Sunteți pe pagina 1din 13

Server Unavailability May Reduce Mean Waiting Time

in Some Batch Service Queueing Systems *

Ho Woo Lee **
Department of Industrial Engineering
Sung Kyun Kwan University
Su won, KOREA 440-746
E-mail: HWLEE@YURIM.SKKU.AC.KR

Dong II Chung
Information Service Dept.
System Integration Division
Hyundai Information Technology Co., Ltd.
Seoul, KOREA

Soon Seok Lee


Switching Method Section
ETRI
Tae Jon, KOREA 305-606

K. C. Chae
Department of Industrial Management
KAIST
Tae Jon, KOREA 305-701

ABSTRACT

This paper shows that in some queueing systems server unavailability is sometimes beneficial both
to the system owner and to the customers. To be more specific, mean queue size and mean waiting
time may decrease even in the presence of server vacations. To this end, we first derive the
probability generating function of the batch service M/ GB /1 system with multiple vacations. We,
then, consider the case of exponential service and vacation times and show that under some parameter
combinations, mean queue size decreases even in the presence of server vacations. We discuss the
implications of this phenomenon.

KEYWORDS
Batch service, vacations, mean queue size

* This is a part of the research supported by the Korea Science and Engineering Foundation(KOSEF),
Grant #921-0900-003-2
** Correspondence
1. INTRODUCTION

It is a general b~lief to the queueing theoreticians and practitIOners that server unavailability
increases the mean queue size and accordingly the mean waiting time. This is true for all single-unit
service systems such as Cli C/l vacation systems, N-policy queues, setup-time systems,
limited-servicesystel11s,to name a few. But we have found that it is no longer true for some batch
service systems. We will show in this paper that there exist some batch service systems in which
server vacations decreases the mean waiting time. If it is true, the system owner may need to force
the server to take a vacation and postpone starting the service until some conditions are met.
As a preview concerning this phenomenon, let us consider the following sample path (Figure 1) of
a batch service queueing system with server vacation. We assume that service times and vacation
times are deterministic in this particular example system. The server can take 2 customers into service
at a time. As can be seen in the figure, the mean system sojourn time (W) reduces from 1.1525
to 1.1025 'I'his ...phe~omenon . occurs because the server, after a vacation, may have a chance to start a
servicewithm()res~tol11ers.than he would have otherwise. Even though the process in the figure is
observed for a finte timedurati()n, we may anticipate the steady-state to sustain the same phenomenon.

(Figure 1. here)

For batch service systems, readers are recommended to see Bailey[I], Downton[7], Jaiswal[9],
Neuts[15], Borthakur and Medhi[2], Powell[16] and Chaudhry et al.[3]. The book by Chaudhry and
Templeton[4] summarizes the general theory concerning batch arrival/service queues.
Fuhrmann and Cooper[8] showed that the stationary queue size distribution of an MI CI 1
queuemg system with generalized server vacation is a convolution of two independent random
variables: one is the queue size of the ordinary MI C/l queue without vacations and the other is the
number of customers at an arbitrary instant at which server is not on service. Lee et al. [12][13][14]
considered batch arrival queues with N-policy and vacations. Readers are recommended to see
Doshi[6] and Takagi [17] for comprehensive surveys of vacation queues.
Very few researches dealt with batch service queues with server vacations. Dhas[5] considered
the Markovian batch service queue with vacations. He obtained the queue size distributions by
B
employing the matrix geometric method. For MI C II queue with single vacation, Lee et al.[lO]
derived the queue size distributions at both random and departure epochs, and other performance
measures. Lee et al. [11] considered the fixed size batch service systems with vacations.
In this paper, we consider a batch service queueing system with following specifications (We will
B
denote the queueing system as MI C I II MV in which 'MV stands for 'multiple vacation'):
I. Customers arrive singly by a Poisson process and are served in the order of arrival by a single
server m batches of maximum size B (B is called the "service capacity".). If less than B
customers are in the queue at a service completion point, all customers are taken into service.
2. As soon as the system empties, the server keeps taking vacations until, on returning from a
vacation, there is at least one customer waiting.
3. Customers who arrive while a group of customers are being served cannot join the on-going

- 1 -
service even if some space is available.

2. QUEUE SIZE DISTIUBUTION AT AN ARBITRARY TIME

In this section, we derive the probability generating functions of the queue size at an aribitrary
time and at a departure· epoch.
Let us define the following notations and probabilities:

A arrival rate

S service time random variable (RV)


sex) service time probability density function (pdt)

S* (S) Laplace-Stieltjes transform (LST) of service time


SR(t) remaining service time at time t

Pre n customers arrive during a service time)

v vacation time RV
vex) vacation time pdf

V*(S) LST of vacation time


R
V ( t) remaining vacation time at time t
pre n customers arrive during a single vacation)
Nit) queue size (excluding the one(s) in service) at time t

Y( t) = {I,
°, if the server is busy
if the server is on vacation

Qn(t,x)dx Pr[Y(t)=O, Nq(t)=n, x:<:;:VR(t):<:;:x+dx], n=O,1,2,'"

Qn(x) = lim Qn(t, x)


t---.'~co

Pn(t,x)dx = Pr[Y(t)=l, Nq(t)=n, x:<:;:SR(t):<:;:x+dx], n=O,1,2,'"

P n(x) = limP net, x)


{·····)co

From the definitions and probabilities defined above, we can easily derive the following
steady-state system equations:

- 2 -
-~ Po(x) = APo(x)- l~[P,(o)+ Qi(O)]S(X),

d
-dtPn(x) = A[Pn(X)-Pn-](X)]-S(X)[Pn+B(O)+Qn+B(O»), (n~1)

Taking Laplace transforms of both sides, we get

eQ~(e) - QO(0) = AQ~ (e) - V* (e)[ Po (0) + Qo(O»), (2.1 )

e Q~ ( e) - Q n (0) = A Q~ ( e) - A Q~ _] (e) , ( n ~ 1) , (2.2)

(2.3)

ep~(e)-Pn(O) = AP~(e)-AP~_] (e)-s*(e)[Pn+B(O)+ Qn+B(O»), (n~ 1).


(2.4)

We define the following generating functions:

= =
P:(z, e) 2; p~(e)zn, 2; Pn(O)Zn,
n=O n=O
= =
Q*(z, e) 2; Q~(e)Zn, 2; Qn(O)Zn.
n=O n=O

Then from eqs. (2.1) and (2.2), we get

eQ*(z, e) - Q(z, 0) = (A - Az)Q* (z, e) - V*(e)[po (0) + Qo (0)]. (2.5)

Similarly, from eqs. (2.3) and (2.4), we have

B
= (A - AZ)P:(Z, e) - s*(e) 2; [Pn
n=]
(0) + Qn (0)]
- s:~e) {[Pq (z, 0) + Q(z, 0)] - ntozn[Pn (0) + Qn(O)]}.

(2.6)
Letting e = A- AZ in eqs. (2.5) and (2.6), we get

Q(z,O) = V* (A - AZ)[ Po (0) + Qo (0)], (2.7)

S* (A - AZ) { [ V* (A - AZ) - 1] [ Po (0) + Qo (0)] + l;] (z B- Z n) [ P n (0) + Q n (0)] }


P/z,O) = ZB _ S*(A-AZ) (2.8)

- 3 -
Substituting eqs. (2.7) and (2.8) into eqs. (2.5) and (2.6), we obtain

[ V* (A - AZ) - V* (e)][Po (0) + Qo(O)]


Q*(z, e) (2.9)
e ~ A + AZ
p~(z, 8) =

[S*(A-AZ)-S*(8)] . {[Po(0)+Qo(0)][V*(A-AZ)-1]+ ntj(ZB_zn)[Pn(O) + Qn(O)]}

[ZB- S*(A-AZ)] (8 - A + AZ)

(2.10)
With e=0 in eqs. (2.9) and (2.10), we obtain

Q*(Z,O) = [1- V* (A - ~Z~o (0) + Qo(O)]


(2.11 )

[1- S* (A - AZ)] .{ [Po (0)+ Qo (0)][ V* (A - AZ) -1] + ntl (ZB - zn)[Pn(O) + Qn(O)]}
P~(z,O)= (A-AZ)[ZB_S*(A-AZ)]

(2.12)

We notice from eqs. (2.8) and (2.12) that

l-S*(A-AZ)
P;(z,O ) = (A _ AZ) S*(A _ AZ) . Pq(z,O) (2.13 )

Then, from the following identify,

Pq(Z) Q*(z,O) + P~(z, 0),


(2.14 )

we obtain the PGF of the queue size distribution,

Pq(Z) =

[Po (0) + Qo(O)][ V* (A- AZ) -1] (1- ZB) + [1- S*(A- AZ)] {rtl(zB - zn)[p n(O) + Qn(O)]}

A(1-z) [ZB_S*(A-AZ)]

(2.15)

Notice that the numerator of Pq(z) contains 2B+2 unkowns, {Pn(O), n=0,1,2, ... ,B} and

{QnCO),n=0,1,2, ... ,B}. But Qn(O) can be expressed in terms of PoCO) as

- 4 -
n=O,1,2,'" (2.16)

Thus the number of unknowns in the numerator of eq. (2.15) reduces to B+ 1. Bailey[l] showed
that for queues without vacations, those unknowns can be obtained by evaluating the numerator with
the zeros of the denominator on Izi < 1.

3. QUEUE SIZE AT A DEPARTURE EPOCH

In this section, we consider the distribution of the queue size at an arbitrary departure epoch.
The reason why we derive the queue size distribution at a departure epoch is that the queue size at an
arbitrary time is closely related to the arbitrary time distribution.

Theorem 3.1. (queue length at departure epoch)


Let p~ be the probability that n customers are left behind by a departing customer. Let

p+ (z) be its PCF defined by

p+ (z)

We then have
B
2; (zB-zn)13n+
n=l
V*('\-'\z)-1
1-130

(3.1 )

Proof. First we notice that

(3.2)

and

p+
n (n=1,2,"')

(3.3 )

Multiplying nth equation by z n and summing over n from 0 to 00, we get the desired result.


The relationship between the arbitrary time queue size distribution and the departure time
distribution is in order.

~ 5 -
Theorem 3.2. (Relationship between P q(z)an1 p+ (z))
We have the following relationship;

S*(.\ - .\z)[1- V*(.\ - .\z)](1- 130) -IPt + [1- S*(.\ - .\z)] p+ (z)
(3.4 )
.\(1- z)S*(.\ - .\z)[E(S) + E( V)Pt (1- 130) -1]

Proof. We know that n customers are left behind by a departing customer if and only if there are

n costomers in the queue just before the departure. Therefore we have

p+ n=0,1,2,·", (3.5)
n

where D is the normalization constant. Then we get

D . Piz,O) (3.6)

From P+(l) = 1 and eq. (2.8), we have

(3.7)

The theorem completes from eqs. (2.8), (2.13), (2.15) and (3.6).

Theorem 3.3.
We have

(1) Preserver is on vacation) = E( V)Po(O) (3.8)


1-130
1- 13 - E( V)Po(O)
(2) Pre server is busy) = --------0 (3.9)
1-130

and (3) Pr(system is empty) = PoCO) (3.1 0)


.\

Proof. Since Q*(z,O) is the PCF of the queue size distribution when the server is on vacation,
ex)

Q*(1, 0) = ~ Qn becomes the probability that the server is on vacation. Eq. (3.8) comes from
n=O

eqs. (2.11) and (2.16). Since P:(z,O) is the PCF of the queue size distribution when the server is

busy, P:(1, 0) becomes the probability that the server is busy which accounts for eq. (3.9). Eq.

(3.10) comes from Q*(O,O) which can be evaluated from eqs. (2.11) and (2.]6).


- 6 -
Remark 3.1. From eqs. (3.5), (3.7) and (3.9), we get

E(S)
D (3.11 )
Preserver is busy)

Relationship (3.10) can also be observed In an ordinary MI CB /1 queue (Chaudhry and

Templeton[4]) and MI CB II queue with single vacation (Lee et aI.[10]).

4. MEAN QUEUE SIZE MAY DECREASE EVEN IN THE PRESENCE OF


SERVER VACATIONS

In this section, we probe into the phenomenon that server unavailability may reduce the mean
queue size.

4.1 The MIMBI1/MV system

Let us consider MI MBI1/ MV system in which vacation times follow the exponential distribution
with mean Ih. The PGFs of queue size distributions at a departure epoch and at an arbitrary time
point are. respectively given by (from eqs. (3.1) and (2.15))

p+ (z) =

[AO - z) - (A - AZ+ Y) n~/zB- zn)13n]O -130) -lPt - (A - AZ+ y) n~l(zB - zn)p~


(A - AZ+ Y)[BpzB+ 1_ (1 + Bp)zB+ 1]

(4.1)
P/z)

{ (A - AZ + 11) (1- ZB) - (A - AZ+ '6) n::tl13n(zB - z") }po(O) (1-130) -1 - (A - AZ+ '6) ,,::t (ZB - z")P ,,(0)
1
(A - AZ+ '6) [BpZB+l_(1 + Bp)ZB+ 1]

(4.2)
A
where p = -l1-B- .

The denominator of eq. (4.1) has B-1 zeros within the unit circle (Izl < 1) and one zero on

Izl = 1 (Bailey[I]). Thus there are two zeros outside the unit circle (izi > 1), one of which is

obviously 1+ y/A. Let the other zero on Izl> 1 be zo0 Let m6 and mO be the zeros of the

numerator of eq. (4.1) and numerator of eq. (4.2) respectively which lie outside the unit circle. For
the zero of the denominator which lies within the unit circle, the numerator must vanish. Thus we

- 7 -
can express eqs. (4.1) and (4.2) as

- l
(z
= z o
z o
)(
)(
11.- 1I.z+)(
)(
z- w +
1- wt ) '
(4.3)

1-Z0 )( z-wo
P/z) ( z - z)(0 (4.4 )
A - 1I.z+ )( )( 1- )w 0 .

We can easily show that Im(zo)=O, Im(wo)=O and Im(wt)=O. Zo and 0)6 can be found

easily by a numerical search. From eqs. (3.2) and (3.3), we have the following expressions

p+ = 13 -1 wt
wt -1 ) '
Zo
o .0
( Zo
)(

(4.5)

p~ =130( Z~l)( CO/-I)' {cot,~Zoi(l-13o)n-i- ~>oi(l-13o)n-i-l}, (n=1,2,''') (4.6)

Also from eq. (4.4), we have

( 4 .7 )

(4.8)

Using equations (3.4), (3.7), (4.5) and (4.7), the following relationships between zo, 0)0 and 0)6 are

easily derived:

wto + Bp)zo (4.9)


wt + Bpzo
0) + - --~ 0) oBpzo
o - (1 + Bp)zo - 0)0 '

(4.10)

(4.11 )

4.2 Decrease in the mean queue size

Let us define .1 1 and .12 as the differences of the mean queue sizes and the variances of queue
sizes as follows:

- 8 -
L11 =L q,M!M"/l- L q,M!MB!1!MV'

L12 = Var q,M!MB/l - Var q,M!MB!1!MV'


where L q, 'I and Varq, 'I are the mean and variance of the queue size of the system Q. Then

by differentiating the PCF of the queue size distribution(eq. 4.4), we get

1 (4.12)
Bpzo + Zo - 1

zo(Bp + 1) (4.13 )
(Bpzo + Zo _1)2

From eqs. (4.12) and (4.13), we get theorem 4.1. without proof.

Theorem 4.1.
The range of CO 0 in which the mean queue size decreases IS given by

(4.14 )

Also the range of CO 0 in which the variance of the queue size decreases is given by

1+V4C+1
l<wo<1+ 2C (4.15)

where


Remark 4.1
Zo and Wo are functions of the system parameters A, 11, ¥ and B. Therefore there is no

way we can interprete !J. 1 and !J. 2 ' or we can tell under what parameter combinations we observe

the mean or variance decrease.

4.3 An example

As an example, we consider a case of B= 5 , A = 1.0. Figure 2 and 3 show L11 and L12

A
respectively as functions of ¥ and 11 (p = -&- ). According to our extensive case studies, the

- 9 -
amount of decrease is very small compared to the mean queue size of the original no-vacation
A
systems. For example, if we have p = -llB- = 0.79 the mean queue size of the ordinary

no-vacation system is 11.73 whereas the maximum decrease is less than 0.25 as can be observed in
Figure 2. The decresase in variance is also seen to be very small. For p = 0.79 , the true
vanance of the queue size of the ordinary queue without vacations is 153.08 whereas the decrease in
vanance is less than 2.5.

(Figure 2. here)

(Figure 3. here)

4.4 Implications and suggestions for further study

The implications of this phenomenon is that, in some cases, it IS more beneficial, quite contrary
to the general belief, to. force the server to delay starting service. The questions are "under what
conditions?" and "how long?". If the mean queue size decreases, the mean waiting time also decreases
from the Little's theorem. The "forced idle time" may be utilized for taking an intermittence or
attending to another task. A further economic study in this respect will benefit the system designers.

- 10 -
REFERENCES

1. Bailey, N.TJ., "On Queueing Process with Bulk Service", 1. Roy. Stat. Soc., B16, 80-87, 1954
2. Borthakur, A. and Medhi, J., "A Queueing System with Arrival and Service in Batches of
Variable Size", Trans. &i., 7, 85-99, 1973
3. Chaudhry, M.L., Madill, B.R. and Briere, G., "Computational Analysis of Steady-state Probabilities
of MI C a, b II and related Nonbulk Queues", Queueing systems: Theory and Application, 2,
93-114, 1987
4. Chaudhry, M.L. and Templeton, J.G.C., A First Course in Bulk Queues, Wiley, New York, 1983
5. Dhas, A.H., Markovian General Bulk Service Queueing Model, Ph.D. thesis, Dept. of Math., PSG
college of Tech., India, 1989
6. Doshi, B.T., "Queueing Systems with Vacations: A Survey", Queueing systems: Theory and
Application, 1, 29-66, 1986
7. Downton, F., "Waiting Time in Bulk Service Queues", 1. Roy. Stat. Soc., B17, 256-261, 1955
8. Fuhrmann, S.W. and Cooper, R.B., "Stochastic Decompositions in the MI C/l Queue with
Generalized Vacations", Opns. Res., 33(5), 1117-1129, 1985
9. Jaiswal, N.K, "Time dependent solution of the bulk service queueing problem", Opns. Res., 8,
773-781, 1960
10. Lee, H.W., Lee, S.S., Chae, K.C., and Nadarajan, R., "On a batch service queue with single
vacation", Appl. Math. Modeling, 16, 36-42, 1992
11. Lee, H.W., Lee, S.S., and Chae, KC., "Fixed-size batch service queue with vacations", 1. Appl.
Math. and Stochastic Analysis, Vol. 9, No.2, 1996

12. Lee, H.W., Lee, S.S., Chae, KC., "Operating Characteristics of MXI C/l Queue with N-policy",
Queueing systems: Theory and Application, 15, pp.387-399, 1994
13. Lee, S.S., Lee, H.W., Chae, K.C., "On a Batch Arrival Queue with N-policy and Single
Vacation", Comput. & Opel'. Res., 22(2), 173-189, 1994
14. Lee, H.W., Lee, S.S., Park Jeong Ok, Chae, KC., "Analysis of MXI C/l Queue with N-policy
and Multiple Vacations", 1. Appl. Prob., 31, 467-496, 1994
15. Neuts, M.F., "A General Class of Bulk Queues with Poisson Input", Ann. Math. Stat., 38,
757-770, 1967
16. Powell, W.B., "The Bulk Service Queue with a General Control Strategy: Theoretical Analysis and
a New Computational Opns. Res., 34(2), 267-275, 1986
Procedure",
17. Takagi, H., Queueing Analysis: A Foundation of Pe~formance Evaluation, Vol 1, Vacation and
Priority Systems, Part 1, North-Holland, 1991

- 11 -
(a)
51 52 53 54

0.71 0.97 1.51 1.89 2.31 2.66 3.11

(b)

Area = 4.41
W = 1.1025

0.71 0.97 1.892.06 2.66 2.86 3.66


.} t .}
vacation return vacation
5 I 52 53

<Figure 1> Sample path of a batch service queue


(a) without vacation
(b) with vacation

S-ar putea să vă placă și