Approximation in an M/g/1 Queueing System with Breakdowns and Repairs

In this paper, we study the strong stability in the M/G/1 queueing system with breakdowns and repairs after perturbation of the breakdown's parameter. Using the approximation conditions in the classical M/G/1 system, we obtain stability inequalities with exact computation of the constants. Thus, we can approximate the characteristics of the M/G/1 queueing system with breakdowns and repairs by the classical M/G/1 corresponding ones.


INTRODUCTION
In classical queueing models, it is common to assume that the service station have not failed.However, in practice we often meet cases where the service station may experience breakdowns.Therefore, a more realistic queueing model is a model which its service station is unable from time to time to deliver service to its customers.Such models are useful to investigate the performance of unreliable computer systems, manufacturing systems, etc.The impact of these service breakdowns on the performance of a given queueing system will depend on the specific interaction between the breakdowns process and the service process.Therefore, such systems cannot, in general, be analyzed as the reliable one.
Unreliable server queues have been developed tremendously since the pioneering work of Thiruvengadam [1] and Avi-Itzhak and Naor [2].Li and al. [3] and Tang [4] investigated the behavior of the unreliable server, and the effect of server breakdowns and repairs in the M/G/1 queueing models from both the queueing and the reliability points of view.Wang and al. [5] investigated the controllable M/H k /1 queueing system with unreliable server.They developed analytic closed-form solutions and provided a sensitivity analysis.Pearn and al. [6] have obtained analytical results for s ensitivity analysis of the optimal management policy for M/G/1 queueing system with removable and non-reliable server.Wang [7] has used a supplementary variable method for analysis of an M/G/1 queue with second optional service and server breakdowns.Wang and al. [8] have used maximum entropy principle to derive the approximate formulas in a single removable and unreliable server in the N policy M/G/1 queueing system with general startup times.Lam and al. [9] has used a supplementary variable method for analysis of a geometric process model for M/M/1 queueing system with repairable service station.Recently, Ke and Lin [10] developed an approach that can provide fuzzy system characteristics for queues which are described and solved by parametric nonlinear programming.
Therefore, analytic closed-form solutions could be obtained only for certain special unreliable models.Nevertheless, even in these cases, the complexity of these analytical formulas does not allow exploiting them in the practice, that is the case of the generating function of the various limiting distributions or Laplace transform (see for example [11]).For this reason, there exists, when modelling a real system, a common technique for substituting the real but complicated elements governing a queueing system by simpler ones in some sens close to the real elements.
The queueing model so constructed represents an idealization of the real queueing one, and hence the "stability" problem arises.
A queueing system is said to be stable if small perturbations in its parameters entail small perturbations in its characteristics [12,13].So, the margin between the corresponding characteristics of two stable queueing systems is obtained as function of the margin between their parameters.
The strong stability method elaborated in the beginning of the 1980s [13,14] permits us to investigate the ergodicity and the stability of the stationary and non-stationary characteristics of Markov chains [12].In contrast to other methods, we assume that the perturbations of the transition kernel are small with respect to some norms in the operator space.Such a strict condition allows us to obtain better stability estimates and enables us to find precise asymptotic expressions of the characteristics of the "perturbed" system [15].
The remainder of this paper is composed as follows: In section 2, we present the considered queueing models and we pose the problem.Notations and preliminaries are introduced in Section 3. The main results of this paper are given in Section 4. After perturbing the breakdown's parameter, we estimate the deviation of its transition operator, and we give an upper bound to the approximation error (Section 4).

MODELS DESCRIPTION AND ASSUMPTIONS
We consider an M/G/1 (F IF O, ∞) queueing system with a repairable service station.Customers arrive according to a Poisson process with rate λ and demand an independent and identically distributed service times with common distribution function B with mean 1/µ.We shall assume that when a server fails, the time required to repair it have exponential distribution with rate r > 0.
In this model, we consider the breakdowns with losses.If an arriving customer finds the server idle-up (i.e. it is ready to serve), it will be immediately occupy the server with probability q > 0 and leaves it after service completion if no breakdown had occurred during this period.If the server is failed, then this blocked customer may leave the system altogether without being served with probability 1 − q.We assume that when the server fails, it is under repair and it cannot be occupied.As soon as the repair of the failed server is completed, the server enters an operating state and continues to serve the other customers immediately.
The state of M/G/1 queueing system with a repairable service station at time t, can be described by the following stochastic process, where X(t) ∈ N is the number of customers in the system at time t and {e(t), K(t)} are two supplementary state variables.e(t) ∈ {0, 1} describes global properties (up or failed) of the server: e(t) = 0, if the server is available for service at time t, 1, if the server is failed at time t.
(2) K(t) is a random variable, which is defined below: • If e(t) = 0 and X(t) = 0, then K(t) is equal to the time elapsed since the time t to the occurrence time of a breakdown, if after t the arrivals are interrupted.
• If e(t) = 0, K(t) is the duration of the remaining service period.
• If e(t) = 1, K(t) is the duration of the remaining repair period.
Clearly, S(t) is a Markov process.The computation of its transitional regime intervenes the integro-differential equations.For that, we use the method of the embedded Markov chain.
We use the following notations: • γ n : are the times of "end" of the n th service, or "end" of the n th repair.
• ∆ n : is a random variable, that represents the number of the arrivals during the n th service time, or during the n th repair time.• X n = X(γ n ): is number of customers in the system immediately after the "end" of the n th service time, or immediately after the "end" of the n th repair time.
We have, Therefore the transition probabilities P ij are written under the following form: • If X n = i = 0, then, • If X n = i = 0, then, Where, the distribution of ∆ n is given by • If e n = 1, then, To calculate the operator transition, we distinguish the following cases: (1) If the system is empty (i.e., X n = i = 0), then, (2) If there are no arrivals during the service duration of (n + 1) th customer, or during the repair duration of the server just after the departure of the n th customer, then, (3) If j > i − 1, then, (4) If j < i − 1, then, Consider also an M/G/1 (F IF O, ∞) system with null breakdown's rate.It behaves exactly as a classical M/G/1 system: Arrivals occur as a Poisson process of rate λ, and it has the same general service time distribution function B. Let P be the transition operator for the corresponding Markov chain X n , in the classical M/G/1 system.We have [16] We propose to obtain the stability inequalities, and this by applying the strong stability criterion [13] for the M/G/1 queueing system with null breakdown's rate.Indeed, the approximation conditions in the classical M/G/1 queueing system have been made [17] at the time of perturbation of the retrial's parameter in the M/G/1/1 retrial queue.By making tend the retrial's rate to ∞ and the resultant system is exactly the M/G/1 system with infinite retrials (or the classical M/G/1 system).Also, in our case, we make tend the breakdown's rate to zero (or we make tend the probability q to one) in the M/G/1 with breakdowns and repairs, so we will have the classical M/G/1 system.Therefore, we can consider the two perturbed systems, the M/G/1 with repairable station and the retrial M/G/1 queue, as two sequences which have the same limit.

PRELIMINARIES AND NOTATIONS
Let M = {µ j } be the space of finite measures on N, and η = {f (j)} the space of bounded measurable functions on N. We associate with each transition kernel P the linear mapping: Introduce on M the class of norms of the form: Where υ is an arbitrary measurable function (not necessary finite) bounded below away from a positive constant.This norm induces in the space η the norm: Let us B, the space of linear operators, with norm: Definition 3.1.The Markov chain X with a transition kernel P and an invariant measure π is said to be strongly υ-stable with respect to the norm .υ , if P υ < ∞ and each stochastic kernel Q on the space (N, B(N)) in some neighborhood {Q : Theorem 3.1.( [14]) The Markov chain X with the transition kernel P and invariant measure π is strongly υ-stable with respect to the norm .υ if and only if there exist a measure α and a nonnegative measurable function h on N such that πh > 0, α1 = 1, αh > 0, and a) The operator Here 1 is the function identically equal to 1. Theorem 3.2.( [18]) Let X be a strongly υ-stable Markov chain that satisfies the conditions of theorem (3.1).If ν is the measure invariant of a kernel Q, then for the norm Q − P υ sufficiently small, we have Under the conditions of Theorem (3.1), we have as υ → 0 corollary 3.2.Under the conditions of Theorem (3.1), for υ < (1−ρ) c , we have the estimate Where

STABILITY CONDITIONS IN THE CLASSICAL M/G/1 QUEUE
The first step consists of determination of the strong υ-stable conditions of the M/G/1 system with null breakdown's rate.In fact, these conditions are already made by [17] at the time of perturbation of the retrial's parameter in the M/G/1/1 retrial queue.So, to apply Theorem (3.1) to the Markov chain X n , we choose the function υ(k) = β k , β > 1, and α j = P 0j .
The embedded Markov chain X n being strongly stable then, the characteristics of the M/G/1 queueing system with breakdowns and repairs can be approximated by the corresponding ones of the classical M/G/1 queue.

STABILITY INEQUALITIES IN THE CLASSICAL M/G/1 QUEUE
In this section we obtain the inequalities of stability.

ESTIMATION OF THE TRANSITION KERNEL DEVIATION
To be able to estimate numerically the margin between the stationary distributions of the Markov chains X n and X n , we estimate the norm of the deviation of the transition kernel P , and we use the following lemma: Lemma 4.1.[19] Let us suppose that ( Theorem 4.1.Let P (respectively P ) be the transition operator of the embedded Markov chain in the M/G/1 queueing system with breakdowns and repairs (respectively in the classical M/G/1 system).Let us suppose that the conditions of the Lemma (4.1) are verified.Then, for all β such that 1 < β < β 0 , we have where 16), we have (i) For k = 0 : (ii) For k ≥ 1: Then, we have: ∞ 0 e λ(β−1)x |B − E|(dx) < βW. (26) From which according to the Lemma (4.1), we will have:

STABILITY INEQUALITIES
To be able to apply the Theorem (3.2), let us use the Lemma below for estimate π υ .Lemma 4.2.[17] Let π (respectively π) be the stationary distribution of the embedded chain of the M/G/1 queueing system with breakdowns and repairs (respectively of the classical M/G/1 system).Then, for all 1 < β < β 0 , we have where c 0 is given by and m = E(ξ), (ξ is the service time).

CONCLUSION
In this work, we use the approximation conditions [17] for which ones it will be possible to approximate the characteristics of the M/G/1 queueing system with breakdowns and repairs by the one of the classical M/G/1 queueing system.We obtain stability inequalities with an exact computation of the constants.These theoretical results allows to estimate numerically the error due to this approximation.