We consider in this paper an M/G/1 type queueing system with the following extensions. First, the server is unreliable and is subject to random breakdowns. Second, the server also implements the well-known N-policy. Third, instead of a Bernoulli vacation schedule, the more general notion of binomial schedule with K vacations is applied. A cost function with two decision variables is developed. A numerical example shows the effect of the system parameters on the optimal management policy.
1. Introduction
Queueing systems where the server uses her/his idle time to perform some secondary job such as maintenance are called systems with server vacations. These systems have received a lot of attention due to their wide applications in different domains such as telecommunications, computer systems, service systems, and production and quality control problems. Survey papers have been written on this subject; the most recent one being that of Ke et al. [1].
Keilson and Servi [2] introduced a class of vacation models called the Bernoulli vacation schedule. When a customer has just been served and other customers are present, the server serves the next customer in line with probability p or takes a vacation of random duration with probability (1-p). The Bernoulli vacation schedule has been extensively considered. Among the most recent references we cite Kumar et al. [3], Choudhury and Ke [4], Gao and Liu [5], Tao et al. [6, 7], and Wu and Lian [8].
Kella [9] generalized the Bernoulli vacation schedule to a more general scheme according to which the server goes on i consecutive vacations with probability pi if the queue upon her/his return is empty. Ba-Rukab et al. [10] propose another generalization. They argued that since the server may attend different activities while idle, a binomial vacation schedule may be more appropriate than a Bernoulli vacation schedule. In that case, instead of taking just one vacation, the server may take many vacations, for a maximum number of, say, K vacations.
Yadin and Naor introduced the N-policy in which, following an idle period, the server resumes his service only when the number of waiting customers reaches the level N. This policy is efficient in that it reduces setup costs. The N-policy too has been extensively studied by researches. We refer the reader to the following recent references: Kumar and Jain [11], Lee and Yang [12], Lim et al. [13], and Wei et al. [14].
Another characteristic of servers in a queueing system is that they may break down while providing service. White and Christie [15] were the first to study a queueing system with an unreliable server. Since then, many authors have incorporated this feature in their studies. We cite the recent papers of Dimitriou [16], Wu and Lian [8], Choudhury and Ke [4], Ke et al. [17], Yang et al. [18], Yarmand and Down [19], Kumar et al. [3], and Zhang and Wang [20].
The goal of this paper is to study an M/G/1 type queueing system where the server implements the notion of a binomial vacation schedule. Under this policy, at a service completion and before serving the next customer, the server takes a series of vacations. The number of vacations follows a binomial distribution with parameters p and K. Each vacation has a random duration and corresponds to some auxiliary activity. Also, we assume the N-policy discipline, so that, following a busy period, the server remains idle and does not resume work until the number of units in the queue reaches the threshold level N. Furthermore, we assume that the server is unreliable and may breakdown while providing service to a customer. In that case, the server is repaired and the repair period has a random duration. Following repair, service is resumed.
There are various reasons for studying this queueing system. First, from a theoretical point of view, although some aspects of this system have been discussed separately, no work has been found that combines N-policy, binomial vacation schedule, and all service interruptions. Hence this paper is an attempt to fill up the gap. Second, from a practical point of view, combining these features gives the decision maker a better control over the system. As pointed out earlier, the N-policy reduces the number of setups and thus reduces setup costs. The binomial schedule allows the server to accomplish more secondary tasks than a regular Bernoulli schedule. Secondary tasks could include actions such as virus scans, maintenance operation, quality control tests, and attending other queues.
The notation used to describe this system is given in the next section. In Section 3, we obtain the probability generating function of the system size in the steady-state. We also derive the main performance measures. These measures are used to prescribe an optimal management policy for the system by finding the optimal values of the thresholds K and N in Section 4. Section 5 summarizes the paper and provides some future research directions.
2. Model Description
We consider an M/G/1 type queueing system with positive arrival rate λ. The service time random variable B has cumulative distribution function (CDF) B(t), Laplace-Stieltjes transform (LST) B*(θ)=E[e-θB], and finite first and second moments bj,j=1,2.
The server is unreliable and the time to failure has an exponential distribution with positive rate α. Following a breakdown, a repair of the server takes place. The repair time random variable R has CDF R(t), LST R*(θ)=E[e-θR], and finite first and second moments rj,j=1,2.
We introduce the modified service time H which includes the actual service time and possible repairs. The random variable H has CDF H(t), LST H*(θ)=E[e-θH], and finite and second moments hj,j=1,2. It is easy to see that the modified service times, actual service times, and repair times are related through the following expression:
(1)H*(θ)=∑n=0∞∫0∞e-θxe-αx(αx)nn![R*(θ)]ndB(x)=B*(θ+α(1-R*(θ))).
From this expression, the first two moments of the modified service time are found as
(2)h1=b1(1+αr1),h2=b2(1+αr1)2+αb1r2.
The server implements the binomial vacation schedule. At the end of a service, if no customer is present in the system, the server takes k(k=0,…,K) vacations of length Vk with probability
(3)yk=(Kk)pk(1-p)K-k,k=0,…,K.
The vacation time random variable V has CDF V(t), LST V*(θ)=E[e-θV], and finite first and second moments vj,j=1,2.
We now introduce the generalized service time G. The random variable G has CDF G(t), LST G*(θ)=E[e-θG], and finite first and second moments gj,j=1,2. It is easy to see that the generalized service times, modified service times, and vacation times are related through the following expression:
(4)G=H+V0+⋯+Vkwithprobabilityyk.
Therefore,
(5)G*(θ)=H*(θ)∑k=0Kyk[V*(θ)]k.
From this expression, the first two moments of the generalized service time are found as
(6)g1=h1+Kpv1,g2=h2+2Kph1+K(K-1)p2.
We are assuming that the server implements also the N-policy discipline. Thus, if there are N or more customers in the queue at a service completion epoch, then the server serves the next customer in line. However, if there are less than N customers in the queue at a service completion epoch, then the server remains idle and waits for the queue to reach the level N.
Finally, we denote by Q(t) the number of customers in the system at any instant of time t≥0. We are first interested in the distribution of this process in the steady-state.
3. Model Analysis
In this section, we derive the PGF of the queueing process Q(t) in the steady-state along with the main performance measures required to develop an optimal management policy of the system.
3.1. Probability Generating Function
Let pi=limt→∞P{Q(t)=i},i=0,1,…, denote the steady-state probability of state i. Since our model is of M/G/1 type with a modified service time, see, for example, Çinlar [21], we can readily generalize Pollaczek-Khinchine formula to obtain the probability generating function P(z)=∑i=0∞pizi as
(7)P(z)=1-ρN×(∑k=0K(zN-1)B*(λ-λz+α(1-R*(λ-λz)))×∑k=0Kyk[V*(λ-λz)]k)×(∑k=0Kz-B*(λ-λz+α(1-R*(λ-λz)))×∑k=0Kyk[V*(λ-λz)]k)-1,
where ρ=λ[b1(1+αr1)+Kpv1] must satisfy the ergodicity condition
(8)ρ<1.
3.2. System Performance Measures
The expected number of customers in the system at an arbitrary instant of time in the steady-state is given by
(9)L=λ[b1(1+αr1)+Kpv1]1-λ[b1(1+αr1)+Kpv1]+(λ2{b2(1+αr1)2+αb1r2+2Kpb1(1+αr1)+K(K-1)p2})×(2{1-λ[b1(1+αr1)+Kpv1]})-1+N-12.
The expected length of an idle period of the server in the equilibrium is given by
(10)I=Nλ.
The expected length of the busy period is
(11)B=N[b1(1+αr1)+Kpv1]1-ρ.
The expected length of the busy cycle C=I+B is given by
(12)C=Nλ(1-ρ).
Using (10)–(12), one can derive the following probabilities.
The probability that the server is idle is as follows:
(13)PI=IC=1-ρ.
The probability that the server is busy is as follows:
(14)PB=BC=ρ.
4. Optimal Management Policy
We now are ready to develop an expression for the total expected cost per unit of time. The decision variables in this expression are K and N. The goal is to find the optimal values of these two parameters. This would optimize the performance of the system, as it would allow the server to know exactly when to end an idle period and how many auxiliary jobs should be performed before moving to the next customer in line.
4.1. Total Expected Cost per Unit of Time
Using a linear cost structure, the total expected cost function per unit time is given by
(15)TC(K,N)=chL+coBC+caIC+cs1C=ch[ρ+λ2g22(1-ρ)+N-12]+coρ+ca(1-ρ)+csλ(1-ρ)N,
where g2 is given by (6), ρ is given by (8), ch is the holding cost per unit for each customer present in the system, co is the cost per unit time for keeping the server on and in operation, ca is the startup cost per unit time for the preparatory work of the server before starting, and cs is the setup cost per busy cycle.
Treating the decision variables as continuous, the optimal values of K and N are found by solving the system of two equations
(16)∂TC(K,N)∂K=0,∂TC(K,N)∂N=0
and checking that the Hessian matrix is positive definite. These calculations are done numerically since expression (15) is nonlinear and closed form expressions for the optimal values K* and N* are difficult to obtain. To determine these values, we present the following procedure. For a given K, the optimal value N*(K) of N is given by the first l such that
(17)TC(K,l+1)>TC(K,l);
that is,
(18)N*(K)=min{N≥1∣I(K,N)>0},
where
(19)I(K,N)=TC(K,N+1)-TC(K,N).
We summarize the procedure to find the optimal values (K*,N*) as follows.
Set K=1. Determine N*(K) using (18) and compute TC(K,N*(K)) using (15).
Compute N*(K+1) using (18) and TC(K+1,N*(K+1)) using (15).
If TC(K+1,N*(K+1))>TC(K,N*(K)), stop. The optimal values are (K*,N*)=(K,N*(K)). Otherwise, go to step 2.
4.2. Numerical Illustration
We now present some numerical computations to show the practicality of the results obtained. For illustration, we will assume that the service, repair, and vacation times are exponentially distributed. We recall that if X follows the exponential distribution with mean E[X], then
(20)P(X<x)=1-e(x/E[X]),B*(θ)=11+θE[X],E[X2]=2E[X]2.
As an illustration, we assume the system parameters listed in Table 1.
Data.
Parameters
Values
Unit costs
ch=2,co=50,ca=50,cs=300
Mean arrival rate
λ=0.05
Mean breakdown rate
α=0.05
Probability of a vacation
p=0.60
Mean service time
b1=0.7
Mean repair time
r1=0.5
Mean vacation time
v1=0.25
Figure 1 shows the variations of the total expected cost per unit of time as K and N vary. The curve is convex and the optimal thresholds are found to be (K*,N*)=(5,4). The optimal cost is TC(K*,N*)=56.67. Therefore, to minimize the total expected cost per unit of time, following an idle period, the server of this system should not be activated until the number of customers waiting in the queue reaches 4 customers. Also, at a service completion, when the queue is not empty, the server may take up to 5 vacations or perform 5 auxiliary tasks, before serving the next customer in line. It is worth noting that the optimal value of K is not K*=2, which would correspond to the well-known and widely used Bernoulli vacation schedule. This shows the effectiveness of the binomial over the Bernoulli schedule.
Variations of the total expected cost per unit of time.
4.3. Sensitivity Analysis
We also performed a sensitivity analysis by looking at the effect of the system parameters on the optimal solution. The sensitivity analyses with the monetary parameters are given in Tables 2–5. We note that as customer holding cost ch increases, both K* and N* decrease while they both increase as operation setup cost cs increases (Tables 2 and 5). However, as server keeping cost co increases, K* decreases, while server idle cost ca has the opposite effect on K* (Tables 3 and 4). It is interesting to see that N* remains unchanged; that is, the amount of server costs does not influence the threshold level of waiting customers.
Effect of the unit cost ch on the optimal solution.
ch
1
2
3
4
5
6
7
8
9
10
K*
11
5
3
1
2
1
1
1
1
1
N*
5
4
3
3
2
2
2
2
2
2
TC(K*,N*)
54.85
56.67
57.92
58.97
59.91
60.46
61.01
61.56
62.11
62.65
Effect of the unit cost co on the optimal solution.
co
50
60
70
80
90
100
110
120
130
140
K*
5
1
1
1
1
1
1
1
1
1
N*
4
4
4
4
4
4
4
4
4
4
TC(K*,N*)
56.67
57.12
57.55
57.98
58.42
58.85
59.29
59.72
60.15
60.59
Effect of the unit cost ca on the optimal solution.
ca
5
6
7
8
9
10
20
30
40
50
K*
1
1
1
1
1
1
1
1
1
5
N*
4
4
4
4
4
4
4
4
4
4
TC(K*,N*)
13.63
14.59
15.55
16.50
17.46
18.42
27.98
37.55
47.12
56.67
Effect of the unit cost cs on the optimal solution.
cs
100
200
300
400
500
600
700
800
900
1000
K*
1
3
5
8
8
11
11
13
15
15
N*
2
3
4
4
5
5
6
6
6
7
TC(K*,N*)
53.49
55.28
56.67
57.81
58.81
59.70
60.55
61.28
62.00
62.69
Concerning the nonmonetary parameters, we note from Tables 6, 7, 8, 9, 10, and 11 that their effect is almost negligible on N*. Also, α has very slight effect on K*. However, K* increases as v1 increases and decreases as either of the other parameters (p, b1, or r1) increases. Finally, the nonmonetary parameters affect very little the optimal cost.
Effect of the arrival rate λ on the optimal solution.
λ
0.010
0.015
0.020
0.025
0.030
0.035
0.040
0.045
0.050
0.055
K*
1
2
8
2
5
6
7
4
5
5
N*
2
2
2
3
3
3
4
4
4
4
TC(K*,N*)
52.50
53.25
53.98
54.49
54.97
55.44
55.91
56.32
56.67
57.01
Effect of the breakdown rate α on the optimal solution.
α
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
0.50
K*
5
5
5
5
5
4
4
4
4
4
N*
5
4
4
4
4
4
4
4
4
4
TC(K*,N*)
56.67
56.67
56.67
56.67
56.67
56.67
56.67
56.67
56.67
56.67
Effect of the probability of a vacation p on the optimal solution.
p
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
K*
26
13
9
7
6
5
4
4
N*
4
4
4
4
4
4
4
4
TC(K*,N*)
56.67
56.67
56.67
56.67
56.67
56.67
56.67
56.67
Effect of the mean service time b1 on the optimal solution.
b1
0.35
0.70
1.05
1.40
1.75
2.10
2.45
2.80
3.15
3.50
K*
6
5
4
3
2
1
1
1
1
1
N*
5
4
4
4
4
4
4
4
4
4
TC(K*,N*)
56.68
56.67
56.65
56.64
56.62
56.61
56.60
56.60
56.60
56.60
Effect of the mean repair time r1 on the optimal solution.
r1
0.125
0.250
0.500
1.000
2.000
4.000
8.000
16.000
32.000
64.000
K*
5
5
5
5
5
4
4
3
1
1
N*
4
4
4
4
4
4
4
4
4
4
TC(K*,N*)
56.67
56.67
56.67
56.67
56.62
56.68
56.67
56.69
56.82
57.45
Effect of the mean vacation time v1 on the optimal solution.
v1
0.10
0.12
0.14
0.16
0.18
0.20
0.22
0.24
0.26
0.28
K*
2
2
2
2
4
4
4
5
5
5
N*
4
4
4
4
4
4
4
4
4
4
TC(K*,N*)
56.69
56.69
56.69
56.68
56.68
56.68
56.67
56.67
56.67
56.66
It is interesting to find that K* is much more sensitive than N* to the parameter changes. Actually, N* remains around 4 customers under most parameter values.
5. Conclusion
We have considered in this paper an unreliable queueing system where the server implements a binomial up to K-vacation schedule and an N-customer waiting policy. The system characteristics are obtained and an optimal management policy is described. The effect of the system parameters on the optimal threshold level N* and the optimal maximum number of auxiliary jobs K* that the server should perform is shown in a numerical example.
It is interesting to find that K* is much more sensitive than N* to the parameter changes. Actually, N* remains around 4 customers under most parameter values. The manager of this queueing system thus should be more concerned with limiting vacations than with increasing the maximum number of customers allowed to wait.
This work can be generalized by assuming, for example, a batch arrival Poisson process. Binomial schedule may also be worth investigating in retrial queueing systems.
Conflict of Interests
The authors declare that there is no conflict of interests regarding the publication of this paper.
Acknowledgment
The authors would like to thank the two referees for carefully reading the paper and for making suggestions to improve it.
KeJ.-C.WuC.-H.ZhangZ. G.Recent developments in vacation queueing models: a short survey20107438KeilsonJ.ServiL. D.Oscillating random walk models for GI/G/1 vacation system with Bernoulli schedules19862337908022-s2.0-002277868710.2307/3214016KumarB. K.RukmaniR.LakshmiS. R. A.Performance analysis of an M/G/1 queueing system under Bernoulli vacation schedules with server setup and close down periods20136611910.1016/j.cie.2013.06.005ChoudhuryG.KeJ.-C.A batch arrival retrial queue with general retrial times under Bernoulli vacation schedule for unreliable server and delaying repair20123612552692-s2.0-8005182637910.1016/j.apm.2011.05.047ZBL1236.90031GaoS.LiuZ.An M/G/1 queue with single working vacation and vacation interruption under Bernoulli schedule20133731564157910.1016/j.apm.2012.04.045TaoL.WangZ.LiuZ.The GI/M/1 queue with Bernoulli-schedule-controlled vacation and vacation interruption20133763724373510.1016/j.apm.2012.07.045ZBL1277.60162TaoL.ZhangL.XuX.GaoS.The GI/Geo/1 queue with Bernoulli-schedule-controlled vacation and vacation interruption20134071680169210.1016/j.cor.2012.12.019WuJ.LianZ.A single-server retrial G-queue with priority and unreliable server under Bernoulli vacation schedule2013641849310.1016/j.cie.2012.08.015KellaO.The threshold policy in the M/G/1 queue with server vacations1989361111123Ba-RukabO.TadjL.KeJ.-C.Binomial schedule for an M/G/1 queueing system with an unreliable server201333-420621810.1504/IJMOM.2013.058326KumarK.JainM.Threshold F-policy and N-policy for multi-component machining system with warm standbys2013928910.1186/2251-712X-9-1LeeD. H.YangW. S.The N-policy of a discrete time Geo/G/1 queue with disasters and its application to wireless sensor networks201337239722973110.1016/j.apm.2013.05.012LimD.-E.LeeD. H.YangW. S.ChaeK.Analysis of the GI/Geo/1 queue with N-policy20133774643465210.1016/j.apm.2012.09.037WeiY.YuM.TangY.GuJ.Queue size distribution and capacity optimum design for image-policy image queue with setup time and variable input rate2013575-61559157110.1016/j.mcm.2012.12.032WhiteH.ChristieL. G.Queueing with preemptive priority or breakdown195861799610.1287/opre.6.1.79DimitriouI.A mixed priority retrial queue with negative arrivals, unreliable server and multiple vacations20133731295130910.1016/j.apm.2012.04.011KeJ.-C.HsuY.-L.LiuT.-H.ZhangZ. G.Computational analysis of machine repair problem with unreliable multi-repairmen201340384885510.1016/j.cor.2012.10.004YangD.-Y.ChiangY.-C.TsouC.-S.Cost analysis of a finite capacity queue with server breakdowns and threshold-based recovery policy201332117417910.1016/j.jmsy.2012.06.002YarmandM. H.DownD. G.Server allocation for zero buffer tandem queues2013230359660310.1016/j.ejor.2013.05.029ZhangF.WangJ.Performance analysis of the retrial queues with finite number of sources and service interruptions201342111713110.1016/j.jkss.2012.06.002ÇinlarE.1975Englewood Cliffs, NJ, USAPrentice-Hall