Académique Documents
Professionnel Documents
Culture Documents
Abstract
Lot streaming is the process of splitting a production lot into sublots, and then scheduling the sublots in overlapping
fashion on the machines, in order to improve the overall performance of the production system. Simulation-based and
industry-based reports have con"rmed that substantial bene"ts are possible via lot streaming. In this paper, we present,
for the "rst time, analytical results pertaining to the potential bene"ts of lot streaming in #ow-shop systems. The results
are developed using three common performance measures. These measures are (a) makespan (i.e., the total completion
time of all the lots), (b) mean #ow time, and (c) average WIP level. For each, an expression of the ratio of the measure
under lot streaming to the measure without lot streaming is developed. These expressions can be used to evaluate the
bene"ts of lot streaming under certain operating conditions. It is further shown that, in special extreme cases, these
expressions purely depend upon the problem parameters (i.e., the number of machines, the number of lots, the lot-sizes,
etc.) 2000 Elsevier Science B.V. All rights reserved.
Keywords: Flow shop; Lot streaming; Potential bene"ts
1. Introduction
Lot streaming is the process of splitting a production lot into sublots, and then scheduling the sublots in
overlapping fashion on the machines, in order to accelerate the progress of orders in production [1,2], and to
improve the overall performance of the production system.
Truscott [3] mentions several potential bene"ts of lot streaming. These include (a) reduction of production
lead times (thus, better due-date performance), (b) reduction of WIP (work-in-process) inventory, and
associated WIP costs, (c) reduction of interim storage and space requirements, and (d) reduction of material
handling system (MHS) capacity requirements.
The literature on lot streaming contains many simulation studies that have con"rmed that bene"ts can
indeed be achieved via lot streaming in various batch production environments, such as job-shops [4],
132
#ow-shops and serial production systems [5], Flexible Assembly Systems (FAS) [6], Group Technology
(GT) cells [7], and Cellular Manufacturing Systems (CMS) [8]. Simulation is a viable tool in studying system
performance. However, it is limited by the fact that it is data speci"c. Analytical results, on the other hand,
have general appeal. In this paper, we present, for the "rst time, such results pertaining to the potential
bene"ts of lot streaming in #ow-shop systems. The results are developed using three common performance
measures. These measures are (a) makespan (i.e., the total completion time of all the lots), (b) mean #ow time,
and (c) average WIP level.
The analysis in this paper is done for the case of a single lot as well as for the case of multiple lots. It is
assumed, however, that the transfer times and the setup times are negligible. If this is not the case, closed-form
formulae for the performance measures may not be available. For the analysis of special cases in which
transfer and setup times are not negligible the interested reader is referred to Kalir and Sarin [9].
Table 1
Processing time (in min) and lot size data for the case study
Lot
M
M
M (BN)
M
M
1
2
3
4
4
1
5
10
2
6
8
5
6
1
9
10
4
6
8
4
5
8
2
2
7
15
11
11
133
Table 2
Summary of results for the case study
I
II
III
IV"I/II
V
Makespan (min)
MFT (min)
318
590
46
0.54
0.50
224
476
53
0.47
0.31
2.81
3.23
13
0.87
0.62
To be able to fully take advantage of the bene"ts of lot streaming, setup and transfer times must be kept
signi"cantly small compared to processing times. For such cases where this is possible, it is useful for
managers to know the extent of the potential bene"ts of lot streaming under various operating conditions.
We address this issue for #ow-shop systems in this paper.
p
H
M
MFT
WIP
lot size
number of sublots streamed through the #ow shop, i"1,2, n
number of machines in the #ow-shop, j"1,2, m
sublot size
processing time of a single item on machine j, p ,max
+p ,
WHWK H
makespan
mean #ow time
average WIP level
A superscript &LS' will be added to a measure, such as the makespan, to denote that the analysis applies to
the lot streaming case. A superscript &*' will be added to denote optimality.
All the results are based on the following assumptions:
E
E
E
E
E
E
In this section, three performance measures are considered. These include makespan, mean #ow time, and
average WIP. An expression is derived to evaluate the potential bene"ts by utilizing lot streaming, under
negligible setup and transfer times, for each one of them.
134
3.2. Makespan
If n equal sublots are to be utilized, Baker (1988) has shown that the makespan value is as follows:
Q K
.
(3.1)
M"
p #((n!1)p
H
n
H
To assess the possible reduction in the makespan via lot streaming, note that, without the use of lot
streaming, the makespan for the lot would be
K
M"Q p .
(3.2)
H
H
Denoting by M*1 the makespan under lot streaming, given by (3.1) with nH"Q, the ratio of the two
makespans is as follows:
M*1 (Q!1)p # K p
H H .
"
Q K p
M
H H
(3.3)
Two extreme cases are of interest. First, consider the case of a small lot-size, processed on a large number of
machines, such that the following holds:
K
(Q!1)p ; p ,
(3.4a)
H
H
i.e., the total processing time per item is signi"cantly larger than the product of the number of items and the
maximum processing time per item. In this case, the ratio approaches the value 1/Q. On the other hand, if Q is
large, such that
K
(Q!1)p < p ,
(3.4b)
H
H
the ratio approaches the value p / K p . Note that if p "p("p ) j, i.e., a perfect balance of work
H H
H
across the machines, then this limiting value admits its lower bound, which is 1/m. The maximal makespan
reduction (MMR) is, therefore, as follows:
m!1
1
.
MMR"1! "
m
m
(3.5)
Hence, the maximal makespan reduction is attained when all the processing times are equal and Q is
su$ciently large. Under these circumstances, the makespan reduction depends purely on the number of
machines according to (3.5). Note that even for a relatively small number of machines, the reduction is
substantial (e.g., for m"5, 80% reduction).
We now develop similar results for the two other performance measures, namely, the mean #ow time
(MFT) and the average WIP.
3.3. MFT
Since portions of the lot become available to the downstream machines before the last portion, the mean
#ow time of the entire lot under lot streaming is expected to be smaller than what it would be without the use
135
of lot streaming. Let FT be the #ow time of the kth sublot (k"1,2, n, where n"Q/, and the sublot
I
size.) Assume that n is integer. Then,
K
FT "(k!1) p # p , k"1,2, n.
I
H
H
(3.6)
(3.7)
K
L
L
FT " (k!1)p # p
H
I
H
H
H
" p
"Q
L
K
(k!1)#n p
H
I
H
n!1
K
p # p .
H
2
H
(3.8)
(3.9)
Although, in this analysis, it was assumed that n is integer, expression (3.9) is in fact valid for every n value
(i.e., continuous as well as integer). As in the makespan case, the optimal mean #ow time is obtained for the
largest n possible, i.e., when sublots of unit-size are utilized. Thus, for optimality we have nH"Q. Substituting
this in (3.9), we get
Q!1
K
p # p .
MFT*1"
H
2
H
(3.10)
On the other hand, when lot streaming is not utilized, the mean #ow time is simply Q K p and therefore,
H H
the ratio of the two is as follows:
MFT*1 ((Q!1)/2)p # K p
H H .
"
MFT
Q K p
H H
(3.11)
For the same two extreme cases considered earlier for the makespan criterion, (3.4a) and (3.4b), the MFT
ratio results in the same value as for the case when lot size is relatively small while the number of machines is
relatively large. For the case when lot size is large, the ratio approaches the value p /2 K p (half the value
H H
obtained under the makespan criterion), which, under perfect balance (i.e., p "j), becomes 1/2m.
H
136
(3.12)
The numerator in expression (3.12) represents the area lying under the WIP level as a function of time. The
denominator in expression (3.12) represents the total time period over which the WIP level changes.
Expression (3.12) can be simpli"ed as follows:
+Q K p #p [(((Q/)!1)#((Q/)!2)#2#1)],
H H
WIP*1"
[ K p #(n!1)p ]
H H
Q K p #p ((n!1)#(n!2)#2#1)
H H
"
K p #(n!1)p
H H
Q K p #p (n(n!1)/2)
H H
"
K p #(n!1)p
H H
which leads to
WIP*1"Q
( K p #((n!1)/2)p )
H H
.
K p #(n!1)p
H H
(3.13)
From (3.13), it is clear that the minimum WIP is attained for the largest possible n, i.e., nH"Q. The average
WIP without lot streaming is simply Q units (since the entire lot stays in the system.) Hence, the ratio of the
two is as follows:
WIP*1 K p #((Q!1)/2)p
.
" H H
WIP
K p #(Q!1)p
H H
(3.14)
Note that the above ratio approaches 1.00 under the extreme condition of (3.4a), thereby indicating that no
reduction in the average WIP is expected via lot streaming. This is indeed anticipated. If the lot size is
relatively small while the processing times on the machines are relatively large, the time di!erence between
the completion of the "rst sublot and the completion of the last sublot becomes insigni"cant with respect to
reducing the average WIP. Thus, streaming sublots throughout the system would not have a signi"cant
impact on the average WIP.
For the condition speci"ed by (3.4b), the ratio reduces to . This implies that the maximal reduction in the
average WIP level under lot streaming is 50%.
3.5. Upper bounds on the benexts via lot streaming of a single lot
Table 3 presents a summary of the above "ndings. These pertain to the bene"ts of lot streaming. The limits
on the bene"ts with respect to each of the three performance measures, namely, the makespan, the mean #ow
137
Table 3
Upper bounds on the bene"ts via lot streaming
Measure
General ratio
If (Q!1)p < K p
H H
If (Q!1)p ; K p
H H
Makespan
(Q!1)p # K p
H H
Q K p
H H
p
K p
H H
1/Q
((Q!1)/2)p # K p
H H
Q K p
H H
p
2 K p
H H
1/Q
Work-in-process
K p #((Q!1)/2)p
H H
K p #(Q!1)p
H H
1/2
time, and the average work-in-process are provided in terms of the ratio of the values of the performance
measures obtained with lot streaming to that obtained without lot streaming.
(4.3)
138
where I is the total idle time between the lots on the bottleneck machine (without lot streaming). Therefore,
the ratio of the two is
,\p # , Q p
# K
p
#I*1
M*1
H
H
G G
G
,
H ,> ,
H
"
.
,\
M
Q p # , Q p
# K
Q p
#I
H
H
G G
G
,
H ,> ,
,
H
(4.4)
Consider the extreme case of the lots being sequenced under lot dominance, such that there does not exist any
idle time between the lots in both cases. That is, the following condition holds:
p *p
G
H
G>
H\
i, j.
(4.5)
This condition results in I*1"I"0 (which can easily be veri"ed.) If, in addition, we use equal lot sizes, i.e.,
Q "Q; expression (4.4) reduces to
G
M*1
Q , p
# ,\p # K
p
G G
,
H
H
H ,> ,
H .
"
,\
M
Q(
p # , p
# K
p
)
H
H
G G
,
H ,> ,
H
(4.6)
(4.7)
Under the case of perfect balance, i.e., p "p i, j (which still satis"es conditions (4.1) and (4.5)), the above
GH
expression reduces to the following:
M*1/M"N/[N#(m!1)].
(4.8)
Note that expression (4.8) is purely dependent on the problem parameters. For a relatively small number of
lots, and a large number of machines, lot streaming is expected to be very bene"cial. For example, for N"5,
and m"10, the reduction in the makespan is: 1! "64.3%. This, however, assumes, as mentioned earlier,
that unit-sized sublots are utilized and that setup and transfer times are negligible. On the other hand, note
that the above analysis assumes zero idle time on the bottleneck machine, even without lot streaming.
However, it is expected that in many instances I<I*1, which further enhances the gains via lot streaming.
We also note that, based on expression (4.8), the bene"ts are likely to decrease as the number of lots
considered, N, increases.
The other extreme case, that is considered, is the case in which the number of machines is large while the lot
sizes and the number of lots are relatively small, such that
,
,\
K
Q p
< p # p
.
G
,
G
H
,
H
G
H
H ,>
In this case, from expression (4.6), we get
M*1
,\p # K
p
H
H
H ,> ,
H
"
M
Q( ,\p # , p
# K
p
)
H
H
G G
,
H ,> ,
H
(4.9)
(4.10)
139
Expression (4.10) suggests even greater improvements to the makespan, via lot streaming, than expression
(4.8) does. However, note that N must be su$ciently small in this case. Also, note that when N"1, i.e.
a single lot is considered, both expression (4.8) and (4.10) reduce to their respective ratios in the single-lot case
(i.e., to 1/m and 1/Q).
4.2. MFT
The mean #ow time (MFT) for multiple lots, where sublots of the same lot follow one another,
can be obtained by dividing the sum of the MFTs of the individual lots by the number of lots. Thus,
we have
,
MFT*1" MFT*1 N.
(4.11)
G
G
Expression (4.11) represents the lot-MFT, not to be confused with the sublot-MFT. Using expression (3.10),
for the "rst lot in the sequence, we have
K
Q !1
p
.
(4.12)
MFT*1 " p #
,
H
2
H
Assuming that conditions (4.1) and (4.5) are in force, then for the ith lot in the sequence, the mean #ow time
can be computed as follows:
K
Q !1
MFT*1" p # G
p
.
G
G
H
G
,
2
H
Here, it is assumed that the system starts at time zero, and no interference of #ow occurs due to the schedule
of previous lots (because conditions (4.1) and (4.5) are in force). However, the ith lot actually starts not at time
zero, but at time G\ Q p . Therefore, the MFT of the ith lot is
P P
P
G\
K
Q !1
MFT*1" Q p # p # G
p
#I
(4.13)
G
P
P
G
H
G
,
G
2
P
H
where I is the slack time present in the processing of job [i] on machines before BN. Substituting (4.13) in
G
(4.11) we get
, G\ Q p #( , K p )#( , ((Q !1)/2)p
)# , I
G H G
H
G G
,
G G
.
MFT*1" G P P
P
N
(4.14)
Utilizing the same approach as for the case without lot streaming, we have
G\
MFT " Q p ,
G
H
H
G\
K
MFT " Q p # Q p #I .
G
P
P
G
G
H
G
P
H
I is the counterpart of I for this case. Therefore,
G
, MFT
, G\ Q p # , K Q p # , I
G " G P P
P
G H G
G
H
G G
.
MFT" G
N
N
(4.15)
(4.16)
140
G
,
G G
.
" G P P
P
MFT
, G\ Q p # , K Q p # , I
G P P
P
G H G
G
H
G G
(4.17)
For the extreme case of equal lots and a perfect balance (Q "Q, p "p i, j), expression (4.17) reduces to the
G
GH
following:
QN#2m!1
MFT*1 (N(N!1)/2)Qp#Nmp#N((Q!1)/2)p
"
"
.
MFT
(N(N!1)/2)Qp#NmQp
Q(N#2m!1)
(4.18)
From expression (4.18), if the lot size is su$ciently large, the ratio approaches the value N/(N#2m!1),
which is signi"cant as long as N is relatively small. Note that for N"1 (i.e., a single lot), the above value
reduces to 1/2m, in agreement with the results for the single lot case (see Section 3.2). Also note that if the
number of lots becomes very large, the value approaches 1.00, as in the case of the makespan.
4.3. Average WIP
Following the analysis for the single case under the WIP objective, note that at time zero, all the lots are
waiting to be processed, and therefore, the WIP level is , Q . The "rst sublot of the "rst lot leaves the
G G
shop after K p time units. The next (Q !1) unit-sized sublots of the "rst lot each leave the shop after
H G
H
an additional p
time units. For each unit-sized sublot of the ith lot, the inter-departure time is p
. If,
,
G
,
for simplicity, we assume at this stage that the lots are of equal size and that the processing times are equal
(i.e., perfect balance), the average WIP level under lot streaming is as follows:
(NQ)(mp)#p ,/\(NQ!i) NQ(m#(NQ!1)/2)
G
WIP*1"
"
.
mp#(NQ!1)p
m#NQ!1
(4.19)
On the other hand, for the regular case (i.e., without lot streaming), we have
(NQ)(Qmp)#(Qp)Q ,\(N!i) NQ(m#(N!1)/2)
G
WIP"
"
.
Qmp#(N!1)Qp
m#N!1
(4.20)
(4.21)
Although expression (4.21) suggests that for large N (such that N<m), the value approaches 1.00, i.e., no
bene"ts via lot streaming, consider the following instance: N"10, Q"10, m"5. The resulting ratio is
77.2%, i.e., a reduction of 22.8% in the average WIP level. Moreover, for a reasonable N, and a large Q, the
ratio becomes
1 m#N!1
.
2 m#(N!1)/2
This ratio admits its lower bound, of , when m<N. This lower bound is identical to the one obtained for the
single lot case.
141
Table 4
Approximate upper-bounds on the bene"ts via lot streaming in a multiple-lot #ow-shop
Measure
Makespan
Ratio
Q , p
# ,\p # K
p
G G
,
H
H
H ,> ,
H
Q( ,\# , p
# K
p
)
H
G G
,
H ,> ,
H
Large number of
machines, small
lot sizes
N
[N#(m!1)]
m!1
Q[N#(m!1)]
N
, G\ Q p #( , K p )#( , ((Q !1)/2)p
)# , I
G P P
P
G H G
H
G G
G
,
K G
N#2m!1
, G\ Q p # , K Q p # K I
G P P
P
G H G
G
H
G G
1/Q
Work-in-process
m#(NQ!1)/2 m#N!1
m#(N!1)/2 m#NQ!1
(perfect balance)
1 m#N!1
2 m#(N!1)/2
Acknowledgements
The authors thank the reviewers for several helpful comments.
142
References
[1] K.R. Baker, D. Jia, A comparative study of lot streaming procedures, OMEGA International Journal of Management Science 21 (5)
(1993) 561}566.
[2] J. Chen, G. Steiner, Lot streaming with detached setups in three-machine #ow shops, European Journal of Operational Research
96 (3) (1997) 591}611.
[3] W. Truscott, Production scheduling with capacity constrained transportation activities, Journal of Operational Management
6 (1986) 333}348.
[4] T.L. Smunt, A.H. Buss, D.H. Kropp, Lot splitting in stochastic #ow shop and job shop environments, Decision Science 27 (2) (1996)
215}238.
[5] C.T. Wu, P.J. Egbelu, Control of material #ow in serial production systems, in: Proceedings of the 1994 International Mechanical
Engineering Congress, Chicago, IL, USA, 1994, pp. 1}8.
[6] G. Sohlenius, A. Arnstrom, P. Grondahl, Solution to #exible, productive assembly as part of the total manufacturing system.
Proceedings of Manufacturing International '88, Vol. 1, ASME, Atlanta, GA, USA, 1989, pp. 223}230.
[7] S. Vembu, G. Srinivasan, Heuristics for operator allocation and sequencing in just-in-time #ow line manufacturing cell, Computers
and Industrial Engineering 29 (1}4) (1995) 309}313.
[8] R. Logendran, P. Ramakrishna, Manufacturing cell formation in the presence of lot splitting and multiple units of the same
machine, International Journal of Production Research 33 (3) (1995) 675}693.
[9] A.A. Kalir, S.C. Sarin, Optimal and heuristic solutions to the single and multiple batch #ow shop lot streaming problems with equal
sublots, International Journal of Production Research, in press.