Journal of Manufacturing Systems 50 (2019) 119–134

Contents lists available at ScienceDirect

Journal of Manufacturing Systems

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

Technical Paper

An extensive evaluation of CONWIP-card controlled and scheduled start T

time based production system designs
Debjit Roy , Vasudha Ravikumaran
⁎

Indian Institute of Management, Vastrapur, Ahmedabad, 380015, India

ARTICLE INFO ABSTRACT

Keywords: Order release policies such as a card-controlled CONWIP policy aim to improve system responsiveness and
Manufacturing system minimize system-wide inventory levels. It is not clear if order release policies (without card control) can be
Tandem stations equally effective under certain settings of the production system design parameters. In this research, we study
Order release policies the performance of alternate order release (material control policies) such as CONWIP and scheduled start time
Queuing network models
policy under a variety of design parameter settings such as number of stations, station service time character-
Throughput matching
istics, availability of raw material, and the location of the bottleneck station using queuing network models.
Using an exhaustive numerical study, we observe that throughput times in CONWIP systems are about 2–50%
less compared to open control systems; however, for a balanced system, the Deterministic scheduled start time
policy performs equally well for a large set of production system designs. We further identify that the number of
stations and the location of the bottleneck station affect the choice of the order release policy. Additionally, we
also analyze the system for a variety of demand inter-arrival times and check its effect on the expected number of
backlogs and system-wide expected waiting times.

1. Introduction
Companies today must be responsive and agile enough to react
quickly to the fluctuations in customer demand and manage fulfillment
operations with little inventory. Fulfilling customers’ orders by main-
taining large piles of raw-material, work-in-progress and finished goods
inventory not only ties up cash for working capital but also causes long
throughput times and reduces market competitiveness. In a survey
conducted by R. Michael Donovan & Co., 82% of senior executives who
responded said that inventory reduction was a major concern [1].
An effective manufacturing system design includes robust produc-
tion plans and efficient order release policies (also referred to as ma-
terial control policies). As discussed by Lödding and Lohmann [2], the
goal of production planners is to effectively manage capacity and in-
ventories, which includes identifying the optimal capacity investments
in terms of machine and workforce (capabilities/skillsets, timing,
sizing) in the short, medium and long term under deterministic as well
as stochastic inputs (for example, see Van Mieghem [3], Wu et al. [4]
and Alp and Tan [5] for a review of literature). The key material control
policies that are practised on the shop floor are:
• Card controlled system: It is a closed manufacturing system in which
the flow of jobs in the system is controlled by cards. A card is
attached to the job during the manufacturing process. If jobs are
waiting to be processed, a new job enters the system as soon as the
processing of the previous job is completed. Since a few cards may
be waiting to be paired with incoming jobs, the number of cards
provides an upper bound on the total number of parts in progress.
There are mainly two types of card controlled systems in practice,
namely:
○ Kanban system: This is a pull-based system where cards (known
as Kanbans) signal the replenishment of used material from in-
ternal or external suppliers to the buffer of inventories. Kanbans
reduce outages and shortages of materials and supplies, which
improves customer service levels. Kanbans support pull produc-
tion and continuous flow as the material is not produced at the
supplier until a signal to replenish materials is received from the
customer. Using this approach, Kanbans reduce overall inventory
levels (refer to Sugimori et al. [6], Buzacott and Shanthikumar
[7], Krieg and Kuhn [8], Wang and Wang [9], Gstettner and Kuhn
[10]).
○ CONWIP system: CONWIP (constant work-in-progress) is also a
pull-oriented production control system, where the start of each
product manufacturing process is triggered by the completion of
another at the end of the production line. CONWIP is a kind of
single-stage Kanban system. While Kanban systems maintain

⁎
Corresponding author.
E-mail addresses: debjit@iima.ac.in (D. Roy), vasudha.ravikumaran@gmail.com (V. Ravikumaran).

https://doi.org/10.1016/j.jmsy.2018.12.006
Received 14 April 2018; Received in revised form 5 December 2018; Accepted 6 December 2018
Available online 19 December 2018
0278-6125/ © 2019 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
D. Roy, V. Ravikumaran
tighter control of system WIP through the individual cards at
each workstation, CONWIP systems are easier to implement and
adjust since only one set of system cards is used to manage system
WIP. No part is allowed to enter the system without a card (au-
thority). After a finished part is completed at the last workstation,
a card is transferred to the first workstation and a new part is
pushed into the sequential process route. (Refer to Spearman
et al. [11]).
• Scheduled start times policy: This system behaves as an open system
where the jobs enter the line and depart after one pass. Releases into
the line are triggered by the material requirements plan without
regard to the number of jobs in the line. Hence, the number of jobs
in the system varies over time. The throughput (λ) of the system is
based on anticipated orders. A job enters the system at a rate, λ.
There are two types of scheduled start time policies, namely:
• Poisson policy: In this job scheduling policy, the product arrivals
follow a Poisson process (the jobs have exponentially distributed
inter-arrival times and the CV of job inter-arrival times is 1). (The
coefficient of variation (CV), also known as relative standard de-
viation (RSD), is defined as the ratio of the standard deviation (σ)
to the mean (μ) of the variable.) This policy mimics the pattern in
which the jobs arrive at the shop floor (Refer to Borthakur and
Medhi [12]).
• Deterministic policy: In this job scheduling policy, the product
inter-arrival times are constant and the CV of inter-arrival times is
0. Even though the jobs arrive at the shop floor randomly, they are
released for production at equally spaced inter-arrival times.
• Card controlled system with authorized start times: Paired-cell
overlapping loops of cards with authorization (POLCA) is a variant
of the Kanban card control policy and is typically suitable for
companies that produce high-mix and low-volume products. POLCA
divides the work content into flexible and multidisciplinary staffed
work-cells. The production cell only makes semi-finished products
for receiving work-cells which have free capacity. To ensure this
protocol, POLCA-cards circulate between the work-cells, which
signal if and where there is a capacity for further processing (refer to
Krishnamurthy and Suri [13]).
• Workload regulated system: In this process, a job is released into the
network whenever the total amount of remaining work at the bot-
tleneck station falls below a threshold level (Refer to Lozinski and
Glassey [14] and Wein [15]). Also, clearing function based optimi-
zation model that determines the size and the time phasing of the
individual production lots to balance the workload and minimize
the makespan have been proposed (Refer to Graves [16], Günther
[17]).
Though a card controlled system has advantages over an open
system, managing the flow of cards requires more coordination among
the stations and adds to the overhead costs. We expect that for a certain
range of design parameters, the performance of open system policies
(modelled as an open queuing network with tandem stations) with
deterministic start times and arrival rate equal to current customer
demand rate could be comparable to a closed system (modelled as a
closed queuing network) with N CONWIP cards, such that the
throughput of the closed system is equal to the current customer de-
mand rate. In cases where raw materials are always available, there is
no delay in production due to waiting time for raw material arrival. But
raw materials required for production may not be available all the time.
Hence, the availability of raw materials becomes one of the factors
affecting the output of the system.
In this research, we aim to study the three material control policies –
CONWIP, Poisson scheduled start time and Deterministic scheduled
start time – under different sets of design parameters. The experimental
setup has two model variants – (a) raw materials are always available
and (b) raw materials arrive at some rate. Through this research, we
aim to analyze the effect of raw material arrival on expected WIP and
120
Journal of Manufacturing Systems 50 (2019) 119–134
expected system throughput times. We also analyze the effect of de-
mand inter-arrival time parameters on the system performance. The
synchronization of the finished goods with the demand arrivals is
modelled using a join/ fork station. We study the effects of demand
arrival rate and coefficient of variation (CV) of demand inter-arrival
times on the expected number of backlogs and the expected waiting
time for the backlogs to clear. Backlog refers to any order that is waiting
to be fulfilled, which is hence delayed.
The rest of this paper is organized as follows. We review the relevant
literature in Section 2. The system description and formulation of the
research questions are given in Section 3. The raw materials always
available case is discussed in Section 4. The raw materials arriving at
some rate case is discussed in Section 5. The policy conclusions are
derived in Section 6 and the conclusions are drawn in Section 7.
2. Literature Review
We restrict the scope of the literature review to CONWIP and MRP
material control policies only. In particular, we review two areas: (i)
modeling CONWIP and MRP material control policies, and (ii) com-
parison of alternate material control polices.
2.1. Modeling CONWIP and MRP material control policies
Spearman et al. [11] introduce a pull alternative to push known as
the CONWIP system. In this paper, the practical advantages of a pull
system over push are outlined. Since then, several researchers have
modeled the CONWIP-controlled manufacturing system under different
conditions. Economopoulos and Kouikoglou [18] study a single-stage,
constant work-in-process (CONWIP) production system that produces
one product to stock to cope with random demand. The objective is to
determine the CONWIP level and the base backlog that maximize the
mean profit rate of the system. Numerical results show that managing
inventories and backlogs jointly achieved higher profit than other
control policies. Gstettner and Kuhn [10] analyze the performance of a
CONWIP system with a single product flow, exponential service time
distributions, and unlimited demand at the final buffer (saturated
lines). Lee and Seo [19] apply a max-plus algebra based solution
method to compare the performance of three types of WIP-controlled
line production systems (m-node tandem queues with finite buffers)
with constant processing times such as Kanban, CONWIP (constant
work-in-process) and DBR (drum-buffer-rope). Using a decomposition
approach, Lagershausen et al. [20] analyze closed assembly (or dis-
assembly) queuing networks with converging and diverging material
flow, finite buffer sizes, generally distributed processing times, and a
constant number of workpieces. Huang et al. [21] propose a heuristic
algorithm of searching work-in-process (WIP) upper bound for a
CONWIP system. Satyam and Krishnamurthy [22] analyze the perfor-
mance of the CONWIP-controlled manufacturing system using a closed
queuing network with synchronization stations. Satyam and Krishna-
murthy [23] extend the model for performance analysis of multi-pro-
duct CONWIP-controlled manufacturing systems by explicitly con-
sidering the effect of batch size constraints on production performance
using parametric characterizations and traffic process approximations.
Heragu and Gupta [24] argue why a semi-open queuieng network re-
presentation of a CONWIP policy is a better modeling representation
than a closed-queuing network for the manufacturing system (also refer
Roy [25]). While the CONWIP system is typically modeled in the lit-
erature using a closed queuing network, the MRP system is typically
modeled as an open queuing network (see Krishnamurthy et al. [26]).
2.2. Comparison between CONWIP and MRP material control policies
Wein [27] assesses the impact of scheduling on the performance of
semiconductor wafer fabrication facilities. The performance measure
considered here is the mean throughput time for a lot of wafers. The
D. Roy, V. Ravikumaran
goal is to compare the performance of Poisson, Workload regulated,
CONWIP and Deterministic arrivals for a specific design data. He con-
clude that when compared to Poisson policy, Deterministic scheduled
start time policy performs better (16–20% throughput time reduction)
than the CONWIP policy (8–16%). The Workload regulated system
performed better (21–26% throughput time reduction) than the De-
terministic scheduled start time policy. Whitt [28] investigates the re-
lation between open and closed queueing network models. He con-
cludes that open queueing network models are analytically more
tractable but closed queueing network models are more realistic.
Spearman and Zazanis [29] aim to show the benefits of a pull system in
relation to a push system such as less congestion, bounded WIP, and less
throughput time variability. They conclude that, in CONWIP, with in-
creasing N, the throughput increases and then reaches an asymptote. In
a push system, if the throughput is increased beyond a certain point, the
utilization increases and the WIP builds up without a bound. It was also
observed that controlling WIP was more robust than controlling
throughput. The goal of Duenyas [30] is to compare CONWIP (generic
cards vs. product specific cards) with the Workload regulating policy.
He concludes that throughput time with CONWIP (product specific
cards) exceeds the throughput time with generic cards by 3–110% and
throughput time with CONWIP (product specific cards) exceeds the
throughput time with Workload regulating policy by 10–40%. The
order of performance for the specific design data: CONWIP (Specific
cards) performs better than the Workload regulated policy and the
Workload regulated policy performs better than CONWIP (generic
cards). Zahorjan [31] examines the effect of mean system performance
measures on the workload representations chosen. He compares the an
open and closed representations under equivalent constraints such that
that they result in identical system throughput or mean system popu-
lation level for the job classes. It is shown that an open system has a
greater response time than an equivalent closed system.
2.3. Gap in literature and contributions of this research
There have been two recent review articles focused on CONWIP
material control strategies [32,33]. In Table 3, Jaegler et al. [32]
conclude that “it appears that determining which type of system is best
in which type of environment remains quite difficult and requires more
research effort.” We found that inspite of a large number of papers
present in this area, each study analyzes a specific scenario and the
larger implications are difficult to generalize. For instance, Framinan
et al. [33] also discuss that CONWIP performs better than UNIF [34],
FIXED [35] or clock-work (Gilland 2002); however, they state that
there are limited studies to assess the impact of manufacturing condi-
tions on the performance of the CONWIP system. Roderick et al. [35]
include comparison of material control strategies for a range of varia-
bility in processing times on performance measures and show that the
FIXED release strategy yields a higher WIP and a lower throughput in
the system in comparison to CONWIP for exponential service times.
Likewise, [44] state that lot release rule, UNIF, which does not consider
the system status, offers lower performance than CONWIP. However,
we suggest that with appropriate inter-order release times and with
appropriate setting of CONWIP cards, both FIXED and CONWIP systems
can achieve similar throughput. Further, most studies assume that raw
materials are always available. We compare different order release
strategies where raw materials arrive at a certain rate i.e., how does
uncertainty in the raw material availability affect system performance.
It could happen that the orders are released but the raw materials re-
quired for production are not yet available. Studies on comparison of
order release strategies with uncertain raw material availability are
scarce. Bahaji and Kuhl [36] show that the push system outperforms
CONWIP in all experiments involving low mix and high volume. We
show that this result depends on both the location of the bottleneck and
the number of stations. For system with large number of stations and
the last location of the bottleneck, CONWIP clearly outperforms push
121
Journal of Manufacturing Systems 50 (2019) 119–134
based order release system in terms of throughput time performance.
We contribute to this domain by performing a systematic and ex-
tensive experimentation of both balanced and unbalanced system. We
address the concern by providing an analytical modeling framework for
both open and closed system. For fair comparison of performance, we
use the same system operating parameters and perform an exhaustive
simulation. To summarize, our study is different in comparison with the
existing literature in the following ways:
• We study the effect of raw material availability on open and closed
system performance.
• We develop a throughput matching algorithm to match the
throughput performance between an open and a closed system.
• We perform an extensive numerical experimentation to compare
open and closed system performance with various design parameter
setting in particular the location of the bottleneck, number of sta-
tions, and the material availability rate.
3. System description and modeling approach
We consider a tandem manufacturing system where the jobs pro-
ceed to the subsequent station sequentially during the manufacturing
process. We also assume that there exists sufficient buffer space for the
jobs to queue at the stations without any delays. We consider two
variants of the manufacturing system: a balanced system where all
stations have the same utilization and an unbalanced system where the
workload varies among the stations.
We use queuing network models to analyze alternate material
control strategies with a single product. Each manufacturing station is
considered as a node and is characterized by the mean and CV of its
service time. Product arrivals to the queue are characterized by the
mean and CV of inter-arrival times. As a product arrives, it is sent to the
subsequent station on an FCFS basis. If the station is busy, the product
waits in a queue till it enters the station. The expected waiting time
depends on the mean and CV of the service time at the station, utili-
zation of the station and arrival characteristics to the station. The ex-
pected service time and the expected waiting time in the queue together
constitute the expected throughput time of the station. The summation
of the expected throughput times at all stations gives the system-wide
expected throughput time in a single class system. The throughput time
of a manufacturing process is the amount of time required for a product
to pass through that process, thereby being converted from raw mate-
rial into finished goods in the end of the process. The relationship be-
tween the expected WIP, throughput and the expected throughput time
of the system is given by Little's Law [37].
The design parameter set consists of the mean and CV of station
service times, station utilization, the mean and CV of demand inter-
arrival times. In addition to these parameters, we also consider the
mean and CV of raw material inter-arrival times. For modeling pur-
poses, the order-fulfillment system is divided into two systems – System
I and System II. System I denotes the manufacturing system whereas
System II denotes the demand fulfillment process.
System I: System I (manufacturing system) with CONWIP card
control policy is modeled using a closed queuing network whereas for
Scheduled Start Time Policy, it is modeled with a tandem open queuing
network. It consists of K service stations, where the products are sent to
the subsequent stations starting from station 1 to K on a first-come first-
served (FCFS) basis. Each station k is characterized by the mean (Dk)
and CV (cs,k) of the service times. The finished goods, after exiting the
last station, proceed to System II. In the CONWIP policy, the CONWIP
card is released from the last station and reaches the first station after
the finished goods exit the last station. System I is analyzed for different
sets of design parameters and the performance measures such as station
expected throughput times and expected work in progress are obtained.
With respect to availability of raw materials, two analytical model
variants of System I are possible: (1) raw materials are always available:
D. Roy, V. Ravikumaran
Fig. 1. Modeling framework and organization of this paper.
in this case, the raw materials required for production are always
available without any delay. In such situations, raw materials are
purchased in bulk. Hence, the orders do not wait for the raw material to
arrive, and (2) raw materials arrive according to a stochastic arrival
process, characterized by the mean and CV of inter-arrival times ( p,11
and cp,1). In such situations, raw materials are purchased in small
quantities. The synchronization between raw materials arrival buffer
(p1) and order arrival buffer (f1) is modeled as a join/fork station. The
order waits for the raw materials to arrive. The changes in the CV of
raw material inter-arrival times affect the performance measures of the
system.
System II: For both CONWIP policy and scheduled start time policy,
System II models the demand fulfillment process using a fork/join
station. It consists of a finished products arrival buffer and a demand
arrival buffer. Note that both buffers have finite space limits. The fin-
ished goods arrive at a rate equal to the throughput of System I and the
CV of inter-arrival times is equal to the CV of inter-departure times from
the last station in the manufacturing network. The order arrival is
characterized by the mean demand inter-arrival times ( p,21 ) and its CV
(cp,2). System II is analyzed for different values of CV of order inter-
arrival times and buffer capacities. The performance measures include
the expected number of backlogs, which wait at the buffer and the
expected waiting time for the backlogs to get fulfilled.
Based on the modeling approach and the system description, the
framework for our analysis is summarized in Fig. 1. The two analytical
model variants of CONWIP policy (raw materials always available and
raw materials available at a given rate) are shown in Figs. 2 and 3
respectively. Likewise, the two analytical model variants of Scheduled
Start Time Policy are shown in Figs. 4 and 5 .
3.1. Formulation of research questions
To compare scheduled start time policies (open queuing network)
with CONWIP policy (closed queueing network), we first obtain a result
Fig. 2. Closed queuing network model for the CONWIP policy when raw materials are always available.
122
Journal of Manufacturing Systems 50 (2019) 119–134
that shows that the expected throughput time of an open system is
greater than that of the CONWIP system but diminishes as the number
of stations increase. The following observation provides a theoretical
justification for the current study. The number of CONWIP cards in the
practical system is typically set to an integer value. However, the
throughput requirements of a manufacturing system can be time-
varying. In that case, the manufacturing manager can introduce
CONWIP cards to increase throughput depending on demand require-
ments or remove CONWIP cards to adjust the production throughput to
the demand. A time-average of the number of CONWIP cards in the
system can be a fractional number [38]. If we experiment with integral
values of CONWIP cards, then the throughput from the closed system
will be close to that of the open system, but may not match exactly. For
some special cases, the throughput may match exactly (discussed in
numerical experiments).
Theoretical result: Suppose we have a balanced single-class open
BCMP network (FCFS scheduling) [39] with exponential service times
and a balanced closed BCMP network with identical exponential service
times and if the throughput of the closed network (λc) matches exactly
with throughput of the open network (λo), then:
Ro Rc
=
1
Ro K (1)
where:
• R is the expected throughput time of the open network;
o
• R is the expected throughput time of the closed network;
c
• K is the number of stations in the network (K ≥ 2).
Proof. Refer to Table 1 for notation
Since the closed system is a balanced system, the average number of
jobs at each station in steady state would be N . By using arrival theorem
K
D. Roy, V. Ravikumaran Journal of Manufacturing Systems 50 (2019) 119–134

Fig. 3. Closed queuing network model for the CONWIP policy when raw materials arrive at rate λp,1.

Fig. 4. Open queuing network model for the scheduled start time policy when raw materials are always available.

(K 1) U
N= (b)
1 U

KU
Q= o Ro = (c)
1 U

N
N= c Rc =
Dk (N + K 1) R c

Ro Rc N
=1 (d)
Ro Q
Fig. 5. Open queuing network model for the scheduled start time policy when
raw materials arrive at rate λp,1.
(K 1) U
Ro Rc 1 U K 1 1
=1 KU
=1 =
(see Lazowska et al. [40]), the average throughput time at any station Ro K K
1 U
observed by a tagged order arriving to a station in the closed CONWIP-
controlled system would be [1 + K ] Dk , which can be simplified as
N 1 □
] Dk . Hence, the total average throughput time, Rc, is given by
K+N 1
[ K
(K + N − 1)Dk. Table 1
N Additional notations used in the proof.
For a closed queueing network: using Little's Law, c = D (N + K 1)
k
(a). Symbol Parameter
For both systems (CONWIP controlled as well as open system with
λc Throughput of the closed queueing network
Poisson arrivals) to perform at equal throughput rate levels, we match λo Arrival rate to the open queueing network
their corresponding system throughput. N Number of CONWIP cards
Since λc = λo, D (N + K 1) = o
N
Q Expected queue length (including jobs in service) in open queueing
k
network
N+K 1 1 1 Rc Expected throughput time in closed queueing network
= = Ro Expected throughput time in open queueing network
N Dk o U Dk Expected Service time of station k
U Utilization of the station

123
D. Roy, V. Ravikumaran Journal of Manufacturing Systems 50 (2019) 119–134

Table 2 Table 4
Design parameters. Symbols and description.
S. no. Parameter Value Terms Description

1 Uk 80% K Number of stations

2 Dk (s) 1.0 Uk Utilization of the station k
3 λ (s−1) 0.8 Dk Expected service time of the station k
cs,k CV of service times at station k
ca,k CV of inter-arrival times at station k
cd,k CV of inter-departure times at station k
Table 3 N Number of CONWIP cards
Dunb Expected service time at bottleneck station
Comparison of expected throughput times.
Uunb Utilization of bottleneck station
K Ro Rc Ro Rc 1 cunb CV of service times at bottleneck station
Ro K λp,2 Demand arrival rate
cp,2 CV of demand inter-arrival times
3 15 8.75 0.40 0.33 Kf,2 Maximum queue length of finished goods arrival buffer
10 50 43.75 0.12 0.10 Kp,2 Maximum queue length of demand arrival buffer
λ Throughput of System I
Wk Expected waiting time at station k
Rk Expected throughput time of station k
3.1.1. Numerical validation R Expected throughput time of the entire system
WIP Expected work in progress
Consider a balanced single-class open network and a closed network
Lp,2 Expected number of backlogs
with exponential service times. The expected service time at the sta- Lf,1 Expected queue length of product arrival buffer
tions (Dk) is 1 s i.e., D1 = D2 = D3 = D4 = D5 = D6 = D7 = D8 = D9 Lk Expected queue length at station k
= D10 = 1 s, the station utilization Uk = 80% and the throughput of λp, 1 Raw material arrival rate
the system = 0.8 jobs per second. (See Table 2) λf, 1 Product arrival rate
cp, 1 CV of raw material inter-arrival times
From Table 3, we see that the benefit of the closed card controlled cf, 1 CV of inter-arrival times
system in terms of expected throughput times decreases with the in- Kp, 1 Maximum queue length of raw material buffer
creasing number of stations i.e., oR c with three stations and
R R 1
Kf, 1 Maximum queue length of product arrival buffer
o 3
OP Open Poisson start time policy
with 10 stations. We also observe that oR c due to the
Ro Rc 1 R R 1
Ro 10 o K OD Open deterministic start time policy
fact that we consider only integral values for the number of jobs in the CON CONWIP Policy
system, which is used to obtain the throughput time of the closed Nb Expected number of backlogs
W Expected waiting time for the backlogs to get fulfilled (System II)
network. The throughput matching algorithm is discussed later in Al-
gorithm 1.
Motivated by the above observation, we address the following re-
search questions for both balanced and unbalanced manufacturing
4.1.1. Analysis of Scheduled Start Time Policy
systems.
The single product system with K manufacturing stations operating
under scheduled start time policy is modeled with an open queuing
1. How do the Poisson and deterministic start time policies perform
network (see Fig. 4). The arrival process is either Poisson (ca,1 is 1) or
compared to the CONWIP policy for a particular configuration? How
Deterministic (ca,1 = 0) (refer to Borthakur and Medhi [12]). Each
does the relative performance change as the number of stations in
station is modeled using a GI/G/1 queue and the performance measures
the network varies?
are estimated using results from Marshall [41], Marshall [42] and Whitt
2. How does the change in mean and CV of service times as well as raw
[43]. The expected waiting time at each station is determined using Eq.
material inter-arrival times affect the relative performance of the
(2).
three policies?
3. Which material control policy is best suited for a given set of system Wk Dk Uk (ca2, k + cs2, k ) g /2(1 Uk ) (2)
parameters (such as station utilization, expected service times and
CV of service times, position of the bottleneck station for an un- where
balanced system)? 2 2
2(1 Uk ) (1 ca, k )
4. How do the three start time policies affect the expected number of g = exp 3Uk ca2, k + cs2, k
(ca2, k < 1)
backlogs and expected waiting time to fill the backlog?
= 1 (ca2, k 1)
In the subsequent sections, we analyze System I with CONWIP, The CV of inter-departure times (cd,k) which equals the CV of inter-
Poisson, and deterministic scheduled start time policies for both ba- arrival times for the next station is given by Eq. (3). [43]
lanced and unbalanced systems. We perform separate analyses for both
variants of System I, i.e., when raw materials are always available and cd2, k Uk2 cs2, k + (1 Uk2 ) ca2, k (3)
when raw materials arrive at some rate. The notation used for the The expected throughput time for System I is given by Equation (4).
model discussion are presented in Table 4. The expected WIP is estimated using Little's Law.
K

4. Raw materials are always available R= Rk

k=1 (4)
4.1. Model development and analysis where Rk = Dk + Wk, k = 1, …, K

In this section, we analyze the models where raw materials are al- 4.1.2. Analysis of CONWIP policy
ways available for production. Hence, there is no delay involved in the This section describes the queuing network model of a single pro-
availability of raw materials. duct with K manufacturing stations (labeled k = 1, …, K) operating

124
D. Roy, V. Ravikumaran Journal of Manufacturing Systems 50 (2019) 119–134

under the CONWIP policy (see Fig. 2). The total number of cards in the
system is equal to the number of parts in progress (N). To analyze the
CONWIP policy, we adapt the solution approach developed by Satyam
and Krishnamurthy [22]. The analysis starts with the characterization
of the parameters of each station namely the CV of service times (cs,k),
station utilization (Uk) and the CV of inter-arrival times (ca,k) to the
station and establishing a relation among these parameters (refer to Eq.
(5)). Then the mean queue length is found in each station which de-
pends on the service times of the station (Dk) and the expected waiting
time in that station (Wk) (Eqs. (6)–(10)). Finally, all the station para-
meters are linked together to give the linkage equation which states
that the sum of the mean queue lengths at the stations is equal to the
number of CONWIP cards (N)(refer to Eq. (10)). The equations are
explained now.
1. Characteristic equation:
The CV of inter-arrival times to station k + 1 (ca,k+1) is given as a
function of the station utilization (Uk) and the CV of service times
(cs,k) at station k.
ca2, k + 1 (1 Uk2 ) cs2, k + Uk2 ca2, k
2. Mean queue length equation for station k:
The Mean queue length at each station is given by:
Lk = (Dk + Wk )
where
Dk2 cs2, k + ca2, k
Wk fk
1 Dk 2 (7)
and fk is a correction factor.
4.1.4. Algorithm for throughput matching
The solution algorithm begins with an initial value of throughput
λLB and progressively updates the estimates of throughput and different
traffic process parameters until they converge (within a user-specified
tolerance limit) and are consistent with the input parameters. The
iterative procedure starts by equating the initial throughput estimate,
λLB, to the arrival rate, λ, at station 1. To completely specify the arrival
process at station 1, an initial choice of ca,1 is also made. However, the
initial choice of λ and ca,1 might not be consistent with each other.
Therefore, in the next step, the algorithm updates the initial estimate of
ca,1 till it is consistent with the current choice of λ. To update the es-
timate of ca,1, the algorithm uses the linkage and characterization
equations successively at each station in the routing. Subsequently, the
departure process parameters at station 1 and K are estimated. Since
departures from station K form the arrivals to station 1, a new estimate
of ca,1 is obtained. If the old and the new estimates of ca,1 are not within
a pre-defined tolerance limit (ε3), the procedure is repeated with the
new estimate of ca,1. Once the algorithm obtains a value of ca,1, which is
consistent with the choice of λ, in the next step the algorithm verifies
whether the values of λ and ca, 1 are consistent with the other input
(5)
parameters of the network. In particular, the algorithm verifies whether
the sum of the mean queue lengths at stations sums up to N. If this
equation is satisfied within a user-specified tolerance limit (ε2), the
algorithm proceeds to the next step. Otherwise, the choice of λ is in-
cremented by 0.001 and the above steps are repeated. In the next step,
(6)
the algorithm verifies if the λ obtained after the above step is equal to
the throughput of the open queueing network (λac) and is specified by
the user. If this equation is satisfied within some user-specified toler-
ance limit (ε1), the algorithm terminates. Otherwise N is incremented
by 1 and the above steps are repeated till λ is equal to λac. Finally, we
get the throughput of the closed system with N cards that matches the
arrival rate of the open system, λ.
Algorithm 1. Algorithm for throughput matching

4.1.3. Calculation of fk Given: K, λac, Dk, cs,k

To determine fk, a single-class product-form closed queuing net- Define: Δ1 = |λac − λ|, Δ2 = |L − N|, Δ3 = |ca,1 − ca,K|
Initialize: λ1 = λLB, N = 1, Δ1 = ε1 + 1
work, C is defined. The stations in C correspond to the manufacturing
while Δ1 > ε1 do
stations of the original model. To ensure that C has a product-form Initialize: Δ2 = ε2 + 1
solution, it is assumed that the service times are exponentially dis- while Δ2 > ε2 do
tributed for the manufacturing stations in C. Initialize: ca,1 = cs,1, Δ3 = ε3 + 1
while Δ3 > ε3 do
Under this assumption, network C has a product form solution that
Compute ca,1 and ca,K from Equation 7, compute Δ3 = |ca,1 − ca,K| and set
can be obtained using Mean Value Analysis to determine performance ca,1 = ca,K
measures such as the throughput, (λk), and the expected waiting time, end while
WC,k, at each station k in C. Next, consider an M/M/1 queue with arrival Compute fk from Eq. (10) and mean queue lenghts Lk from Equation 8 and 11,
rate equal to the throughput of the network C, λk, and exponential L=
K
k =1 Lk
service times Dk. Let WO,k be the expected waiting time in this M/M/1 Compute Δ2 = |L − N|
queue. Then, fk is defined as the ratio of the expected waiting time in C λ = λ + 0.001
end while
to the expected waiting time in the M/M/1 queue of station k.
Compute Δ1 = |λac − λ|
Therefore, fk is written as: N ++
end while
WC ,k 1 k Dk
fk = = WC , k 2
WO,k D
k k (8)
4.1.5. Analysis of demand fulfillment station
Therefore from Eq. (7), System II is a join station for both the CONWIP policy and the
scheduled start time policy. System II is analyzed for different values of
Dk2 cs2, k + ca2, k 1 k Dk
Wk WC ,k 2
order-inter arrival times ( p,21 ), its CV (cp,2), and buffer capacities. The
1 Dk 2 k Dk (9) performance measures include the expected number of backlogs and
the expected waiting time for the backlogs to get fulfilled. The syn-
1. Linkage equation: Note that the CONWIP card is sent back to Station chronization between the demand arrival and finished products arrival
1 immediately after production, hence this equation differs from the buffer are modeled as a join station (see Satyam and Krishnamurthy
linking equation proposed by Satyam and Krishnamurthy [22]. [22]).
K
Lk = N 4.2. Analytical model validation
k= 1 (10)
This section describes the set of design experiments conducted to
validate the analytical model results. The analytical results are

125
D. Roy, V. Ravikumaran Journal of Manufacturing Systems 50 (2019) 119–134

Table 5
Results of model validation.
Expected absolute error % Percentage error range
Policy No. of scenarios R W R W

Poisson 12 3.41 3.27 (1.14, 4.98) (0.7, 6.37)

CONWIP 12 1.69 3.33 (0.62, 3.18) (0.68, 6.43)
Deterministic 12 2.77 3.37 (0.88, 3.84) (0.66, 6.45)

validated using the models developed using Simulation with Arena Table 7
software. There are 36 scenarios (12 scenarios each for the three ma- Comparison of expected throughput times when raw materials are always
terial control policies) considered for the case when raw materials are available.
always present. For each material control policy, we have four scenarios Experiment no. λ N Comparison of expected throughput times
(two levels of expected service time, and two levels of CV of service
time) for the balanced system and eight scenarios for the unbalanced OP CON %dec. OD %dec.
system (two levels of expected service time, two levels of CV of service
1 0.8 3 6.50 3.75 −42.31 3.31 −49.08
time, and two levels of location of the bottleneck station). For each 2 0.8 7 15.00 8.75 −41.67 11.66 −22.27
scenario, 20 replications were run for 24 h with a warm up period of 3 0.8 11 14.80 13.75 −7.09 11.44 −22.70
5 h. The expected absolute error percentage is obtained for system 4 0.8 35 50.00 43.75 −12.50 46.51 −7.00
parameters – expected work in progress (WIP) for System I, expected 5 0.94 3 11.6 3.19 −72.5 3.19 −72.5
6 0.94 3 7.13 3.19 −55.25 3.67 −48.53
throughput time (R) for System I, expected number of backlogs in 7 0.94 11 23.11 11.70 −49.37 14.21 −37.50
System II and expected waiting time for the backlogs to get cleared in 8 0.94 11 23.11 11.70 −49.37 19.60 −15.19
|y y|
System II using the formula a y s × 100 , where ya is the result obtained 9 0.94 9 17.68 9.57 −45.87 9.33 −47.22
s
10 0.94 9 12.24 9.57 −21.81 9.93 −18.87
from analytical model and ys is the result obtained from the simulation
11 0.94 34 45.70 36.17 −20.85 37.29 −18.40
model. The distribution of the percentage error of the above mentioned 12 0.94 34 45.70 36.17 −20.85 43.24 −5.38
parameters is given in Appendix A. For the four parameters, the abso-
lute error percentage range and the expected values are given in
Table 5. The frequency distribution of the absolute percentage of error Table 8
for the expected throughput time of System I, expected work in progress Analysis of System II (balanced system).
in System I, expected number of backlogs in System II and expected K Kp,2 Kf,2 cp,2 cs,k Nb W (s)
waiting time for the backlog to get cleared in System II is given in
Appendix A. 3 3 100 0.5 0.3 1.64 1.37
10 7 100 0.5 1.0 5.19 4.32
3 3 100 1.0 0.3 1.59 1.32
4.3. Numerical results 10 7 100 1.0 1.0 5.12 4.26
3 3 100 2.5 0.3 1.45 1.21
10 7 100 2.5 1.0 4.6 3.83
In this section, we numerically analyze the models where raw ma-
3 11 100 0.5 0.3 9.28 7.73
terials are always available for production. For the performance eva- 10 18 100 0.5 1.0 16.23 13.53
luation of System II, the demand arrival rate is taken to be 1.2 jobs per 3 11 100 1.0 0.3 9.15 7.62
second. The effect of changes in the CV of demand inter-arrival times is 10 18 100 1.0 1.0 16.00 13.33
analyzed by changing the CV value to 0.5, 1 and 2.5 for each experi- 3 11 100 2.5 0.3 8.25 6.88
10 18 100 2.5 1.0 14.40 12.00
ment. The experimental details for both balanced and unbalanced sys-
tems are summarized in Table 6. The comparison of expected WIP and
expected throughput times for both systems is given in Table 7. cs,k = 0.3 and 1.0 respectively, the demand arrival rate λp,2 = 1.2 s−1
Balanced system: In this section, we consider that the stations of the the CV of demand inter-arrival times cp,2 = 0.5, 1, and 2.5 (see Table 8).
system have the same expected service times (Dk = 1) and we consider Unbalanced system: In this section, we consider that the system has
two values of the CV of service times (0.3 and 1). The values of the one bottleneck station with higher service times than the rest of the
expected number of backlogs and the expected waiting time for the stations (Dunb = 1 s and Dk = 0.8 s). We consider two extreme positions
backlogs to get fulfilled in a balanced system are calculated for the for the bottleneck station - the first station and the last station. These
number of stations K = 3 and 10, the CV of service times of the station

Table 6 Table 9
Design parameters for the case when raw materials are always available. Analysis of System II (unbalanced system).
Experiment no. K Uk Dk (s) cs,k Uunb Dunb (s) Position of bottleneck K cs,k Kp,2 Kf,2 cp,2 Nb W (s)

1 3 80% 1 0.3 – – – 3 0.3 3 100 0.5 1.15 0.96

2 3 80% 1 1.0 – – – 10 0.3 9 100 0.5 5.96 4.97
3 10 80% 1 0.3 – – – 3 0.3 3 100 1.0 1.14 0.95
4 10 80% 1 1.0 – – – 10 0.3 9 100 1.0 5.87 4.89
5 3 75% 0.8 0.3 94% 1.0 First station 3 0.3 3 100 2.5 1.02 0.85
6 3 75% 0.8 0.3 94% 1.0 Last station 10 0.3 9 100 2.5 5.26 4.38
7 3 75% 0.8 1.0 94% 1.0 First station 3 1.0 11 100 0.5 7.74 6.45
8 3 75% 0.8 1.0 94% 1.0 Last station 10 1.0 34 100 0.5 30.84 25.70
9 10 75% 0.8 0.3 94% 1.0 First station 3 1.0 11 100 1 7.63 6.36
10 10 75% 0.8 0.3 94% 1.0 Last station 10 1.0 34 100 1 30.38 25.32
11 10 75% 0.8 1.0 94% 1.0 First station 3 1.0 11 100 2.5 6.83 5.69
12 10 75% 0.8 1.0 94% 1.0 Last station 10 1.0 34 100 2.5 27.20 22.67

126
D. Roy, V. Ravikumaran Journal of Manufacturing Systems 50 (2019) 119–134

Fig. 8. Expected throughput time vs number of stations (upstream position of

Fig. 6. Expected throughput time vs number of stations when raw materials are bottleneck) (Experiments 7 and 11).
always available (Experiments 8 and 12).

two cases are analyzed separately. The value of the expected number of
backlogs and the expected waiting time for the backlogs to get fulfilled
in an unbalanced system are calculated for the number of stations K = 3
and 10, the CV of service times of the station cs,k = 0.3 and 1.0 re-
spectively, demand arrival rate λp,2 = 1.2 s−1 and the CV of demand
inter-arrival times cp,2 = 0.5, 1, and 2.5 (see Table 9).

4.4. Insights

As expected, from the numerical results, it can be concluded that

both Deterministic start time policy and CONWIP policy provide im-
proved performance over Poisson input, by reducing the expected
Fig. 9. Expected throughput time vs number of stations (downstream position
throughput times and expected work in progress (WIP).
of bottleneck) (Experiments 8 and 12).

4.4.1. Effect of change in the number of stations on expected throughput

time
From the experiments, it is observed that as the number of stations
increases, the expected WIP of the system increases almost linearly. For
both balanced and unbalanced systems, as the number of stations in-
creases, CONWIP policy performs better than both Poisson and de-
terministic scheduled start time policies. Fig. 6 illustrates the variation
in expected the throughput time for Poisson scheduled start time policy,
deterministic scheduled start time policy, and CONWIP policy.
4.4.2. Effect of CV of service time on expected throughput time
From the experiments, it is observed that as the CV of station service
time increases, the expected WIP of the system increases nonlinearly. As
the CV increases, CONWIP policy becomes more efficient than both
open Poisson and deterministic scheduled start time policies. As an il-
lustration, we only show the variation in the expected throughput time
for Poisson scheduled start time policy, deterministic scheduled start
time policy and CONWIP policy for an unbalanced system (see Fig. 7).
In all the cases, both Deterministic scheduled start time policy and
CONWIP policy perform better than Poisson scheduled start time policy
(illustrated in Figs. 8 and 9 ). For raw material arriving at some rate, as
the CV of raw material inter-arrival time increases, the expected
throughput time and expected WIP increases.
4.4.3. Effect of the location of bottleneck on throughput time
From the experiments, it is observed that the Poisson scheduled start
time policy performs better when the bottleneck station is placed in the
downstream location. The deterministic scheduled start time policy
performs better when the bottleneck station is placed in the upstream
location. The position of the bottleneck does not have any effect on the
CONWIP policy.

Fig. 10. Expected throughput time vs number of stations – Poisson scheduled

Fig. 7. Expected throughput time vs CV of service times – unbalanced system
start time policy (Experiments 5, 6, 9, and 10).
(Experiments 5 and 7).

127
D. Roy, V. Ravikumaran
Fig. 11. Expected throughput time vs number of stations – deterministic
scheduled start time policy (Experiments 5, 6, 9, and 10).
Fig. 12. Expected throughput time vs number of stations – CONWIP policy
(Experiments 5, 6, 9, and 10).
Figs. 10–12 illustrate the variation in the expected throughput time
for Poisson scheduled start time policy, deterministic scheduled start
time policy, and CONWIP policy, respectively.
4.4.4. Effect of CV of demand inter-arrival times on expected number of
backlogs
From the analysis conducted on System II, it can be concluded that
when the raw materials are always available, the finished products’
inter-arrival times and their CV does not vary with the changes in the
CV of demand inter-arrival times as they are independent of each other.
Hence, when there are large variations in order arrivals (increasing CV
of demand inter-arrival times), the number of finished products stored
in the inventory increases which reduces the number of backlogs to be
fulfilled. Hence, as the CV of demand inter-arrival times increases, the
expected number of backlogs and the expected waiting time for the
backlogs to get fulfilled decreases. This has been illustrated in Fig. 13.
5. Raw materials arrive at some rate
5.1. Model development and analysis
In this section, it is considered that the raw materials arrive at some
rate. The product arrival buffer and raw material arrival buffer are
taken as a join/fork station. It is analyzed according to the formulas
given by Satyam and Krishnamurthy [22]. In our model, the CONWIP
card is returned to the empty card buffer f, 1 as soon as the production
is complete. We also study the effect of changes in the CV of raw ma-
terial inter-arrival times on the system throughput and station utiliza-
tion.
128
Journal of Manufacturing Systems 50 (2019) 119–134
Fig. 13. Expected number of backlogs vs CV of demand inter-arrival times for
an unbalanced system with CV of service time is 1 and number of stations 10.
5.1.1. Analysis of scheduled start time policy
This section describes the queuing network model of a single pro-
duct system with K manufacturing stations operating under the
scheduled start time policy (see Fig. 5). A station k is described by three
parameters – expected service time (Dk), CV (cs,k) of the service time
and station utilization (Uk). The product arrival process for Poisson and
dDeterministic scheduled start time policies are characterized by the
mean inter-arrival time ( f ,11), and the CV of inter-arrival times, cf,1 = 1
and cf,1 = 0, respectively. The raw material arrival process is char-
acterized by the mean inter-arrival times ( p,11) and CV of inter-arrival
times (cp,1). The raw material arrival buffer and the product arrival
buffer are considered as join stations.
After obtaining the arrival characteristics from the join station, all
the parameters are calculated in the same manner as in the case when
raw materials are always available. However, there is an additional
waiting time involved due to the queue in the product buffer. Hence,
the expected throughput time is given by:
K
R= (Rk ) + Lf ,1/
f ,1
(11)
k=1
5.1.2. Analysis of CONWIP policy
This section describes the queuing network model of a single pro-
duct with K manufacturing stations operating under the CONWIP policy
(see Fig. 3). The total number of cards in the system is equal to the
number of parts in progress (N). The product arrival process is char-
acterized as Poisson if the CV of inter-arrival times (cf,1) is 1 or De-
terministic if cf,1 is 0 and the arrival rate is (λf,1). The raw material
arrival process is characterized by the mean inter-arrival times ( p,11)
and CV of inter-arrival times (cp,1). The raw material arrival buffer and
the product arrival buffer is considered as a join station.
After obtaining the arrival characteristics from the join station, all
the parameters are calculated in the same manner as in the case when
raw materials are always available. In addition to the queue length at
K
the stations ( k = 1 Lk ), the linkage equation also considers the expected
queue length in the product arrival buffer (Lf,1).
K
Lf ,1 + Lk = N
k=1 (12)
The solution approach is identical to the one explained in Section 4.1.
5.2. Analytical model validation
There are 108 scenarios (36 each for three material control policies)
including both the model variants for the case with raw materials ar-
riving at some rate. For the balanced case, we have 12 scenarios, which
D. Roy, V. Ravikumaran Journal of Manufacturing Systems 50 (2019) 119–134

Table 10
Results of model validation.
Policy No. of scenarios Expected absolute error % Percentage error range

R W R W

Poisson 36 4.91 4.49 (3.47, 7.05) (2.98, 6.51)

CONWIP 36 4.03 4.51 (1.66, 7.35) (3.03, 6.53)
Deterministic 36 4.01 4.52 (1.86, 6.54) (3.99, 6.57)

comprise one raw material arrival rate, three levels of raw material
inter-arrival time CV, two levels of buffer limit K, and two levels of
service time CV. For the unbalanced case, we have 24 scenarios, which
comprise one raw material arrival rate, three levels of raw material
inter-arrival time CV, two levels of buffer limit K, two levels of service
time CV, and two levels of CV of the bottleneck station's service time.
The remaining experimental conditions are similar to the cases where
raw materials are always available. The distribution of percentage error
of the above mentioned parameters is given in Appendix A. For the four
parameters, the absolute error percentage range and expected values
are given in Table 10.

5.3. Numerical results
In this section, the numerical experiments are performed for the
case where raw materials arrive at some rate. The card/authorization
signal arrival buffer and raw material arrival buffer are taken as a join/
fork station. We also study the effect of changes in CV for the raw
material inter-arrival times on the system throughput and station uti-
lization. The experiments are described in Table 11. The analysis and
insights obtained for System II are similar to the ones obtained when
raw materials are always available.
Balanced System: In this section, we consider that the stations of the
system have the same expected service times (Dk = 1) and we consider
two values of CV of service times (0.3 and 1). The comparison of the
expected throughput time, the value of expected number of backlogs
and the expected waiting time for the backlogs to get fulfilled are
shown in Table 12.
Unbalanced System: In this section, we consider that the system
has one bottleneck station with higher service times than the rest of the
stations. (Dunb = 1 and Dk = 0.8). We consider two positions for the
bottleneck stations – first station and last station. These two cases are
analyzed separately. The comparison of the expected throughput time,
the value of expected number of backlogs, and the expected waiting
time for the backlogs to get fulfilled has been shown in Table 12.
Fig. 15. Expected throughput time vs number of stations (CV = 0.3)
(Experiments 1 and 7).
5.4. Insights
As expected, from the numerical results, it can be concluded that
both Deterministic start time policy and CONWIP policy provided im-
proved performance over Poisson input, by reducing the expected
throughput times and expected work in progress (WIP). The effect of
number of stations, CV of service times and location of bottleneck, and
CV of raw material inter-arrival times on throughput time and number
of expected backlogs are discussed below.
5.4.1. Effect of CV of raw material inter-arrival times on expected
throughput time
From the experiments, it is observed that as the CV of raw material
inter-arrival time increases, the expected system throughput time in-
creases for all the three policies. As an illustration, we only show this
effect for the Poisson scheduled start time policy (see Fig. 14).
5.4.2. Effect of CV of service times on expected throughput time
For low CV of service time and higher number of stations – CONWIP
performs better than the Deterministic scheduled start time policy and

Fig. 14. Expected throughput time vs Number of stations: CV of raw material

Fig. 16. Expected throughput time vs number of stations (CV = 1)
inter-arrival times for Poisson scheduled start time policy (Experiments 1, 2, 3,
(Experiments 4 and 10).
7, 8, and 9).

129
D. Roy, V. Ravikumaran
Fig. 17. Expected number of backlogs vs CV of raw material inter-arrival times
for an unbalanced system with CV of service times 0.3 and number of stations 3.
Fig. 18. Expected throughput time vs CV of service time for balanced system for
Poisson start-time scheduled policy.
both perform better than the Poisson scheduled start time policy (illu-
strated in Fig. 15). For high CV of service time and lower number of
stations – CONWIP and deterministic scheduled start time policy per-
form in a similar fashion and both perform considerably better than the
Poisson scheduled start time policy. For high CV of service time and
higher number of stations – CONWIP and deterministic scheduled start
time policy perform in a similar fashion and both perform considerably
better than the Poisson scheduled start time policy (illustrated in
Fig. 16).
5.4.3. Effect of CV of raw material inter-arrival time on expected number of
backlogs
From the experiments, it is observed that as the CV of raw material
inter-arrival time increases, the expected number of backlogs and the
expected waiting time for the backlogs to get fulfilled also increases.
This relationship has been illustrated in Fig. 17.
6. Policy conclusions
6.1. Effect of raw material availability
From the experiments conducted on both balanced and unbalanced
systems for various design input parameters (Sections 4.3 and 5.3), it
130
Journal of Manufacturing Systems 50 (2019) 119–134
Fig. 19. Expected throughput time vs CV of service time for balanced system for
CONWIP policy.
Fig. 20. Expected throughput time vs CV of service time for balanced system for
deterministic start-time scheduled policy.
can be concluded that the policies perform considerably better when
raw materials are always available. It is also observed that the policies
follow similar trend under various design parameters irrespective of
raw material availability. In Figs. 18–20, we show the performance of
Poisson, Deterministic and CONWIP policies for different values of CV
of service time for a balanced system with number of stations = 3 and
CV of raw material inter-arrival times for raw material arriving at some
rate case = 1. At low CV of service times, the raw material availability
has less effect of throughput times.
6.2. Choice of start time policies in balanced system
From the experiments conducted on the balanced system for var-
ious design input parameters (Sections 4.3 and 5.3), it can be con-
cluded that for low CV of service time and high number of stations –
CONWIP outperforms the deterministic scheduled start time policy.
Further, both CONWIP and deterministic scheduled start time policies
perform considerably better than the Poisson scheduled start time
policy. For the remaining scenarios, CONWIP performs similar to de-
terministic scheduled start time policy. When raw materials arrive at
some rate, as the CV of raw material inter-arrival time increases, the
expected throughput time and expected WIP increases. The summary
of the performance comparison between CONWIP policy and
D. Roy, V. Ravikumaran
Fig. 21. Performance comparison of CONWIP policy and deterministic sched-
uled start time policy for balanced system (refer to Tables 7 and 12).
Fig. 22. Performance comparison of CONWIP policy and deterministic sched-
uled start time policy for unbalanced system (refer to Tables 7 and 12).
deterministic scheduled start time policy for a balanced system is
shown in Fig. 21.
6.3. Choice of start time policies in unbalanced system
From the experiments conducted on unbalanced systems (Sections
4.3 and 5.3), it can be concluded that CONWIP performs better than the
deterministic scheduled start time policy giving high reduction in ex-
pected throughput times and expected WIP for upstream location of the
bottleneck and lower number of stations as well as for downstream
location of the bottleneck for both higher and lower number of stations.
The performance of the CONWIP policy is approximately equal to that
of the deterministic scheduled start time policy for the upstream loca-
tion of the bottleneck and for a higher number of stations. In all the
cases, both deterministic scheduled start time policy and CONWIP
policy perform better than the Poisson scheduled start time policy. For
raw material arriving at some rate, as the CV of raw material inter-
arrival time increases, the expected throughput time and expected WIP
increases. The summary of the performance comparison between the
CONWIP policy and the deterministic scheduled start time policy for an
unbalanced system is shown in Fig. 22.
7. Conclusions and future work
The contributions of this research lie in performance analysis and
131
Journal of Manufacturing Systems 50 (2019) 119–134
design trade-offs among three material control policies namely
CONWIP policy, Poisson scheduled start time policy, and deterministic
scheduled start time policy under a large set of design parameters. The
design parameters include station utilization (Uk), expected service
times of the station (Dk), CV of service times of the station (cs,k), lo-
cation of bottleneck for an unbalanced system, demand arrival rate,
CV of inter-arrival times, raw material arrival rate, and CV of raw
materials inter-arrival times (for the case where raw materials arrive
at some rate). From the numerical results, various insights were de-
veloped. As the number of stations increases, the expected throughput
time and expected WIP increases almost linearly. With an increase in
the CV of service times, CONWIP performs better than the other two
systems for a fixed number of service stations. As the CV of demand
inter-arrival times increases, the expected number of backlogs and the
expected waiting time for the backlogs to get fulfilled decreases. With
an increase in the CV of raw material inter-arrival times (for the case
where raw materials arrive at some rate), the throughput of the system
decreases and the expected throughput time, the expected number of
backlogs and the expected waiting time for the backlogs to get fulfilled
increases. As expected, the CONWIP and the deterministic scheduled
start time policies perform better than the Poisson scheduled policies
in all the cases. For a balanced system, the CONWIP performs better
than the Deterministic scheduled start time policy only when the CV of
service times is low, and the number of stations is high. For all the
other cases (for low CV of service times and less number of stations,
high CV of service times and any number of stations), their perfor-
mance is similar. In an unbalanced system, the CONWIP policy per-
forms similar to the Deterministic scheduled start time policy only for
upstream location of bottleneck and higher number of stations. For all
the other cases (for upstream location of the bottleneck and lower
number of stations and downstream location of bottleneck and any
number of stations), the CONWIP performs better than the
Deterministic scheduled start time policy. From the above observa-
tions it can be concluded that though CONWIP is a better material
control policy in most cases and could be adopted on the shop floor for
all the design parameters considered in this research, the
Deterministic policy is comparable with CONWIP in some cases and
hence can be adopted as an alternate start time policy. Further,
CONWIP and Deterministic policy performs similar when raw mate-
rials arrive at some rate. The performance of the deterministic policy
deteriorates when raw materials are always available.
This research focuses on the evaluation of a single class system. We
have analyzed the system for various design parameters. In future, a
multiple class system (more than one product class) could be analyzed
for the various design parameters. For an unbalanced system, instead of
just one bottleneck station, a combination of two or more bottleneck
stations could be considered and the impact of its location on the
system parameters can be evaluated. More material control policies
such as POLCA and Workload regulated policies could be considered for
performance analysis.
Acknowledgements:
We would like to thank Prof. Mary Vernon (University of
Wisconsin–Madison) for encouraging us to work in this area. We would
also like to thank the editor and reviewers for their feedback, which
helped us immensely to improve the structure and exposition of this
paper.
D. Roy, V. Ravikumaran Journal of Manufacturing Systems 50 (2019) 119–134

Appendix A

Table 11
Design parameters for the case when raw materials arrive at some rat.
Experiment no. System I parameters System II parameters

λp,1 (s−1) cp,1 K Dk (sec) cs,k Dunb cunb Bottleneck's position λp,2 (s−1) cp,2 Kp,2 Kf,2

1 0.8 0.5 3 1.0 0.3 – – – 0.8 0.5 100 3

2 0.8 1.0 3 1.0 0.3 – – – 0.8 1.0 100 3
3 0.8 2.5 3 1.0 0.3 – – – 0.8 2.5 100 3
4 0.8 0.5 3 1.0 1.0 – – – 0.8 0.5 100 11
5 0.8 1.0 3 1.0 1.0 – – – 0.8 1.0 100 11
6 0.8 2.5 3 1.0 1.0 – – – 0.8 2.5 100 11
7 0.8 0.5 10 1.0 0.3 – – – 0.8 0.5 100 7
8 0.8 1.0 10 1.0 0.3 – – – 0.8 1.0 100 7
9 0.8 2.5 10 1.0 0.3 – – – 0.8 2.5 100 7
10 0.8 0.5 10 1.0 1.0 – – – 0.8 0.5 100 35
11 0.8 1.0 10 1.0 1.0 – – – 0.8 1.0 100 35
12 0.8 2.5 10 1.0 1.0 – – – 0.8 2.5 100 35
13 0.94 0.5 3 0.8 0.3 1.0 0.3 first station 0.94 0.5 200 12
14 0.94 0.5 3 0.8 0.3 1.0 0.3 last station 0.94 0.5 200 12
15 0.94 1.0 3 0.8 0.3 1.0 0.3 first station 0.94 1.0 200 12
16 0.94 1.0 3 0.8 0.3 1.0 0.3 last station 0.94 1.0 200 12
17 0.94 2.5 3 0.8 0.3 1.0 0.3 first station 0.94 2.5 200 12
18 0.94 2.5 3 0.8 0.3 1.0 0.3 last station 0.94 2.5 200 12
19 0.94 0.5 3 0.8 1.0 1.0 1.0 first station 0.94 0.5 200 16
20 0.94 0.5 3 0.8 1.0 1.0 1.0 last station 0.94 0.5 200 16
21 0.94 1.0 3 0.8 1.0 1.0 1.0 first station 0.94 1.0 200 16
22 0.94 1.0 3 0.8 1.0 1.0 1.0 last station 0.94 1.0 200 16
23 0.94 2.5 3 0.8 1.0 1.0 1.0 first station 0.94 2.5 200 16
24 0.94 2.5 3 0.8 1.0 1.0 1.0 last station 0.94 2.5 200 16
25 0.94 0.5 10 0.8 0.3 1.0 0.3 first station 0.94 0.5 200 20
26 0.94 0.5 10 0.8 0.3 1.0 0.3 last station 0.94 0.5 200 20
27 0.94 1.0 10 0.8 0.3 1.0 0.3 first station 0.94 1.0 200 20
28 0.94 1.0 10 0.8 0.3 1.0 0.3 last station 0.94 1.0 200 20
29 0.94 2.5 10 0.8 0.3 1.0 0.3 first station 0.94 2.5 200 20
30 0.94 2.5 10 0.8 0.3 1.0 0.3 last station 0.94 2.5 200 20
31 0.94 0.5 10 0.8 1.0 1.0 1.0 first station 0.94 0.5 200 35
32 0.94 0.5 10 0.8 1.0 1.0 1.0 last station 0.94 0.5 200 35
33 0.94 1.0 10 0.8 1.0 1.0 1.0 first station 0.94 1.0 200 35
34 0.94 1.0 10 0.8 1.0 1.0 1.0 last station 0.94 1.0 200 35
35 0.94 2.5 10 0.8 1.0 1.0 1.0 first station 0.94 2.5 200 35
36 0.94 2.5 10 0.8 1.0 1.0 1.0 last station 0.94 2.5 200 35

Table 12
Comparison of expected throughput times and expected backlogs for the case when raw materials arrive at some rate.
Experiment no. Uunb % Uk % λ (s−1) N Comparison of expected throughput times System II analysis
OP CON %dec. OD %dec. Nb W (sec)

1 – 59.00 0.59 3 4.21 3.03 −28.03 3.01 −28.50 1.35 1.69

2 – 58.00 0.58 3 4.35 3.05 −29.89 3.03 −30.34 1.40 1.75
3 – 58.00 0.58 3 5.41 4.53 −16.27 4.31 −20.33 1.51 1.89
4 – 71.00 0.71 7 8.54 6.77 −20.73 5.21 −38.99 5.35 6.69
5 – 70.00 0.70 7 9 7 −22.22 5.39 −40.11 5.69 7.11
6 – 68.00 0.68 7 10.4 7.68 −26.15 6.87 −33.94 6.32 7.9
7 – 69.00 0.69 11 11.91 10.46 −12.17 10.51 −11.75 2.99 3.74
8 – 68.00 0.68 11 12.02 10.67 −11.23 10.92 −9.15 3.04 3.8
9 – 67.00 0.67 11 14.1 11.92 −15.46 12.7 −9.93 3.2 4
10 – 77.00 0.77 35 42.08 34.755 −17.40 34.965 −16.90 16.22 20.27
11 – 77.00 0.77 35 42.63 35.15 −17.55 35.64 −16.40 16.3 20.40
12 – 76.00 0.76 35 45.15 36.73 −18.65 37.16 −17.70 18.54 23.18
13 90.00 72.00 0.90 12 5.34 3.19 −40.26 3.51 −34.27 3.48 3.70
14 90.00 72.00 0.90 12 4.56 3.19 −30.04 3.71 −18.64 3.48 3.70
15 90.00 72.00 0.90 12 6.66 3.56 −46.55 4.59 −31.08 3.98 4.23
16 90.00 72.00 0.90 12 5.78 3.56 −38.41 4.75 −17.82 3.98 4.23
17 89.00 71.00 0.89 12 11.3 7.94 −29.73 9.01 −20.27 4.80 5.12
18 89.00 71.00 0.89 12 10.71 7.94 −25.86 9.67 −9.71 4.80 5.12
19 90.00 72.00 0.90 16 13.8 10.47 −24.13 11.01 −20.22 5.65 6.01
20 90.00 72.00 0.90 16 13.25 10.47 −20.98 12.1 −8.68 5.65 6.01
21 90.00 72.00 0.90 16 14.56 10.88 −25.27 11.77 −19.16 6.08 6.47
22 90.00 72.00 0.90 16 13.65 10.88 −20.29 12.61 −7.62 6.08 6.47
23 89.00 71.00 0.89 16 17.63 13.58 −22.97 14.93 −15.31 7.68 8.17
(continued on next page)

132
D. Roy, V. Ravikumaran Journal of Manufacturing Systems 50 (2019) 119–134

Table 12 (continued)

Experiment no. Uunb % Uk % λ (s−1) N Comparison of expected throughput times System II analysis
OP CON %dec. OD %dec. Nb W (sec)

24 89.00 71.00 0.89 16 16.04 13.58 −15.34 14.53 −9.41 7.68 8.17
25 91.00 73.00 0.91 20 12.73 9.18 −27.89 9.23 −27.49 5.84 6.21
26 91.00 73.00 0.91 20 11.68 9.18 21.40 10.02 −14.21 5.84 6.21
27 91.00 73.00 0.91 20 13.6 10.25 −24.63 10.2 −25.00 6.97 7.41
28 91.00 73.00 0.91 20 12.73 10.25 −19.48 11.37 −10.68 6.97 7.41
29 89.00 71.00 0.89 20 18.65 15.2 −18.50 15.3 −17.96 8.94 9.51
30 89.00 71.00 0.89 20 17.49 15.2 −13.09 16.03 −8.35 8.94 9.51
31 92.00 74.00 0.92 35 34.72 28.7 −17.34 29.06 −16.30 16.57 17.73
32 92.00 74.00 0.92 35 34.15 28.7 −15.96 32.94 −3.54 16.57 17.73
33 91.00 73.00 0.91 35 34.72 28.67 −17.43 28.87 −16.85 17.81 18.95
34 91.00 73.00 0.91 35 33.83 28.67 −15.25 31.56 −6.71 17.81 18.95
35 88.00 70.00 0.88 35 38.64 32.2 −16.67 32.57 −15.71 19.78 21.04
36 88.00 70.00 0.88 35 36.85 32.2 −12.62 35.2 −4.48 19.78 21.04

Appendix B. Results of model validation

Fig. 23 shows the frequency distribution of the absolute percentage of error for expected throughput time of System I. It also represents work in
progress of System I as it is directly correlated to throughput time. Fig. 24 shows the frequency distribution the expected waiting time for the
backlogs to get cleared in system II. It also represents expected number of backlogs in System II as it is directed correlated to expected waiting time.

Fig. 23. Abs error % (expected throughput time of System I).

Fig. 24. Abs error % (expected waiting time for the backlogs to get cleared in System II).

References
[1] Donovan RM. Reduce inventories and improve business performance. 2005 (Online;
accessed 28.06.17). http://www.reliableplant.com/Read/141/reduce-inventories.
[2] Lödding H, Lohmann S. Incap – applying short-term flexibility to control in-
ventories? Int. J. Prod. Res. 2012;50(3):909–19.
[3] Van Mieghem J. Commissioned paper: capacity management, investment, and
hedging: review and recent developments. Manuf. Serv. Oper. Mange.
2003;5(4):269–302.
[4] Wu SD, Erkoc M, Karabuk S. Managing capacity in the high-tech industry: a review
of literature? Eng. Econ. 2005;50(2):125–58.
[5] Alp O, Tan T. Tactical capacity management under capacity flexibility? IIE Trans.
2008;40(3):221–37.
[6] Sugimori Y, Kusunoki K, Cho F, Uchikawa S. Toyota production system and Kanban
system materialization of just-in-time and respect-for-human system? Int. J. Prod.
Res. 1977;15(6):553–64.
[7] Buzacott J, Shanthikumar J. Stochastic Models of Manufacturing Systems.

133
D. Roy, V. Ravikumaran
Englewood Cliffs, NJ: Prentice Hall; 1993.
[8] Krieg GN, Kuhn H. Performance evaluation of two-stage multi-product kanban
systems? IIE Trans. 2008;40(3):265–83.
[9] Wang H, Wang H-PB. Optimum number of Kanbans between two adjacent work-
stations in a JIT system. Int. J. Prod. Econ. 1991;22(3):179–88.
[10] Gstettner S, Kuhn H. Analysis of production control systems Kanban and CONWIP?
Int. J. Prod. Res. 1996;34(11):3253–73.
[11] Spearman M, Woodruff D, Hopp W. CONWIP – a pull alternative to Kanban? Int. J.
Prod. Res. 1990;28(5):879–94.
[12] Borthakur A, Medhi J. Poisson's input queueing system with start time and under-
control operating policy? Comput. Oper. Res. J. 1987;14(1):33–40.
[13] Krishnamurthy A, Suri R. Planning and implementing POLCA: a card-based control
system for high variety or custom engineered products? Prod. Plan. Control
2009;20(7):596–610.
[14] Lozinski C, Glassey CR. Bottleneck starvation indicators for shop floor control
[semiconductor manufacturing process]? IEEE Trans. Semicond. Manuf.
1988;1(4):147–53.
[15] Wein L. Scheduling netwrok of queues: Heavy traffic analysis of multi-station net-
work with controllable input. Oper. Res. 1990;38:1065L 1078.
[16] Graves SC. A tactical planning model for a job shop? Oper. Res. 1986;34(4):522–33.
[17] Günther H-O. The block planning approach for continuous time-based dynamic lot
sizing and scheduling? Bus. Res. 2014;7(1):51–76.
[18] Economopoulos A, Kouikoglou. Analysis of a simple CONWIP system with im-
patient customers. IEEE Conference on Automation Science and Engineering
2008:291–6.
[19] Lee H, Seo D-W. Performance evaluation of WIP-controlled line production systems
with constant processing times. Comput. Ind. Eng. 2016;94:138–46.
[20] Lagershausen S, Manitz M, Tempelmeier H. Performance analysis of closed-loop
assembly lines with general processing times and finite buffer spaces? IIE Trans.
2013;45(5):502–15.
[21] Huang G, Chen J, Wang X, Shi Y. A simulation study of CONWIP assembly with
multi-loop in mass production, multi-products and low volume and OKP environ-
ments. Int. J. Prod. Res. 2015;53(14):4160–75.
[22] Satyam K, Krishnamurthy A. Performance evaluation of a multi-product system
under CONWIP control. IIE Trans. 2008;40:252L 264.
[23] Satyam K, Krishnamurthy A. Performance analysis of CONWIP systems with batch
size constraints? Ann. Oper. Res. 2011;209(1):85–114.
[24] Heragu SS, Gupta A. CONWIP: closed or semi-open queuing network? Int. J. Oper.
Res. 2015;24(3):356–67.
[25] Roy D. Semi-open queuing networks: a review of stochastic models, solution
methods and new research areas. Int. J. Prod. Res. 2016;54(6):1735–52.
134
Journal of Manufacturing Systems 50 (2019) 119–134
[26] Krishnamurthy A, Suri R, Vernon M. Re-examining the performance of MRP and
Kanban material control strategies for multi-product flexible manufacturing sys-
tems. Int. J. Flex. Manuf. Syst. 2004;16:123L 150.
[27] Wein L. Scheduling semiconductor wafer fabrication? IEEE 1988;1(3):115–30.
[28] Whitt W. Open and closed queue models of networks of queue? Bell Syst. Tech. J.
1984;63(9):1911–79.
[29] Spearman M, Zazanis M. Push and pull production systems: issues and comparisons.
Oper. Res. 1992;40:521L 532.
[30] Duenyas I. A simple release policy for networks of queues with controllable inputs?
Oper. Res. 1994;42(6):1162–71.
[31] Zahorjan J. Workload representations in queueing models of computer systems.
Proc. 1983 ACM SIGMETRICS Conf. on Measurement and Modeling of Computer
Systems 1983:70–81.
[32] Jaegler Y, Jaegler A, Burlat P, Lamouri S, Trentesaux D. The ConWip production
control system: a systematic review and classification. Int. J. Prod. Res. 2017:1–22.
[33] Framinan JM, González PL, Ruiz-Usano R. The CONWIP production control system:
Review and research issues? Prod. Plan. Control 2003;14(3):255–65.
[34] Glassey CR, Resende MGC. Closed-loop job release control for vlsi circuit manu-
facturing? IEEE Trans. Semicond. Manuf. 1988;1(1):36–46.
[35] Roderick L, Phillips D, Hogg G. A comparison of order release strategies in pro-
duction control systems? Int. J. Prod. Res. 1992;30(3):611–26.
[36] Bahaji N, Kuhl ME. A simulation study of new multi-objective composite dis-
patching rules, CONWIP, and push lot release in semiconductor fabrication. Int. J.
Prod. Res. 2008;46(14):3801–24.
[37] Little JDC. A proof for the queuing formula: l = w ? Oper. Res. 1961;9(3):383–7.
[38] Herer YT, Shalom L. The Kanban assignment problem – a non-integral approach?
Eur. J. Oper. Res. 2000;120(2):260–76.
[39] Baskett F, Chandy KM, Muntz RR, Palacios FG. Open, closed, and mixed networks of
queues with different classes of customers. J. ACM 1975;22(2):248–60.
[40] Lazowska E, Zahorjan J, Graham G, Sevcik K. Quantitative System Performance:
Computer System Analysis Using Queueing Network Models. New Jersey: Prentice-
Hall, Inc.; 1984.
[41] Marshall KT. Some inequalities in queuing? Oper. Res. 1968;16(3):651–68.
[42] Marshall KT. Some relationships between the distributions of waiting time, idle
time and interoutput time in the GI/G/1 queue. SIAM J. Appl. Math.
1968;16(2):324–7.
[43] Whitt W. The queueing network analyser? Bell Syst. Tech. J. 1983;62(9):2779–814.
[44] Yoon Sungwook, Kim Jihyun, Sukjae Jeong. The optimal decision combination in
semiconductor manufacturing. Sustainability 2017;9(10):1788. https://doi.org/10.
3390/su9101788.