Research Article  Open Access
Siew Khew Koh, Ah Hin Pooi, Yi Fei Tan, "New Approach for Finding Basic Performance Measures of Single Server Queue", Journal of Probability and Statistics, vol. 2014, Article ID 851738, 11 pages, 2014. https://doi.org/10.1155/2014/851738
New Approach for Finding Basic Performance Measures of Single Server Queue
Abstract
Consider the single server queue in which the system capacity is infinite and the customers are served on a first come, first served basis. Suppose the probability density function and the cumulative distribution function of the interarrival time are such that the rate tends to a constant as , and the rate computed from the distribution of the service time tends to another constant. When the queue is in a stationary state, we derive a set of equations for the probabilities of the queue length and the states of the arrival and service processes. Solving the equations, we obtain approximate results for the stationary probabilities which can be used to obtain the stationary queue length distribution and waiting time distribution of a customer who arrives when the queue is in the stationary state.
1. Introduction
The GI/G/1 queue has been given a lot of attention due to its applications to a variety of manufacturing systems, communication systems, computer systems, and other related areas. The steadystate waiting time distribution and queue length distribution are important measures of performance in this model.
When the interarrival and service times are discrete random variables, the GI/G/1 queue is referred to as a discretetime queue. Several authors determined the steadystate waiting time distribution in the discretetime GI/G/1 queue by using the WienerHopf factorization (see [1–5]), the polynomial factorization (see [6, 7]), and the matrixanalytic method (see [8, 9]).
Steadystate waiting time distribution in the continuoustime GI/G/1 queues can be found by numerical approximations based on the theory of Fredholm integral equations [10].
In Kim and Chaudhry [11], a discretetime version of the distributional Little’s law is established. Based on this law, the queue length distribution for the discretetime GI/G/1 queue may be obtained from its waitingtime distribution.
Neuts [12] approximated the general distributions of service time and arrival time with phasetype distributions and solved the steadystate distribution using the matrixgeometric approach.
In this paper, we consider a single server queue in which the service time and interarrival time are assumed to have constant asymptotic rates. Abbreviating “constant asymptotic rate” as “CAR,” we may denote the resulting queue as a CAR/CAR/1 queue and any distribution with a constant asymptotic rate as a CAR distribution. In practice many distributions such as exponential, Erlang, hyperexponential, and gamma satisfy this requirement. However the heavytailed arrival time distributions which are often used in queuing analysis are not CAR. In order to be still able to apply the results in this paper, we may attempt to approximate the nonCAR distributions by CAR distributions. For example, in the case when there is a value such that the probability of the given random variable larger than is negligible, we may approximate the distribution of by a CAR distribution by changing the rate of the distribution of at time for all to be all equal to the rate at time .
To analyze the CAR/CAR/1 queue, we may first discretize it by segmenting the time axis into a sequence of equal intervals of length . We next derive a set of equations for the stationary probabilities of the queue length and the states of the arrival and service processes. By solving the equations, we obtain approximately the stationary probabilities. Each of the probabilities may be found more accurately by an extrapolation of the probability on the values of .
From the stationary queue length distribution, we have also obtained the waiting time distribution and sojourn time distribution of a customer who arrives when the queue is in a stationary state.
The approach which is closely related to the methods in this paper is the polynomial factorization used in [7]. Instead of taking note of the three variables given by queue length and the states of the arrival and service processes at the end of the time interval of length , Haßlinger inspected the queue length and the time until the next arrival or next departure, only at time slots when an arrival occurs or a service is completed. Thus Haßlinger only needed to use a 2dimensional state representation instead of the 3dimensional state representation used in this paper. An important assumption used by Haßlinger is that the stationary probability of is given by where and are parameters, and is another parameter which depends on .
A major contribution of this paper is the introduction of an easily implementable algorithm for solving the equations for the stationary probabilities of the queue length and the states of the arrival and service processes. This algorithm may be used to investigate other queuing systems which cannot be represented as continuoustime Markov chains. For example, Yang et al. [13] studied a model given by an unreliable M/M/1 queue with a multistate server whose service rate deteriorates due to the shocks which occur randomly with random magnitudes. This algorithm has been successfully used to investigate the model given in Yang et al. [13] with the distribution of the interarrival time or service time changed to a fairly general distribution (see [14]).
2. Derivation of Equations for the Stationary Probabilities
In this section, a set of equations for the stationary probabilities of the queue length and the states of the arrival and service processes in the discretized CAR/CAR/1 queue is derived. First let be the probability density function (pdf) of the service time and the interval for . Furthermore let where is large enough such that
Suppose a service starts at time . Then, for sufficiently small , the probability that the service will be completed in the interval is approximately , and, given that the service is not completed in , the probability that the service will be completed in is approximately , where for .
For the arrival process, let be the pdf of the arrival time. Furthermore let where is large enough such that
Suppose a customer has arrived at time . Then the next customer will arrive in the interval with an approximate probability , and, given that the next customer does not arrive in the interval , the probability that the next customer will arrive in will be approximately for where for .
Given that a service starts at a time in , we may define the state number of the service process at the end of as
Next, given that a customer arrives at a time in , we may define the state number of the arrival process at the end of as
Let be the queue length at the end of and . We may refer to as the vector of characteristics of the queue at the end of .
Let be the probability that, at the end of , the number of customers in the system is (including the customer that is being served), the service process is in state , and the arrival process is in state , where , and . Assume that exists. To find the , we first make the following observations.
Suppose at the end of the system is not empty, and the service and arrival processes are in state and at the end of , respectively. Then only one of the following events can occur in the next time interval :(a) a customer enters the system with the arrival rate , and, at the end of , the vector of characteristics becomes ;(b) a customer leaves the system with the departure rate , and ;(c) no customers enter or leave the system, and .
Here, and . However if the system is empty at the end of , and the arrival process is in state , then either one of the following events may occur in ;(d) a customer enters the system with arrival rate , and ;(e) no customers enter the system, and .By setting , and letting event (b) occurs in , we get When , (8) yields, In general, for a given value of , we can likewise find the combinations of and the event in which lead to , and obtain an equation similar to (9). Thus we can obtain the following equations: When the queue length is , When the queue length is , the expressions for , , are the same as those given by (13), (14), and (15). Other when the queue length is can be computed from the equations below:
When the queue length is , the values of all the (except ) can be computed using (21) to (24) and (13) to (15), whereas those of can be computed using the following equations:
3. Stationary Queue Length Distribution
Before solving (9) to (27) in Section 2, to obtain the stationary queue length distribution, we may first let , , , and be constants and introduce the following notations:(a); (b) denotes the set of equations of the form (c) denotes the equation of the form
With the above notations, (12) to (18) in the case when can be represented as and (19) to (24) together with (13) to (15) in the case when may be represented as Furthermore (25) to (27) together with (21) to (24) and (13) to (15) in the case when may be represented as It can be shown that, from the set of equations given by (30), we can get By substituting the expression of given by (33) into (31), and solving for , we get By substituting the expression of given by (34) into (32) when and solving for , we get Next for , repeat the process of substituting the expression of given by into (32) and solving for to get When is large enough, we may set all the in (37) to be zero and obtain Substituting (38) into (37) when , we get Similarly, for , we may perform the substitution of into (37) and obtain When , (40) yields . By using the results given by and (9) to (11), we get the following system of equations: An inspection of (41) reveals that, among the equations, only of them are linearly independent. Hence, we need to include another linearly independent equation so that the resulting system of equations has a unique solution. Equating the sum of the left sides of the equations given by (40) to the sum of the right sides of (40), we get an equation of the form, where the are constants.
As , we get from (42) an equation involving only , . The equation derived from (42) and equations chosen from (41) constitute a system of equations which can be solved to yield numerical answers for , . Then using (40), we can get numerical answers for where , , and . The stationary probability that the queue length is can then be obtained as
4. Waiting Time Distribution
Suppose a customer arrives at the system which is in the stationary state. Let the time of arrival of the customer be denoted as . Furthermore, let be the time the customer needs to wait before being served and the cumulative distribution function (cdf) of the waiting time .
To find the waiting time distribution , we first note that when the system is in the stationary state, an arrival of a customer at time which is inside an interval of length may occur with an approximate probability if the arrival process is in state at the beginning of the interval . Meanwhile at the beginning of , the service process may be in state where . Thus the probability that(i)the queue length at the beginning of is ,(ii)the service process is in state at the beginning of ,(iii)the arrival process is in state at the beginning of ,(iv)a customer arrives in is given approximately by Let be the pdf of the service time given that the service process is in state at the beginning of , and the fold convolution of . The customer who arrives in (see (iv)) under the conditions given by (i), (ii), and (iii) above will have a waiting time of zero if , and a waiting time in which the pdf is given by the convolution if . The cdf is then given approximately by The cdf may also be computed approximately by a simulation procedure described below.
Suppose a customer arrives at time and the next th customer arrives at time , where are independent and identically distributed with pdf . Next let the service time of the next th customer be of which are independent and identically distributed with pdf . For a chosen large integer , the value of is generated and the following waiting times are obtained: where is the waiting time of the th customer. Then
5. Numerical Examples
Let denote a gamma distribution of which is the shape parameter and is the scale parameter. The related probability density function is then given by . Consider an example in which the service time () has a gamma distribution with the parameter vector , and the interarrival time () has another gamma distribution with the parameter vector . The utilization factor will then be . The reason for considering gamma distribution with fractional values of the shape parameter is that the term appearing in the pdf and will usually make the existing analytical methods for finding queue length distribution fail. The reason behind such failure is that when we set , the function cannot be expressed as a finite sum of products of a function of alone and a function of alone.
The following is a procedure to find the values of and (or ). Initially we find the value of such that the rates at time exhibit small variations. A small fractional value (e.g., 0.1 or 0.05) is assigned to and is then obtained as the integer which is approximately equal to . If is very large (e.g., ), then a bigger unit is chosen for until . It can be shown that when , suitable values of and are, respectively, and . By using the proposed numerical method, the stationary queue length distribution is found. The stationary queue length distribution may also be computed using the simulation procedure in the software “QtsPlus” (accompanying software for Gross and Harris [15]) when the number of runs is . The results obtained are shown in Table 1.

From Table 1, we see that the stationary queue length distribution obtained using the proposed numerical method is close to that obtained from the software “QtsPlus.”
Table 2 shows the stationary queue length probabilities found by the numerical method using . The reason for choosing is that, for each given value of , the value of will then be of the order of about which is small enough for us to get, for , a value of which is accurate up to the 6 decimal points. For a given value of the queue length, we may use the values in the columns under and in Table 2 to fit a polynomial of low degree in to the queue length probability. The value given by the fitted polynomial when will represent the final result based on the numerical method for the queue length probability. It can be shown that, for all the values of considered, the fitted polynomials are all very satisfactory. Thus the final results at would be quite accurate. Table 2 shows that the numerical results for queue length , and 4 agree quite well with those based on “QtsPlus.” Meanwhile the numerical results for queue length , and 7 are somewhat more accurate than those based on simulation.

Table 3 shows that the stationary waiting time distribution obtained by using the numerical method in Section 4 is close to that obtained by the simulation procedure.

6. Discrete Time GI/G/1 Queue
The stationary queue length and stationary waiting time distributions of a discrete time GI/G/1 can also be found by using the proposed numerical method in Sections 2 to 4 after some modifications of the equations for the stationary probabilities given in Section 2. An explanation of why the above modifications are necessary is as follows.
First we note that the values of the (or ) for the discrete service time (or arrival time) distribution are such that all the (or ) are zero except for the cases when coincides with the service time (or arrival time) which has a nonzero probability of occurrence. Let the values of such be denoted by . The value of a typical will be such that is a constant. This means that when is made very small, the value of will have to be inflated correspondingly. Thus, if the system is not empty at time , the simultaneous occurrence of the events that(A)a customer arrives within the interval , and(B)a service is completed within the interval may not tend to zero when tends to zero. Thus the equations for stationary probabilities given in Section 2 need to be modified by taking into account of the simultaneous occurrence of events (A) and (B). The modified version of the equations in Section 2 is as follows.
When the queue length is , the values of the can be found from (9)–(11). When the queue length is , the expressions for , , are the same as those given by (16)–(18), whereas can be computed from the equations as follows: When , the values of the , can be computed using (50)(51), while (19)(20) can be used to find the values of , . The values of the and can be obtained from (22) and (23), respectively. All the other values of can be computed using the following equations: For , the values of all the (except ) can be computed using (22)(23) and (50)–(54), whereas those of can be computed using the following equations:
We may solve the above equations by using the proposed numerical method in Section 3 to obtain all the values of and subsequently the stationary queue length distribution. From the values of the stationary probabilities, we can find the stationary waiting time distribution by using the method proposed in Section 4. The cdf for the discrete time GI/G/1 queue is now given approximately by
Table 4 shows the results of the stationary queue length distribution computed using the proposed numerical method. The table also shows the results given in [11] where the authors found the stationary queue length distribution from the sojourn time distribution using the distributional Little’s law. The functions and appearing in Tables 4−6 and 89 are, respectively, the probability generating functions of the discrete service time and interarrival time.

From Table 4, we can see that the queue length probabilities obtained by using the proposed numerical method are close to those given in [11]. When , the values of the (or ) are able to capture all the details of the discrete distribution of the service time (or arrival time). Thus the results given in columns 2 and 3 in Table 4 are identical.
Tables 5 and 6 show the stationary queue length distribution in three other examples of discrete queue.
