Download
Download
Abstract
A queueing system (M/G1 ,G2 /1/K ) is considered in which the service time of a customer
entering service depends on whether the queue length, N (t), is above or below a threshold
L. The arrival process is Poisson and the general service times S1 and S2 depend on whether
the queue length at the time service is initiated is < L or L, respectively. Balance
equations are given for the stationary probabilities of the Markov process (N (t); X (t)) where
X (t) is the remaining service time of the customer currently in service. Exact solutions for the
stationary probabilities are constructed for both infinite and finite capacity systems. Asymptotic
approximations of the solutions are given, which yield simple formulas for performance
measures such as loss rates and tail probabilities. The numerical accuracy of the asymptotic
results is tested.
1 Introduction
1.1 Background
Queueing systems arise in a wide variety of applications such as computer systems and communi-
cation networks. A queueing system is a mathematical model to characterize the system, in which
the arrivals and the service of customers (users, packets or cells) occur randomly. The customers
arrive at the facility and wait in the queue (or buffer) if the server is not available. If there are many
customers in the queue, they may suffer long delays which cause poor system performance. Thus,
the arrival rate or the service rate may need to be controlled to reduce the delays. These systems may
This research was supported in part DOE Grant DE-FG02-96ER25168. D.I. Choi was supported by the Korea
Research Foundation. Current address of D.I. Choi: Department of Mathematics, Halla Institute of Technology, San
66 HeungUp-Li HeungUp-Myon, WonJu-Shi, KangWon-Do, 220-712, South Korea
1
be represented by queueing systems with queue-length-dependent arrival rates or service times.
That is, if the queue length exceeds a threshold value, the arrival rate may be reduced (e.g. overload
control), or the service rate may be increased (e.g. the cell discarding scheme [2]). Many schemes
of traffic control in ATM (asynchronous transfer mode) networks have been analyzed using such
threshold-based queueing systems ([2],[5],[8]-[12]).
In this paper, we analyze a queueing system with queue-length-dependent service times.
Customers arrive at the queue by a Poisson process, and there is only one server. The service times
of customers depend on the queue length. Concretely, we specify a threshold value L for the queue.
If the queue length at service initiation of a customer is less than the threshold L (respectively,
greater than or equal to the threshold L), the service time of the customer follows a distribution with
probability density function b1 () (respectively, b2 ()). We believe that our analysis can be extended
to the case of multiple thresholds. Both infinite (M/G1 ,G2 /1) and finite capacity (M/G1 ,G2 /1/K )
queues are considered.
The analysis of this queueing system was directly motivated by the cell-discarding scheme
for voice packets in ATM networks (see [2, 9, 11]). In [11], a system with deterministic service
times was proposed, in which voice packets are divided into high and low priority ATM cells. The
cells arrive as a concatenated pair (i.e. two cells per arrival) and the threshold L and the capacity
K are measured in terms of cell pairs. The cell discarding occurs at the output of the queue and
immediately prior to transmission, based upon the total number of cell pairs in the queue. If this
number is less then L then the next pair of cells is transmitted (no cells are discarded). However, if
the queue length is greater than or equal to L then only the high priority cell is transmitted. Thus,
the low priority cell is discarded. The analysis in [11] is based on numerically solving an embedded
chain formulation of the problem. The system is studied only at service time completions. Our
results, when specialized to Gi = Di (deterministic service), are directly applicable to this model.
We analyze this queueing system using the supplementary variable method. We first consider
the case of an infinite capacity queue, and obtain an explicit formula for the steady-state queue
length distribution pn . When the service times have different exponential distributions, the queue
length distribution has a simple, closed form. We also compute asymptotic approximations to the
queue length distribution pn for various choices of the system parameters. Next, we examine the
finite capacity queue, and again obtain explicit expressions for pn and in particular the probability
pK that the queue is full and the probability p0 that it is empty. Also, we investigate asymptotic
approximations for the queue length distribution for large values of the threshold L and the queue
capacity K . We show that this queueing system has very different tail behavior (and hence loss
probability) than other M=G=1 type models, and that the service tails can sometime determine the
tail of the queue length.
There has been some previous analytical work on queueing systems with queue-length-
2
dependent service times [1, 3, 4, 6, 7]. C. M. Harris [6] considered the M X =G=1 queue with
queue-length-dependent service times. In particular, if there are i customers in the queue, the
service time of the customer starting service has a general distribution depending on i. By using the
embedded Markov chain method, C. M. Harris [6] derived the probability generating function for
the queue length at the departure epochs. However, the obtained probability generating function
contains infinitely many unknown constants. A closed form was obtained only for some special
service times of two types. Fakinos [4] analyzed the G=G=1 queue in which the service discipline
is last-come first-served and the service time depends on the queue length at the arrival epoch of
each customer. Abolnikov, Dshalalow and Dukhovny [1] considered queues with bulk arrivals
(i.e. compound Poisson input) and state-dependent arrivals and service. Assuming that the state-
dependence applies only when the queue length is below a critical level, the authors characterize
the generating function (for both transient and steady-state cases) of the queue length probabilities
in terms of the roots of a certain equation. Ivnitskiy [7] also considers a model with bulk, state-
dependent arrivals and state-dependent service. Using the supplementary variable method, he
obtains a recursion relation for the Laplace transforms of the transient queue length probabilities.
For a very good recent survey of work on state-dependent queues, we refer the reader to Dshalalow
[3].
The model here is a special case of that studied in [7]. However, we are able to give more
explicit analytical expressions, from which we can easily obtain asymptotic expansions for tail
probabilities and loss rates. These clearly show the qualitative dependence of these performance
measures on the arrival rate and the service distribution(s). In particular, we show that the tail
behavior of the model with threshold L is much different than the tail behavior of the standard
M/G/1 and M/G/1/K models.
3
The stationary probabilities are denoted by
p0 = tlim
!1
Pr[N (t) = 0]: (1.2)
For finite capacity systems (K < 1), these limits clearly exist. For infinite capacity systems
(K = 1), we assume the stability condition
Z1
2 = M1 = t b2 (t) dt < 1: (1.3)
0
? dp
dx
n
= ?pn(x) + pn?1(x) + b1 (x)pn +1 ( 0); n = 2; : : : ; L ? 1 (1.6)
dp
? dxL = ?pL (x) + pL?1(x) + b2(x)pL +1 ( 0) (1.7)
? dp
dx
n
= ?pn(x) + pn?1(x) + b2 (x)pn +1 ( 0); n = L + 1; : : : ; K ? 1 (1.8)
dp
? dxK = pK ?1(x): (1.9)
In the following sections, we construct exact solutions to the infinite capacity model and
then the finite capacity model. As we will show, the solution of the infinite capacity model can be
used to construct the solution of the finite capacity model. Then we obtain simple formulas for the
performance measures by constructing asymptotic approximations to the exact solutions.
4
M/M1 ,M2 /1 Queue: To illustrate the important characteristics of the solution, we first consider
the case in which the service times, S1 and S2 , are exponential with probability density functions
The solution of (1.4)-(1.8) can be constructed in a straightfoward manner as follows. For n < L,
we assume a solution of (1.4)-(1.6) in the form pn (x) = e?1 x Pn which leads to the difference
equation
Pn 1 ? (1 + 1)Pn + 1 Pn?1 = 0
+ (2.2)
P2 = 1 P1; P1 = p0 (2.3)
where
1 = : (2.4)
1
The solution of (2.2)-(2.3) is
Pn = p01n?1 ; n = 1; : : : ; L ? 1 (2.5)
so that
pn(x) = p0 1n?1e? x; n = 1; : : : ; L ? 1:
1
(2.6)
We now compute pL (x) by first finding the value of pL (0) using (1.6) with n = L ? 1 and the
known functions pL?1 (x) and pL?2 (x), which leads to
To compute pL+1 (0), we substitute (2.8) into the local balance result (1.11) with n = L. This leads
to
pL+1(0) = p02 1L?1 + 2 ; 2 = + 1 2
which when used in (2.8) gives
" #
pL(x) = p01L?1 1 e? x + e? x : (2.9)
+ 1 + 1
1 2
5
which leads to the system of difference equations
( + 1)An = An?1
( + 2 )Bn = Bn?1 + 2[An +1 + Bn 1]:
+
where c and AL are to be determined. By setting n = L in (2.10), with An and Bn defined above,
and equating the result to (2.9), we find that
AL = p01L?1 +1
1
c =
p01L?1 + ? :
1 2
Z1
The constant p0 is determined by normalization using (1.10) and the marginals pn = pn(x)dx
0
are summarzied below.
Theorem 1 M/M1 ,M2 /1 Queue: Let bi (x) = ie?i x and i = =i for i = 1; 2 and assume that
the stability condition 2 < 1 is satisfied. The marginal probabilities are given by
pn = p0n1 ; n = 0; 1; : : : ; L ? 1; (2.11)
8 !n?L 1 9
< ? = +
;; n L
pn = p0 1L?1 : + ? 2n?L 1 + +1 ?2 1 +1
+
(2.12)
1 2 1 2 1
where " #
1 ? L1 L1 :
p0 = 1= 1 ? 1
+
1 ? 2
(2.13)
An interesting aspect of the result is the tail behavior as n ? L ! 1. The tail probability
has a different form depending on the parameters:
8
>
> n?L+1
< + 1 ? 2 2
>
> if + 1 > 2
pn p01 > 2 ? 1
L? 1
!n?L+1 (2.14)
>
>
>
1
if + 1 < 2 :
: ?? 1+
2 1 1
6
M/G1 ,G2/1 Queue: We now consider the system in which the service times S1 and S2 have
general distributions. For n < L, we solve (1.4)-(1.6) with L = 1 by introducing the generating
function 1
X
G(z; x) = pn(x)zn (2.15)
n=1
into (1.5)-(1.6) to obtain
where b̂1 (s) is the Laplace transform of b1 (t). Combining the result (2.18) for G(z; 0) in (2.17) and
simplifying we obtain
We invert G(z; x) to find an integral representation of the stationary probabilities in the form
p
pn(x) = 2i n 0
Z z ? 1 Z1
e(z ? 1)(t ? x)b1 (t)dt dz:
C z [z ? b̂1 ( ? z )] x
(2.20)
The contour C is a loop in the complex z -plane which encircles the origin but excludes any other
poles of the integrand. By setting n = 1 and x = 0 in (2.20), we find that (1.4) is satisfied. Based
on the above calculation, we see that (2.20) is in fact a solution of (1.5)-(1.6) for L < 1. Thus,
we assume that for n < L, the solution is of the form
k Z
pn(x) = 2i n z ? 1 Z1
e(z ? 1)(t ? x) b1 (t)dt dz; n = 1; : : : ; L ? 1;
C z [z ? b̂1 ( ? z )] x
(2.21)
Hx(z; x) + (z ? 1)H (z; x) = ?pL?1(x) + b2 (zx) [pL(0) ? H (z; 0)]: (2.23)
7
The solution of (2.23) is
The numerator of (2.25) vanishes as z ! 1? using the local balance result (1.11) with n = L ? 1,
i.e. Z1
pL(0) = pL?1(t)dt: (2.26)
0
We use the known solution for pL?1 (x) given by (2.21) to compute the integral in the numerator of
(2.25) as
Z1
pL?1 (t)e
k
(z?1)t dt =
Z w?1 ?b̂1 ( ? w) + b̂1 ( ? z) dw (2.27)
0 2i C w [w ? b̂1 ( ? w)]
L? 1 (z ? w)
and using (2.26) we find that for L 3
pL(0) = 2k
Z w?1
i C wL?1[w ? b̂1 ( ? w)] dw: (2.28)
Thus, H (z; 0) is determined by (2.25) using (2.27) and (2.28) up to the constant k. We now use
H (z; 0) and pL(0) to construct H (z; x) using (2.24). After some algebra, we find that
R 1 ? z?1 x?u ? e? w?1 x?u )b (u)du
k
H (z; x) = 2i
Z w?1 x (e
( )( ) ( )(
1
)
dw
C w [w ? b̂1 ( ? w)]
L?1 z?w
" #
k Z w ? 1 b̂ 1 ( ? z ) ? b̂1 ( ? w )
? 1 dw
+
2i(z ? b̂2 ( ? z )) C w L?1 [w ? b̂1 ( ? w)]
Z1
z ? w
e? z?1 x?u b2 (u)du
x
( )( )
(2.29)
Thus, for n > L, the stationary probabilities can be found by inverting (2.22) using
1 Z
pn(x) = 2i zn?1L+1 H (z; x)dz: (2.30)
C
The quantities of interest are the marginal probabilities pn which are obtained by removing
the dependence on x. We first compute
Z1 X
H (z) = H (z; x)dx = pnzn?L:
0 nL
8
The following identity is useful in calculating the marginal probabilities
Z 1 Z 1
eA(x?u)
bj (u)du dx = 1 ? Ab̂j (A) :
0 x
We find that by integrating H (z; x) and simplifying that
In addition, we assume that L 3 which is needed for the above simplification and to avoid a
degenerate problem. We invert (2.31) to obtain the marginal probabilities for n = L; L + 1; : : :,
Z Z b̂ ( ? z ) ? z w?1
k
pn = (2i)2 1 dwdz :
? 2 ( ? z ) w [w ? b̂1 ( ? w )](z ? w )
(2.32)
jzj<jwj z b̂ L? 1 z n?L+1
The contours for the double complex integral are such that jwj > jz j on the w-contour and both are
small loops about the origin.
The marginals for n < L are obtained by integrating (2.21) with respect to x which leads to
k Z ?1 + b̂ ( ? z ) 1
pn = 2i 1
dz; n = 1; : : : ; L ? 1;
C z ? b̂1 ( ? z ) z n
(2.33)
Again the contours are small loops inside the unit circle.
To complete the solution, we find the constant k using the normalization condition (1.10).
First, we compute
1
X k b̂0 (0) + 1 Z dw
pn = H (1) = 2i 1 0
n= L 1 + b̂2 (0) C wL?1[w ? b̂1 ( ? w)]
+
where on the contour C + we choose jwj > 1. The only poles inside C + are at at w = 0 and w = 1
if 1 < 1. Since ?b̂01 (0) = m1 and ?b̂02 (0) = M1 , we find that
pn = 2ki 11 ? 1 Z
X1 dw
n L
=
? 2 C wL?1[w ? b̂1 ( ? w)]
+
(2.35)
where on the contour C + we have jwj > 1 and jwj < 1 + . In 1 < jwj < 1 + , the function
w ? b̂1 ( ? w) is assumed non-zero.
9
PL?1
For n < L, we must compute p0 + n=1 pn . We simplify (2.33) as follows
pn =
k Z b̂1 ( ? z) ? z + z ? 1 1 dz
2i jzj<1 z ? b̂1 ( ? z) zn
?kn1 + 2ki jzj<1 z ? 1 z1n dz
Z
=
z ? b̂1 ( ? z)
which leads to
?1
LX Z
p0 + k
pn = 2i 1
1 ? L?1 dz:
1
jzj<1 z ? b̂1 ( ? z )
(2.36)
n= 1 z
Combining the two sums, we find that
1 11 1 ? 1 Z dw
p0 k 2i 1 ? 2 C + w [w ? b̂1 ( ? w)]
= =
L? 1
1 Z
1 ? L?1 dw:
1 1
+
2i jwj<1 w ? b̂1 ( ? w)
(2.37)
w
We simplify (2.37) by noting that if 1 < 1 then
1 Z dw 1 1
= Resw =
1 ? m1
2i C + w ? b̂1 ( ? w) =1
w ? b̂1( ? w)
so that " #
1 ? 1 1 Z dw
1
=
1
=
1
1 ? 1
+
1 ? 2
? 1 2i C + :
wL?1[w ? b̂1 ( ? w)]
(2.38)
p0 k
By shifting the integration contour, we can remove the restriction that 1 < 1, and obtain the
alternate form for p0 in (2.43).
The final results for the marginals are summarized below.
Theorem 2 (M/G1 ,G2 /1) Let bi (x) for i = 1; 2 be the density functions for the general service
times with moments defined by
Z1 Z1
mi = x b1 (x)dx; Mi = xib2 (x)dx
i (2.39)
0 0
and let 1 = m1 and 2 = M1 . We assume the stability condition 2 = M1 < 1 is satisfied and
that L 3. The marginal probabilities are given by
Z ?1 + b̂ ( ? z ) 1
pn = 2i k 1
dz; n = 1; : : : ; L ? 1;
C z ? b̂1 ( ? z ) z n
(2.40)
and for n = L; L + 1; : : :,
k Z Z b̂ ( ? z ) ? z w?1 dwdz
pn = (2i)2 1
:
jzj<jwj z ? b̂2 ( ? z ) [w ? b̂1 ( ? w )](z ? w )
(2.41)
w z
L? 1 n?L 1 +
10
The probability that the system is empty is
p0 = k (2.42)
where
1 1 2 ? 1 1 Z dw :
1 ? 2 1 ? 2 2i C w L?1[w ? b̂1 ( ? w)]
= + (2.43)
k
All the integration contours are small loops about the origin.
To illustrate the formulas in Theorem 2, we consider the case of exponential service times
in which the density functions are defined in (2.1). The Laplace transforms of the density functions
are defined by
b̂i (s) = +i s ; i = 1; 2 (2.44)
i
so that for n < L, (2.40) reduces to
For n L, we must compute the double contour integral. For the exponential case, the formula
(2.41) becomes
?k
pn = (2i)2
Z Z 2 + ? z z ? 1=1 1 1 + ? w 1 1 dzdw
+ ? z z ? 1=2 (?) w ? 1=1 w z L 1 z ? w :
jzj<jwj 1
L?1 n? +
(2.46)
The remaining integral can be evaluated by observing that the integrand has poles at z = 0, 1=2 ,
and 1 + 1=1 . Since the only pole inside of C is at z = 0, we have the equality: ?(Residue at 0) =
(Residue at 1=2 ) + (Residue at 1 + 1=1 ) so that (2.48), after some simplification, reduces to
8 !n?L+1 9
< ? =
pn = k1L?1 : + ? 2n?L+1 + + ? 1 + ; ; n L:
1 2 1
(2.49)
1 2 1 2 1
The above results agree with (2.11) and (2.12) in Theorem 1. Also, the formula for p0 in Theorem
2 reduces to (2.13).
11
Asymptotic Approximations: Simple formulas can be obtained by asymptotically expanding the
exact integral representations for the stationary probabilities given in Theorem 2. An important
limit is L 1 and n ? L 1. The asymptotic expansion in this limit depends on the location of
the zeros of w ? b̂1 ( ? w), w ? b̂2 ( ? w), and poles of b̂1 (w). Specifically, we need to locate
the singularities of the integrand which are closest to the origin. Let A be the non-zero solution of
Z1
+ A = eAz b (z) dz:1 (2.50)
0
We assume that b̂i () are analytic for some with <() < 0, which insures the existence of a
> <
unique solution to (2.50). Clearly, A > ? and satisfies A = 0 if 1 = 1. We let B be the solution
< >
of Z1
+ B = eBz b (z) dz 2 (2.51)
0
which satisfies B > 0 when 2 < 1 and B # 0 as 2 " 1. The Laplace transform b̂1 () may have
singularities in the half-plane <() < 0. We assume the singularity with the largest real part is at
= ?C and that
b̂1() ( +DC ) ; ! ?C (2.52)
for some constants ; D > 0. For example, if b1 is exponential then b̂1 () =
1 so that
1 +
1 ? zr r r ?
D = 1, C = 1, and = 1 in (2.52). When b1 (z) = Γ(r) e (1r) z , r > 0 , i.e. r-stage
1 1
!r
Erlang, then b̂1 () =
1r
1r + so that D = (1r) , C = 1r and = r in (2.52). If the service
r
We define Z1
Il (A) = zl eAz b1( ) z dz (2.54)
0
so that I0 (A) = 1 + A= and
!L?1
pn A + A k 1 Z b̂1 ( ? z) ? z dz :
2i I1 (A) ? 1 jzj " [z ? (1 + A=)][z ? b̂2 ( ? z )] z L 1
n? +
(2.55)
= 1
12
z-plane
0
1 1+B/λ 1+C/λ
ε1
Figure 1: A sketch of the z -plane indicating the location of the poles of the integrand.
z-plane
Γ
1+C/λ
where Z1
Jl (B ) = zl eBz b2 (z)dz; J0(B ) = 1 + B :
0
13
w-plane
Γ′
n ? L ! 1:
p L?1 Z n?L 1
0 A ( A) D +
pn 2i I (A) ? 1 C ? A C
+ 1 1
dz:
1 + ? b̂2 (?C ) Γ ( ? z + C ) z
(2.57)
1
dz ? Γ0 1 wD 1 + C (n ? L + 1)
?1
Γ ( ? z + C )
=
z
?n+L?1
w
1? n?L+1C + 1
dw
?n+L?1 Z ew=(C +)
D
? 1+ C (n ? L + 1)
?
w dw
1
Γ0
where Γ0 is shown in Figure 3. Using the substitution w C + ) in the integral, we find that it
= (
can be expressed in terms of the Gamma function, so that for C < B ,
A
pn p 0 C ? 1 1 D 1
A I1 (A) ? 1 1 + ? b̂2 (?C ) (C + )?1
C
n?LΓ 1 ? A L?1 C n?L 1 ; n ? L 1:
( + )
( )
1
+ +
+
We note that if b̂1 () has a simple pole at = ?C then = 1 and the algebraic factor (n ? L + 1)?1
disappears.
We derive an asymptotic formula for L 1, n 1 and n < L by using (2.40) and noting
that
p Z ?1 + b̂ ( ? z ) 1 p Z z?1
pn = 2i0 1
C z ? b̂1 ( ? z ) z n dz 0 1
2i C z ? b̂1 ( ? z ) z n
dz
?2ip0 Resz 1 A=[ ]
= +
!n
p A
I1(A) ? 1 + A :
0
=
14
We now compute asymptotic approximations to p0 for L 1. The constant p0 is defined
in Theorem 2 as
1 1 2 ? 1 1 Z dw
p0 1 ? 2 1 ? 2 2i C wL?1[w ? b̂1( ? w)] :
= + (2.58)
The poles of the integrand closest to w = 0 are at w = 1 and w = 1 + A=. We approximate 1=p0
when L 1 as
2 L?1 ! 3
2 ? 1 4
1
p0 1 ? 2
1
+
1 ? 2
? 1
1 ? 1
? +A
1
1 ? I1 (A)
5 (2.59)
!L?1
1 1 ? 2 1
=
1 ? 1
+
1 ? 2 1 ? I1 (A) + A
:
Now if 1 = m1 < 1 then A > 0 and
p0 1 ? 1 : (2.60)
The final case is when 1 1. We set 1 = 1 + La and find that (2.50) gives
A ? m2a L
2
Theorem 3 Asymptotic approximations for M/G1,G2 /1 queue: For L 1, and A, B , and C defined
by (2.50)-(2.52), the stationary probabilities have the following asymptotic approximations:
n ? L 1 and C > B :
!L?1 !n?L+1
pn p 0 + A A b̂1 (?B ) ? (1 + B ) ?1
+B B ? A 1 ? J1(B ) I1 (A) ? 1 (2.63)
15
where Z1
Jl (B ) = zl eBz b2 (z)dz; J0(B ) = 1 + B :
0
n ? L 1 and C < B :
A
pn p 0 C ? 1 1 D 1
A I (A) ? 1
1 1 + C ? b̂2 (?C ) (C + ) ?1
(2.64)
! !
n ? ΓL(+ )1)
( ?1 L?1 n?L 1 ; n ? L 1: +
+A +C
For n 1 and n < L, !n
p A
pn I (A) ? 1 + A
0
(2.65)
1
and 8
>
>
>
> 1 ? 1 ; 1 < 1
>
< 1 ? I1 (A)
> A L?1
p0 >> 1? 2 (1 ?2 ) 1 + ; 1 > 1
>
>
> a
: = exp 2
> 2a
?1 ; 1 ? 1 = a=L = O(L?1):
L m2
For L 1 and n ? L fixed (just above the threshold), the asymptotic result is given by (2.55) with
k = p0 given above.
We specialize the above formulas to the exponential service case with bi (t) = i e?i t for
i = 1; 2. For this case, we find that m1 = 1=1, m2 = 2=21, M1 = 1=2, and M2 = 2=22. We
explicitly solve (2.50) and (2.51) to obtain
A = 1 ? ; B = 2 ? :
In addition, from (2.52), we find that
C = 1 ; D = 1; = 1:
Using the above results, we obtain
which agrees with the result in (2.14). We note that C < B is equivalent to 1 < 2 ? . A similar
reduction occurs for the case when C > B.
16
Finally, we examine the case when C B , more precisely B ? C = O(L?1 ). For L 1,
n 1 and n < L, we have the approximation (2.65). We let C = B + =L and note that the
integral in (2.55) now has singularities (at 1 + B= and 1 + C=) that are close to each other. We
consider the integral
Integral =
1 Z b̂1 ( ? z) ? z dz :
2i jzj="1 [z ? (1 + A=)][z ? b̂2 ( ? z )] z L+1
n?
We make the change of variables
z= 1+
B 1 + u
L
and use the approximations
1
1 ? J1 (B ) z ? 1 1? B=
z ? b̂2 ( ? z) 1
17
We set n = K ? 1 in (2.30) and use the result in (3.1) to obtain
pK (x) = (2p 0
Z Z 1 1 w?1 1
i) jzj<jwj z w w ? b̂1 ( ? w) z ? w
2 K ?L L ?1
(3.2)
(R 1 R 1 (1?z)(x?u) R1 R 1 (1?w)(x?u)
b 1 (u)du ? x e b 1 (u)du b 1 (u)du ? x e b1 (u)du
x
1?z
? x
1?w
R1 R 1 (1?z)(x?u) )
b̂ 1 ( ? z ) ? b̂1 ( ? w ) ? z + w x b2 (u)du ? x e b 2 (u)du
+
1?z
dzdw:
z ? b̂2 ( ? z)
Here we have used the fact that
Z 1Z 1
e 1?z ( s?u) b
)(
u duds =
1( ) (3.3)
x s Z Z1
1
1?z
1
b1 (u)du ? e(1?z)(s?u) b1 (u)du :
x x
The main quantities of interest are again the marginal probabilities. The marginals for pn ,
1 n < L and L n < K are given by (2.40) and (2.41) in Theorem 2, respectively. Again, the
constant k = p0 must be re-calculated. The marginal probability pK must be computed using
Z1
pK = pK (x)dx:
0
18
and
m1 ? 1 Z Z 1 ? 1 w ? 1 1 1
dzdw
(2i)2
jzj<jwj
w ? 1 z ? 1 z K ?L w L?1 z ? w w ? b̂1 ( ? w)
1 ? 1 Z 1 1
= dw:
2i C w L?1
w ? b̂1 ( ? w)
Substituting (2.40), (2.41) (with k replaced by p0), and (3.5) into (3.6), we find after simplifying
that
Z
1
=1+
1 1 1
dw (3.7)
p0 2i C w L? 1
b̂1 ( ? w) ? w
2 Z Z 1 w?1 1 b̂1 ( ? z) ? z
+
(2i)2 z K ?L wL?1 z ? 1 z ? w [w ? b̂ ( ? w)][b̂ ( ? z ) ? z ] dzdw:
jzj<jwj 1 2
(3.8)
19
The normalization constant p0 , i.e. the probability that the system is empty, is
1
=1+
1 Z 1 1
dw
b̂1 ( ? w) ? w
(3.12)
p0 2i C w L ?1
ZZ 1 w?1 1 b̂ ( ? z) ? z
i)2 jzj<jwj zK ?LwL?1 z ? 1 z ? w [w ? b̂1 ( ? w)][b̂2 ( ? z) ? z] dzdw:
2 1
+
(2
On all the contours it is assumed that the origin is the only singularity within the contour.
2. 1 < 1 and 2 1
p0 1 ? 1 (3.15)
3. 1 1 and 2 < 1
" #
p0 La exp(2a=( 2
1
m2)) ? 1 ; 1 ? 1 = La (3.16)
20
4. 1 > 1 and 2 < 1 L?1
1 ? 2 A
p0 ? [1 ? I1(A)] 1+
(3.17)
1 2
b
12)
; (3.18)
1 ? 1 = La ; 2 ? 1 = Lb ; = KL?L
6. 1 > 1 and 2 > 1
L?1 K ?L
p0 1 + A B B (B ? A)[1 ? I1 (A)][J1(B ) ? 1]
1+
2 A[I0(B ) ? J0(B )] + (3.19)
7. 1 1 and 2 > 1
p0 + (3.20)
8. 1 > 1 and 2 1
L?1
A
p0 1 + [1 ? I1 (A)]
1 b 1
1 ? 1 L exp(2b =(2M2 )) ? 1 (3.21)
+ + ?
Note that in the last case pn has a bimodal behavior as K; L ! 1, with peaks near n = 0 and
n = K.
The asymptotic expansion of pK as K; L ! 1 is computed by using (3.5) for pK =p0 and
the expansions of the integrals which were derived for the infinite capacity case in Section 2. The
results are summarized below.
Theorem 5 Asymptotic expansions for the M/G1 ,G2 /1/K queue: For K; L 1 and A, B , and C
defined by (2.50)-(2.52), the stationary probabilities pn have the asymptotic expansion (2.65) for
n < L and (2.63)- (2.64) for n ? L 1. The asymptotic expansion for pK is
8
pK
!L?1
A(2 ? 1) >< [I0(B)?J0 (B)] K ?L ; B<C
B(B?A)[J
? 1 (B )?1] +B
p0 +A 1 ? I1 (A) >
: K ?L 1
D 1 K ?L ; B>C
C + C (C ?A)[1+C=?J0 (C )] Γ( ) +C
(3.22)
where the asymptotic expansions of p0 are given by items 1-9 above for various values of 1 and 2 .
21
4 Discussion and Numerical Results
We now demonstrate the usefulness of our results for a finite capacity system, namely the
M/D1,D2 /1/K queue. For this model, the density functions of the service times are
1 ?1 1 1 z j
?1 X
z?w =
w 1 ? z=w =
w w:
j =0
Using this result in the integrand in (3.10) allows the double integral to be separated into
nX?L 1 Z
1 Z ?1 dw ;
+L?n?1
z j dz 2i C
j =0 2i C wj L+
where can be identified from (3.10). The sum truncates at j = n ? L since for j > n ? L the z
integral vanishes. Finally, we must evaluate both integrals by computing the residues at z = 0 and
w = 0 for each value of j . Again the calculation is only feasible for L and n ? L moderate in size.
As before, we use Maple to perform the calculation.
The calculation of the asymptotic approximation requires computing the constants A and B
by solving (2.50) and (2.51), respectively. For deterministic service times, there is no singularity
C so that C = 1 > B . Given the constants A and B , we evaluate the formulas (2.65) and (2.63)
for pn if n < K and (3.22) for pK . The constant p0 is given by (3.14)-(3.21), depending on the
values of 1 and 2 .
22
In Table 1, we consider a queue with the threshold L = 5 and capacity K = 10. The
arrival rate is = 1 and the service times are m1 = 1=2 and M1 = 1=4 so that 1 = 1=2 and
2 = 1=4. Thus, when the queue length exceeds the threshold, the service time of the jobs entering
service is half of the original service time. For these values of i , the asymptotic value of p0 is
given by (3.14). The solutions of (2.50) and (2.51) are A 2:512 and B 9:346, respectively.
Table 2
L = 5; K = 15; 1 = 1 =2 ; 2 = 1=4
As we see from Table 1, the asymptotic expansion is quite accurate for n = 0, n = 2?4
and n > 8.It is not accurate for n = 1 and n = 5 ? 8. This is consistent with our results in
Theorem 5 since the asymptotic result (2.65) is valid for n 1 and (2.63) is valid for n ? L 1.
n and n ? L are only moderate in size. We
It is remarkable that our results are this accurate since
cannot expect the asymptotic solution to be accurate for n = L since it assumes that n ? L ! 1.
We have chosen values for L and K that are quite small. In reality, we would expect them to be of
the order 102 (see [11]). For such large values of L and K , calculation of the exact solution would
prove difficult while the evaluation of the asymptotic solution is straightforward.
In Table 2, we consider the same queue as in Table 1 but now with K = 15. The values
for A and B are the same as for Table 1 since these constants are independent of L and K . We
see that the exact and the asymptotic results are numerically close to those in the previous example
for n < 10, as expected. For n 10, the relative error starts at about 4% and rapidly decreases to
23
under 1%.
The example in Table 3 is a queue with threshold L = 8 and capacity K = 15. The arrival
rate is = 1 and the service times are m1 = 2 and M1 = 1=2 so that 1 = 2 and 2 = 1=2. In
this example, 1 > 1 so that in the absence of the threshold, the queue length distribution would be
peaked near the capacity K . Now A ?0:7968 and B 2:512. The asymptotic results are quite
accurate when n < L. However, when n L the results are not accurate.
Table 4
L = 8; K = 25; 1 = 2 ; 2 = 1 =2
n exact asympt. rel. err.
0 .28292e-5 .28292e-5 0.000003
1 .18076e-4 .18690e-4 0.033972
Table 3 2 .91756e-4 .91986e-4 0.002507
L = 8; K = 15; 1 = 2 ; 2 = 1=2 3 0.000453 0.000453 0.000104
n exact asympt. rel. err. 4 0.002228 0.002228 0.000002
0 .28305e-5 .28292e-5 .00044 5 0.010966 0.010966 0.000003
1 .18084e-4 .18690e-4 .03351 6 0.053968 0.053968 0.000003
2 .91796e-4 .91986e-4 .00206 7 0.265605 0.265604 0.000003
3 .00045286 .00045271 .00033 8 0.291673 4.492908 14.403916
4 .0022290 .0022280 .00044 9 0.198693 1.278988 5.436998
5 .010970 .010965 .00044 10 0.104037 0.364087 2.499576
6 .053991 .053967 .00044 11 0.045645 0.103644 1.270637
7 .26572 .26560 .00044 12 0.01759 0.029504 0.677286
8 .29180 4.4929 14.397 13 0.00615 0.008399 0.365690
9 .19878 1.2789 5.4341 14 0.002001 0.002391 0.195078
10 .10408 .36408 2.4980 15 0.000618 0.000681 0.100900
11 .045665 .10364 1.2696 16 0.000185 0.000194 0.049898
12 .017598 .02950 .67654 17 .53894e-4 .55154e-4 0.023365
13 .0061526 .0083988 .36508 18 .15540e-4 .15700e-4 0.010298
14 .0020014 .0023908 .19455 19 .44505e-5 .44694e-5 0.004260
15 .00043941 .00047573 .0826589 20 .12702e-5 .12723e-5 0.001652
21 .36196e-6 .36218e-6 0.000601
22 .10308e-6 .10310e-6 0.000205
23 .29348e-7 .29350e-7 0.000064
24 .83549e-8 .83550e-8 0.000017
25 .16624e-8 .16624e-8 0.000001
In Table 4 we retain L = 8 and increase K to 25. Now the asymptotic results are accurate
to within 5% for 16 n 25. In Tables 3 and 4, the results are not accurate for L = 8 n 14.
24
Apparently, n ? L = 6 is too small a value for (2.63) to be useful, for these particular parameter
values.
In Table 5, we consider the same case as in Table 3 except that now M1 2=3 and hence
=
2 = 2=3. Now we have A ?0:7968 and B 1:144. The results are to within 7% if n ? L 4
(i.e. 12 n 15). In Table 6 we increase K to 25. The error is at most 7% for all n ? L 4 (i.e.
12 n 25).
Table 6
L = 8; K = 25; 1 = 2 ; 2 = 2=3
n exact asympt. rel. err.
0 .21219e-5 .21219e-5 0.000005
1 .13557e-4 .14017e-4 0.033970
Table 5 2 .68817e-4 .68989e-4 0.002505
L = 8; K = 15; 1 = 2 ; 2 = 2=3 3 0.00034 0.00034 0.000103
n exact asympt. rel. err. 4 0.001671 0.001671 0.000004
0 .21329e-5 .21219e-5 .0051 5 0.008224 0.008224 0.000005
1 .13627e-4 .14017e-4 .02865 6 0.040476 0.040476 0.000005
2 .69173e-4 .68989e-4 .00265 7 0.199204 0.199203 0.000005
3 .00034125 .00033953 .00504 8 0.258429 0.859833 2.327155
4 .0016796 .0016710 .00514 9 0.210249 0.401035 0.907433
5 .0082667 .0082241 .00514 10 0.134169 0.187047 0.394119
6 .040685 .040475 .00514 11 0.074339 0.087241 0.173549
7 .20023 .19920 .00514 12 0.0379 0.04069 0.073622
8 .25976 .85983 2.3100 13 0.01844 0.018978 0.029203
9 .21133 .40103 .8976 14 0.008758 0.008852 0.010655
10 .13486 .18704 .3869 15 0.004114 0.004129 0.003551
11 .074723 .087239 .1675 16 0.001924 0.001926 0.001078
12 .038095 .040690 .0680 17 0.000897 0.000898 0.000297
13 .018535 .018978 .0239 18 0.000419 0.000419 0.000073
14 .008803 .0088516 .0054 19 0.000195 0.000195 0.000014
15 .0025867 .0025790 .00295 20 .91124e-4 .91124e-4 0.000001
21 .42501e-4 .42501e-4 -
22 .19823e-4 .19823e-4 -
23 .92457e-5 .92457e-5 -
24 .43123e-5 .43123e-5 -
25 .12564e-5 .12564e-5 -
Our results should be very accurate for K > 25, but it then becomes difficult to evaluate the
exact solution. These numerical comparisons show that the asymptotic approximations are quite
25
robust, and are accurate even for relatively small values of L and K . Tables 4 and 6 show that
when n = 25 = K , the two results agree to five decimal places. Loss rates can thus be calculated
to a very high precision using our asymptotic formulas.
References
[1] L. ABOLNIKOV, J.H. DSHALALOW AND A. DUKHOVNY, On stochastic processes in a multilevel
control bulk queueing system, Stoch. Anal. and Appl. 10, 1992, 155-179.
[2] B. D. CHOI AND D. I. CHOI, Queueing system with queue length dependent service times
and its application to cell discarding scheme in ATM networks IEE Proc.-Commun. 143, Feb.
1996, 5-11.
[3] J.H. DSHALALOW, Queueing systems with state-dependent parameters, in: Advances in
Queueing (ed. by J.H. Dshalalow), 61-116, CRC Press, Boca Raton, FL, 1995.
[4] D. FAKINOS, The G/G/1(LCFS/P) queue with service depending on queue size, Eur. J. Oper.
Res. 59, 1992, 303-307.
[5] W. B. GONG, A. YAN AND C. G. CASSANDRAS, The M/G/1 queue with queue-length dependent
arrival rate, Stochastic Models 8, 1992, 733-741.
[6] C.M. HARRIS, Some results for bulk-arrival queue with state-dependent service times, Manage.
Sci. 16, 1970, 313-326.
[7] V.A. IVNITSKIY, A stationary regime of a queueing system with parameters dependent on the
queue length and with nonordinary flow, Eng. Cyb. ,1974, 85-90.
[8] S. Q. LI, Overload control in a finite message storage buffer, IEEE Trans. Commun. 37, 1989,
1330-1338.
[9] K. SRIRAM, Methodologies for bandwidth allocation, transmission scheduling and congestion
avoidance in broadband ATM networks, Comput. Netw. ISDN Syst. 26, 1993, 43-59.
[10] K. SRIRAM AND D. M. LUCANTONI, Traffic smoothing of bit dropping in a packet voice
multiplexer, IEEE Trans. Commun. 37, 1989, 703-712.
26
[12] H. TAKAGI, Analysis of a finite capacity M/G/1 queue with resume level, Performance
Evaluation 5, 1985, 197-203.
27