Stationary Analysis of Geo/geo/1 Queue with Two-speed Service and the Optimal Switching Threshold for the Service Rate

This paper considers a Geo/Geo/1 queueing system with infinite capacity, in which the service rate changes depending on the workload. Initially, when the number of customers in the system is less than a certain threshold L, low service rate is provided for cost saving. On the other hand, the high service rate is activated as soon as L customers accumulate in the system and such service rate is preserved until the system becomes completely empty even if the number of customers falls below L. The steady-state probability distribution and the expected number of customers in the system are derived. Through the first-step argument, a recursive algorithm for computing the first moment of the conditional sojourn time is obtained. Furthermore, employing the results of regeneration cycle analysis, the direct search method is also implemented to determine the optimal value of L for minimizing the long-run average cost rate function.


Introduction
In the classical queueing literature, the server is usually assumed to work at constant speed as long as there is any work present.However, we know that this assumption may not always be appropriate when the system's workload affects the server's efficiency in some real world situations.To better understand this fact, we can cite some practical examples to illustrate this point.In a manufacturing system, the decision-maker is responsible for deciding the service speed of the production equipment according to the level of market demand.If the current X.D. Lin tomer?
The rest of this paper is organized as follows.In Section 2, we describe the mathematical model for the problem under consideration.The steady-state analysis of the model is presented in Section 3 and some important system performance measures are derived in this section.Using the first-step argument, we develop an analytical scheme for the customer's sojourn time.Furthermore, we also carried out regeneration cycle analysis to find the expected length of two types of busy periods.In Section 4, a long-run average cost rate function is established based on the system characteristics to determine the optimal switching threshold for the service rate.Section 5 concludes the research and suggests some future topics.

Model Formulation
We consider a discrete-time queue with single server or machine, whose service rate may be affected by the number of customers or jobs present in the system.In our model, inter-arrival times 1 2 , , , A A  are independent and identically distributed (i.i.d.) random variables with probability mass function (p.m.f.) Arriving customers form a single waiting line based on the order of their arrival.Initially, when the number of customers in the system is less than the given threshold level ( ) 1 L L ≥ , the server serves customers with low service rate.The high service rate is activated at the in- stant when the number of customers in the system becomes equal to L, and it will be preserved until the long queue empties.Here, we assume that the two types of service times, namely low S and high S , are independent and geometrically distributed with respective p.m.fs = − and 1 2 µ µ < .λ , 1 µ and 2 µ denote the customer arrival rate, low service rate and high service rate, respectively.In discrete-time queueing system, the time axis is divided into equal intervals called slots and all queueing activities occur at the slot boundaries.Traditionally, there are two types of systems in the discrete-time case (see [16] and [17]), one is the late arrival with delayed access (LAS-DA) and the other is the early arrival system (EAS).In this paper, we consider the model for the late arrival system with delayed access and therefore, a potential arrival occurs in ( ) , and a potential departure takes place in ( )  .To make it clear, the various time epochs at which events occur are shown in a self-explanatory figure (see Figure 1).

Steady-State Analysis
In this section, we first apply the Markov process theory to obtain the steady-state difference equations governing the system.Next, the generating function technique and a recursive method are employed to develop the analytical solutions in a neat close-form.Toward this end, we need to define some commonly used notations to analyze the queueing system as follows: ( ) N t ≡ the number of customers in the system at time t − , ( ) Y t ≡ the server speed state at time t − , where X. D. Lin ( ) 0, the single server provides service to the customers at a low speed, 1, the single server provides service to the customers at a high speed.
 is the Markov chain for queueing system with state space   Furthermore, let us define the following stationary probability distributions for the Markov chain:

Steady-State Equation
From the state-transition-rate diagram for the Geo/Geo/1 queue with service rate switching threshold (see Figure 2), we can set up steady-state equations for ,0 i P and ,1 i P in the following: ( ) ( ) ( ) ( ) ( )  2, the state space of this queueing system is a single communicating class.Thus, the Markov chain ( ) t ξ is irreducible.Under assumption that 1 2 µ µ < , it is clear from the standard Geo/Geo/1 theory that 2 1 λ µ < is a necessary and sufficient condition for ergodicity of the system.Intuitively speaking, if, on aver- age, arrivals happen faster than service completions the queue will grow indefinitely long and the system will not have a stationary distribution.
Remark 2. Relating the state probabilities at epochs t and 1 t + or t + and ( ) , and letting t → ∞ , we can easily obtain a set of difference equations, which have exactly the same mathematical form as Equations ( 1)- (8).So the stationary probabilities of the system state at epochs t and t + are identical with the one at epoch t − .Furthermore, in LAS-DA, since an outside observer's observation epoch falls in a time interval after a po- tential departure and before a potential arrival, the probability that outside observer sees i customers in the system and the server in state j is also the same as ( ) , 0,1, 2, , 0,1 . For these reasons, only the system state probability at time point t − is considered in this paper.

Two Relationships between P0,0 and PL−1,0
In this subsection, we first derive two important relationships between 0,0 P and 1,0 L P − .On this basis, we also give the explicit expression for the stationary probability 0,0 P .In later section, we will see that this quantity is very useful when the cost structure will be introduced in our model.To this end, let us first define the following probability generating functions: Multiplying Equations ( 1)-( 4) by n z and summing over n, 0 n ≥ , we obtain If we add the right hand sides and left hand sides of the Equations ( 1)-( 4) and cancel the common terms, the following equality holds: Remark 3. As shown in Figure 2, according to the different service rates provided by the server, the state space of the queueing system is divided into two macro-states, namely 0 and 1.They are accessible to each other.
Thus, the mean transition rate from state ( ) is equal to the mean rate from state (1,1) to state (0,0).This fact is properly reflected in the above Equation (10).
Substituting Equation (10) into Equation ( 9) and after some algebraic manipulation, we have ( ) ( ) Using a method similar to the derivation of Equation ( 11), with Equations ( 5)-( 8), we get ( ) ( ) Thus, we can rewrite ( ) ( ) Since ( ) is the probability generating function of the queue length distribution, we have ( ) . By taking into account the normalization condition, and letting 1 z = in Equation ( 13), we have Based on Equation ( 14) the first relationship between 0,0 P and 1,0 1 .

X. D. Lin
On the other hand, with the help of Equations ( 2)-( 4), another relationship between 0,0 P and 1,0 L P − can be obtained by a backward recursion procedure.
Substituting Equation ( 16) into Equation ( 15), it follows that: Remark 4. Obviously, the queueing system for 1 2 µ µ = coincides with the classic Geo/Geo/1 queue studied by Hunter [16].Additionally, under such an assumption, after some brief algebraic manipulations, the Equation ( 17) can further be simplified as follows: 0,0 1 P ρ = − , where The formula derived here agrees with the one given by Hunter [16], and it also shows the correctness of our analysis presented above.
Having computed the stationary probabilities 0,0 P and 1,0 L P − , a recursive algorithm for computing the other steady state probabilities can be established.To demonstrate the working schemes of the recursive method, we describe the solution algorithm in the following Table 1.

Explicit Expression for the Expected Number of Customers in the System
Once the explicit expressions for 0,0 P and 1,0 L P − are given, the expected number of customers in the system can be determined from them.Let N be the number of customers in the system in steady state.We have , , , , the base case stops the recursion using Equations ( 17), ( 16) and (4).for 3 :1: P and 2,1 P using Equations ( 1) and ( 5).
for 2 :1: ) Remark 5.As a matter of fact, the explicit expression for the expected number of customers in the system has been given by Equation (18).Just because the explicit expressions 0,0 P and 1,0 L P − are slightly cumbersome to write, we do not indent to substitute Equations ( 16) and ( 17) into Equation (18).

Sojourn Time Performance
In this subsection, we deal with the customer's sojourn time W, defined as the time between the arrival epoch of a customer till the instant at which his service request is satisfied.Here, our aim is to determine the first order moment of the sojourn time.To achieve this goal, we need to introduce some auxiliary random variables., r i j W : Customer's sojourn time given that he finds the queueing system at state ( ) , i j just before his arrival and the residual service time of customer that the server is currently processing is greater than or equal to one time slot.( ) W : Conditional sojourn time of a customer who arrives at the system when its state is ( ) We also denote the corresponding z-transforms of W,  respectively.Furthermore, because of the BASTA (i.e.Bernoulli arrivals see time averages) property, we have that By differentiating Equation (19) with respect to z, and evaluating at 1 z = , we arrive at For determining the unknowns  , we apply a first-step argument and set up the fol- lowing equations.
Pr Pr Pr Pr Pr Pr Assume that a customer arrival will occur in ( ) , t t − .If prior to this arrival there are ( ) customers in the system and the server is busy with low service rate, then the departure of the customer that the server is currently processing will take place in ( ) , t t + with probability 1 µ , thus 1 µ is the probability that the above event does not occur.Hence we can easily get the following relationships For the same reason as mentioned above, when a customer arrives at the system during a busy period with high service rate, we conclude that ( ) Alternatively, we can use the memoryless property of the geometric distribution to find the z-transform of , r i j W .For For 0 j = , we have Pr Pr Similarly, for 2,3, , Differentiating both sides of Equation ( 21) and Equations ( 25)-( 28) with respect to z and evaluating at 1 z = , we can obtain the following equations for the first moment of the conditional sojourn time. 1 1 1 , 1, 2, , Therefore, from the above results and Equations ( 22)-( 24), we obtain ,0 Thus, the problem of computing the mean conditional sojourn times  can be consi- dered solved.Consequently, with the help of stationary probability , i j P , we can evaluate the expectation of the unconditional sojourn time by using Equation (20).
To demonstrate the feasibility and efficiency of the proposed algorithm, a numerical experiment is carried out on a personal computer implementing an Intel Core i5 CPU (2.7 GHz) and 4.0 GB RAM.In this example, we L = and let λ vary from 0.1 to 0.16. Figure 3 illustrates the effect of cus- tomer's arrival rate on the mean value of the unconditional sojourn time.Also, on putting 1 2 0.18 , the queueing system under consideration can be regarded as the classic Geo/Geo/1 queue with constant service rate.From Figure 3 we can conclude that setting the switching threshold for the service rate can greatly reduce the customer's average sojourn time, for example, when the customer arrival rate is 0.16, the gap between the two average sojourn times is about 25 time units.

Regeneration Cycle
Regeneration cycles are models of stochastic phenomena in which an event (or combination of events) occurs repeatedly over time, and the times between occurrences are independent and identically distributed.Models of such phenomena typically focus on determining limiting averages for costs or other system parameters.In this paper, the reason for performing regeneration cycle analysis is to determine the optimal switching threshold value L, where the high service rate is activated.
A regeneration cycle of our current model consists of a server's idle period and a server's busy period.As regeneration points, we choose the points at which the system becomes empty.There are two types of cycles depending on whether there is a change in service rate during the server's busy period.A cycle is called "type-1" if it does not include switching of the service rate; otherwise it is of "type-2" cycle.To better understand the structure of regeneration cycle, examples of the type-1 and type-2 cycles are shown in Figure 4 and Figure 5, respectively.
We denote ( ) z θ as the probability generating function of the busy period for classical Geo/Geo/1 queue.If customer arrival occurs according to a Bernoulli process with parameter λ , and the service times provided by a   .According to the model assumptions, the busy period with high service rate is only activated by L customers waiting in the queue (including the one in service).By conditioning on the duration of the remaining service time for the customer currently being served, we get where swtich P denotes the probability that the service rate does switch in a regeneration cycle.Thus, from Equation (37), we have On the other hand, let [ ] E C be the unconditional expected length of the regeneration cycle, the mean dura- tion of busy period with high service rate can also be obtained from a result of renewal theory.Using [ ] [ ] 0,0 , Comparing the right hand sides of Equations ( 38) and we see that and [ ] E N , we can try to con- struct the cost structure of this queueing system in the next section.

Optimal Switching Threshold for the Service Rate and Numerical Examples
In manufacturing process management, managers are always interested in minimizing the long-run average cost per unit time of the system.In this section, based on the performance measures that we obtained in the previous section and the renewal reward theorem, we first construct an expected cost rate function ( ) TC L for the Geo/Geo/1 queue with switching threshold for the service rate, in which a key decision variable L is considered.Here, our objective is to determine the optimal threshold value * L under some cost structure, so as to minimize the long-run average cost rate.
Let us consider the following cost elements: Utilizing the definition of each cost element listed above, the long-run average cost rate minimization problem can be illustrated mathematically as As shown in Figure 4 and Figure 5, the switching cost is incurred at most only once in a regeneration cycle, and the switching occurs with probability switch P .This is the reason why we multiply switch C by switch P in our cost structure.On the other hand, we also note that it is rather difficult to develop analytic results for the optimal value of L because the long-run average cost rate function is highly non-linear and complex.In spite of that, since L is a discrete variable, the optimal value * L may be found by using direct substitution of successive val-  As can be seen in Figure 6, we observe that this function is convex and a single relative minimum exists.The optimal value * L and the corresponding long-run average cost rate ( ) * TC L are tabulated in Table 2. From Table 2, it appears that the minimum average cost per unit time of 24.2347 is obtained with * 7 L = .

Conclusion
In this paper, we have carried out an analysis of a discrete-time infinite-buffer Geo/Geo/1 queuing system under  X.D. Lin a modified service rate switching policy that has potential applications in modeling manufacturing and telecommunication systems.We have developed a recursive method to find the steady-state queue size distribution.The recursive method is powerful and easy to implement.Further, we obtain the analytically explicit expressions for the expected number of customers in the system.Using the first-step argument, a simple algorithm for calculating the customer's mean sojourn time has been proposed.Moreover, we also performed regeneration cycle analysis of the queue to find the optimal service rate switching threshold L. Our current model is useful and significant to engineers or managers who design an efficient system with economic management.It should be pointed out that the economic importance of this model resides in the multiple applications to manufacturing processes, since most of them operate on a discrete time basis.Furthermore, the optimal control of service rate switching policy is also a main objective from the enterprise point of view.For future studies, the present investigation can be extended by incorporating bulk input or bulk service.Another area of interest may be expanding our model into Geo/G/1 type, because there will be a significant improvement inapplicability to real world system.

Figure 1 .
Figure 1.Various time epochs in late arrival system with delayed access (LAS-DA).

Figure 2 .
Figure 2. State-transition-rate diagram for the Geo/Geo/1 queue with switching threshold for service rate. for

Figure 3 .
Figure 3.The effect of λ on the mean value of the unconditional sojourn time.
the mean value of the busy period for classic Geo/Geo/1 queue.Furthermore, let I, low B and high B respectively denote the length of server's idle period and the length of busy periods with low and high service rates in a regeneration cycle.It is obvious that I follows geometric distribution, thus [ ] 1 E I λ = .Next, we derive the probability generating function of busy period with high service rate

sC
≡ setup cost per cycle; switch C ≡ switching cost for changing the service rate in a regeneration cycle; h C ≡ holding cost per customer per unit time; low C ≡ running cost per unit time when the service provides low speed service; high C ≡ running cost per unit time when the service provides high speed service.

6 .
we can obtain the results presented in Table2 and FigureThe curve representing the long-run av- erage cost rate function ( ) TC L is plotted in Figure 6 for different values of L.

Figure 6 .
Figure 6.The plot of TC(L) against the switching threshold L.

Table 1 .
Comutation of the stationary distribution , X. D. Linues of L into the long-run average cost rate function until the minimum value ( ) TC L is achieved.To illustrate the direct search algorithm described above, a numerical example is provided by considering the following cost parameters:

Table 2 .
The long-run average cost rate against the values of L.