BATCH ARRIV AL QUEUE WITH
N- POLICY AND SINGLE VACATION Soon Seok Lee ETRI Tae Jeon, KOREA 305-606 Ho Woo Lee, Seung Hyun Y oon Department of Industrial Engineering Sungkyunkwan University Su Won, KOREA 440-746 K. C. Chae of Management Science KAIST Tae Jeon, KOREA 305-701
Department
ABSTRACT We consider an as the system becomes When he returns value
MXI
ell
N-policy
he begins to serve the customers.
N
system size distribution
decomposes
into two random variables
ell
queue.
of random length
We derive the system size distribution
The interpretation
one of which is the system size of
of the other random
We also derive the queue waiting time distribution
and show that the
variab1e will also be provided.
of an arbitrary customer.
Finally we develop a
procedure to find the optimal stationary operating policy under a linear cost structure.
Key Words: MXI operating policy
ell
queue,
11:
If not, the server waits in the system until
or exceeds
MXI
As soon
if the system size is greater than or equal to a pretermined
the system size reaches
ordinary
and single vacation.
empty, the server leaves the system for a vacation
from the vacation,
N(threshold),
queueing system with
N-policy,
single
vacation,
system
Correspondence: Ho Woo Lee, Dept. of Industrial Engineering, Sung Kyun 440-746, (TEL) 82-331-290-5516, (FAX) 82-331-291-4502 (E-mail)
[email protected]
size, waiting
Kwan
Univ.,
time, optimal
Su Won,
Korea
SCOPE AND PURPOSE
This paper concerns the modeling of a production
system in which the
production does not start until some specified number of raw materials, say are accumulated during an idle period.
N,
We assume that the machine undergoes
some extra operations (for example, machine repair, preventive maintenance, etc) when there are no raw materials to process. assume that the raw materials modeled by
To be more realistic, we also
arrive in batches.
This production
MX/ G/l/ N- policy queue with single vacation.
system is
We derive the
probability distributions of the number of materials and the waiting times of each material.
We
also suggest
a procedure
to obtain
the
optimal
minimizes the long-run average cost under a linear cost structure.
N* which
1.
INTRODUCTION Consider
a manufacturing
system
in which the production
does not start until some
specified number of raw materials, say N, are accumulated during an idle period.
We assume that
the operator of the machine performs some extra operations (for example, machine repair, preventive maintenance, etc) when there are no raw materials to process. assume that the raw materials arrive in batches.
To be more realistic, we also
This production system can be modeled by an
MX I GIl queue with N-policy and single vacation. We consider a queueing
system m which customers
X.
Poisson process with random arrival size leaves for a vacation of random length
arrive according to the compound
As soon as the system becomes empty the server
V(vacation period).
When he returns from the vacation and
the system size is greater than or equal to a predetermined value
N(threshold), the server begins to
serve the customers until there is no customers to serve (busy period).
If he finds fewer customers
N, he waits in the system until the system size reaches or exceeds
than
Thus, in our system,
N (dormant period).
a vacation period, a dormant period (the length of which is zero if the
returning server finds N or more customers) and a busy period constitute a cycle. this queueing system as MXI Gill N- policyl VAC(l) vacation'.
queue in which
We will denote
'VAC(1)'
implies 'single
The system is depicted in Figure 1. Vacation queues have attracted many attentions from numerous researchers since Levy and
Yechiali [11].
Fuhrmann
vacation queues.
and Cooper
[4] proved
the well-known
"decomposition
property" for
For application of vacation models to polling systems, see Takagi [13].
For
comprehensive survey on vacation queues, see Doshi [2] and Takagi [12]. Hofri [5] studied the two the
MI G/l queue with N-policy
problem.
N-policy
queues attended by a single server.
and vacations.
Kella [6] studied
He considered the optimal policy as a stopping
Their works both dealt with single-unit arrival systems. Batch arrival queues with threshold and with/without multiple vacations were first studied
by Lee and
Srinivasan[7].
They
derived
the mean
performance
measures
procedure to find the optimal operating policy under a linear cost structure. procedure to calculate the system size probabilities.
and deveoped
the
Lee [8] developed a
Lee et al. [10] studied the same queuemg
system and found that the system size decomposes into two random variables: one is the system size of the ordinary
MX I G/l
N-l
given by
period.
l~
queue and the probability generating function
(PGF) of the other is
. N-l 'J[
j Zl
I
l~
'J[
j
where
'J[
j
is the probability that the system state visits j during an idle
They also derived the condition under which the optimal threshold
N* is found under a
linear cost structure.
While Lee and Srinivasan [7] concentrated on the mean measures such as the
mean system size, and on the calculation of state probabilities with
N-policy, this paper concentrates on the development
[8] of the multiple vacation system
of the system size and waiting time
transform solutions. and its probabilistic interpretation of the single vacation system with threshold. In this study we derive the system size distribution and show that it decomposes into two random
variables
one
of which
is the system
size of the stationary
interpretation of the other random variable will also be provided. distribution of an arbitrary customer.
MXI G/l
queue.
We also derive the waiting time
We obtain other performance measures and the condition under
which the optimal operating policy is achieved under a linear cost structure. [Figure 1. 2. THE
here]
SIZE DISTRIBUTION
In this section we derive the system equations and the distribution.
N A
X gk ak
X(z) S V sex) vex) S' (e)
V'(e)
SoU) V°Ct) N(t)
yCt) RnCt)
PGF of the system size
Let us define following notations and probabilities:
threshold group arrival rate arrival size random variable Pr(X = k) probability that k customers arrive during a vacation the probability generating function( PGF) of X service time random variable vacation time random variable the probability density function of S the probability density function of V the Laplace - Stieltjes transform (LST) of S the LST of V remaining service time of the customer in service at time t remaining vacation time of the server in vacation at time t system size at time t ==
j
The
0 if server is on vacation, 1 if server is in dormancy, 2 if server is busy,
= Pr[ NCt) = n, yCt) = 1J.
n = 0, 1, ... , N - 1
P n(x, t) dt
=Pr[NCt)=n,
x~S°U)~x+dt,
YCt)=2],
n=1,2,
Qn(x, t) dt
=Pr[NCt)=n,
x~V°Ct)~x+dt,
YCt)=O],
n=O,1,2,
... ...
Then, we can easily set up the following steady-state system equations, (2-1)
0
- ARO + QO(O),
(2-2)
0
-ARn+
n
d
- -dx PI (x)
(2-3)
Qn(O) +A'f;lRn-k·gk'
= - API (x)
(n=1,2,
.. ,N-l),
+ Pz(O)s(x),
(2-4)
d --dxPm(X) (2-5)
m-l
= -APm(X) d
- -d-x P n(X)
+Pm+JO)S(x)
=
-'-APn(X)
+A'f;lPm-k(X)gk,
(m=2,3,"',N-l),
n-l
+ Pn+l(O)s(x)
+ A 'f;IPn-k(X)gk
N-I
+ AS(X) (2-6) (2-7)
- - d Q (x) dx 0 d
(2-8)
(2-9)
- AQO(X) -AQm(X)
- -dx Qm(x)
Taking the LSTs
'f;o Rk . g n- k + Qn(O) s(x),
(n~N)
+ PI (O)V(X), +
m
A'f;IQm-ix)gk
(m~l).
of both sides of the eqs. (2-3)-(2-7), we have
8~(8)
= AP~(8)
- PI (0)
- Pz(O)S* (8), m-l
1
= AP*",(8) - P m+ (O)S* (8) - A
8P*",(8) - Pm (0)
'f;1 P:,-k(8)gk>
(m= 2, 3, ''', N-l), n-l
(2-10)
8P:(8)
- Pn (0)
= AP:(8)
- Pn+ l(O)S* (8) - A 'f;l P:-A,(8)gk N-l
(2-11)
8Q~(8)
- Qo (0)
- AS*(8)
'f;o Rk'
gn-k-
= AQ~(8)
- Pl(O)
V' (8),
Qn(O)S* (8),
(m~l).
(2-12)
Let us define the following probability generating functions, 00
P*(z,8)
L p:(8)zn,
n=l
00
P(z,O)=
00
Q*(z,8)
L
n=O
L Pn(O)zn,
n=l 00
Q:(8)zn,
Q(z, 0)
=
L
n=O
Qn(O)zn.
(n~N),
After some manipulations with eqs. (2-11) and (2-12), it follows that
= Q(z,O) - PI(0)V'(9).
[9-J\+J\X(z)]Q*(z,9)
(2-13)
Letting 9 = J\ - J\X(z),
we have
Q(z, 0)
(2-14)
= PI(0) V'[J\ - J\X(z)].
Thus, eq. (2-13) becomes (2-15)
Q*(z,9)
PI(O){ V'[J\-J\X(z)] - V'(9)} 9 - J\ + J\X(z)
Similarly, from eqs. (2-8), (2-9), and (2-10), we have
(2-16)
[9-J\+J\X(z)]P*(z,9)
= P(z,0)-S*(9)[
P(~.O)
-PI(O)+Q(z,O)
From eqs. (2-1) and (2-2), we have
(2-17)
=-J\[X(z)-1]R(z) N-I
where R(z)
= ~
Rnzn.
n=O
Thus, eq. (2-16) becomes (2-18)
[9- J\+J\X(z)]P*(z,9) = P(z, 0) - S*(9) [ P(~.
0)
+ PI (O){ V'[J\
- J\X(z)] -l}
+ J\R(z)[X(z)
-1]].
e = 1\ -
Letting
1\X(z) in eq. (2-18),
(0) + 1\[X(z) -l]R(z)}
0) = z S*[1\ - 1\X(z)]{ [ V' [1\ - 1\X(z)] -l)PI
P(
(2-19)
we have
z,
z-
S*[1\-1\X(z)]
.
P*(z, e) becomes
Thus from eq. (2-18), (2
2
p*(z,e)=
z{S*[1\-1\X(z)]-s*(e)}·
0
{[ V'[1\-1\X(z)]-l)PI(O)+A[X(z)-l)R(z)}
{z-S*[1\-1\X(z)]}·
[e-1\+1\X(z)]
Finally the PGF of the system size distribution in steady-state becomes
= P*(z, 0) + Q* (z, 0) + R(z)
P(z)
(2-21)
_(_z_-~l~)S_*~r 1\_-~1\X_(~z~)1 _l_-_V'~r 1\_-_1\~X~(~z)~l . [. PI (0) z-s*[1\-1\X(z)] 1\-1\X(z)
+ R(z) ] .
n
gi
]I n = ~
• ]I n ~
]I n
i'
z= I
the
MX/
is the probability
that the system state visits
G/I! N- policy queue without vacations (Lee et al. [9]).
n during an idle period in
Then the system size given by
eq. (2-21) becomes
P(z)
(2-22)
where
p
= 1\E(X)E(S),
= (I-p)(z-I)s*r1\-1\x(z)l z-
\If
~z)
= P(z, ordinary MX/ G/l) . \IfN(Z)
,
and
n=O
=
nZ
n
+
N-I
+~
\If n
(2-14),
we
1\E( 11)
From
eq.
1- V'r1\-1\X(z)l
1\E(V) N-I
1\E( V)
n=O
PROOF.
(z) N
N-I "" \If
(2-23)
. \If
s*[1\-1\X(z)]
+~
\If n
E( V)[1\ - 1\X(z)]
.
n=O
have
Q n (0) = P I (0) a n'
First
we
show
that
n
1\R n = PI (0) ,~ eq. (2-2).
ai
. 1T
n_ ,= PI (0) II' n by mathematical
aH I
+
For
n = 1, it is obvious from
n = 2,3, ... ,k, let us see the case of n = k+ 1:
Assuming that it holds for
= PI (O){
induction.
k+Ik+I ~I
i
a j 1TH I -
J~
i- j
g ,}
k
=PI(O){aHI'
,~ai'
1TO+
1THI-,}
Multiplying both sides of the above equation by
1
R(z)
=
N-I
J: PI (0) n~o
II' nzn.
z n and summing over n from 0 to N- 1 yields
From eq. (2-21) and
PO)
=
1,
1\0- p) N-I
1\E( 11)
+ 2: n=O
II' n
Then the theorem follows. REMARK random
2.1.
From THEOREM
variables,
interpretation
of
THEOREM
2.2.
l1li
one
of which
II' j'
j
f 1,
Let
Conditioning
is the system
the other term (II' N(Z»
Ij =
size
of the
system
ordinary
will be given in THEOREM
= 0, 1, 2, ... , N - L is the probability
customers in the system) visits
PROOF.
2.1 we see that the stationary
size is the sum of two
MX / GIl
queue.
2.5 and REMARK
The
2.2.
that the system state (number of
j during a dormant period.
if state j is visited during a dormant period,
l0,
o/w.
on the arrival size during the vacation, we have j - I
Pr[ Ij where
Pr(Ij=l)=1Tj'
= 1] = a j
Letting
+ ~O
PrCIj=l)=lI'j
a k . Pr(Ij and
- k = 1) .
1I'0=ao.
we have
II'j=
toak'
1Tj-k"
Then
from
THEOREM
2.1,
the
result
follows.
N-l
THEOREM
2.3.
\IIj is the mean number of arrival groups during a dormant period.
J~
N-l
PROOF.
From THEOREM
2.2, we observe that
~ J=O
Ij
is the number
of states visited during a
dormant period which is is equal to the number of arriving groups during the same period. N-l
E(J~)=
N-l
N-l
N-l
J~E(Ij)=
J~Pr(Ij=l)=
J~
Then
\IIj•
II1II
THEOREM
2.4.
Let
dormant period).
PROOF.
2.3,
THEOREM
be the idle period
random variable
(idle period
vacation period
staying time in a state during an idle period is
expected
2.5.
length
of
a
dormant
E I be the event that the server is idle.
Let
either on vacation or in dormancy.
-}
period
and from THEOREM
1
becomes
I
J\E( V) +
Then,
= Pr(given
N-l
Ef, server zs zn dormancy with j customers),
\II j
J~
j=0,1,2, J\E(V)
= pr(given
N-l
J\E( V)
+
J~
\II j
J:
Note that the server is idle
\II.
and
+
Then
The expected
the
YON
... ,N-l
Ef, server zs idle due to vacation) .
N-l J~
\II j.
if
he is
From THEOREM
PROOF.
is the proportion
2.4,
of time that the system
state is j during a dom1ant period.
From the renewal reward theorem, the results follow.
REMARK
2.5, we can interprete
2.2.
From THEOREM
\If
Jlz)
given by eq. (2-23). A group \If
arriving during a dom1ant period finds
J customers with probability
J\E( V) PGF acconnts for the first tem1 of eq. (2-23). server
is
the
1- V'rJ\~J\X(z)l E( V)[1- J\X(z)]
,
define
vacation
number which
of occurs
customers with
l1li
1
+
Then its
N-l l~
\If j
The mcrease of the number of customers due to that
arrive
during
a
residual
J\E(V)
probability
J\E( V)
+
N-l l~
vacation
period,
Notationally,
let us
\If j
A j as the event that a group amvmg during a dormant period finds j customers and A v
as the event that a group aniving during an idle period finds that the server is idle due to vacation. The PGF of the system size distribution can then be rewritten as
.
P(z)
P(z,
This tells us that
X
ordznary M
\If
N(Z)
I G/l)·
IS the
N-l. { l~
Pr(A)
Zl
+
Pr(A
v)·
1-- V'TJ\-J\X(z)l
E( V)[J\ - J\X(z)]
} .
PGF of the conditional system SIze distribution during the server
idle period.
THEOREM
2.6.
Let
pt(z)
be the
PGF of the departure point system size distribution.
(1- p) (z- l)s*rJ\-J\X(z)l z- S*[J\-J\X(z)]
1- X(z) E(X)( 1- z)
(2-24)
pt(z)
PROOF.
We follow the arguement of Chaudhry and Templeton [3].
.
\If
Then,
N(Z)
A departing customer will see
j customers in the system just after a departure if and only if there were
j+ 1 customers in the
system just before the departure.
Thus we have
pt Let pt(z) be the
PGF of {Pt, j =
P-I-(z)=D· N
p(z.a)
=D'
we
a, 1, 2, ...
j'~
}.
a.
Then
il.O-p)s*ril.-il.X(z)l
z- S*[iI.-il.X(z)]
Z
ptO) = L
From
= D . P j -I- 1 (a) ,
1
have
D= il.E(X)
.
[X(z)--l]
and
the
. qiJz).
theorem
follows.
II1II
3. THE WAITING TIME In this section, we derive the
LST of the waiting time of an arbitrary customer.
Let us
define the following notations,
Tv,;(8):
the
LST of the queueing waiting time of an arbitrary customer,
WB( 8):
the
LST of the waiting time of an arbitrary customer who arrives when the server is
busy,
Wy(8):
the
LST of the waiting time of an arbitrary customer who a:rrives while the server is
on vacation, the
LST of the waiting time of an arbitrary customer who arnves when the server is
dormant,
WA(8)
:
the
LST of the sum of the serVIce times of those who precede the test customer m
the same group.
The waiting time of an arbitrary customer who arrives while the server is busy and sees customers in the system is composed of, 1) the residual service time of the customer in service, 2) the sum of the service times of those
n - 1 customers in the queue, and
3) the sum of the service times of those customers who precede him in the same group.
Thus after some laborious algebraic manipulation, we have
n ( :z 1)
f P~(8)
(3-1)
[5* (8)]
n~
WA(8)
1 •
n=l
F*rS*(8).81 5*(8)
l-X(S*(8)) E(X)[l5*(8)]
(1- p)r A ~ 8 - A
f
l
REMARK
~l
1- X( 5*(8)) E(X)[l5*(8)]
AX( 5*(8))1
+ AX(S*(8))
q1 j[
5* (8)]
AE( 11)+
~l
AE( V)
j
q1 j
+
AE( 11)+
~l
q1 j
•
11 ).
E( 11)[A - AX( 5* (8)]
1- X( 5* (8lL.
(1- p)rA-AX(S*(8))1 8 - A + AX( 5* ( 8))
3.1. In eq. (3-1),
of an arbitrary customer
11*r A - AX( 5* (8)
1-
E(X)[l-
5*(8)]
... 1S the wa1tmg time
who arrives while the server is busy in the ordinary
without N - policy and vacations.
MXI Gil
queue
But in our system, an arbitrary customer who arrives while the
server is busy would wait longer due to those customers who have arrived before our test customer but are in the waiting line because of server idleness. the vacation (this occurs with probability
Those customers, if they have arrived during
VZ _ 1
AE( AE( 11) +
J~
their countribution
),
to the additional
q1 j
waiting time of the test customer is the sum of the service times of those who have arrived during a residual vacation
time
(This property
was proved
AE(V)
accounts for the terms
AE( 11)
+
N-l J~
q1 j
in many
ordinary
1 - 11*r A - AX( 5* (8)) E( 1I)[A - AX( 5* (8))]
1
vacation
queues).
in the bracket.
This
It is easily
seen that the contribution of those customers who have arrived while the server is in dormancy is explained by the first term in the bracket. The waiting time of the customer who arrives while the server is dormant can be obtained from
the
waiting
time
of
the
customer
who
arrives
MXI Gill N- policy queue without vacations (Lee et al. [9]).
while
the
server
Thus we have,
is
idle
in
the
+ 'f r~,=]
:t [S'
e 8)] i -]
•
l . _~_g_r r
EeX)
}
N-]
(l-p)
~ n=O
Wn[S*(8)]n
,\Ee 11) + N-
n -]
{ r~]
where
r; (8)
is the
N-]
~
n=O
Wn
- e S* e 8» r] EeX)[ 1- S*(8)] g r[l
[TN-n-r(8)-1J+
1- XeS*(8» EeX)[l-S*(8)]
},
LST of the idle period of the MX I GIl queue with threshold j and without
vacations. The most complicated part of the waiting time comes from the case where the test customer arrives while the server is on vacation. server is on a vacation of length
Suppose that at the arrival epoch of the test customer, the
V = x and the residual vacation time is y (see Figure 2).
test customer belongs to a group of size
If the
r and sees k customers in the queue, his waiting time
consists of 1) the total service time of those
k customers in the queue,
2) the residual vacation time y, 3) the total service time of those who precede the test customer in the same group, and 4) the dormant period (the dormant period is not zero if vacation time for
m customers arrive during the residual
k+ r+ mI
Pr(Ax,)dydx= Then from eq. (3-5),
j
l
xvCx). E( 11)
•
dydx.
x
we have
Wv(9)
(3-6 a)
(1- p)"AE(X) "AE( 11) +
~coI Wve
=
9lx, y) Pr(Ax,y)
_0-p)"AEC1I) -
•.
+
"AE(1I)
+
lo le
+ :%1l[J
-eyN-l * D (9 k, ~
o - p) "AEC11)
{
----N--~l~·
y
x
x
1-XCS*(9))1 E(X)[l-
j
:~.:r r;, ~
xvCx).l dd E( 11)
Y)
=0
-{eyH.(x-y)[l-X{S'{e))]}.
co
+
l[J j
J~
II e
0-p)"AEC1I) "AE( 11)
"AE( 11)
N-l
dydx
5*(9)] ~dYdx E(1I)
(9) -1] . CP~-n(9)
}
l[J j
J~
+ 0- p)"AEC 11) N-l "AE( 11) + l[J j J~
V* r "A- "AXC5* C9)) 1- 11*C9)
1- XCS*(9))1 E(X)[l5*(9)]
E( 11)[9 -"A
+ "AX(S'
(9))]
where (3-6 b) cp* 9 - N-n N-n( )~l
We note that
cp~-n(9)
gJ1-(S*(9))r] E(X)[l-S*(9)]
is the
(CO
Jo
I
_eyN-n-r e
J~
. qj (x-y)[S
*
j
(9)] zN-n-r-/y)dy
LST of the time length that consisits of the residual vacation time,
the total service time of those customers in the queue and the total service time of those who precedes the test customer in the same group. Finally the
(3- 7 a)
LST of the waiting time is given by
dVCx) E(V)
0- p )8 8-A+AX(S*(8))
1- X( S*(8)) E(X)[ 1- S* (8)]
+
AEC V)
AE( V)
. 1-
+ N~l\If
V*(8)
8E( V)
n
)
n=O
where
From eqs. (3-7 a) and (3-7 b),
we have the mean waiting time (see Appendix), N-l
~ n=O
(3-8)
n\Ifn
+
N-l
AE( 11)
+~
\If
n=O
n
E( V2) 2E( V) .
AE(V) AE( 11)1 +
N-l
~
\If n
n=O
Combining eq. (3-8) with eqs. (2-22) and (2-23), we see that the celebrated
L q = AE( X) VVz
4.
Little's
formula
works.
OTHER PERFORMANCE
MEASURES
In this section we derive the distributions of the number of customers at a busy period initiation epoch (i.e., idle period termination point), the busy period distribution and the idle period distribution.
THEOREM
4.1.
PGF of Q~.
Let
Q~
be the queue size at a service
Then,
(4-1)
The first and second factorial
moments become
initiation
epoch.
Let
Q~(z)
be the
+
= E( X) [ "AE( V)
E( Q'Jy)
E(X)
j]
V2) + E(X2
="A 2 E2(X)E(
E( Q'Jy(Q'Jy-l»
~11JJ
-
Q'Jy, it is worthwhile to note that if
n(
j]
+ 2E(X)
~>IJJ
j
Conditioning on the vacation length and the number of customers that arrive during the
vacation, we have
if k = 0, 1, ... , N - 1, r;(8
I V=x, if k:? N.
Pr[N(x)
But
=
kl V=
k
x]
=
e- AX(J\X) igii) i!
= ,~
where
gii)
the
IS
i-fold convolution of
= 1. Just after the termination of the vacation, if k( < N) customers are
with itself and g~O)
{ g n}
k(X)
in the queue, the remaining idle period is stochastically identical to the idle period of
N - k.
queue with threshold
Finally d
5.
unconditioning f
f
on e
MX / GIl
Thus unconditioning on k, we have
x, we have the desired result. r
e
n
E( a
r
N)
IS
easily o
derived
by
n
OPTIMAL OPERATING POLICY
In this section, we find the optimal threshold
N* that minilnizes the average operating cost
under a linear cost structure. Let us consider following costs:
Cs
Turn - on cost per cycle.
Co
Operating cost per unit time.
Ch
Holding cost per unit time.
R
Reward due to vacation Per unit vacation. E( T c), is the sum of the expected idle period and the expected
Since the expected cycle length, busy period,
(5-1)
from COROLLARY 4.1 and THEOREM 4.2, we have
E( T c) = E(
r + E(B N)
N)
+ ~
E( V)
~lW
i+
EC~E~
S)
. [ l\E( V)
+
i]
~lW
N-l
l\E( V)
+
l~
Wi
1\(1- p) Then the total average cost with threshold
m becomes
l\ 2 E(X)E( V2)
2 . [l\E( V)
m-l
A
+ Ch l~
l\E( V)
+
i1¥ i m-l l~
1¥
i
+
7~11¥ i]
)
MX/ GIl queue.
size of the ordinary
THEOREM
5.1.
N* be the optimal threshold value that minimizes
Let
cost under the linear cost structure.
I
[AE( V) Note that
l~
V)
+ 1m]~E(
k such k[AE(V)+lk]-Mk>
first
i) \Iii > -C- ]. h
(m-
m.
\Ii i and M m=
+ 1m]{~E(
[AE( V)
A
m-I
mAE( V) + z~
m.
Let 1m=
PROOF.
operating
N* is given by:
Then
N* = min [ m?:.l
(5-2)
the stationary
+ 1m-I] V)
l~
i\Ii i.
Then
. {Ch[m(AE(
+ 1m-I] > 0 A
-C- . h
and
V)
+ 1m)-
m[AE( V)
Mm] - A}
+ 1m]-
Mm> O.
Let
m be the
Then m+ I
= (m+1)AE(V)+
(m+1)[AE(V)+lm+I]-Mm+I
1~(m+1-i)1IJi m
(m+ 1)AE( V)
+ l~(m-
nt·
i)w i+
m[ AE( V) + 1m]- M m+ AE( V) Thus
6.
TC( n) > TC(m)
NUMERICAL
for
n>m
and
the
z~
\Ii i
A + 1m> -C. h theorem
follows.
EXAMPLE
As an example,
suppose
N= 2, A = 1, gI = gz = 1/2, service time=O.1 (constant) and vacation
time
=
1.0 (constant).
E(X) = 1.5,
From a k
E(X2
= pr( k
-
Then
5*(8)
we have
X) = 1,
= e-O.1S,
V"(8)
E( V2) = 1,
p = 0.15,
E(52)=O.Ol,
-1 •
g2)+_e
-ii=
6
-1
. ,rl+_e -(2g1 2 From the definition of
TI: k
-1
•
g3)+_e6-(3~
in THEOREM 2.1,
3=gl'
we get
TI:o=3/4,
2+g2' TI:l=5/8,
TI:
TI:4= gl . TI:3+ g2 . TI:2= 11/16. Also from THEOREM 2.1, we obtain
+a1'1[O=e-1,
W1=aO
'TI:1
W2=aO
'1[2+al'1[I+a2
W3 = ao . 1[3+a1
W4=aO
6.1.
'1[4+al'
. 1[0=
13e-1/8,
. 1[2+ a2 . 1[1+ a3 . TI:o= TI:3+a2 '1[2+a3
Mean Waiting Time
'1[1+a4'
1ge-1/12 TI:o=
-1
. gZ)+_e
TI:1= gl . TI:o= 1/2,
1t
~(8)=
13e-1/48,
TI:o= 1,
TI:1+g2'
X(z) = z(1 +z)/2,
arrivals during a vacation), we have
. (2g1
TI:2=gl'
= e--s,
68ge-1/384.
-gi=73e-1/384.
24
1\
8+1\'
From eq. (3-2),
_ (1 - p)III0 - AE(1I)+IIJo+lIJ) (1- p)IIJ) S*(8) + AE(1I)+IIJo+lIJ)
(8) -1]
g) . [ ~
• {
+
E(X)
1 - X( S* (8» E(X)[1-S*(8)]
}
1- X*(S* (8» E(X)[1-S*(8)] (1-p). AE(1I)+IIJo+lIJ)
1-X*(S*(8» E(X)[1-S*(8)]
(1-p)IIJo g + AE( 11)+ IIJ0 + IIJ) . -E-(b-
[IIJ + IIJ)S*(GI)] o
(8) -1].
. [~
From eq. (3-6 a),
Wv(8)
_ (1- p)AE( 11) . [~(8)-1]· - AE( 11)+ IIJ0 + IIJ)
+
cD~(8)
1- X( S' (8»
(1- p)AE( V) AE( 11)+ IIJ0 + IIJ)
E(X)[1-
v*rA-AX(S* UJ»lE(1I)[8-A+AX(S*(8»]
S*(8)]
V*(8) .
From eq. (3-6 b),
""*(8) )
g)
=-E-(X-) ---g)-
-
I r= J Ie e r= -Ax.l l_
g_)_
E(X)
Jo e
gj
*(. ) 0 =
-8y
-
-qo ( x-y -A(X-Y)
o
= -E-(X-) . cD)
e
-8y
E(X)
__
Jr= o
8
V*(A)-
(
e
)
e
dV(x) . Zo ( Y )d Y E(V) dV(x) d Y E( 11)
-AY
-AX) dV(x) E( 11)
V*(8+A) 8E( 11)
Q)
AE(X)E(
11) .
Thus, we have (1- p)AE( 11) . cD~(8) AE( 11)+ IIJ0 + IIJ)
+
(1- p)AE( V) AE( 11)+ IIJ0 + IIJ)
Then from eq. (3-7 a),
. [~(8)-1]
1- X(S*(8» E(X)[1S*(8)]
r
V* A- AX( S* ( 8»1 - V* ( 8)
E( 11)[8 -A+AX(S*(8»]
(1-- p)8 1- X(5*(8)) 8 - A + AX(5*(8)) E(X)[1- 5*(8)] \If 0 +\If p$'* ( 8) . AE( 11) .•{ AE( 11) + \If0+ \If 1 + AE( 11) + \If0 + \If 1
1- V*(8)
8E(1I)
}
+ WN(8)
Now we use eqs. (3-7 a) and (3-7 b) (see also eqs. (A-I) and (A-2) of Appendix) for the mean waiting time rather than directly using eq. (3-8).
Using the values of aj,
6.2
1Ij,
and
\Ifj
we already obtained,
we easily see that
~=O.4774.
Mean System Size
From eq. (2-22) and (2-23),
Thus
we
see
that
0.8661=(1)(1.5)(0.4774+0.1)
which
confinns
the
Little's
formula
L= :\E(X) [ ~
6.3
+ E(5)].
Optimal operating policy
Suppose
we
have
costs,
C s = 10. 000.
C h = 1000. R = 500.
Then
we
get
A = 8825 and
A
-C- = 8.825. h
Using
l]J /s,
5.4372 for
we calculate
m
mJ\E( 11) +
m-l o~
(m - i)l]J
1 =
1.3679
for
m
==
L 3.1036 for
= 3. 8.3532 for m = 4 and 13.0635 for m = 5. Thus we get N* = 5.
m
= 2,
REFERENCES 1. Burke, P.l,
"Delays in Single-server Queues with Batch Input", Opns. Res., 23, 830-833, 1975
2.
Doshi, B.T., "Queueing Systems with Vacations: A Survey", Queueing Systems,
3.
Chaudhry, M.L. and Templeton, lG.C., A first Course in Bulk Queues, Wiley, 1983
4.
Fuhrmann, S.W. and Cooper, R.B., "Stochastic Decompositions Generalized Vacations", Opns. Res., 1117-1129, 1985
5.
Hofri, M., "Queueing Systems with a Procrastinating Server", Performance ACMSIGMETRICS 1980, Performance Evaluation Review, 14(1), 245-253, 1986
6.
Kella, 0., "The Threshold Policy in the MI G/l Logistics, 36, 111-123, 1989
7.
Lee, H.S. and Srinivasan, M.M., "Control Policies for the Sci., 35(6), 708-721, 1989
8.
Lee, H.S, "Steady state Probabilities for the Server Vacation Model with Group Arrivals and under Control-operating Policy (in Korean)", J. of the Korean ORjMS Soc., 16(2), 36-48, 1991
9.
Lee, H.W., Lee, S.S. and Chae, K.c., "Operating Characteristics of -Policy", To appear in Queueing Systems, 1994
MXI G/l
11. Levy, Y, and Yechiali, Y., "Utilization of Idle Time in an Sci., 22, 202-211, 1975
13. Takagi, H. Analysis
MI G/l
Queue with '86
and
Queue with Server Vacations", Naval Research
10. Lee, H.W., Lee, S.S., Park lO. and Chae, K.C., "Analysis of and Multiple Vacations", To appear in J. of Applied Prob., 1994
12. Takagi, H., Queueing Analysis: A Foundation Priority Systems, Part I, North-Holland, 1991
in the
1, 29-66, 1986
of Performance
of Polling Systems, the MIT Press, 1986
Queueing System", Mgmt.
MXI G/l
Queue with
N
MXI G/l Queue with N-policy
MIG/I
Queueing Systems", Mgmt.
Evaluation,
Vol I, Vacation and
Appendix (Derivation of mean waiting time
~
of eq. (3-8))
From eq. (3-7 a), (A-I)
~
~*(O)
=-
IGll!
~(MX
VAC(l»
N-l
p2;nWn
+
n=O
l\E( V) +
l\E(V)
+
N-l
2;
Wn n=O
N-l
l\E( V)
+ n=O 2; W n
From eg. (3-7 b),
=
W;;(O)
WN = -
(A-2)
(1 - p) N-I l\E( V)
N- l
+ 2;
~ W
N- n {
E(Tn)'
n n-l
[
~ r-l
rg r
•
- - -) WN-n-r+l\E(V)cDN-n(O) E (X
]}
n=O
By successive evaluation of
(A-3)
• _.k cDk(O)-~IE(X)
rgr
cD~(O)
fl oo
0
kak
= l\E(X)E(V)'
given by eg. (3-6 b), we have k-r l~Zi(X-Y)Zk-r-!(Y)dYE(V)
k=1,2,
dV(x)
....
Thus we have
By successive evaluation, it is easily seen that (A-5)
N-n [ ~_lrgrWN-n-r+ r
n-l (N-n)aN-n
] .
'5""'Jti=
l~
N-I
2;
n=O
nWn·
.
Thus we have N-l
(1- p) ~
n=O
(A-6)
nW
n
.
N-l
AE( V)
+ n=O ~ Wn
Finally the mean waiting time of an arbitrary customer becomes A
7
AE( V) AE( V)
+ ~lW n=O
E( n
V2)
2E( V)
III