This paper analyzes a discrete-time Geo a/Geo b/1 queuing system with batch arrivals of fixed size a, and batch services of fixed size b. Both arrivals and services occur randomly following a geometric distribution. The steady-state queue length distribution is obtained as the solution of a system of difference equations. Necessary and sufficient conditions are given for the system to be stationary. Besides, the uniqueness of the root of the characteristic polynomial in the interval (0, 1) is proven which is the only root needed for the computation of the theoretical solution with the proposed procedure. The theoretical results are compared with the ones observed in some simulations of the queuing system under different sets of parameters. The agreement of the results encourages the use of simulation for more complex systems. Finally, we explore the effect of parameters on the mean length of the queue as well as on the mean waiting time.
Queuing systems are being studied from the beginning of the XX century, trying to model fluctuations arising in a queue, such as number of clients, waiting time, service time, etc. In fact, its origin can be found in the study of congestion problems on telephone networks. The first work on these problems was published in 1909 by Agner Krarup Erlang [
Besides the queues in the telephone exchanges that were the origin of the theory of queues, nowadays, batch queues are common in transports, distribution and telecommunications. Consequently, the words “customers” or “clients” are used in general terms, whether it is about people or, say, printer jobs waiting their turn. Chaudhry and Templeton [
In particular, models in discrete-time have received much attention in the last years, mostly focusing on the calculus of the length queue distribution and waiting time. In pursuing these goals, different approaches are used and with different models that arise when considering varying characteristics such as the service policy, or distribution of arrivals or of inter-arrivals time.
The present paper focuses on discrete time queuing systems for batch arrivals and services, both following a geometric distribution and both with fixed size, though different. Following the notation introduced by Kendall [
A general discrete time GIX/GY/1 model was studied by Alfa and He [
In [
For arrivals of variable size following a Poisson distribution, Claeys et al. [
A related model, the GI/Geo/1 queue under N-policy, is analyzed by Lim et al. [
Bruneel et al. [
On the other hand, Yao et al. [
The present paper shows how the specific characteristics of the model being studied make it possible to obtain the theoretical solutions without the need of finding all the roots of A ( z ) − z s . In particular, we prove the existence of a unique root of the characteristic polynomial in (0, 1), which suffices to obtain the solution, provided the system is stationary. Conditions for the system to be stationary are also given.
The model under study is a discrete-time model denoted as Geoa/Geob/1: customers arrive in batches of fixed size a, and they are served in batches of fixed size b. If there is not at least b customers at the end of a service, the server does not start a new service until the number of waiting customers reaches b.
It is assumed that arrivals are independent, service times are independent, and services are independent of arrivals unless the number of waiting customers were less than b. Batch sizes a and b are supposed to be relative prime integers; otherwise, we would take their greatest common divisor as unit.
At each time slot, an arrival of a customers can happen with probability α, or not, with probability 1 − α. Furthermore, if there is a service running, it can finish with probability β, or not, with probability 1 − β. Besides, inter-arrival times are geometrically distributed with parameter α, and service times are geometrically distributed with parameter β. Then, traffic density is ρ = a α / ( b β ) .
Therefore, if there are n customers in the system, Equations (1) to (4) describe the different probabilities.
P ( A ¯ E ¯ | n ) = P ( no arrival and no exit | n ) = { 1 − α if n < b ( 1 − α ) ( 1 − β ) if n ≥ b (1)
P ( A E ¯ | n ) = P ( arrivalandnoexit | n ) = { α if n < b α ( 1 − β ) if n ≥ b (2)
P ( A ¯ E | n ) = P ( noarrivalandexit | n ) = { 0 if n < b ( 1 − α ) β if n ≥ b (3)
P ( A E | n ) = P ( arrivalandexit | n ) = { 0 if n < b α β if n ≥ b (4)
If the system reaches an equilibrium, and q n denotes the probability of “n customers in the system”, then the following Equation (5) holds,
q n = P ( A ¯ E ¯ | n ) q n + P ( A E ¯ | n − a ) q n − a + P ( A ¯ E | n + b ) q n + b + P ( A E | n − a + b ) q n − a + b (5)
where all q m with m < 0 are 0.
Note that, according to Equations (1)-(4), the probabilities of exit or no exit are different depending on whether n is greater than b or not. Consequently, all equations with any subscript less than b are particular equations.
Since the lowest subscript in Equation (5) is n − a , then there are a + b particular equations, from n = 0 to n = a + b − 1 , the ones detailed in Equations (6)-(11), that also should take into account the relation between a and b.
If a < b the particular equations are:
q n = ( 1 − α ) q n + ( 1 − α ) β q n + b if 0 ≤ n < a (6)
q n = ( 1 − α ) q n + α q n − a + ( 1 − α ) β q n + b + α β q n − a + b if a ≤ n < b (7)
q n = ( 1 − α ) ( 1 − β ) q n + α q n − a + ( 1 − α ) β q n + b + α β q n − a + b if b ≤ n < a + b (8)
While, if a > b , the equations become:
q n = ( 1 − α ) q n + ( 1 − α ) β q n + b if 0 ≤ n < b (9)
q n = ( 1 − α ) ( 1 − β ) q n + α q n − a + ( 1 − α ) β q n + b if b ≤ n < a (10)
q n = ( 1 − α ) ( 1 − β ) q n + α q n − a + ( 1 − α ) β q n + b + α β q n − a + b if a ≤ n < a + b (11)
Remember that a and b are supposed to be relative prime integers. In any case, if a = b , the problem would reduce to a = b = 1 , since each block can be treated as a unit.
The general equation for n ≥ a + b in Equation (12) is obtained by replacing Equations (1)-(4) into Equation (5).
q n = ( 1 − α ) ( 1 − β ) q n + α ( 1 − β ) q n − a + ( 1 − α ) β q n + b + α β q n − a + b (12)
Rearranging Equation (12), for n ≥ a + b we have
α ( 1 − β ) q n − a + [ ( 1 − α ) ( 1 − β ) − 1 ] q n + α β q n − a + b + ( 1 − α ) β q n + b = 0 (13)
whose characteristic polynomial is
P ( x ) = α ( 1 − β ) + [ ( 1 − α ) ( 1 − β ) − 1 ] x a + α β x b + ( 1 − α ) β x a + b (14)
The steady-state distribution, if it exists, must be a particular solution of Equation (13). The general solution of Equation (13) is
q n = ∑ k a k x k n ( n ≥ a + b ) (15)
where x k are the roots of the characteristic polynomial in Equation (14).
Then, with the aim of studying the roots of Equation (14), we distinguish two cases:
Case a < b :
If a < b , exponents in Equation (14) are already in ascending order. As a polynomial, P(x) is continuous and differentiable in R. Its derivative:
P ′ ( x ) = a [ ( 1 − α ) ( 1 − β ) − 1 ] x a − 1 + b α β x b − 1 + ( a + b ) ( 1 − α ) β x a + b − 1 = x a − 1 ( − a [ 1 − ( 1 − α ) ( 1 − β ) ] + b α β x b − a + ( a + b ) ( 1 − α ) β x b ) = x a − 1 Q ( x ) (16)
Since a, b are positive integers and α, β lay in ( 0 , 1 ) , Q(x) is an increasing polynomial and Q ( 0 ) < 0 . Consequently, Q(x) has a unique root in ( 0 , ∞ ) , and so has P ′ ( x ) . Therefore P(x) has at most two roots in (0, ∞).
Case a > b :
If a > b , Equation (17) shows P(x) with exponents in ascending order
P ( x ) = α ( 1 − β ) + α β x b + [ ( 1 − α ) ( 1 − β ) − 1 ] x a + ( 1 − α ) β x a + b (17)
Its derivative:
P ′ ( x ) = b α β x b − 1 + a [ ( 1 − α ) ( 1 − β ) − 1 ] x a − 1 + ( a + b ) ( 1 − α ) β x a + b − 1 = x b − 1 ( b α β − a [ 1 − ( 1 − α ) ( 1 − β ) ] x a − b + ( a + b ) ( 1 − α ) β x a ) = x b − 1 S ( x ) (18)
By differentiating again the polynomial S(x) in Equation (18) is
S ′ ( x ) = − a ( a − b ) [ 1 − ( 1 − α ) ( 1 − β ) ] x a − b − 1 + a ( a + b ) ( 1 − α ) β x a − 1 = x a − b − 1 ( − a ( a − b ) [ 1 − ( 1 − α ) ( 1 − β ) ] + a ( a + b ) ( 1 − α ) β x b ) (19)
Like in the previous case, the second factor in Equation (19) is an increasing polynomial in the interval ( 0 , ∞ ) and negative at 0. Then S ′ ( x ) has a unique root in ( 0 , ∞ ) . This implies that S(x) has at most two roots in ( 0 , ∞ ) , and so has P ′ ( x ) .
As S ′ ( x ) must be negative until its unique root in ( 0 , ∞ ) and positive after the root, S(x) is at first decreasing, it reaches a relative minimum and turns to be an increasing function. Because S ( 0 ) > 0 , either S(x) remains above the horizontal axis in ( 0 , ∞ ) (maybe reaching it at the relative minimum), or S(x) crosses the horizontal axis at two different points. However, the former case would imply, Equation (18), that P ′ ( x ) is never negative and therefore P(x) is never decreasing, which is impossible because P ( 0 ) > 0 and P ( 1 ) = 0 .
We conclude that S(x), and thus P ′ ( x ) , crosses the horizontal axis at two different points s 1 and s 2 ( s 1 < s 2 ) in ( 0 , ∞ ) with P ′ ( x ) > 0 if x ∈ ( 0 , s 1 ) , P ′ ( x ) < 0 if x ∈ ( s 1 , s 2 ) and again P ′ ( x ) > 0 if x > s 2 . Accordingly, P(x) is increasing in ( 0 , s 1 ) ∪ ( s 2 , ∞ ) and decreasing in ( s 1 , s 2 ) . In the first interval ( 0 , s 1 ) there are no roots of P(x), since P ( 0 ) > 0 . Consequently, as the only remaining root of P ′ ( x ) is in ( s 1 , ∞ ) , in this interval there is at most two roots of P(x).
This proves that, in both cases (a < b and a > b), the characteristic polynomial has at most two roots in the interval ( 0 , ∞ ) . One of these roots is 1. Let’s suppose that there is another root, r, in ( 0 , 1 ) . Then P(x) must be decreasing up to r and then increasing until 1. In particular, P ′ ( 1 ) must be positive but
P ′ ( 1 ) = a [ ( 1 − α ) ( 1 − β ) − 1 ] + b α β + ( a + b ) ( 1 − α ) β = a − a α − a β + a α β − a + b α β + a β − a α β + b β − b α β = − a α + b β (20)
Thus, P ′ ( 1 ) > 0 ⇔ a α < b β .
This implies that if a α ≥ b β there are no roots of P(x) in ( 0 , 1 ) , then we cannot construct any particular solution as defined in Equation (15), which should generate a probability distribution.
On the other hand, if a α < b β , the root of P(x) in ( 0 , 1 ) is unique, let x 0 be this root. Therefore, for any u > 0 , the series in Equation (21)
∑ n = a + b ∞ u x 0 n (21)
is a convergent series with positive terms.
We can choose u > 0 such that
∑ n = 0 ∞ q n = ∑ n = 0 a + b − 1 q n + ∑ n = a + b ∞ u x 0 n = 1 (22)
that is,
u = ( 1 − ∑ n = 0 a + b − 1 q n ) 1 − x 0 x 0 a + b (23)
Inserting Equation (23) into Equation (21), the probability of n customers in the system, qn, for n ≥ a + b , is written as
q n = ( 1 − ∑ n = 0 a + b − 1 q n ) 1 − x 0 x 0 a + b x 0 n (24)
depending on the first a + b probabilities of the succession. We determine these initial probabilities by replacing q n ( n ≥ a + b ) in the a + b particular equations and solving the resulting linear system.
This provides the distribution of the total number of customers. Equation (25) gives the rule to obtain the number of customers waiting before service
P ( n customerswaiting ) = { q n + q n + b if n < b q n + b if n ≥ b (25)
Also from these equations, the average number of customers in the system is calculated by summing up the series in the following Equation (26)
∑ n = 0 a + b − 1 n q n + ∑ n = a + b ∞ n ( 1 − ∑ n = 0 a + b − 1 q n ) 1 − x 0 x 0 a + b x 0 n (26)
which gives
∑ n = 0 a + b − 1 n q n + ( 1 − ∑ n = 0 a + b − 1 q n ) ( x 0 1 − x 0 + a + b ) (27)
and the average number of customers waiting is computed as:
∑ n = 0 b − 1 n ( q n + q n + b ) + ∑ n = b ∞ n q n = ∑ n = 0 b − 1 n q n + ∑ n = 0 a + b − 1 n q n + b + ( 1 − ∑ n = 0 a + b − 1 q n ) ( x 0 1 − x 0 + a ) (28)
The equations in the previous section can be easily programmed so they provide a straightforward procedure to compute the probabilities. To illustrate the procedure as well as comparing the probabilities against relative frequencies computed with simulated results (for the cases where the probabilities were difficult to compute), we simulated the system for some given values and for around 10,000 cycles. For the sake of simplicity, the case studies we chose were with small numbers for a and b.
For the first simulation study, fixed values of a = 2, b = 3 and β = 0.5 were chosen. By varying α, the traffic density ρ is also varied, so that, for instance, α = 0.3 gives ρ = 0.4, whereas for α = 0.6, a heavier traffic density, ρ = 0.8, is obtained.
In both cases, the first five probabilities { q 0 , q 1 , q 2 , q 3 , q 4 } come from the linear system obtained by substituting and rearranging Equations (6)-(8), and (12) (the details can be found in [
( − α 0 α 1 − α 2 0 0 − α 0 0 1 − α 2 α 0 − α α 2 0 0 α 0 − 1 + α 2 α 2 0 0 α 0 − 1 + α 2 ) ( q 0 q 1 q 2 q 3 q 4 ) − 1 − α 2 ( 1 − x 0 ) ( 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 x 0 x 0 x 0 x 0 x 0 x 0 2 x 0 2 x 0 2 x 0 2 x 0 2 ) ( q 0 q 1 q 2 q 3 q 4 ) − α 2 ( 1 − x 0 ) ( 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 ) ( q 0 q 1 q 2 q 3 q 4 ) = − 1 − α 2 ( 1 − x 0 ) ( 0 0 1 x 0 x 0 2 ) − α 2 ( 1 − x 0 ) ( 0 0 0 0 1 )
where x0 is the root of the characteristic polynomial in ( 0 , 1 ) . The remaining probabilities obey to the geometric formula in Equation (24) with a + b = 5 .
For the total number of customers, relative frequencies obtained when simulating the systems with the same parameters are shown in
Total number of customers | Probability | Waiting customers | Probability | ||
---|---|---|---|---|---|
α = 0.3 | α = 0.6 | α = 0.3 | α = 0.6 | ||
(ρ = 0.4) | (ρ = 0.8) | (ρ = 0.4) | (ρ = 0.8) | ||
0 | 0.1520 | 0.0362 | 0 | 0.2823 | 0.1449 |
1 | 0.1742 | 0.0476 | 1 | 0.3235 | 0.1904 |
2 | 0.2789 | 0.1220 | 2 | 0.3320 | 0.2161 |
3 | 0.1303 | 0.1087 | 3 | 0.0286 | 0.0778 |
4 | 0.1493 | 0.1428 | 4 | 0.0155 | 0.0643 |
5 | 0.0530 | 0.0941 | 5 | 0.0084 | 0.0531 |
6 | 0.0286 | 0.0778 | 6 | 0.0045 | 0.0439 |
7 | 0.0155 | 0.0643 | 7 | 0.0024 | 0.0363 |
8 | 0.0084 | 0.0531 | 8 | 0.0013 | 0.0300 |
9 | 0.0045 | 0.0439 | 9 | 0.0007 | 0.0248 |
10 | 0.0024 | 0.0363 | 10 | 0.0004 | 0.0205 |
11 | 0.0013 | 0.0300 | 11 | 0.0002 | 0.0170 |
12 | 0.0007 | 0.0248 | 12 | 0.0001 | 0.0140 |
13 | 0.0004 | 0.0205 | 13 | 0.0001 | 0.0116 |
14 | 0.0002 | 0.0170 | 14 | 0.0000 | 0.0096 |
15 | 0.0001 | 0.0140 | 15 | 0.0000 | 0.0079 |
16 | 0.0001 | 0.0116 | 16 | 0.0000 | 0.0065 |
17 | 0.0000 | 0.0096 | 17 | 0.0000 | 0.0054 |
18 | 0.0000 | 0.0079 | 18 | 0.0000 | 0.0045 |
19 | 0.0000 | 0.0065 | 19 | 0.0000 | 0.0037 |
20 | 0.0000 | 0.0054 | 20 | 0.0000 | 0.0031 |
21 | 0.0000 | 0.0045 | 21 | 0.0000 | 0.0025 |
22 | 0.0000 | 0.0037 | 22 | 0.0000 | 0.0021 |
23 | 0.0000 | 0.0031 | 23 | 0.0000 | 0.0017 |
24 | 0.0000 | 0.0025 | 24 | 0.0000 | 0.0014 |
25 | 0.0000 | 0.0021 | 25 | 0.0000 | 0.0012 |
26 | 0.0000 | 0.0017 | 26 | 0.0000 | 0.0010 |
27 | 0.0000 | 0.0014 | 27 | 0.0000 | 0.0008 |
28 | 0.0000 | 0.0012 | 28 | 0.0000 | 0.0007 |
29 | 0.0000 | 0.0010 | 29 | 0.0000 | 0.0006 |
30 | 0.0000 | 0.0008 | 30 | 0.0000 | 0.0005 |
Expectation: | 2.4319 | 6.4891 | Expectation: | 1.2473 | 4.1065 |
In the case of lighter traffic density, ρ = 0.4, the first a + b = 5 probabilities, up to 4 customers in the systems, are comparatively large in the system, which is very clear in
probabilities greater than 0.13 for less than 5 customers; while the remaining ones are very small, graphically inappreciable in
The magnitude of q2 stands out clearly that reflects the fact that there must be at least 3 customers for the system to start, but they arrive two by two with less probability. Accordingly, the expectation is near 2 customers (2.4319 as can be seen in the last row of
This same reason explains the behavior of the probabilities for the number of customers that are waiting their turn, fourth column in
Comparatively, the probabilities of the number of customers waiting for their turn are much larger than the probabilities of n customers in the system just up until 2 customers waiting. Then, the probabilities rapidly decrease and are obviously less than their counterparts, and the expectation drops to 1.2473 customers.
The difference between the first five probabilities and the remaining ones is less pronounced in the second system, with ρ = 0.8, as we can see in both
In
Like in the previous case, when looking at the probability of having n customers waiting for their turn, the largest probability is for 2 customers waiting, and again the largest probabilities are indeed for 0, 1 or 2 customers waiting but now the cumulative probability is much smaller, 0.5514, and slowly decreases from 3 customers waiting their turn.
In this case, larger traffic density due to the larger probability of arrival causes larger probabilities for more customers in the system, and thus for the ones waiting because the probability of service is the same.
For illustrative purposes,
As in the previous case, the results for the number of total customers in the system are similar to each other, although in this case, shorter simulation times showed rather high differences between the theoretical probabilities and the observed frequencies.
The largest values are in the first three rows, up to 2 customers in the system, especially for exactly 2 customers. This is due to the needed 3 = b customers to start the service. Then, from 3 to 7 = a + b − 1 customers, both the probabilities and observed frequencies increase and, at that point, they start to slowly decrease from 8 customers in the system, although the high traffic density diffuses the differences among these three block of probabilities (or frequencies). The slow decrease is apparent in the fact that there is still 0.8 in a thousand rate of having 70 customers in the system, with almost 16 customers on average, 15.72 according to
With the aim of studying the effect of the parameters on the waiting time and on
Total number of customers | Relative frequency | Probability | Total number of customers | Relative frequency | Probability |
---|---|---|---|---|---|
0 | 0.0446 | 0.0470 | 36 | 0.0038 | 0.0064 |
1 | 0.0472 | 0.0544 | 37 | 0.0051 | 0.0061 |
2 | 0.0781 | 0.0729 | 38 | 0.0038 | 0.0057 |
3 | 0.0296 | 0.0294 | 39 | 0.0049 | 0.0053 |
4 | 0.0308 | 0.0340 | 40 | 0.0062 | 0.0050 |
5 | 0.0476 | 0.0455 | 41 | 0.0047 | 0.0047 |
6 | 0.0476 | 0.0477 | 42 | 0.0055 | 0.0044 |
7 | 0.0593 | 0.0553 | 43 | 0.0061 | 0.0042 |
8 | 0.0378 | 0.0373 | 44 | 0.0054 | 0.0039 |
9 | 0.0390 | 0.0350 | 45 | 0.0039 | 0.0037 |
10 | 0.0333 | 0.0329 | 46 | 0.0040 | 0.0034 |
11 | 0.0343 | 0.0309 | 47 | 0.0066 | 0.0032 |
12 | 0.0316 | 0.0290 | 48 | 0.0038 | 0.0030 |
13 | 0.0310 | 0.0273 | 49 | 0.0043 | 0.0029 |
14 | 0.0234 | 0.0256 | 50 | 0.0028 | 0.0027 |
15 | 0.0265 | 0.0240 | 51 | 0.0031 | 0.0025 |
16 | 0.0258 | 0.0226 | 52 | 0.0020 | 0.0024 |
17 | 0.0228 | 0.0212 | 53 | 0.0021 | 0.0022 |
18 | 0.0233 | 0.0199 | 54 | 0.0022 | 0.0021 |
19 | 0.0213 | 0.0187 | 55 | 0.0005 | 0.0020 |
20 | 0.0157 | 0.0176 | 56 | 0.0017 | 0.0018 |
21 | 0.0195 | 0.0165 | 57 | 0.0013 | 0.0017 |
22 | 0.0174 | 0.0155 | 58 | 0.0004 | 0.0016 |
23 | 0.0159 | 0.0146 | 59 | 0.0011 | 0.0015 |
24 | 0.0139 | 0.0137 | 60 | 0.0004 | 0.0014 |
25 | 0.0144 | 0.0128 | 61 | 0.0004 | 0.0013 |
26 | 0.0121 | 0.0121 | 62 | 0.0002 | 0.0013 |
27 | 0.0093 | 0.0113 | 63 | 0.0002 | 0.0012 |
28 | 0.0126 | 0.0106 | 64 | 0.0005 | 0.0011 |
29 | 0.0076 | 0.0100 | 65 | 0.0001 | 0.0010 |
30 | 0.0095 | 0.0094 | 66 | 0.0003 | 0.0010 |
31 | 0.0066 | 0.0088 | 67 | 0.0003 | 0.0009 |
32 | 0.0055 | 0.0083 | 68 | 0.0001 | 0.0009 |
33 | 0.0078 | 0.0078 | 69 | 0.0002 | 0.0008 |
34 | 0.0046 | 0.0073 | 70 | 0.0004 | 0.0008 |
35 | 0.0048 | 0.0069 | Expectation | 14.4072 | 15.72 |
the length of the queue, in the two possible situations a < b and b < a, we maintain the size of the blocks, both entering and exiting the system (the same as in section 3.1), and modify the rate service from high (β = 0.2), medium (β = 0.5), to low (β = 0.7). Systematic simulations were conducted by varying the arrival rate α with the purpose of obtaining traffic density ρ from 0.1 to 0.9.
In all cases, Equations (27) and (28) are used to compute the average number of customers in the system and the average number of customers waiting for service. The average waiting time is obtained from simulations. From it, we also computed the percentage of this waiting time due to the insufficient number of customers in the system, i.e. the time spent waiting for the system to reach the minimum b customers to start a service.
The first situation corresponds to the first system studied in Section 3.1, and thus the simulations and computation are made by keeping a = 2 and b = 3. The resulting waiting time is depicted in
It is observed that, for all values of β, the average waiting time is rather high for low traffic density, probably because customers must wait until the number reaches the minimum b. It decreases slowly until medium traffic density and then grows up again for high traffic density due to congestion.
As expected, irrespective of ρ, the time spent waiting decreases with β but not in the same way. The differences among mean times for each r are larger when comparing the slow service (open circles) to the medium one (crosses), with smaller differences between the later and the fast one (filled circles). These distances are remarkably different above all for extreme traffic densities, notice the huge increase in the mean for ρ = 0.9 when β = 0.7 as compared to the medium or the fast one.
This behavior is related to the one depicted in
waiting time that is due to inactivity of the server because there are not enough customers waiting for service, i.e., the server is free but the customers are waiting. At low traffic density, irrespective of the probability of the service, most of the waiting time is due to the insufficient number of customers. For instance, when ρ = 0.1, more than 90% of the total waiting time is spent with the server idle (thus, the increasing mean value of the total waiting time already observed in
Between these two extreme values,
In relation to the number of customers in the system,
More interestingly, the rate service β does not seem to have a high influence on the mean values for small traffic densities, regardless of whether we look at the number of total customers or at the number of those that are waiting. In fact, all the mean values are graphically undistinguishable until approximately ρ =
0.4. From then, the mean values continue their increase with ρ, but the differences due to β become apparent, more importantly for largest traffic densities and also with more dispersion, that is, larger differences among the averages correspond to larger values of ρ, though this dispersion is the same for the total number of customers or when we look only to the customers that are waiting.
Similar conclusions can be drawn with systems with a = 5 and b = 3 (then a > b, as opposed to the preceding case).
The pattern in
It might seem obvious because customers arrive in groups of a = 5 instead of a = 2 but, in fact, it was not so predictable because the rate of arrival is also varying so that the product remains constant; in other words, there are more clients arriving but they do so more spaced in time to maintain traffic density. When comparing to
traffic density increases (up to a point).
Comparing to the case a < b, the decreasing pattern for all values of β is now less linear than it appeared to be in
Comparing with the preceding system, we see that the number of customers increases with ρ but more rapidly than in
and, consequently, there are more customers (both total customers and the number of them that are waiting). It is remarkable the sharp increase of the mean number of customers when passing from traffic density 0.8 to 0.9.
In this paper, we present a discrete-time queuing system where both arrivals and services occur in batches of fixed size, a for arrivals and b for services, following geometric distributions with probabilities α and β, respectively. In this situation, analytical solutions are obtained by explicitly establishing the steady-state equations, and proving that that the system is stationary if and only if a α < b β .
The specific features of the system allow us avoiding the search of all roots of A ( z ) − z s when computing probabilities, since it suffices to take into account the unique real root of the characteristic polynomial in (0, 1), if the stationary requirements hold.
Simulations of the system conducted under different sets of parameters show the similarity between the theoretical probabilities and the relative frequencies obtained in the numerical results. This provides a viable alternative to avoid solving the whole set of equations for each particular case, and then compute the remaining probabilities.
Finally, we explore the effect of parameters on the mean queue length and the mean waiting time. For the cases studied, on average, the waiting time is observed to be high for low traffic density, due probably to the time spent waiting for a group of b customers that starts the server. Then the time decreases as the traffic density increases up to a point where the congestion of the system gives rise again to longer waiting times.
Special thanks are due to Dr. M.ª Cruz Valsero, former professor at the University of Valladolid (Spain) for her encouragement, and useful discussions.
Lorente, A. and Sánchez, M.S. (2018) Effect of Parameters on Geoa/Geob/1 Queues: Theoretical Analysis and Simulation Results. Applied Mathematics, 9, 153-170. https://doi.org/10.4236/am.2018.92011