International Journal of Production Research

ISSN: 0020-7543 (Print) 1366-588X (Online) Journal homepage: https://www.tandfonline.com/loi/tprs20

Determination of batch size at a bottleneck

machine in manufacturing systems

Pyung-Hoi Koo , Robert Bulfin & Shie-Gheun Koh

To cite this article: Pyung-Hoi Koo , Robert Bulfin & Shie-Gheun Koh (2007) Determination of
batch size at a bottleneck machine in manufacturing systems, International Journal of Production
Research, 45:5, 1215-1231, DOI: 10.1080/00207540600675793

To link to this article: https://doi.org/10.1080/00207540600675793

Published online: 22 Dec 2006.

Submit your article to this journal

Article views: 709

View related articles

Citing articles: 6 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tprs20
International Journal of Production Research,
Vol. 45, No. 5, 1 March 2007, 1215–1231

Determination of batch size at a bottleneck machine in

manufacturing systems

PYUNG-HOI KOO*y, ROBERT BULFINz

and SHIE-GHEUN KOHy

yDepartment of Systems Management and Engineering, Pukyong National University,

San 100, Yongdang, Namgu, Busan, 608–739, Korea
zDepartment of Industrial and Systems Engineering, Auburn University,
207 Dunstan Hall, Auburn University, AL 36849–5346, USA

(Revision received February 2006)

In manufacturing systems, there often exists a bottleneck machine whose capacity

is equal to or less than the market demand. Any idle or waste time at the
bottleneck machine directly impacts the output of the entire plant because
it results in a loss of throughput. In order to maximize the capacity utilization
by less setup losses at the bottleneck machine, the parts are often produced in
batches. Traditionally, most batch sizing decisions are made based on the
economic order quantity model where setup and inventory holding costs are
considered. This paper presents an alternative method to determine batch size
at a bottleneck machine. We present a new objective function and cost factors
for batch sizing and investigate queuing and throughput models. A linear search
algorithm is introduced to find the optimal throughput rate and batch size
at the same time. Numerical examples are examined to see how the batching
algorithm works.

Keywords: Batch sizing; Lead time; Throughput; Bottleneck machine

1. Introduction

In unbalanced manufacturing systems, very common in industries, there often exists

a bottleneck machine whose capacity is equal to or less than the market demand
placed upon it. In order to realize the maximum throughput levels, the bottleneck
machine should be fully utilized. Several different production control approaches
have been presented to maintain high production utilization at the bottleneck
machine with less work-in-process (WIP) inventories. They include quick line change
in Toyota Production Systems, WIP-based control in CONWIP (Constant WIP:
Spearman et al. 1989), DBR (drum–buffer–rope) in theory of constraints (TOC:
Goldrett and Cox 1984), shifting bottleneck scheduling (Adams et al. 1988) and
order release control (Glassey and Resende 1988, Lu et al. 1994), to name a few.
In this paper, we focus on batch sizing at the bottleneck machine.

*Corresponding author. Email: [email protected]

International Journal of Production Research

ISSN 0020–7543 print/ISSN 1366–588X online ß 2007 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/00207540600675793
1216 P.-H. Koo et al.

Throughput rate may be the most important performance measure in a

manufacturing system when the production facility is a bottleneck resource
preventing a firm from gaining more revenue. In unbalanced manufacturing
systems, the bottleneck machine determines the throughput rate of the entire
system. The available time at any machine is spent in several different ways:
processing time, setup time and idle time (Srikanth and Umble 1997). Any idle or
waste time at the bottleneck machine directly affects the entire plant because it
results in a permanent loss of throughput. Thus, all available time at the bottleneck
machine should be allocated to either processing or setup time. However, the setup
itself is not a value-added operation. When we reduce setup time, the additionally
available time can be also allocated to processing to increase production. Quick die
change in Toyota Production Systems is a typical example to make a special effort
to reduce the setup time. Another way to reduce the setup time during a planning
horizon is batch production. Given setup times for products, the total setup time
spent at a machine depends on the batch size. As batch size increases, the frequency
of setups decreases, resulting in less total setup time.
Another important performance measure in today’s manufacturing environment
is lead time. The lead time is an index to see how fast a manufacturing system
responds to customer demands. Hence, the lead time is widely recognized as a
competitive factor in the modern mass customization era. The major objective of
lean manufacturing is to minimize lead time with minimum WIP levels while
maintaining a desirable production volume. Little’s law describes a direct relation-
ship between WIP inventory and lead time: L ¼
W, where L is the WIP inventory,

is the throughput rate and W is the lead time. The production lead time is largely
composed of queuing delay and processing time. In many manufacturing facilities,
parts spend more time waiting in queues due to competition for available processing
capacity rather than in actual processing. There are many cost factors associated
with lead times. Long lead times force schedules to be made mostly by using a
forecast instead of actual demand. The accuracy of demand forecasts diminishes
greatly as the planning horizon increases. This may result in a critical error in
production plans or increase the chance of schedule changes. In addition, Little’s law
says that long lead times lead to proportionally larger WIP levels. The higher WIP
inventory covers the problems in manufacturing facilities, which can otherwise be
resolved by taking appropriate actions. With less WIP, there is better control of
the work and less opportunity for working on the wrong parts. When manufacturing
lead time is short, quality problems are recognized more rapidly and actions can be
taken to solve the problems in an earlier stage. Shorter lead times usually result in
smaller variability in the lead time, which improves on-time delivery performance
and decreases safety stocks. In a make-to-order manufacturing environment, the
short lead time is important for sample products. Customers often ask several firms
to make some sample products and then select a firm as a supplier for the products.
When quality is acceptable, they often select the firm that provides the sample
products first. In some manufacturing facilities such as semiconductor fabrication
systems, the lead time affects the yield rate (Glassey and Resende 1988, Leachman
et al. 2002, Kenyon et al. 2005). A shorter lead time will result in increased yield rates
due to a smaller chance of particle contamination. Finally, a shorter lead time allows
quicker response to changing market environment and customer needs. It leads
directly to shorter customer delivery time and higher due-date performances.
 Determination of batch size 1217

The batch size affects the lead time (and WIP level) as well as production rate.
As described earlier, a larger batch size increases the production rate. However, the
larger batch size also results in longer lead time and larger WIP inventory level.
Hence, there should be a tradeoff between the opportunity costs involved in large
batches and the productivity losses from making too many small batches. This paper
presents an alternative method to determine batch size at a bottleneck machine.
We will present a new objective function and cost factors for batch sizing and
investigate queuing and throughput models.

2. Literature review

Traditionally, the batch sizing decision is usually made based on the EOQ (economic
order quantity) model. The objective of the EOQ batch sizing approach is to
minimize the sum of fixed setup costs and inventory holding costs. However, this
representation of costs often fails to capture the nature of the batching problem in
manufacturing flow lines in the following respects. (1) The conventional EOQ batch
sizing generally considers inventory holding costs for finished goods not for WIP
inventories. At a bottleneck machine in an unbalanced manufacturing, the WIP
inventories are more important than the parts completed through the bottleneck
machine which often flow smoothly to the end of the line with little delay. (2) Most
EOQ batch sizing models, with some exceptions, e.g. Karmarkar et al. (1992) and
Kuik and Tielemans (1998), ignore lead time by emphasizing inventory. Srikanth and
Umble (1997) argue that the inventory carrying cost is insignificant compared to the
cost associated with the lead time. (3) The EOQ batch sizing model also assumes that
inventory levels are linearly proportional to the batch size, which is not true in most
manufacturing facilities, especially at the bottleneck facility. (4) The EOQ model
assumes that the throughput rate or demand rate is given. However, batch size has
an effect on the facility utilization that in turn affects throughput rate at the
bottleneck machine. Increased batch sizes reduce the frequency of setups, resulting in
less setup times. The saved time can be used to produce more products. (5) The EOQ
batch sizing model is static in that it ignores any variability in the demand and
processing data. The variability in a manufacturing system is caused from such
factors as demand fluctuation, disturbances in up-stream operations, machine break
down, product mix and order patterns. Note that the stochastic features especially
affect the WIP levels and lead time. (6) The EOQ batch sizing emphases minimizing
costs. The batch sizing decisions should be made in the context of what is best for
the performance of the whole firm. The objective of a firm is to maximize profit.
(7) The EOQ model assumes that the setup cost is fixed. However, as indicated in
Karmarkar (1987) and Srikanth and Umble (1997), the setup cost is often not fixed
but the consequence of the solution.
During the last few decades, a few research works present batch sizing models that
explicitly take lead time into account in a stochastic manufacturing system. In these
models, the manufacturing facility is usually modeled by a queuing system. The
general queuing and congestion problems in stochastic manufacturing systems have
been addressed extensively in Solberg (1980), Shanthickumar and Buzacott (1981)
and Whitt (1983). Karmarkar (1987) examines the relationships between manufactur-
ing lead times, WIP inventories and batch size. He emphasizes the importance of lead
1218 P.-H. Koo et al.

time in determining batch size. He introduces a batch sizing model analogous to the
EOQ formulation. WIP inventories and finished goods inventories are included while
setup cost is not included in his model. He insists that the WIP inventory is much more
costly than finished goods inventory at high machine utilization. It is argued that the
appropriate batch sizes should be quite different from those of EOQ batching model.
Karmarker et al. (1992) present a multi-item batching heuristic with the objective of
minimizing the queuing delays. It is assumed that the throughput is given in their
study. They develop upper and lower bounds on the optimal batch size. Based on the
bounds, three batch sizing heuristics are presented and tested. Benjaafar (1996)
extends Karmarkar’s work by differentiating between production and transfer
batches. He presents a model to obtain optimal transfer batch sizes and examines their
relationship to the size of production batches. The effects of material handling
systems and multiple machines on batching decisions are also evaluated.
Tielemans and Kuik (1996) introduce an optimal batch model that minimizes the
lead time for a single-item case. They insist that an Erlang distribution may be more
appropriate than exponential distribution as a distribution of the inter-arrival time of
the batches at the production unit. Their experiments show that large differences
in outcomes between Poisson and Erlang arrivals of the batches occur. Kuik and
Tielemans (1998) present a batch sizing model that minimizes the average queuing
delay for a multi-item, single-machine work center, using a queuing model. Later,
they investigate the relationship between batch size and lead time variability (Kuik
and Tielemans 1999). They conclude that minimizing lead time would not necessarily
result in minimum lead time variability. However, they insist that for high machine
utilization levels, as the average queuing delay is minimized, the variance will also
be approximately minimized. Kuik and Tielemans (2004) further investigate the
relationship between batch size and lead time at low machine utilization.
Meng (2004) presents a batch sizing model in a multi-item manufacturing system
using an open queuing network. In his model, the batch sizes are operation-specific;
i.e. the same product is allowed to have different batch sizes at each of the machines
it visits. The model is based on the QNA (queuing network analyzer) developed by
Whitt (1983). A batch/burst process is introduced in addition to the departure-
split-superposition processes in the QNA. Kenyon et al. (2005) discuss a batch sizing
problem in a multiple operation system where both discrete and batch processing is
utilized to produce parts of multiple types. They investigate the impact that batch
sizes and material release rates can have on net profits and lead times. Here, net
profit is defined as throughput minus operating expenses as defined in Goldratt
(1990). It is shown through simulation experiments for a semiconductor fabrication
line that with the selection of the best combination of lot sizes, system output can be
increased, and both WIP and cycle time can be reduced.
Most analytical research work on batch sizing take either WIP levels or lead time
as a performance measure. Both performance measures relate to the cost factors that
should be minimized. When a throughput is given, the cost minimization may result
in larger profit for a firm. However, when the batch decision affects the throughput
rate, just minimizing cost does not mean maximum profit. It is believed that the
batch sizing decision should be viewed from the perspective of its effect on the entire
system. The common goal of the system is to maximize profit by maximizing
throughput (revenue) and minimizing cost. This paper takes both the throughput
and lead time into account as performance measures for batch size decisions.
 Determination of batch size 1219

3. Batch sizing model

3.1 Batch size model to minimize lead time

A few batch sizing models attempt to determine batch size to minimize the lead
time (Karmarkar 1987, Tielemans and Kuik 1996). We assume that the parts
arrive at the machine in batches of size b where they wait for service. When the
machine completes its current batch, it is setup for the next batch and starts
processing parts one by one from the batch. The machine is assumed to produce
the batches in a first-come first-served rule. At the completion of the entire batch,
the finished parts are released. It is assumed at this point that an M/M/1 queuing
model applies. The M/M/1 queuing system assumes Poisson arrival processes,
a single machine with exponential process times and a first-come first-served
protocol. This assumption may not represent most manufacturing facilities.
However, it offers basic insight into more complex systems and serves as a
building block for more general systems. Furthermore. if arrivals at the machine
are due to many overlapping and uncoordinated production flows,
the assumption may not be unreasonable (Whitt 1983, Karmarkar 1987).
The following notation is used:
x ¼ production rate (units day
1)
b ¼ batch size
p ¼ processing time (days unit
1)

¼ average arrival rate of batches ¼ x/b (batches day
1)
s ¼ setup time per batch (days)
 ¼ processing and setup time per batch (days) ¼ bp þ s
 ¼ processing rate at machine (batches day
1) ¼ 1/ ¼ 1/(bp þ s)
W ¼ the queuing delay (days)
T ¼ the time in the system, i.e. lead time (days)
The machine utilization, , can be stated as
x xs
 ¼
 ¼ ðbp þ sÞ ¼ xp þ ð1Þ
b b

The first term in the last expression represents machine utilization due to part
processing (productive utilization) and the second term represents machine
utilization due to setup. It can be seen that the machine utilization decreases
as batch size increases.
Since we must have  5 1, the production rate is upper bounded by
x
¼ bb=ðbp þ sÞc. Here, bac is the greatest integer less than or equal to a. When we
have a very large batch size, the production rate is upper-bounded by


x
¼ 1=p ð2Þ

Similarly, given x, the lower bound of the batch size, b, can be obtained from
equation (1) as follows:
 
xs
b¼ : ð3Þ
ð1
xpÞ
1220 P.-H. Koo et al.

The M/M/1 queuing system gives the average lead time, T, as follows:
1 s þ bp
T¼ ¼ ð4Þ
ð1
Þ 1
xp
xs=b
As the batch size, b, becomes smaller and approaches its lower bound, T grows
sharply without bound since  approaches 1. As the batch size becomes larger, the
advantages of batching are eventually eroded and the lead time starts to increase
approximately linearly in b. Since T is a convex function of b (see Karmarkar
et al. (1992) for proof ), the first-order condition of optimality gives the batch size,
b*, that leads to the minimal attainable average lead time, T *, as follows
(Karmarkar et al. 1992, Tielemans and Kuik 2004).
pffiffiffiffiffiffiffiffiffiffi
 xsð1 þ 1=xpÞ
b ¼ ð5Þ
ð1
xpÞ
pffiffiffiffiffiffi 2
 sð1 þ xpÞ
T ¼ ð6Þ
ð1
xpÞ2
As described above, the assumptions of the M/M/1 queuing system are not
satisfied in most manufacturing systems. Process times and inter-arrival times are
seldom exponential. To address the systems with non-exponential inter-arrival and
process time distributions, we consider a G/G/1 queuing system for which there is no
exact queuing model. We will adopt a two-moment approximation model, which
makes use of only the mean and standard deviation of the inter-arrival and service
time distributions. Let Wg be the waiting time and Tg be the lead time in the G/G/1
queue while W is the waiting time and T is the lead time in M/M/1 queue. Then,
T ¼ W þ  and Tg ¼ Wg þ . In the two-moment approximation model, Wg is
expressed in terms of W, as follows:
c2a þ c2s
Wg ¼ W ð7Þ
2
Here ca is the coefficient of variation (CV ) of the interarrival time distribution
and cs is the CV of the service time distribution. The CV is defined as standard
deviation divided by the mean. Let c^ ¼ ðc2a þ c2s Þ=2. From equation (7), we have
following relationship between the lead time in a M/M/1 queue and the lead time in
a G/G/1 queue.
^ þ  ¼ cðT
Tg ¼ Wg þ  ¼ cW ^
Þ þ  ¼ cT
^ þ ðbp þ sÞð1
cÞ
^ ð8Þ
Since T is a convex function of b, Tg is also a convex function. Then, as in
equation (5), the first-order condition of optimality gives the batch size, b*, in the
G/G/1 queue that minimizes the average lead time as follows.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 ^
xpÞ þ cp
xsðpð1
cÞð1 ^ þ ðcpð1^
cÞð1^
xpÞ þ c^2 pÞ=xÞ
b ¼ ð9Þ
^
cpð1 ^
xpÞ2

xpÞ þ pð1
cÞð1

The minimally attainable average lead time in G/G/1 queuing system, Tg, is then
obtained from equations (8) and (9).
 Determination of batch size 1221

Figure 1. Relationship between batch size and lead time with different CVs, (ca, cs).

Consider a simple example with s ¼ 0.02, p ¼ 0.01. The production rate is upper-
bounded by b1=pc, i.e. 100. Suppose 80 parts should be produced during a unit time.
The lower bound of the batch size is 8, i.e. if we have batch size less than or equal
to 8, the lead time will be infinity. Figure 1 shows the relationship between batch size
and lead time for M/M/1 and G/G/1 cases with different CV values. It is seen that as
the batch size increases, the lead time sharply decreases initially. However, as the
batch size increases further, the advantages of batching eventually disappear and the
lead time starts to increase. We have b* ¼ 16.94 and T * ¼ 1.79 from equations (5)
and (6) for M/M/1. Since the batch size should be an integer and the lead time is a
convex function, we can determine the optimal integer batch by comparing the lead
time from the largest integer less than or equal to b* and the smallest integer larger
than or equal to b*. In this example, we have less lead time with batch size of 17 so
that the optimal batch size is 17. The figure also shows that smaller CV values yield
lower lead times. For example, when the CV values of the interarrival time and
service time are 0.4, we have b* ¼ 14.25 and T * ¼ 0.43 from equations (8) and (9).
The optimal batch size in this case is 14. Note that the M/M/1 queue is a special case
of G/G/1 with ca ¼ 1.0, cs ¼ 1.0. It is observed that the smaller the CV values are, the
smaller the optimal batch size is.
In an unbalanced manufacturing system, machines may be classified into two
categories, bottleneck machines and non-bottleneck machines. As discussed earlier,
since the throughput of the system is determined by the bottleneck machine, all
available time at the bottleneck machine should be utilized to process products.
On the other hand, at non-bottleneck machines, process and setup should not
consume all of the available time. There should still be available time left over since
there is extra capacity. The extra capacity obtained from batching may not be used
productively at the non-bottleneck machines, but only to increase idle times.
The concept of bottleneck and non-bottleneck machines has been extensively
discussed in Goldratt (1990).
The batch sizing model discussed above to minimize lead time may be a
promising method for the non-bottleneck machine. However, at the bottleneck
machine, the increased machine capacity may be more important than the lead time.
The next section presents a batch sizing method to maximize profit, considering
throughput rate as well as batch size.
1222 P.-H. Koo et al.

100

90

maximum throughput
80

70

60

50

40
1
5
9
13
17
21
25
29
33
37
41
45
49
batch size

Figure 2. Relationship between batch size and maximum throughput rate.

3.2 Batch sizing model to maximize profit

When the demand is larger than the manufacturing capacity, the throughput rate
may also be a decision variable. A company would attempt to increase throughput
with some acceptable lead time. Consider the example with s ¼ 0.02, p ¼ 0.01 again.
Figure 2 shows the effect of the batch size on maximum throughput rate. It is seen
that as the batch size increases, the throughput rate increases sharply first and then
slowly with the batch size larger. As discussed earlier, the production rate can not
exceed the upper bound, 100.
In order to obtain a maximum production rate, the batch size should be large.
However, the larger batch size may lead to a longer lead time and larger WIP
inventory level. Hence, there should be a tradeoff between the productivity losses
from small batches and the opportunity costs involved in large batches. Here, we
define the profit, Z, as the revenue obtained from throughput (x) minus cost
related to an additional lead time, T þ. The additional lead time is defined as the
difference of an expected lead time, T, and a predefined target lead time, T  ,
i.e. T þ ¼ T
T  . The expected lead time can be obtained by equations (4) or (8)
while the predefined target lead time may be determined by considering product
characteristics and market environment. The objective of the batch size decision is
to maximize profit. Let  be the unit revenue per product and  be the unit cost
per lead time. Now the problem of choosing a batch size b so as to maximize the
profit can be written as

Maximize Z ¼ maxðx
xT þ , 0Þ ð10Þ

xs
s:t xp þ <1 x, b  0, integers ð11Þ
b
Here, the first term of the objective function represents the revenue obtained
from throughput (x) while the second term the cost related to lead time (T ). In the
constraints above, the first expression indicates the capacity limitation. Consider the
 Determination of batch size 1223

80

70

Z (profit)
60

50

40

30
90
10 80 te
20 70 r a
Batch 30 60 tion
size 40 duc
50 50 Pro

Figure 3. Effect of batch size and production rate on profit.

Step 1: x = 1 / p , Z * = −∞
Step 2: Calculate b* and Z by using equations (5) or (9), and equation (10), respectively
Step 3: If Z > Z * then
x
←x, b
←b*, Z *←Z
Step 4: If x = 1 then Stop
Else x ← (x− 1), goto Step 2.

Figure 4. Linear search algorithm to find the optimal production rate and batch size.

previous example with s ¼ 0.02, p ¼ 0.01, with M/M/1 queue discipline. Suppose we
have T  ¼ 1.0, and  ¼ 1 and  ¼ 0.05 (i.e. when the lead time of a part is delayed by
a time unit, it will be panelized by 5% of the product value.) Figure 3 shows the
relationship between batch size, throughput and profit. It is found that the batch size
and throughput that maximize the profit are 12 and 85, respectively, in this example.
With two decision variables, x and b, we cannot have an optimal solution directly
from the first order derivative. However, the unique feature of Z can be used to find
the optimal solution. As mentioned earlier, T is a convex function in b when the
throughput rate, x, is given. Then, given x, Z is a concave function in b. In addition,
the production rate is discrete and upper bounded as in equation (2). With these
features, we propose a linear search algorithm to find an optimal production rate,
x} , and optimal batch size, b} , at the same time. The new algorithm is shown in
figure 4.
The algorithm starts with the production rate (x) with the upper bound and
the current maximum profit (Z} ) of a very large negative number. In step 2, given x,
the batch size that leads to the lowest lead time is calculated by using equation (5) for
1224 P.-H. Koo et al.

the M/M/1 queue or equation (9) for the G/G/1 queue. This batch size maximizes the
profit with the given production rate. In step 3, the new profit is compared to the
current maximum profit. If the new profit is larger than the current maximum profit,
the solution is updated with the new one and the search continues with one less
throughput rate by going back to step 2. Note that, since the production rate is
discrete, the algorithm provides the optimal solution with a finite number of
iterations.

3.3 Discussion on batch sizing at bottleneck and non-bottleneck machines

All of the machines in an unbalanced manufacturing system can be classified into
bottleneck and non-bottleneck machines. A bottleneck machine is defined as a
machine whose capacity holds down the throughput that a manufacturing system
can produce. Bottleneck and non-bottleneck machines should be managed from
different perspectives. The production demand at a non-bottleneck machine is often
determined by the capacity of the bottleneck machine, not by its own capacity.
If a non-bottleneck machine produces more products than a given amount, it would
possibly result in increased WIP inventory on the shop floor without a corresponding
increase in throughput for the manufacturing system. Then, the production at the
non-bottleneck machine produces may not be an issue in determining an appropriate
batch size. It is widely recognized that the major goal of the flow control at the non-
bottleneck machines is to pass the products to the next machines with a minimum
delay. Hence, the algorithm presented in section 3.1 that attempts to minimize the
lead time with a given production rate is a reasonable method to determine a batch
size at the non-bottleneck machines. As discussed in the previous sections, the batch
size affects the throughput rate as well as the lead time. Both the throughput rate and
lead time are important performance criteria for the bottleneck machine.
The capacity of the non-bottleneck machine determines the system throughput
rate, and waiting time at the bottleneck machine often accounts for most of the
system lead time. The algorithm presented in section 3.2 that takes both throughput
rate and lead time into account is a promising approach in determining a batch size
at the bottleneck machine.

4. Numerical examples and discussions

This section explains how the proposed batch sizing algorithm works. Consider the
previous example with s ¼ 0.02, p ¼ 0.01,  ¼ 1,  ¼ 0.05, T  ¼ 1.0, and M/M/1
queuing system. The upper bound of the production rate is 100 (¼|1/p|). Now, the
algorithm starts with Z* ¼
1, x ¼ 100. Table 1 shows the results of each iteration
of the algorithm. The profit increases as the production rate decreases from 100
to 81. When the production rate is reduced to 80, the profit starts decreasing.
The algorithm finds the optimal production rate of 81 and optimal batch size of 18.
At the optimal solution, we have the profit of 76.95. As can be seen in the table, as
the production rate changes, the optimal batch size also changes. Note that the lead
time of the solution is 2.00 which consists of batch (consisting of 18 parts) processing
time of 0.18, setup times of 0.02, and waiting time at the queue of 1.80.
 Determination of batch size 1225

Table 1. Results from the iterative batch sizing algorithm.

Production Optimal Lead Machine

rate batch size time Profit utilization

99 397 795.9 0.0 99.5

98 197 197.9 0.0 99.0
97 130 87.5 0.0 98.5
96 97 48.9 0.0 98.0
95 77 31.1 0.0 97.5
94 64 21.5 0.0 97.0
93 54 15.7 24.4 96.4
92 47 11.9 41.4 95.9
91 41 9.4 52.6 95.4
90 37 7.5 60.3 94.9
89 33 6.2 65.6 94.3
88 30 5.2 69.4 93.8
87 28 4.4 72.1 93.3
86 26 3.7 74.0 92.7
85 24 3.2 75.3 92.2
84 22 2.8 76.1 91.7
83 20 2.5 76.6 91.1
82 19 2.2 76.9 90.6
81 18 2.0 76.9 90.0
80 17 1.7 76.8 89.4
79 16 1.6 76.5 88.9
78 15 1.4 76.1 88.3
77 14 1.3 75.7 87.7
76 14 1.2 75.1 87.2
75 13 1.1 74.5 86.6
74 12 1.0 73.9 86.0
73 12 0.9 73.2 85.4
72 11 0.8 72.4 84.9
71 11 0.8 71.6 84.3
70 10 0.7 70.8 83.7

The long waiting time is common at the bottleneck machine. Figure 5 shows the
relationship between production rate and profit.
The ratio of setup time to processing time, s/p, can affect the optimal batch
size. In our example, s/p ¼ 0.02/0.01 ¼ 2.0. To see the effect of s/p on optimal
production rate and batch size, we perform experiments with various setup times
ranging from 0.001 to 0.1, i.e. with s/p ratio ranging from 0.1 to 10.0. Table 2
shows that the optimal production rate and the optimal batch size change as s/p
increases. As can be expected, the optimal batch size increases as the set up time
(and s/p value) becomes larger. This result highlights the importance of setup time
reduction: i.e. if a setup time is reduced, we can produce more products with
shorter lead time.
When the service time does not follow an exponential distribution, the optimal
throughput rate and batch size are different from the ones obtained in the previous
experiments. Table 3 shows the optimal throughput rate and batch size for M/M/1
queue and G/G/1 queue with ca ¼ cs ¼ 0.5. It is observed that we have a larger profit
1226 P.-H. Koo et al.

80

75

70

Z (profit)
65

60

55

50
99 94 89 84 79 74 69 64 59 54
production rate

Figure 5. Relationship between production rate and profit.

Table 2. Solution over different s/p ratios.

Optimal Optimal Lead

s/p production rate batch size time Profit Utilization

0.1 93 3 0.7 93.99 96%

0.2 91 4 0.9 91.26 95%
0.5 88 8 1.3 86.66 94%
1.0 85 12 1.6 82.27 92%
2.0 81 18 2.0 76.95 90%
3.0 79 24 2.4 73.36 89%
4.0 77 29 2.6 70.59 88%
5.0 75 32 2.7 68.30 87%
10.0 69 49 3.4 60.42 83%

Table 3. Optimal solution for different service time distributions.

M/M/1 queue G/G/1 with ca ¼ cs ¼ 0.5

Production Optimal Lead Optimal Lead

rate batch size time batch size Profit time Profit

90 37 7.59 60.32 35 3.87 80.18

89 33 6.24 65.67 31 3.14 80.92
88 30 5.22 69.45 28 2.63 81.24
87* 28 4.42 72.12 26* 2.22 81.26*
86 26 3.79 74.00 23 1.92 81.07
85 24 3.28 75.30 22 1.65 80.73
84 22 2.87 76.15 20 1.45 80.27
83 20 2.53 76.66 19 1.27 79.71
82 19 2.24 76.91 17 1.14 79.08
81* 18* 2.00 76.95* 16 1.01 78.40
80 17 1.79 76.82 15 0.91 77.67
79 16 1.62 76.56 14 0.82 76.90
78 15 1.47 76.18 13 0.75 76.10
77 14 1.33 75.72 13 0.67 75.27
76 14 1.22 75.18 12 0.62 74.43
75 13 1.11 74.57 11 0.57 73.56
 Determination of batch size 1227

with a higher production rate and smaller lead time under the less variation in inter-
arrival and service distributions.

5. Multi-product batch sizing

This section addresses batch sizing for the system where different kinds of part types
are manufactured. The notation given in section 3.1 is reused in this section. The only
difference is that a subscript i (i ¼ 1, . . . , n) is added to each notation to describe each
item. Karmarkar (1987) indicates that the multi-item batch sizing problem has two
important implications. First queuing delays are externalities in that changing one
item’s batch size potentially has an impact on delays of all other items. Second,
queuing is a joint effect so that changing the lead time for an item cannot be
independently achieved by changing only its own batch sizes—all other batch sizes
must be changed and all lead times will be affected. In addition, heterogeneity of
items leads to variability in processing time. For part i, the number of batches is
given by
i ¼ xi/bi and the processing time  i for P each batch is given by  i ¼ si ¼Pbipp.
The mean processing time is given by EðÞ ¼ i
i ð i þ pi bi Þ=
where
¼ i
i .
Although the processing time for a batch of a certain type of part can be
deterministic, the processing time of a batch becomes stochastic, depending on the
item involved, giving a mix of processing times. Then, the facility may be modeled as
an M/G/1 queue. The M/G/1 queue allows for more than one class of traffic (part) to
arrive, where the classes have differing service requirements. This paper focuses on
the M/G/1 queue to obtain insight on the effect of batch size under a multiple
product case. The average waiting time (W ) for this case is given by the
Pollaczeck–Kintchin formula (see Karmarkar et al. 1992).

Eð 2 Þ
ðVðÞ þ EðÞ2 Þ

W¼ ¼ , ð12Þ
2ð1
Þ 2ð1

EðÞÞ

where  ¼
E().
The average lead time for a batch of item i, Ti, is W þ  i. When the machine
utilization is high, the average waiting time dominates the total lead time: Ti  W
(Kuik and Tielemans 2004). Now the problem of choosing a batch size bi and
production rate xi to maximize profit can be written as:
!
X X
Maximize Z ¼ max i xi
i Tiþ xi , 0 , ð13Þ
i i

X xi si

s:t: xi pi þ < 1 xi , bi  0, integers ð14Þ
i
bi

where i is the unit profit for product i, i is the unit cost related to lead time and Tiþ
is the difference of an actual lead time for product i, Ti, and a predefined target
lead time, Tio . Because of its multivariable characteristics, this problem cannot be
treated easily.
1228 P.-H. Koo et al.

Step 1: Z * = −∞, X = 1/Σi (ri pi)

Step 2: Calculate b*i , Z by using equations (13) and (16)

Step 3: If Z > Z * then

X
←X, x



i = riX , bi ← bi , Z ← Z
* *

Step 4: If X = 1 then Stop

Else X ← (X − 1), goto Step 2

Figure 6. An algorithm to find the optimal production rate and batch size for multiple
products.

In this paper, in order to have some insights on the ralationship among the batch
size, lead time and throughput in multi-product cases, a batch sizing procedure will
be introduced in a special case where all products are homogeneous with respect to
the setup time (i.e. si ¼ s), and the product mix is given. For this case, exact closed-
form expressions can be found for the optimal batch sizes to minimize the waiting
time (Kuik and Tielemans 1999). Let W} be the minimal time in queue achieved by
a vector of optimal batch sizes b ¼ ðb} }
1 , . . . , bn Þ. Kuik and Tielemans (2004) show

2us
W} ¼ ð15Þ
ð1
uÞ2

sð1 þ uÞ
bi} ¼ ð16Þ
pi ð1
uÞ
P
where u ¼ i xi pi . Here, u is a total productive utilization.
P Let X be the total production rate and ri be the productionP mix for part i where
r
i i ¼ 1. Then, x i ¼ ri X. The total machine utilization is X i ðri s=bi þ ri pi Þ and
should be less than one. Hence, the upper bound of X, X, is given as follows,
 

1
X¼ P : ð17Þ
i ðri pi Þ

Since W in equation (12) is a convex function in bi, Z is a concave function in bi.

In addition, the second derivative of Z in equation (9) shows that Z is also a concave
function in X on 1 5 X 5 X. Similar to the single product case, a linear search
algorithm can be constructed as in figure 6 to find the optimal production rate, X} ,
and the optimal batch size, b}i .
A numerical experiment has been performed for a multi-product system to see
how the algorithm works. The system under consideration is shown in table 4 that
consists of five product types with the setup time of 0.02.
The upper bound of total production rate, X, is b90:9c ¼ 90, which is obtained
from equation (17). Then, the batch sizing procedure in figure 6 is performed with
production rate of 90. Table 5 shows the results of the iterative procedure for each
production rate. For each production rate, the optimal batch size is calculated for
 Determination of batch size 1229

Table 4. Multi-product system under consideration.

Product types

Part1 Part2 Part3 Part4 Part5

Processing time 0.010 0.012 0.008 0.014 0.006

Product mix 0.3 0.2 0.1 0.3 0.1
 1.00 2.00 1.00 0.50 1.00
 0.05 0.10 0.05 0.15 0.05

Table 5. Experimental results for a multi-product system.

Batch size

Production Lead Machine

rate time Profit utilization Part1 Part2 Part3 Part4 Part5

90 400.0 – 99.5% 398 332 498 284 663

89 90.70 – 98.9% 188 157 236 135 314
88 21.63 – 98.4% 123 103 154 88 205
87 21.63 – 97.8% 91 76 114 65 152
86 13.72 – 97.2% 72 60 90 51 120
85 9.47 24.47 96.6% 60 50 74 43 99
84 6.93 43.41 96.0% 51 42 63 36 84
83 5.28 55.14 95.5% 44 37 55 31 73
82 4.16 62.74 94.8% 39 32 49 28 65
81 3.37 67.80 94.2% 35 29 43 25 58
80 2.78 71.20 93.6% 31 26 39 22 52
79 2.33 73.49 93.0% 29 24 36 20 48
78 1.98 74.99 92.4% 26 22 33 19 44
77 1.71 75.94 91.7% 24 20 30 17 40
76 1.49 76.47 91.1% 22 19 28 16 37
75* 1.31 76.68 90.4% 21 17 26 15 35
74 1.16 76.66 89.7% 20 16 24 14 33
73 1.03 76.45 89.1% 18 15 23 13 31
72 0.92 76.09 88.4% 17 14 22 12 29
71 0.83 75.61 87.7% 16 14 20 12 27
70 0.76 75.04 87.0% 15 13 19 11 26

each product type by using equation (16). It is seen that the processing time dirrectly
affects the batch size for each product. The part with a larger processing time has
a smaller batch size. As the production rate decreases, the profit increases. However,
after the profit reaches the highest level, it begins to decrease. In this example,
we have the largest profit, 76.68, when the production rate is 75. Suppose there is
enough demand and a manager decides to maintain high machine utilization of
at least 95%. Then, he would produce 83 products. However, this leads to a longer
lead time, 5.28, which results in much lower profit, 55.14, than the profit with the
production rate of 75. This makes it clear that production rate should be considered
at the same time that a batch size is determined.
1230 P.-H. Koo et al.

6. Conclusions

In manufacturing systems, there often exists a bottleneck machine whose capacity is

less than the demand rate. An increased production rate at the bottleneck machine
results in an increase in the output of the entire system. The batch size affects the
facility capacity. In order for the bottleneck machine to produce more products, the
batch size should be large. However, a large batch size may increase production lead
time which is another important performance measure. This paper is motivated by
recognizing that most existing batch sizing models are based on some assumptions
which are often not proper at a bottleneck machine in manufacturing systems.
This paper addresses the importance of throughput at the bottleneck machine and
lead time. Some queuing models have been discussed to determine the batch size to
realize the lowest lead times. A linear search based batch sizing algorithm has been
presented where production rate and batch size are determined at the same time.
We have discussed that the performance criteria at bottleneck and non-bottleneck
machines should be different from each other. A few numerical examples are
presented to show how the algorithm works. This paper also discusses batch sizing
for a multi-product case. Because of the complexity of the case, it is hard to solve the
batching problem for multi-product case. A waiting time model is discussed and
a batch sizing algorithm for a specific multi-product case has been introduced.
This paper shows that throughput rate as well as lead time should be considered
in determining batch size at a bottleneck machine. Some assumptions made in this
paper may be relaxed to represent real-world manufacturing systems more closely.
We focus our attention on a bottleneck machine; there may be also interactions
between the bottleneck and non-bottleneck machines. Batch size may differ for
processing and material handling as discussed in Benjafarr (1996) and Goldratt
(1990). Additional works will be done to address these problems. We are also
interested in extending the current work for more general multi-product cases.

Acknowledgement

This work was supported by Korea Research Foundation Grant (KRF-2004-042-

D00226).

References

Adams, J., Balas, E. and Zawack, D., The shifting bottleneck procedure for job shop
scheduling. Manag. Sci., 1988, 34(3), 391–401.
Benjaafar, S., On production batches, transfer batches and lead times. IIE Trans., 1996, 28,
357–362.
Glassey, C.R. and Resende, M.G.C., Closed-loop job shop release control for VLSI circuit
manufacturing. IEEE Trans. Semicond. Manuf., 1988, 1(1), 36–46.
Goldratt, E.M. and Cox, J., The Goal: A Process of Ongoing Improvement, 1984 (The North
River Press: MA, USA).
Goldratt, E.M., The Haystack Syndrome, 1990 (The North River Press: MA, USA).
Karmarkar, U.S., Lot sizes, lead time and in-process inventories. Manag. Sci., 1987, 33(3),
409–418.
 Determination of batch size 1231

Karmarkar, U.S., Kekre, S. and Kekre, S., Multi-item batching heuristics for minimization of
queue delays. Eur. J. Oper. Res., 1992, 58, 99–111.
Kenyon, G., Canel, C. and Neureuther, B.D., The impact of lot-sizing on net profits and
cycle times in the n-job, m-machine job shop with both discrete and batch processing.
Int. J. Prod. Econ., 2005, 95(3), 263–278.
Kuik, R. and Tielemans, P.F.J., Analysis of expected queuing delays for decision making
in production planning. Eur. J. Oper. Res., 1998, 110, 658–681.
Kuik, R. and Tielemans, P.F.J., Lead-time variability in a homogeneous queuing model of
batching. Int. J. Prod. Econ., 1999, 59, 435–441.
Kuik, R. and Tielemans, P.F.J., Expected time in system analysis of a single-machine multi-
item processing center. Eur. J. Oper. Res., 2004, 156, 287–304.
Leachman, R.C., Kang, J. and Lin, V., SLIM short cycle time and low inventory in
manufacturing at Samsung electronics. Interfaces, 2002, 32(1), 61–77.
Lu, S.C.H., Ramaswamy, D. and Kumar, P.R., Efficient scheduling policies to reduce mean
and variance of cycle-time in semiconductor manufacturing plants. IEEE Trans.
Semicond. Manuf., 1994, 7(3), 374–388.
Meng, G., Batch size model in a multi-item, discrete manufacturing system via an open
queuing network. IIE Trans., 2004, 36(8), 743–753.
Solberg, J.J., CAN-Q user guide: the optimal planning of computerized manufacturing
systems. Report No. 9, School of Industrial Engineering, Purdue University, USA,
1980.
Shanthickumar, J.G. and Buzacott, J.A., Open queuing network models of dynamic job
shops. Int. J. Prod. Res., 1981, 19(3), 255–266.
Spearman, M.L., Woodruff, D.L. and Hopp, W.J., CONWIP: a pull alternative to Kanban.
Int. J. Prod. Res., 1989, 28(5), 879–894.
Srikanth, M.L. and Umble, M., Synchronous Management: Profit-Based Manufacturing for the
21st Century, 1997 (The Spectrum Publishing Company: CN, USA).
Tielemans, P.F.J. and Kuik, R., An explanation of models that minimize leadtime through
batching of arrived orders. Eur. J. Oper. Res., 1996, 95, 374–389.
Whitt, W., The queuing network analyzer. Bell Syst. Tech. J., 1983, 62(9), 2779–2815.