗ (ξ)[ξ − λ(1 − b∗ (ξ))]  λ(1 − b∗ (ξ)) (α + ξ)[λ(1 − b∗ (ξ)) − ξ B −2cR λ2 ξ(1−b∗ (ξ))fR (θ. Subsequently.

69 18.0 1.2 Numerical Examples We now illustrate the solution procedure in two distinct scenarios.85 Uniform on (0. The remaining parameters. Table 1: Optimal repair and retrial rates for two numerical examples.2/µ) 5. Service time distribution λ µ ξ D θ∗ α∗ C(θ∗ . µ+s which gives µ b∗ (ξ) = (µ + ξ) and λµ Bˆ 0 = .5.5 30. the Laplace transform of the service time distribution is given by µ b∗ (s) = .0 11. as well as the optimal solutions (indicated by ∗). In the first case.31 5.0 10. are specified in Table 1.77 5. In this case.18 22 . and cR = cQ = 100. In this case. (µ + ξ)2 In the second scenario.5 30. 2/µ). we assume the service times are exponentially distributed with positive rate parameter µ. α∗ ) Exponential (µ) 5.23 18.0 10.0 1. we assume the service times are uniformly distributed on the interval (0. the Laplace transform of the service time distribution is µ(1 − e−2s/µ ) b∗ (s) = 2s so that µ(1 − e−2ξ/µ ) b∗ (ξ) = 2ξ and · µ ¶µ ¶¸ λµ −2ξ 2ξ Bˆ 0 = 2 1 − exp 1+ . 2ξ µ µ The cost coefficients in both cases are: cθ = cα = cS = 1.0 11.

Queueing Systems. A retrial queue with redundancy and unreliable server. 23 . 33 (14). 72 (2). Microelectronics and Reliability. On the M/G/1/1 queueing system with repeated orders and unre- liable server. Diffusion approximation of systems with repeated calls and an unreliable server. 48 (1). (1994). (1997). J. and J. 309-321. 3032-3034. 23-36. A. [3] Aissani. Queueing Systems. we have θ∗ + α∗ = D = 30. R. 24 (6). On the single server retrial queue subject to breakdowns. 98-123 (in French). Analysis of an M/G/1 queue with constant repeated attempts and server vacations. This research was sponsored by the Air Force Office of Scientific Research (F1ATA0634J001). 6. Acknowledgements. (1988). L. (1993). (1994). Artalejo (1998). [4] Aissani. 17 (3-4). [5] Anisimov. 493-504. [7] Artalejo. A. R. J. Computers and Operations Research. V. 2093-2106. Statistica Neerlandica. Atadzhanov (1994). [6] Artalejo. References [1] Aissani. A. Journal of Technology. Journal of Mathematical Sciences.As expected. and K. R. We wish to express our gratitude to an anonymous referee and Pro- fessor Mark Lewis for several useful comments. In particular. V. with cθ = cα = 1. 431-449. [2] Aissani. A. Table 1 verifies that the budget constraint is binding at the optimal solution.0 in both examples. New results in retrial queueing systems with breakdown of the servers. 30 (3-4). Unreliable queuing with repeated orders.

Madheswari. A game theoretic model for two types of customers competing for service. The M/G/1 retrial queue with Bernoulli schedule. 219-228. London. V.. 191-208. Analysis of multiserver queues with constant retrial rate. 162 (2). Retrial queues with server subject to breakdowns and repairs. Zhang. [14] Kulkarni. and Q. [8] Artalejo. 7 (2). D. 855-880. On the M/G/1 retrial queue subjected to breakdowns. Chapman and Hall. Retrial Queues. Choi (1990). R. Y. 135 (3). 2 (3). A single-server retrial queue with general retrial times and Bernoulli schedule.I. [10] Choi. and A. 1057-1075. European Journal of Operational Research. [17] Lam. Queueing Systems. Applied Mathematical Modelling. Liu (2006). Park (1990). R. Y. Neuts (2001). (2002). Queueing Systems. [9] Atencia. N. [13] Hassin. On optimal and equilibrium retrial rates in a queueing system. B. 7 (2). A.C. K. F. S. The M/G/1 retrial queue with feedback and starting failures. C. 341-368. and J. P. V. V. 569-581. RAIRO Operations Research. priorities and server vacations. G. 24 . B. Queueing Systems. 100-121. [15] Kulkarni. 119-122. European Journal of Operational Research. and K. Moutzoukis (1995). (1983). A geometric process model for M/M/1 queueing system with a repairable service station. Templeton (1997). 20 (3-4). Gomez-Corral. G. 299-310. and B. and E. 223-227. [12] Falin. Probability in the Engineering and Informational Sciences. [16] Kumar. Applied Mathematics and Computation. 26 (11). [11] Djellab. and P.G. 10 (2). (1996). Vijayakumar (2002). J. G. I. [18] Langaris. Operations Research Letters. K. L. 36 (4).. and M. A retrial queue with structured batch arrivals.. Moreno (2005). D. 168 (1).

and C. S. 22 (5). Yang (1995). 85 (1). [22] Liang. Brill. and J. Chen (2004). 203-218. In Applied Probability and Stochastic Processes: International Series in Operations Re- search and Management Science. A single-server retrial queue with server vacations and a finite number of input sources. M. 16 (1-2). 543-552. 1 (2). The M/G/1 retrial queue with the server subject to starting failures. 12 (3). 21 (2-3). Zhao (2005). G. H. 25 . Optimal control of the MX /G/1/K queue with multiple server vaca- tions. Li (1994). 325-339. European Journal of Operational Research. Hlynka. 30-51. [27] Yang. [24] Wang. and G. T.. X. An M/G/1 retrial queue with balking and retrials during service. 531-550. [25] Wang. Computers and Operations Research. 149-160. Optimal routing control in retrial queues. and T. [21] Li. H. server break downs and impatient customers.. Kulkarni (1999). Vol. H. Wang (2005). Reliability analysis of the retrial queue with server breakdowns and repairs. Kao. A retrial queue with a constant retrial rate. and Y. H.G. and H. E. 19. H. and Q. and V. Non-preemptive priorities and vacations in a multiclass retrial queueing system. Q. Cao. Queueing Systems. Langaris (1996). Communications in Statistics: Stochastic Models. Quality Technology and Quantitative Management. P. Reliability analysis of the retrial queue with server breakdowns and repairs. 363-380. International Journal of Operational Research.). M. T. eds. J. Li (2001). Stochastic Models. J. H. (1995). [23] Moutzoukis. [26] Wu. Shanthikumar and Ushio Sumita.[19] Lee.. 1 (1-2). K. 83-96. Queueing Systems. 38 (4). 455-472. [20] Li. (J.