See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/46487265

Analysis of a general Markovian two-stage

continuous-flow production system with a
finite buffer
Article in International Journal of Production Economics August 2009
DOI: 10.1016/j.ijpe.2008.05.022 Source: RePEc

CITATIONS

READS

55

109

2 authors, including:
Baris Tan
Koc University
71 PUBLICATIONS 847 CITATIONS
SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Analysis of the Impact of Correlated Service and Arrival Times in Manufacturing Systems View project

Modeling and Analysis of Network Organizations for Inter-Firm Cooperation View project

All content following this page was uploaded by Baris Tan on 15 April 2015.
The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.

This article appeared in a journal published by Elsevier. The attached

copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elseviers archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright

Author's personal copy

ARTICLE IN PRESS
Int. J. Production Economics 120 (2009) 327339

Contents lists available at ScienceDirect

Int. J. Production Economics

journal homepage: www.elsevier.com/locate/ijpe

Analysis of a general Markovian two-stage continuous-ow

production system with a nite buffer
Bars- Tan a,, Stanley B. Gershwin b
a
b

Graduate School of Business, Koc- University, Rumeli Feneri Yolu, Saryer, Istanbul, Turkey
Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139-4307, USA

a r t i c l e in fo

abstract

Article history:
Received 5 April 2007
Accepted 6 May 2008
Available online 20 January 2009

Fluid ow models are used in the performance evaluation of production, computer, and
telecommunication systems. In order to develop a methodology to analyze general
Markovian continuous material ow production systems with two processing stages
with an intermediate nite buffer, a general single-buffer uid ow system is modelled
as a continuous time, continuous-discrete state space stochastic process and the steadystate distribution is determined. Various performance measures such as the production
rate and the expected buffer level are determined from the steady-state distributions.
The exibility of this methodology allows analysis of a wide range of models by
specifying only the transition rates and the ow rates associated with the discrete states
of each stage. Therefore, the method is proposed as a tool for performance evaluation of
general Markovian continuous-ow systems with a nite buffer. The solution
methodology is illustrated by analyzing a production system where each machine has
multiple up and down states associated with their quality characteristics.
& 2008 Elsevier B.V. All rights reserved.

Keywords:
Continuous-ow systems
Fluid ow models
Markovian production systems
Level crossing analysis
Performance evaluation

1. Introduction
In this study, we consider a two-stage continuous-ow
system separated by a nite capacity buffer (Fig. 1). The
dynamics of each stage is described by a continuous time,
discrete-state Markov chain where a different ow rate is
associated with each state. This model can represent a
wide range of systems. For example, it may represent a
portion of a factory in which a stage represents an
unreliable machine that may have any one of a wide
variety of up- and down-time distributions; or a machine
with variable quality; or multiple machines in series or in
parallel without intermediate buffers. For another example, it can represent a communications network in which
message ow rates change according to Markov processes.

 Corresponding author.

E-mail addresses: btan@ku.edu.tr (B. Tan),

gershwin@mit.edu (S.B. Gershwin).
0925-5273/$ - see front matter & 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.ijpe.2008.05.022

In the following, we use the terms stage and machine

interchangeably.
There is a vast literature on continuous material ow
models of unreliable production lines. Two station single
buffer unreliable continuous-ow production systems
have been analyzed in various studies, (e.g. Wijngaard,
1979; Gershwin and Schick, 1980; Dubois and Forestier,
1982; Yeralan et al., 1986; Yeralan and Tan, 1997, among
others). In most of these studies, each unreliable machine
has two states: a single up state that represents the
condition of a fully productive machine and a single down
state that represents the condition where the machine is
not productive due to a failure and the failure and repair
times are exponential random variables.
More detailed models of production systems where
each stage is modelled by using more than two states have
been used to approximate general processing, failure,
and repair time distributions by using phase-type dis zdogru and Altok, 2003;
tributions (Altok, 1985; O
Dallery, 1994; Tan, 1998); to study qualityquantity
interactions (Tempelmeier and Burger, 2001; Poffe and

Author's personal copy

ARTICLE IN PRESS
328

B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

Mu

Md

Fig. 1. A single-buffer uid ow system with two stages.

Gershwin, 2005); or to develop new approximation

methods with multiple up and down states (Levantesi
et al., 2003). Similarly, analysis of production lines with
series or parallel structures (Mitra, 1988; Patchong and
Willaeys, 2001), or merge structures (Tan, 2001; Helber
and Jusic, 2004; Diamantidis et al., 2004) also received
attention.
Although a variety of models are used to evaluate the
performance of continuous-ow production systems,
currently there exists no unied methodology to analyze
these systems. In the analysis of continuous-ow models,
once the state space is determined based on the underlying assumptions, the steady-state distribution is determined by analyzing the continuous time-continuous and
discrete state space Markov process. In order to analyze
this process, a set of differential equations that describe
the behavior of the system is derived and then solved
subject to boundary and normalization conditions. Without a general methodology, this process is repeated for
each new model and considerable effort is required to
model and to analyze any given system. This study is
motivated by the need to develop a unied methodology
to analyze all Markovian single-buffer continuous-ow
production systems. With a similar objective, Gershwin
and Fallah-Fini (2007) recently proposed a method to
analyze general discrete-time, discrete-material-ow production lines with single buffer and identical processing
rates. For a multiserver queue with Coxian arrival and
service times and innite waiting space, Bertsimas (1990)
presents an algorithm to determine the system-size
distribution and related performance measures. This
approach can be used to analyze a special class of twostage discrete-material, continuous-time production lines
with an innite buffer. In this line, the processing time of
the rst machine is a Coxian random variable. The second
stage has a number of identical machines in parallel and
the processing time of each machine is also a Coxian
random variable.
Fluid ow models with a single buffer are also used to
evaluate the performance of computer and telecommunication systems, (e.g. Anick et al., 1982; Elwalid and Mitra,
1991). Recently, different methodologies are proposed to
analyze general uid ow models of computer and
telecommunication systems with a nite buffer, (e.g.
Serucola, 2001; Ahn and Ramaswami, 2003; Ahn et al.,
2005; Soares and Latouche, 2006). Although the uid ow
models developed for production and computer/telecommunication systems are similar, the methods developed
for telecommunication and computer systems cannot be
used to analyze production systems directly. The main
difference between the models of telecommunication and
computer systems and the models of production systems
is the operation dependent failures that are observed in
production systems. When the failures are operation

dependent, an idle machine that is blocked or starved

cannot fail. If a machine is partially blocked or partially
starved and operating at a reduced rate, its failure rate will
be lower than its rate when the buffer is partially full. As a
result, the boundary processes when the buffer is empty
or full are not the same as the interior process and all
three processes must be analyzed accordingly.
In this paper, we present a methodology to analyze
general Markovian continuous-ow production systems
with a nite buffer. The dynamics of the process when the
buffer is partially full is determined by solving a set of
rst-order differential and algebraic equations. The unknown coefcients of the solution are determined by
using a level crossing analysis. Namely, we rst determine
the probabilities of entering and exiting the full- and
empty-buffer processes while the machines are in specic
states by using a level crossing analysis. Then we link the
entry and exit probabilities by using the conditional
probabilities that are derived from the boundary processes. The only inputs of the model are the transition
rates of each stage, the processing rates associated with
the discrete states of each stage, and the buffer size.
Therefore, our model is quite general and allows analysis
of a wide range of models by determining the required
inputs. We illustrate our methodology by using a detailed
example of a production system with multiple up and
down states. We also discuss how different models can be
analyzed by using our methodology.
The organization of the remaining part of the manuscript is as follows: In Section 2, we present a specic
model where each stage has multiple up and down states
corresponding to their quality characteristics to describe
the types of models that can be analyzed with our
methodology. In Section 3, we give a description of the
general model, its assumptions, and introduce the variables used in the model. In Section 4, we present our
methodology to analyze the general model and determine
the performance measures of interest. In Section 5, the
methodology is illustrated by analyzing the system
described in Section 2. Finally, conclusions are given in
Section 6.

2. Example
Before presenting the methodology to analyze general
Markovian continuous-ow systems, we rst introduce a
specic example to illustrate the type of models and also
to show the generality of our methodology. In this
example, we consider a production system with two
unreliable machines with multiple up and down
states and a nite buffer studied by Poffe and Gershwin
(2005).
In the system we consider, the rst stage, Mu , has two
up (State 1 and State 1) and three down states (State D1 ,
D1 , and DQ ). We refer to states of M u as down when the
processing rate in that state is 0. In State 1, the machine
produces products with no quality problems but when it
is in State 1, the quality of the products produced is not
perfect. Furthermore, the machine is subject to two
different failures: operational failures (State D1 and

Author's personal copy

ARTICLE IN PRESS
B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

State D1 ) and quality failures (State DQ ) and they have

different mean times to repair. Since these failures are
different in nature, they cannot be modelled with a single
down state. The failure rate is reduced proportionally
when the processing rate of the machine is reduced due
to starvation and blockage. The operational dependent
failure mechanism is described in detail in Section 5.
The second stage, M d , has one up (State 10 ) and one down
state (State 00 ). Similar to the previous case, we refer to
State 00 of the second stage as down because the
processing rate of Md is 0 in that state. More detailed
representation of quality and unreliability characteristics
of machines allows us to investigate quality and quantity
issues jointly in the design and operation of production
systems.
The processing rates of the upstream stage in both of
the up states are equal to mu ; the processing rate of the
downstream stage in its up state is md ; and the processing
rates of all the down states for both stages are equal to 0.
Fig. 2 depicts the state transitions for M u and M d for this
model.
In order to analyze this system, a set of differential
equations that describe the dynamics of the system must
be derived and then solved subject to the boundary
conditions. For example, when mu amd , this model yields
seven rst-order differential equations and three algebraic
equations. In order to determine the steady-state probability distributions, seven equations must be derived
from the boundary processes. Once the steady-state
distribution is obtained all the performance measures of
interest can be determined from the distribution. Poffe
and Gershwin (2005) derive these equations and solve
them explicitly.
In order to analyze another system, e.g., an extension of
this system where the second station is also modelled
with three up and two down states, the same procedure
must be repeated to derive all the equations (for this
model when mu amd , sixteen differential and nine algebraic equations) and then they must be solved.
In the next section, we present a methodology to
analyze general Markovian continuous-ow systems with
a nite buffer. The specic model presented in this section

DQ

rQ

g
u

d

u
1

-1

is analyzed in Section 5 in order to illustrate the

methodology in detail.
3. General model
We consider a continuous-ow system with two stages
separated by a buffer with capacity N (Fig. 1). The state of
the system at time t is st X; au ; ad where 0pXpN is
the buffer level, au 2 f1; . . . ; Iu g is the state of the upstream
stage M u and ad 2 f1; . . . ; Id g is the state of the downstream stage M d . There are Iu Id discrete states in the state
space au ; ad 2 SM . Fig. 3 shows a sample realization of the
system described in Section 2.
The maximum processing rate of M u in state i is mui X0
and the maximum processing rate of M d in state j is mdj X0.
The machines operate at their maximum rates unless they
are starved or blocked. With these denitions, states need
not be classied as up or down states as is the case in
most of the other studies in the literature. A state with a
maximum processing rate equal to zero can be considered
as a down state.
When the buffer is empty in the machine state
au ; ad i; j with mui 0 and mdj 40 then Md is said to
be completely starved and it is forced to stop. However,
when the buffer is empty and mdj 4mui 40, M d is said to be
partially starved and it can continue its production at a
reduced rate of mui . When the buffer is full in machine
state au ; ad i; j with mui 40 and mdj 0 then M u is said
to be completely blocked and the ow into the buffer is
stopped. However, in the same state if mui 4mdj 40, M u is
said to be partially blocked and it can continue its
production at a reduced rate of mdj . We assume that M u
is never starved and M d is never blocked.
We partition the discrete states of the system into
three sets depending on whether the buffer level goes up
(U), down (D), or stays the same (Z) in that state
8
>
i; j 2 U if mui 4mdj ;
>
>
<
i; j 2 D if mui omdj ;
>
>
>
: i; j 2 Z if mu md ;
i

D1

D-1

0
0

Fig. 2. State transition diagram of the system of Sections 2 and 5 with

multiple up and down states.

and SM U [ D [ Z. The number of states in each of these

sets are IU jUj, ID jDj, and IZ jZj, respectively, and
IU ID IZ Iu Id .
For M u, when 0oXoN, the transition time from state i
u
to state i0 is an exponential random variable with rate lii0 .
Similarly for M d, the transition time from state j to state j0
d
is an exponential random variable with rate ljj0 . When M u
is partially blocked, the transition time from state i to
state i0 is also an exponential random variable with rate
cuii0 . Similarly, when Md is partially starved, the transition
d
rate from state j to state j0 is cjj0 .
The time-dependent probability density while the
buffer is partially full is
f x; i; j; t

329

q
probXtpx; au t i; ad t j for 0oxoN.
qx

We assume that the process is ergodic and the steadystate probabilities exist. A formal proof of ergodicity is not
given in this paper. However, it is possible to show

Author's personal copy

ARTICLE IN PRESS
330

B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

u
DQ
D1
D1
1
1
0

20

40

60

80

100
d

120

140

160

180

200

20

40

60

80

100

120

140

160

180

200

20

40

60

80

100
t

120

140

160

180

200

0
1

10
5
0

Fig. 3. Sample path for the system of Sections 2 and 5 with multiple up and down states (mu 1:2, md 1, p 0:01, r 0:1, p0 0:05, r 0 0:10, g 0:05,
h 0:10, r Q 0:10).

ergodicity by demonstrating that all of the states in the

Markovian model constitute a single communicating class
and they are aperiodic. The steady-state density functions
are dened as
f x; i; j lim f x; i; j; t
t!1

for 0oxoN,

(1)

and arranged in column vectors as

f S x ff x; i; jg

for i; j 2 S; S U; D; Z.

(2)

The probability of state 0; i; j at time t when the buffer

is empty is denoted by p0; i; j; t and the probability
of state N; i; j at time t when the buffer is full is denoted
by pN; i; j; t. The steady-state probabilities at the
empty and full-buffer states when au ; ad i; j are
p0; i; j limt!1 p0; i; j; t and pN; i; j limt!1 pN; i; j; t,
respectively.
4. Analysis of interior and boundary processes
In this section, the steady-state distribution is determined by analyzing the continuous time, continuous and
discrete state space Markov process. First, the differential
equations that describe the dynamics of the system when
the buffer is in the interior (0oXoN) and when the buffer
is at the boundary, i.e. when the buffer is empty (X 0) or
full (X N), are derived. Then a solution technique is
developed.
4.1. Interior process
State transition equations: Relating the probability
density of the state at time t h to the probability density

of the state at time t yields

1

0
B
f x; i; j; t h f x  mui  mdj h; i; j; tB
@1 
0

Iu
X
i0 1
i0 ai

luii0 hC
A

Id
X
B
C
B
1

ldjj0 hC
@
A
j0 1
j0 aj

Iu
X

f x  mui0  mdj h; i0 ; j; tli0 i h

i0 1
i0 ai

B
B
@1 

Id
X

Id
X
j0 1
j0 aj

ldjj0 hC
A
d

f x  mui  mdj0 h; i; j0 ; tlj0 j h

j0 1
j0 aj

B
B
@1 

Iu
X
i0 1
i0 ai

luii0 hC
A;

i; j 2 SM .

(3)

The above equation can also be written in differential

form by setting h ! 0 as

qf x; i; j; t
qf x; i; j; t
mui  mdj
qt
qx
0
1
Id
Iu
X
BX
C
u
0
l
ldjj0 C
f x; i; j; tB
ii
@
A
i0 1
i0 ai

j0 1
j0 aj

Author's personal copy

ARTICLE IN PRESS
B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

Iu
X

f x; i

u
; j; tli0 i

i0 1
i0 ai

Id
X

f x; i; j

d
; tlj0 j ;

i; j 2 SM .

(4)

j0 1
j0 aj

In steady state, the above equation yields Iu Id equations

given below
1
0
mui  mdj

Id
Iu
X
C
BX
qf x; i; j
u
0
 f x; i; jB
l
ldjj0 C
ii
A
@
qx
i0 1
j0 1
i0 ai

Iu
X

f x; i

j0 aj

i0 1
i0 ai

Id
X

f x; i; j0 lj0 j ;

i; j 2 SM .

(5)

j 1
j0 aj

Solution of the internal equations: Note that the coefcient of qf x; i; j=qx in Eq. (5) can be positive, negative, or
zero. Then the internal equations given in Eq. (5) can be
written in matrix form as
22
33
qf U x
2"
#3
6 6 qx 7 7 "
#
f U x
7
66
A1 A2 6
7
6 4 qf x 7
57
6 f D x 7,
(6)
D
7
6
5
A3 A4 4
7
6
qx
5
4
f Z x
0
where A1 is a square matrix of size IU ID  IU ID , A4
is a square matrix of size IZ  IZ , A2 is a matrix of size
IU ID  IZ , A3 is a matrix of size IZ  IU ID , and 0 is a
column vector of length IZ . These matrices are determined
by the parameters of the system.
Expanding Eq. (6) gives the following set of equations:
3
2
qf U x
"
#
f x
6 qx 7
7 A1 U
6
(7)
A2 f Z x,
4 qf D x 5
f D x
qx
and
"
0 A3

f U x
f D x

#
A4 f Z x.

where L A1  A2 A1
4 A3 . The solution of this rst-order
matrix differential equation is
"
#
f U x
(11)
eLx w,
f D x
where eLx is a matrix exponential determined by matrix L
and w is a column vector of length IU ID .
Inserting the solution for f U x and f D x given in
Eq. (11) into Eq. (9) yields the solution for f Z x
f Z x O eLx w.

u
; jli0 i

(8)

Eq. (8) is a set of algebraic equations and it can be

solved directly to express f Z x in terms of f U x and f D x
as
"
#
f U x
f Z x O
,
(9)
f D x
where O A1
4 A3 . Since the square matrix A4 has nonzero diagonal elements where each diagonal element is
the sum of the transition rates from state i; j 2 Z, A1
4
always exists.
Now, inserting Eq. (9) into Eq. (8) yields a rst-order
matrix differential equation given below
3
2
qf U x
"
#
f U x
6 qx 7
7
6
(10)
4 qf D x 5 L f D x ,
qx

331

(12)

When vector w is determined, all the density functions are

determined by Eqs. (11) and (12). Since the length of w is
IU ID , IU ID equations are needed to determine the
weights uniquely. We determine these equations by
analyzing the boundary processes in the following. We
rst discuss important concepts and results of level
crossing analysis.
4.2. Level crossing
In order to relate the densities of the partially-full
buffer process and the boundary buffer processes when
the buffer is empty or full, we use a level crossing analysis
similar to the one utilized in Yeralan and Tan (1997). With
this approach, the entry and exit probabilities into the
empty- and full-buffer processes are determined from the
density functions.
In order to explain this approach, rst note that since
the buffer is nite and the process is ergodic, any process
realization must visit any given buffer level an innite
times in the long run. Equivalently, at a buffer level x, the
number of upward crossings in a given time period is
equal to the number of downward crossings in the same
time period in the long run.
In order to dene level crossing at the buffer level x
formally, consider the following event:
Gx; i; j; h fau i; ad j; 0pxpXpx hpNg.
A level crossing at x when au ; ad i; j is dened as
event Gx; i; j; h as h ! 0 when i; j 2 U [ D.
The rate of change of buffer level in state i; j is
jmui  mdj j. Note that jmui  mdj ja0 only when i; j 2 U [ D.
Once event Gx; i; j; h occurs, it will last for h=jmui  mdj j
time units. Therefore, when h ! 0, the probability that
au ; ad changes its state in this time period goes to zero.
The probability of this event can be determined as
Z xh
f x0 ; i; j dx0 f x; i; jh oh. (13)
probGx; i; j; h
x

Let Lx; i; j; T denote the number of level crossings in state

x; i; j in the time interval t; t T. If the system is observed
for T units of time as T ! 1, the probability that event
Gx; i; j; h occurs is also the fraction of time the system
spends in set Gx; i; j; h in the long run. Consequently,
f x; i; jh lim

T!1

Lx; i; j; T
h
oh.
u
T
jmi  mdj j

(14)

Therefore, the expected number of level crossings per

unit time in the long run is determined by the densities

Author's personal copy

ARTICLE IN PRESS
332

B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

and the ow rates as

Iu
X

lim

T!1

Lx; i; j; T
jmui  mdj jf x; i; j.
T

(15)

In state i; j with mui 4mdj , mui  mdj f x; i; j is the

expected number of upward crossings at buffer level
0oxoN per unit time. Similarly in state i; j with mui omdj ,
mdj  mui f x; i; j is the expected number of downward
crossings per unit time. Since at any given buffer level, the
expected number of upward and downward crossings are
equal in the long run, we can also write
Id
Iu X
X
mui  mdj f x; i; j 0.

(16)

i1 j1

The above result can also be derived by adding all the

transition equations given in Eq. (5) for i; j 2 SM that
yields the level crossing equivalence in differential form.
For other results in level crossing analysis, the reader is
referred to Blake and Lindsey (1973) and Brill (1978).
In order to complete the analysis, we must determine
p0; i; j, the steady-state probability that the buffer is
empty and the machine states are au ; ad i; j; and
pN; i; j, the steady-state probability that the buffer is full
and the machine states are au ; ad i; j. Next we focus
on the x 0 boundary.

p0; i0 ; j; tlui0 i

i0 1
i0 ai
i0 ;j2S0
Id
X

p0; i; j0 ; tcdj0 j ;

i; j 2 S0 .

(17)

j0 1
j0 aj
i;j0 2S0

Eq. (17) can be written in matrix form as

dp0S0 t
dt

A0 p0S0 t,

(18)

where p0S0 t fp0; i; j; tg for i; j 2 S0 and A0 is a IS0  IS0

square matrix.
The empty-buffer process ends with a transition into a
state where the buffer level starts increasing. Let
q0; i; j; t 0k t be the rate at which the process enters into
the state i; j 2 U at time t0k t given that the buffer
became empty at time t 0k . This rate can be determined as
q0; i; j; t

Iu
X

p0; i0 ; j; tlui0 i

i0 1
i0 ai

i0 ;j2S0

Id
X

p0; i; j0 ; tcdj0 j ;

i; j 2 U.

(19)

j0 1
j0 aj
i;j0 2S0

Eq. (19) can be written in matrix form as

q0U t B0 p0S0 t,

4.3. Empty-buffer process

(20)

Now we derive the equations that describe the

dynamics of the system when the buffer is empty, or
becomes empty, or stops being empty. As the buffer level
decreases in states i; j 2 D, the buffer eventually becomes
empty if no other transition occurs rst. Once the buffer
becomes empty, it stays empty until the system makes a
transition to a state i; j 2 U. When the buffer is empty, the
set of states where the buffer stays empty is S0 D [ Z
and IS0 jS0 j ID IZ .
State transition equations: Let t 0k be the kth time the
buffer becomes empty. We calculate p0; i; j from the state
transition rates that determine p0; i; j; t 0k t, the probability that X 0 and au ; ad i; j at time t 0k t given
that the buffer became empty at time t0k and has been
empty during t0k ; t 0k t.
In order for the buffer to become empty at time t0k the
machine state au ; ad must have been in set D at that
time. For it to stay empty during t 0k ; t 0k t, au ; ad must
be in set S0 D [ Z during that interval. For it to become
non-empty at time t 0k t, au ; ad must make a transition
into U at that time.
The dynamics of the system during an interval when
the buffer stays empty are given by the following
equations:
1

0
Iu
BX

dp0; i; j; t
 p0; i; j; tB
@
dt

i0 1
i0 ai

luii0

Id
X
j0 1
j0 aj

cdjj0 C
A

where qU t fq0; i; j; tg for i; j 2 U and B0 is a IU  IS0

matrix.
Entry and exit probabilities: In order to link the interior
and the empty-buffer process, we rst analyze how
the buffer becomes empty and then how it exits the
empty-buffer states. Let us dene a discrete time random
enter
; k 1; 2; . . .g sampled from the process
process ffk
fst; tX0g at the instances t 0k , k 1; 2; . . . when the buffer
enter
consists of
becomes empty. The random variable fk
states of the machines at the instant when the buffer
becomes empty for the kth time. That is, if Xt 0k  h 0
enter
and Xt0k 0 as h ! 0 then fk
au t 0k ; ad t 0k . The
enter
in steady
subscript k is dropped to represent fk
state.
The probability that the buffer becomes empty while
the machines are in state i; j is the ratio of the number of
downward crossings in this particular state to the number
of all possible downward crossings at X 0
L0 ; i; j; T=T
enter
i; j lim P
probf
0 0
T!1
i0 ;j0 2D L0 ; i ; j ; T=T
P

mui  mdj f 0 ; i; j

m  mdj0 f 0 ; i0 ; j0

u
i0 ;j0 2D i0

i; j 2 D.

,
(21)

Similarly, let us dene another discrete time random

exit
; k 1; 2; . . .g sampled from the process
process ffk
fst; tX0g at the instances t0k , k 1; 2; . . . when the buffer
level starts increasing following being empty. The random
exit
describes the states of the machines at the
variable fk

Author's personal copy

ARTICLE IN PRESS
B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

instant when the buffer level starts increasing after being

empty for the kth time. That is, if Xt0k  h 0 and Xt0k
exit
au t0k ; ad t0k . Then the
0 as h ! 0 then fk
probability that the process exits the empty-buffer state
with a transition into state i; j 2 U is given as
probf

exit

i; j P

mui  mdj f 0 ; i; j

i0 ;j0 2U

d f 0 ; i0 ; j0
j0

u
i0

m m

i; j 2 U.

(22)
The empty-buffer process relates the probabilities given
in Eqs. (21) and (22). More specically
exit
i; j
probf
X
exit
enter

i; jjf
probf
i0 ;j0 2D

enter
i0 ; j0 probf
i0 ; j0 ;

i; j 2 U.

(23)

Inserting Eqs. (21) and (22) into Eq. (23) and using the
equivalence of the upward and downward crossings given
in Eq. (16) yields
X
exit
enter
mui  mdj f 0 ; i; j
i; jjf
probf

333

given as
p0; i; j

L0 ; i ; j ; TET 0i;j;i0 ;j0 

0

lim

i0 ;j0 2D

T!1

(26)

By using Eq. (15), we can write p0; i; j in terms of the

densities, processing rates, and expected sojourn times as
p0; i; j

mdj0  mui0 f 0 ; i0 ; j0 ET 0i;j;i0 ;j0 .

(27)

0 0

i ;j 2D
0

The i; j; i ; j element of matrix A1

0 determined from
Eq. (18) gives the expected sojourn time in state i; j 2 S0
given that au ; ad starts in state i0 ; j0 2 S0 . Since the
empty-buffer process can start only in states i; j 2 D, we
dene ET 0  to be an IS0  ID matrix that is obtained by
eliminating the columns of A1
0 corresponding to states
in S0 that are not in D, i.e., S0 nD.
Using the solution of the density functions given in
Eq. (11) in Eq. (27) gives
p0 ET 0 0ID IU diagmD w,

(28)

0 0

i ;j 2D

i0 ; j0 mdj0  mui0 f 0 ; i0 ; j0 ;

i; j 2 U (24)

exit
enter
i; jjf
where conditional probabilities probf
0 0
i ; j  are determined from Eqs. (18) and (20).
is the conditional
The i; ji0 ; j0 element of B0 A1
0
probability that the empty-buffer process exits in a
particular state i; j 2 U given that it starts in one of the
states i0 ; j0 2 S0 where the buffer stays empty. Since the
empty-buffer process can start only in states i; j 2 D, let
G0 be a IU  ID matrix that is obtained by eliminating
corresponding to states S0 nD.
the columns of B0 A1
0
Accordingly, by using the solution of the density
functions given in Eqs. (11), (24) can be written in matrix
form as
diagmU 0IU ID w G0 0ID IU diagmD w,
u
i

d
j ji; j

(25)

where
mU fm  m
2 Ug
and
mD fmdj  mui ji; j 2 Dg. We use diag(a) to represent a
diagonal matrix formed with the elements of vector a and
0kl is a k  l matrix of zeros.
P
exit
i; j 1, Eq. (25) gives
Since
i;j2U probf
IU  1 linearly independent equations that will be used
to determine w.
Steady-state probability distribution: Due to ergodicity
of the process, the probability that X 0 and au ; ad
i; j is also the fraction of the total time the process stays
in this state in a given time period the long run.
We can determine the total time the process stays in
state i; j 2 S0 while X 0 in a given time period by
determining the number of times the buffer becomes
empty and the time the process stays in this state for
each time the buffer becomes empty in the same time
period.
Given that the machine states au ; ad i0 ; j0 2 D at
the time the buffer becomes empty, the expected time
that the machine states au ; ad stay in i; j 2 S0 before
exiting to a state au ; ad 2 U is denoted by ET 0i;j;i0 ;j0 .
Then, the steady-state probability of state 0; i; j, p0; i; j is

where p0 fp0; i; jg.

Now, we can also determine the probability that the
buffer is empty as
probX 0

X
i0 ;j0 2S0

p0; i; j uIS0 p0 ,

(29)

where uk 1; 1; . . . ; 1 is a row vector of ones of length k.

4.4. Full-buffer process

The last step is the analysis of the full-buffer process.
As the buffer level increases in states i; j 2 U, the buffer
eventually becomes full if no other transition occurs rst.
Once the buffer becomes full, it stays full until the system
makes a transition to a state i; j 2 D. When the buffer is
full, the set of states where the buffer stays full is SN
U [ Z and ISN jSN j.
State transition equations: Let tN
k be the kth time the
buffer becomes full. We calculate pN; i; j from the state
transition rates that dene pN; i; j; t N
k t, the probability
that X N and au ; ad i; j at time t N
k t given that the
and
has
been full during
buffer became full at time t N
k
N
;
t

t
.
tN
k k
In order for the buffer to become full at time t N
k the
machine state au ; ad must have been in set U at that time.
N
For it to stay full during tN
k ; t k t, au ; ad must be in set
SN U [ Z during that interval. For it to become non-full
at time t N
k t, au ; ad must make a transition into D at
that time.
The dynamics of the system when the buffer stays full
in state i; j 2 SN are given as
0
B
dpN; i; j; t
 pN; i; j; tB
@
dt

1
Id
X
j0 1
j0 aj

ldjj0

Iu
X
i0 1
i0 ai

cuii0 C
A

Author's personal copy

ARTICLE IN PRESS
334

B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

Id
X

pN; i; j ; t


N
XtN
as h ! 0 then jexit
au tN
k N
k ; ad tk . Then
k
the probability that the process exits the full-buffer state
with a transition into state i; j 2 D is given as

d
lj0 j

j0 1
j0 aj
i;j0 2SN
Iu
X

pN; i ; j; t

u
ci0 i ;

i; j 2 SN .

(30)

i0 1
i0 ai
0
i ;j2SN

m  mdj0 f N ; i0 ; j0

u
i0 ;j0 2D i0

pNSN t
AN pN
(31)
SN t,
dt
where pN
SN t fpN; i; j; tg for i; j 2 SN and AN is a ISN 

ISN square matrix.

The full-buffer process ends with a transition into a
state where the buffer level starts decreasing. Let
qN; i; j; t N
k t be the rate at which the process enters
into the state i; j 2 D at time t N
k t given that the buffer
became full at time t N
k . We can determine this rate as
Id
X

The full-buffer process relates the probabilities given in

Eqs. (34) and (35). More specically,
probjexit i; j

probjexit i; jjjenter

0 0

i ;j 2U

i0 ; j0 probjenter i0 ; j0 ;

i; j 2 D.

(36)
Inserting Eqs. (34) and (35) into Eq. (36) and simplifying
by using the equivalence of the upward and downward
crossings given in Eq. (16) gives
X
probjexit i; jjjenter
mdj  mui f N  ; i; j
i0 ;j0 2U

d
lj0 j

i0 ; j0 mui0  mdj0 f N  ; i0 ; j0 ;

pN; i; j ; t

j 1
j0 aj
Iu
X

i; j 2 D,
(37)

i;j0 2SN

i; j 2 D.

(35)

The above equation can be written in matrix form as

qN; i; j; t

mui  mdj f N  ; i; j

probjexit i; j P

u
ci0 i ;

pN; i ; j; t

i; j 2 D

(32)

i0 1
i0 ai

i0 ;j2SN

or in matrix form
N
qN
D t BN pSN t,

(33)

where qN
D t fqN; i; j; tg for i; j 2 D and BN is a ID  ISN
matrix.
Entry and exit probabilities: In order to link the interior
and the full-buffer process, we rst analyze how the
buffer becomes full and then how it exits the full-buffer
states and the buffer level starts decreasing. Let us dene
; k 1; 2; . . .g
a discrete time random process fjenter
k
sampled from the process fst; tX0g at the instances
where the buffer becomes full. The random variable
jenter
describes the states of the machines at the
k
instances tN
k , k 1; 2; . . . when the buffer becomes full for

and XtN
the kth time. That is, if Xt N
k  h N
k N as
N
N

a
t
;
a
t
.
The
subscript
k is
h ! 0 then jenter
u k
d k
k
dropped to represent the random variable in steady state.
The probability that the buffer becomes full while the
process has been in a specic state is the ratio of the
number of upward crossings in this particular state and
the all possible upward crossings at X N 
probjenter i; j P

mui  mdj f N  ; i; j

m  mdj0 f N  ; i0 ; j0

u
i0 ;j0 2U i0

i; j 2 U.
(34)

Similarly, let us dene another discrete time random

; k 1; 2; . . .g sampled from the process
process fjexit
k
fst; tX0g at the instances tN
k , k 1; 2; . . . when the buffer
level starts decreasing following being full. The random
describes the states of the machines at the
variable jexit
k
instant when the buffer level starts decreasing following
being full for the kth time. That is, if XtN
k  h N and

where conditional probabilities probjexit i; jjjenter

i0 ; j0  are determined from Eqs. (31) and (33).
More specically, the i; ji0 ; j0 element of matrix
BN A1
N is the conditional probability that the full-buffer
process exits in a particular state i; j 2 D given that it
starts in one of the states i0 ; j0 2 SN where the buffer stays
full. Since the full-buffer process can start only in states
i; j 2 U, let GN be a ID  IU matrix that is obtained by
corresponding to
eliminating the columns of BN A1
N
states SN nU.
Using the solution of the density functions given in
Eq. (11) in Eq. (37) yields
(38)
0ID IU diagmD  eLN w GN diagmU 0IU ID eLN w.
P
Since i;j2D probjexit i; j 1, Eq. (38) gives ID 
1 linearly independent equations that will be used to
determine w.
Steady-state distribution. Ergodicity of the process
ensures that the probability that X N and au ; ad
i; j is also the ratio of the total time the process stays in
this state in a given time period the long run.
The total time the process stays in state i; j 2 SN while
X N in a given time period can be determined by
multiplying the number of times the buffer becomes full
and the time the process stays in this state once the buffer
becomes full in the same time period.
0 0
Given that the machine states au ; ad i ; j 2 U at
the time the buffer becomes full, the expected time
that the machine states au ; ad stay in i; j 2 SN before
exiting to a state au ; ad 2 D is denoted by ET N
i;j;i0 ;j0 .
Then, the steady-state probability of state N; i; j, pN; i; j
is given as
pN; i; j

X
i0 ;j0 2U

lim

T!1

LN ; i0 ; j0 ; TET N
i;j;i0 ;j0 
T

(39)

By using Eq. (15), we can write pN; i; j in terms

of the densities, processing rates, and expected sojourn

Author's personal copy

ARTICLE IN PRESS
B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

335

4.6. Performance measures

times as
X

pN; i; j

mui0  mdj0 f N  ; i0 ; j0 ET N
i;j;i0 ;j0 .

(40)

0 0

i ;j 2U

The i; j; i0 ; j0 element of matrix A1

N determined from
Eq. (31) gives the expected sojourn time in state i; j 2 SN
given that au ; ad starts in i0 ; j0 2 SN . Since the full-buffer
process can start only in states i; j 2 U, we dene ET N  to
be an ISN  IU matrix that is obtained by eliminating the
columns of A1
N corresponding to states in SN that are not
in U, i.e., SN nU.
Inserting the solution of the density functions given in
Eq. (11) in Eq. (40) yields
pN ET N diagmU 0IU ID eLN w,

(41)

where pN fpN; i; jg.

Then the probability that the buffer is full can be
calculated as
X
probX N
pN; i; j uISN pN .
(42)
i;j2SN

When the probability densities are determined, all

performance measures of interest can be calculated. In a
production setting, the main performance measures of
interest are the production rate and the expected buffer
level.
The production rate is the amount of material
processed per unit time in the long run. The production
rate of the rst stage and the second stage are the same
due to the conservation of ow. Therefore, we give the
production rate of the rst stage without loss of generality. The production rate in the internal states can be
determined in a straight-forward way. Since the rst stage
can be forced to produce at a reduced rate due to partial
blocking and the second stage can be forced to produce at
a reduced rate due to partial starvation, this must be taken
into consideration. The following equation gives the
production rate of the rst stage:

X
i;j2S0

4.5. Solution of the probability densities

X Z

mui p0; i; j

i;j2SM

mui f x; i; j dx

mdj pN; i; j.

(47)

i;j2SN

Once the weight vector w is determined, all the steadystate probabilities are also determined. Since there are
IU ID weights and Eqs. (25) and (28) give a total of IU
ID  2 equations, two additional equations are required to
uniquely determine w.
The rst equation is the equivalence of the total
upward and downward crossings in the interior region.
Integrating Eq. (16) from 0 to N yields
Z

N
0

Id
Iu X
X

mui f x; i; j dx

i1 j1

N
0

Id
Iu X
X

mdj f x; i; j dx,

(44)

Id
Iu X
X
p0; i; j pN; i; j
i1 j1

Id
Iu X
X

i1 j1


xf x; i; j dx NpN; i; j .

(48)

Once the steady-state distribution is determined, other

performance measures of interest can also be evaluated
directly.

5. Analysis of the example

The second equation is the normalization equation

I d Z
Iu X
X

(43)

EX

i1 j1

or in matrix form
Z N

eLx dx w 0.
mU  mD 

The expected buffer level is determined as

f x; i; j dx 1.

(45)

i1 j1

By using Eqs. (11), (12), (29) and (42), the normalization

equation can be written in matrix form as

Z N

eLx dx
uIS0 ET 0 0ID IU diagmD  n
0

N
LN
w 1,
(46)
uISN ET diagmU 0IU ID  e
where n uIU ID uIZ O.
Now Eqs. (25) and (38) with Eqs. (44) and (46) give
IU ID linearly independent equations that uniquely
determine w. Therefore, all the steady-state probability
distributions that describe the dynamics of the system are
determined by these equations.

In this section, we analyze the specic system

described in Section 2 by using our methodology. In order
to explain the methodology, all the variables dened in
Section 3 are given explicitly for this model. We also
evaluate the performance of the system as some of the
system parameters change.
Before dening the variables, we rst discuss modelling operation dependent failures in this setting. Although
our methodology is developed to work with arbitrary
u
d
values of cii0 and cjj0 , in this example and in the
production examples analyzed in Tan and Gershwin
(2007), a specic case where the reduction in the
transition rates at the boundaries is proportional to the
reduction in the processing rate is considered similar to
other papers in the literature (e.g. Gershwin and Schick,
1980). That is when the buffer is empty and M d is
d
d
producing at a reduced rate of mui , cjj0 mui =mdj ljj0 . This
d
u
setting implies that when mi 0, cjj0 0 and therefore it
is not possible to make a transition when Md is completely
starved. Similarly, when the buffer is full and M u is
u
d
producing at a reduced rate of mdj , cii0 mdj =mui lii0 .
u
d
Similar to the previous case, when mj 0, cii0 0

Author's personal copy

ARTICLE IN PRESS
336

B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

and therefore a transition is not possible when M u is

completely blocked.

mu 4md , IU 4, ID 3, and IZ 3. In this case

5.1. Model inputs

mU mu  md mu  md mu mu ;
mD md md md ;
u

Our solution methodology requires only matrices l

u

d
j g,

m fm

and the buffer size N as its inputs. In this specic

u

mZ 0 0 0:

flii0 g, l fljj0 g, c fcii0 g, c fcjj0 g, vectors mu fmui g,

d

There are 10 discrete states in the state space. When

For this specic case, the submatrices A1 , A2 , A3 , and A4

are

example, since cii0 mdj =mui lii0 and cjj0 mui =mdj ljj0 , c
d

and c are dened by the other inputs.

We rst order the states of M u as f1; 1; D1 ; D1 ; DQ g
and number them from 1 to Iu 5. According to the state
transitions for M u and M d given in Fig. 2, the transition
rate matrix of M u is given as
3
2
g  p
g
p
0
0
6 0
p  h 0
p
h 7
7
6
7
6
u
r
0
r 0
0 7.
(49)
l 6
7
6
7
6 0
r
0
r
0
5
4
rQ
0
0
0 r Q

mu mu mu 0 0.

r0

m m

6 u d
g
6
6 mu md
6
6 p0
6 mu
6
6
0
A1 6
6
6
6 p
md
6
6
6
0
6
4
0

php0
mu  md

pgr 0

mu

mu  md

r0

m u  md

mu

mu  md

phr0

mu

mu

mu  md

rp0

 mp

rp0

 mh

md

0
md

rQ

m u  md

7
0 7
7
7
0 7
7
7
7
0 7,
7
7
0 7
7
7
0 7
7
5
0

r Q r

md

(52)
2

0
6
6
A2 6 0
4
0

The processing rates in states f1; 1; D1 ; D1 ; DQ g are

2 pgp0

0
0

rQ

 mr

mu

mu

mu

3T

7
0 7
7 ,
5
0
 mr

 mr

(53)

Similarly, the states of M d are ordered as f1 ; 0 g and

numbered from 1 to Id 2. The transition rate matrix of
M d is given as


p0 p0
.
(50)
ld
r
r0

p0

6
A3 4 0
0

p0

07
5,

In states f10 ; 00 g the processing rates of M d are given as

r  r 0

5.2. Analysis of the model

Once these inputs are given, we can specify matrices
A1 , A2 , A3 , A4 , A0 , B0 , AN , BN and vectors mU , mD , and mZ
directly. Once these matrices and vectors are specied, the
methodology outlined in the preceding sections yields the
desired performance measures directly.
The table given in (51) lists the states, the corresponding processing rates, and the classication of each state in
sets U, D, and Z depending on mu and md . In this section
only the case mu 4md is discussed in detail.

r Q  r

(54)

p0

r  r 0

6
A4 4

md md 0.

3
0

7
5.

(55)

The submatrices for the cases mu md and mu omd can

be written similarly.
When mu amd , the buffer level does not change when
both stages are in down states. Since these states cannot
be reached when the buffer is empty or full, S0 D and
SN U. Therefore IS0 ID 3 and ISN IU 4.
For the empty-buffer process, since Md is completely
starved in all transient states, the matrices A0 and B0 for
the empty-buffer process are
2

r

6
A0 4 0
0

r

rQ  r

3
0

7
5,

(56)

and
2
.

(51)

rQ

60
6
B0 6
40

r
0

0 7
7
7.
0 5

(57)

1
Since S0 D, ET 0  A1
0 and G0 B0 A0 .
For the full-buffer process, M u is partially blocked
in states 1; 10 and 1; 10 and completely blocked in

Author's personal copy

ARTICLE IN PRESS
B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

states 1; 10 and 1; 10 . Then the matrices AN and BN are

3
2
0
r0
0
p mmd  g mmd  p0
u
u
7
6
6
g mmd
p mmd  h mmd  p0
0
r0 7
7
6
u
u
u
AN 6
7,
6
p0
0
r 0
0 7
5
4
g mmd r0
0
p0

matrices and vectors into Eqs. (25) and (38) with Eqs. (44)
and (46) yields a system of equations that determine the
weight vector w. Then Eqs. (47) and (48) yield production
rate and the expected buffer level.
All the results, in this section, are validated by
simulation. Each model is simulated by using both a
continuous ow and also a discrete event simulation
model. When the continuous simulation is run for 106
events, the percentage error between the analytical
production rate and the simulated production rate is less
than 105 . The time required to determine the performance measures by using the general methodology is very
short and not affected by the buffer level. All the models
analyzed in this paper and in the technical report (Tan and
Gershwin, 2007) are solved in less than 0.005 s on a Intel
Core 2 computer with 1.83 GHz processor and with 1 Gb
RAM by using Matlab 7.4.0.
Figs. 4 and 5 show that increasing the processing rate
of each stage increases the production rate until it reaches
its limit. However, the expected buffer level increases with
the processing rate of the rst stage and it reaches its
capacity and decreases with the processing rate of the

(58)
2 m
p md
6 u
6 0
BN 6
4
0

p mmd

h mmd

7
07
7.
5
0

337

(59)

1
Since SN U, ET N  A1
N , GN BN AN .

5.3. Performance evaluation

Now, since all the input matrices and vectors are
determined, the solution methodology outlined in the
preceding sections yields the probability densities and the
performance measures directly. Namely, inserting these

20

0.8

15

[X]

0.6
0.4

10

0.2
0

0
0.5

1
u

1.5

0.5

1
u

1.5

Fig. 4. Effect of the processing rate of the upstream station in the model with multiple up and down states (md 1, p 0:005, r 0:15, p0 0:015,
r 0 0:15, g 0:01, h 0:20, r Q 0:15, N 17).

20

0.8

15

[X ]

0.6
0.4

10

0.2
0

0
0

0.5

1
d

1.5

0.5

1
d

1.5

Fig. 5. Effect of the processing rate of the downstream station in the model with multiple up and down states (mu 1, p 0:005, r 0:15, p0 0:015,
r 0 0:15, g 0:01, h 0:20, r Q 0:15, N 17).

Author's personal copy

ARTICLE IN PRESS
B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

0.9

12

0.89

10

0.88

8
E [X]

338

0.87

0.86

0.85

0.84

0
0

10
N

15

20

10
N

15

20

Fig. 6. Effect of the buffer capacity in the model with multiple up and down states (m1 1, m2 1, p 0:005, r 0:15, p0 0:015, r 0 0:15, g 0:01,
h 0:20, r Q 0:15).

second stage and it approaches zero. Fig. 6 shows that

increasing the buffer level increases the production rate
and the expected buffer level as expected.

6. Conclusion
We presented a general methodology to analyze
continuous-ow material ow two stage-single buffer
production systems. The method handles general Markovian transitions and different processing rates associated
with each state for both stages. The run time of the
method is very fast and not affected by the buffer size.
A wide range of models can be analyzed by our
methodology directly by determining the transition rates
of each stage and the ow rates associated with the
discrete states of each stage. Tan and Gershwin (2007) use
the methodology presented in this study to model and
analyze various single-buffer continuous-ow systems
including the systems where each stage has a number of
identical machines in parallel or in series, systems where
the up- and down-times of each station are Erlang random
variables with different number of stages, and a model of
a three station merge system with a shared buffer.
Analysis of series production lines with limited buffer
capacity is one application domain where this model can
be used. If the failure and repair times of the machines in a
series production line have phase-type distributions, the
upstream and downstream processes of a buffer can be
captured by a two stage model as analyzed in this paper.
By using this model as a building block, a decomposition
method can be used to evaluate the performance of the
system approximately.
In addition to the production systems, our methodology
can also be used in performance evaluation of computer
and telecommunication systems. Since the operation
dependent failure mechanism differentiates the models of
production and computer/telecommunication models, setting the operation dependent failure rates equal to the
original rates in our methodology allows us to use the
same tool in the performance evaluation of computer and
telecommunication systems.

Therefore, we propose our model as a general tool to

model and analyze single-buffer uid ow systems.

Acknowledgments
BA
The rst author acknowledges the support from TU
BITAK. The second author acknowledges the
and TU
support from the Singapore-MIT Alliance and the General
Motors Research and Development Center.
References
Ahn, S., Jeon, J., Ramaswami, V., 2005. Steady state analysis for nite uid
ow models using nite QBDs. Queueing Systems 49, 223259.
Ahn, S., Ramaswami, V., 2003. Fluid ow models and queues: a connection
by stochastic coupling. Stochastic Models 19 (3), 325348.
Altok, T., 1985. On the phase-type approximation of general distributions. AIIE Transactions 17, 110116.
Anick, D., Mitra, D., Sondhi, M., 1982. Stochastic theory of a data handling
system with multiple sources. Bell System Technology Journal 61,
18711894.
Bertsimas, D., 1990. An analytical approach to a general class of G/G/s
queueing systems. Operations Research 38 (1), 139155.
Blake, I., Lindsey, W., 1973. Level-crossing problems for random
processes. IEEE Transactions on Information Theory 19 (3), 295315.
Brill, P.H., 1978. An embedded level crossing technique for dams and
queues. Journal of Applied Probability 16 (1), 174186.
Dallery, Y., 1994. On modeling failure and repair times in stochastic models
of manufacturing systems using generalized exponential distributions.
Queueing Systems Theory Applications 15 (14), 199209.
Diamantidis, A., Papadopoulos, C., Vidalis, M., 2004. Exact analysis of a
discrete material three-station one-buffer merge system with
unreliable machines. International Journal of Production Research
42 (4), 651675.
Dubois, D., Forestier, J.P., 1982. Productivite et en-cours moyens dun
ensemble de deux machines separees par une zone de stockage.
RAIRO Automatique 16, 105132.
Elwalid, A., Mitra, D., 1991. Analysis and design of rate-based congestion
control of high speed networks, i: stochastic uid models, access
regulation. Queueing Systems Theory Applications 9, 1964.
Gershwin, S.B., Fallah-Fini, S., 2007. A general model and analysis of a
discrete two-machine production line. In: Analysis of Manufacturing
Systems 2007.
Gershwin, S.B., Schick, I.C., 1980. Continuous model of an unreliable twomachine material ow system with a nite interstage buffer. Report
LIDS-R-1039, MIT Laboratory for Information and Decision Systems.
Helber, S., Jusic, H., 2004. A new decomposition approach for non-cyclic
continuous material ow lines with a merging ow of material.
Annals of Operations Research 125, 117139.

Author's personal copy

ARTICLE IN PRESS
B. Tan, S.B. Gershwin / Int. J. Production Economics 120 (2009) 327339

Levantesi, R., Matta, A., Tolio, T., 2003. Performance evaluation of

continuous production lines with machines having different processing
times and multiple failure modes. Performance Evaluation 51, 247268.
Mitra, D., 1988. Stochastic theory of a uid model of multiple failuresusceptible producers and consumers coupled by a buffer. Advances
in Applied Probability 20, 646676.
zdogru, U., Altok, T., 2003. Analysis of two-valve uid ow systems
O
with general repair times. In: Analysis and Modeling of Manufacturing Systems. Kluwer Publications, Dordrecht, pp. 255288.
Patchong, A., Willaeys, D., 2001. Modeling and analysis of an unreliable ow
line composed of parallel-machine stages. IIE Transactions 33, 559568.
Poffe, A., Gershwin, S., 2005. Integrating quality and quantity modelling in a
production line. Technical Report ORC-377-05, Massachusetts Institute
of Technology Operations Research Center Working Paper Series.
Serucola, B., 2001. A nite buffer uid queue driven by a Markovian
queue. Queueing Systems 38, 213220.
Soares, A.D.S., Latouche, G., 2006. Matrix-analytic methods for uid
queues with nite buffers. Performance Evaluation 63, 295314.
Tan, B., 1998. Effects of variability on the due-time performance of a
continuous materials ow production system in series. International
Journal of Production Economics 54, 87100.

339

Tan, B., 2001. A three station continuous materials ow merge system

with unreliable stations and a shared buffer. Mathematical and
Computer Modelling 33 (89), 10111026.
Tan, B., Gershwin, S.B., 2007. Modelling and analysis of Markovian
continuous ow production systems with a nite buffer: a general
methodology and applications. Technical Report ORC-381-07, Massachusetts Institute of Technology Operations Research Center
Working Paper Series.
Tempelmeier, H., Burger, M., 2001. Performance evaluation of unbalanced
ow lines with general distributed processing times, failures and
imperfect production. IIE Transactions 33, 293302.
Wijngaard, J., 1979. The effect of interstage buffer storage on the output
of two unreliable production units in series with different production rates. AIIE Transactions 11 (1), 4247.
Yeralan, S., Franck, W., Quasem, M.A., 1986. A continuous materials ow
production line model with station breakdown. European Journal of
Operational Research 27, 289300.
Yeralan, S., Tan, B., 1997. A station model for continuous materials ow
production. International Journal of Production Research 35 (9),
25252541.