Вы находитесь на странице: 1из 12

Customer Service in Pull Production Systems Author(s): Mark L. Spearman Source: Operations Research, Vol. 40, No. 5 (Sep.

- Oct., 1992), pp. 948-958 Published by: INFORMS Stable URL: http://www.jstor.org/stable/171820 Accessed: 22/01/2009 11:24
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=informs. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Operations Research.

http://www.jstor.org

CUSTOMERSERVICE IN PULL PRODUCTION SYSTEMS


MARKL. SPEARMAN
Northwestern University, Evanston, Illinois

(ReceivedMarch1990;revisionsreceivedSeptember1990,June 1991;acceptedOctober1991) In this paper,we considerthe issue of customerservicein pull productionsystems.We first discussthe meaningof customerservicein a pull systemand contrastit to push systems.We also discussboth pureand hybridpull systemsthat arefoundin the literature. We then investigate the effectsof changinginventorylevelsand processing time characteristics on customerservicein a pure kanbansystem.Finally,we show that a hybridsystemknown as CONWIPnot only has betterservicethan a pure kanbansystem,but also solvescertainimplementation problems.

successof Toyotaand otherJapanese manufacturers during the 1970shasattracted the attention of production researchers to a set of techniques knowncollectively as just-in-time or JIT.An integral featureof JIT systemsis the use of pull shop floor control systems, the most common of which is havegainedpopularity kanban. Although pullsystems amongpractitioners, therearea numberof issuesthat require study.One of theseis customer service. To begin, we must have defintionsand measures. In a push system,the notion of customerserviceis well developed.If the time to completea job is less than or equal to the allotted lead time, the job is consideredto be on-time. Due dates are explicitly consideredin the operationof such systems.Order releasedatesare typicallycomputedby subtracting a knownlead time from the desireddue date. Typical measuresinclude fractionof jobs on-time and the tardiness. average In a pull systemthe veryconceptof customer takes on different Ohno(1988),generally meaning. credited as the developer of kanban,statesthatthe ideaforthe systemcamefromhis observations of American supermarkets. He describes a supermarket as a placewhere "acustomer can get: 1) whatis needed,2) at the time needed,and 3) in the amount needed."At Toyota, each processbecameboth a supermarket for downstream processes and a customer to preceding processes. Consequently, theleadtimein a pullsystem is zero.Eitherthe item is available or it is not. Appropriateservicemeasures for pull systemsarethe probabilityof stockout,the expectedtime to fill demand, and the expected backlogof orders. Thisdifference betweenpushandpullis one reason why a kanbansystem,usedto controlthe shop floor,
Areaof review. MANUFACTURING,
PRODUCTION AND SCHEDULING.

The

cannotbe compared directlyto an MRP systemthat is used to controlworkrelease.In the MRP system, production is controlled the use of the Master through Production whichultimately determines Schedule, the release rate to the system. In many such systems a capacity check is provided in the form of "input/outputcontrol" (Wight 1977). This check trackswork-in-process (WIP) levels to detect overloadedconditions(WIPincreasing) or underutilized processes (WIPdecreasing). a kanban Conversely, system will operate withoutlinkageto due datesif left to run unimpeded. Althoughthis linkageis beyondthe scopeof kanbanitself,it is worthnotinghere. Typically, in a pull due dateintegrity is maintained some sortof production systemby establishing quota. As Schonberger (1986) states,"the day is not done until everyjob has been completed." In this way,the due dateof a newjob is established the by considering current production quotaand the current backlogof jobs. So, unlikepushsystemsthatcontrolthroughput and measureWIP, pull systems control WIP and measure throughput. It has been arguedthat a fundamental reasonfor the successof pullsystems lies not in the factthatthey pull, but in the way they preventWIP buildupby directly controlling WIPand throughthe earlydetection of capacitydeficiencies (Spearman and Zazanis 1992).We will returnto this importantpoint when we discuss differentmeans for implementingpull production systems. The purpose of this paper is to discuss various of customer in pullproduction aspects service systems. To do this, we first discuss and compare various implementations of kanbanand otherpull strategies. We then investigate the effectsof changing inventory

Subject classification. Inventory/production: operating characteristics; stochastic models. Operations Research0030-364X/92/4005-0948 Vol. 40, No. 5, September-October 1992

948

$01.25 0 1992 Operations Research Society of America

Customer Servicein PullSystems / 949


levels and processingtime characteristicson customer service in pull systems. Finally, we show that a hybrid push/pull system has better service than the pure pull system representedby kanban. This last result rationalizes the common industrial practice of implementing less than "pure"forms of kanban. 1. BACKGROUND Although implementations in this country are relatively new, kanban itself is not. Ohno states that Toyota had adopted this system plant wide as early as 1962. He provides the underlying philosophy of the system but gives few operational details. Detailed descriptions of the system are provided by Sugimori et al. (1977), Monden (1983), Hall (1983), and Karmarkar(1986a). Kimura and Terada (1981) provide an early mathematical description and use simulation to evaluate the impact of stochastic variation. Other simulation studies include Philipoom et al. (1987) and Krajewskiet al. (1987). Because of difficulties in modeling systems with blocking, there have been relatively few quantitative studies of kanban systems. Notable exceptions include Bitran and Chang (1987) who provide a means of setting card counts in a deterministic kanban assembly system. Buzacott (1988) uses linked queueing networks to formulate models of both kanban and MRP systems, but does not solve these models. Deleersnyder et al. (1989) model a kanban system using a Markov chain and then provide several numerical examples to show the benefits of variabilityreduction. Because each of these formulations of kanban is slightly different, it may be useful to describe a typical installation in some detail. Figure 1 is similar to one in Hall and describes what is called the two-card kanban system. The two cards are move cards and production cards and are attached to standardcontainers of parts. A standardcontainer contains a fixed quantity of parts of a certain type. The cards contain information on both the identity of the parts as well as the source and destination of the standard container. At each process center, standardcontainers are stored at both inbound stockpoints (on whose containers are attached a move card) and outboundstockpoints (equipped with production cards). A material handler from a downstream station arrives with the demand in the form of a move card. The material handler removes the production card from the standard container and places it in a card stand, then attaches the move card to the container and brings it to the inbound stock point of the demanding station. As containers are used at a station, the move card is removed and placed in a different stand from which the material handlers remove them and fetch the replenishing parts. Meanwhile, the accumulation of production cards in their stand authorizes the production center to transforminbound stock into outbound stock. In some instances, these cards are removed from the stand and placed on a kanban board that is used to schedule production. The benefits of such a system are several. One is that all workers can immediately see what work needs to be done by looking at the various kanban boards. Another is that, unlike a push production system, excessive WIP will not be pushed into the system whenever the system capacity is optimistically overestimated. Also, it appears that it is inherently easier to control WIP than to control throughput(i.e., small errors in setting card counts have less proportional impact on system performance than do errors in setting release rates) (Spearman and Zazanis). Finally, when the system works well there is no need to schedule production because parts are always available when needed and in the quantity needed. Although kanban was first implemented in Japan

Figure 1. A two-card kanban production system.

950 / SPEARMAN
in the 1960s, it has roots that go back to the 1950s and the United States. Simpson (1958) describes a "base stock system" that is very similar to kanban. He also attributes the system to George Kimball (1988) in an unpublished Arthur D. Little report. (This insightful work is no longer unpublished thanks to Professor Uday Karmarkarwho obtained permission to publish the report in 1988.) Simpson describes the system as follows: Whenan orderis placed,it is filled from inventoryif the inventory is not zero. If the inventoryis zero,the orderis placedin a backorder file,to be filledwhenan item arrives. In any event,a manufacturing orderis immediately placed with the preceding manufacturing operation to produce an item to replacethe item that has been consumed.The manufacturing operator,in turn, immediatelyplaces an orderfor the required raw materials againstthe preceding inventory, and as soon as this orderis filled[i.e., he hasthe on it to produce neededinventory], proceedsto 'operate' the requireditem. In this way an order againstthe last inventory fora finished item is immediately transmitted all the waybackalongthe line to all the manufacturing operations,eachof whichis galvanized into production. The main difference between the base stock system and kanban lies in the words "in any event." Kanban would not place an order for more parts if a demand (in the form of a move card) had arrived when there was no stock in the outbound stock point. Instead, there would be one or more production cards already in process and whenever one of these was completed, the waiting move card would immediately be attached to the recently completed container of parts and the production card would be sent back into production. In this way, a kanban system bounds the amount of WIP there can be in the system while the base stock system does not. Unfortunately, kanban is not for everybody. A kanban system takes as axiomatic that setups will always be short so that a process center can switch production on parts as frequently as needed to meet the demand dictated by the production cards. Likewise, the product mix must be highly stable and the production level must be extremely smooth (Monden spends an entire chapter on how to smooth production). To see the importance of this, consider what would happen if there were suddenly demand for a part for which there had been no demand for six months. Either the parts would be in the outbound stockpoint or they would not. If they were there they would have been there for six months or more. If not, the kanban system failed to meet its customer service goal of 100%and the demand would have to percolate back through the system. Other problems can arise from the sheer number of active part numbers. We have observed cases with over 10,000 active part numbers (i.e., parts for which there are orders in the foreseeable future) with many of these part numbers representingdemand for "onetime-only" jobs that will not be needed again for months. When one such company wanted to implement some sort of pull system, it became apparent that kanban was not the answer. Indeed, if kanban had been implemented, inventories would have increased. This is because there must be WIP in standardcontainers for each active part number. In recent years, a number of hybrid push/pull systems have been proposed (Kamarkar 1986b). However, before discussing hybrid systems further, let us provide a working definition of pull and push. We shall say a system employs push if it schedules the release of work a priori. A pull system, on the other hand, authorizes the release of work based on current plant conditions. A hybrid system involves aspects of both. A pull system similar to kanban was proposed by Denardo and Tang (1992) in which moves occur at the end of each production period and where the amount of production authorized during the period is set so that the ending inventory is equal, in expectation, to a specified target. Groenevelt and Karmarkar (1988) describea hybridsystem that was implemented in a textile plant. In this system, kanbans are dynamically adjustedin response to changes in the forecasted product mix. A different hybrid system known as CONWIP is discussed in more detail in Section 3, where we compare it to kanban. 2. EFFECT OF SYSTEM CHARACTERISTICS ON CUSTOMERSERVICE We now examine a simplified (one card instead of two) version of the system, as reviewed above (see Figure 2). At each station, work can commence whenever there is both stock and a card available. The equivalence of this to a two-card kanban system is shown in Figure 3. In this case, the inbound stockpoint for an operation is directly associated with that

Figure 2. A kanban production system subject to external demand.

CustomerService in Pull Systems /

951

Figure 3. Equivalence of one and two-card kanban systems.

operation (squarebox), while the outbound stockpoint is associated with the next operation (rounded box). This next operation represents the performance of material movement. Hence, if station zero is a move operation (e.g., shipping)and demands are move cards (from outside the system), then each odd numbered station representsa process operation, while each even numbered station represents a move operation. The card flows are then exactly the same as discussed in Section 1. Without loss of generality,we assume that the cards move instantly (e.g., electronic kanbans) so that when a demand arrives to an idle system (i.e., all stockpoints full), production will commence at each station immediately. We also assume that the last station always has parts (raw materials). In keeping with JIT philosophy, we assume that the standard containers are small so that all demands are integer multiples of the container size (Deleersnyder et al.). The service measure of interest will be the expected time required to fill demand. We desire this measure to be small and it should be equal to zero if stock is always available whenever there is a demand. For this reason we do not consider the time to perform the last (move) operation when computing our service measure. (This move might representthe time to pick the order from inventory, the time to ship the order, etc.) We choose this measure because, in a pull system, the longer a stockout occurs, the more likely it is to result in the disruption of downstream processes. Thus, the probability of stockout alone contains insufficient information. Likewise, the expected length of the backlog is an indirect way to measure the same performance. 2.1. The Model In this paper, we will make stochastic comparisons of various queueing systems. Similar work has been done with respect to other systems and performance measures. Shanthikumarand Yao (1987, 1988a, b, 1989a, b) describe relationships between throughput, queue lengths, and service times in closed queueing networks

with and without blocking. Our approach follows that taken in 1989a. Hopp and Simon (1989) make similar stochastic comparisons of kanban systems involved in assembly operations. Also, Adan and Van der Wal (1989) have shown that throughput is monotonic in the number of jobs in a closed queueing network. We model the system with N single server stations in series producing a single product. The service times for station i are representedby the random variable Si - Fi and the set of service times lSij} representsa stochastic process. Demands in the form of kanbans from an external source are received at times Di, i = 1, 2, .... Demand for more than one standard container at a time is modeled as multiple demands occurringat the same instant. This, of course, implies that the time between demands need not be iid. We start the system with mi > 0 kanbans attached to the standard containers residing in the stockpoint of station i and with no free kanbans in the system. Work will commence immediately at all stations upon receipt of the first external demand, i.e., at time D,. We denote the time the jth container is sent from station i as T(i, j) and the particularkanban system as K(S, m) 2.2. Definitions The relations between the completion times, service times, and the demands in the kanban system are: + Soj, T(O, j) = max IDj, T( 1,j - mo), T(O, j - 1)}I j= 1, 2, ... (1) T(i-1I, j) -Si- ~j, T(i + I,j -mi), T(i,j) = maxlI T(i, j -1)}+ Sij, j = 1, 2, .. ; i= 1, 2, ...,N, (2) where T(k, 1) = 0 for k > N and I < 0. The max term representsthe time that work will commence on the jth container of parts. This is either the time of arrival of the last kanban (or demand, in station zero), the time of arrival of the standard container of parts

952

SPEARMAN

needed for the jth demand, or the time the operation last became idle, whichever is later. The jth completion time is simply the time that work starts plus the jth service time. The following definitions provide our measures of customer service. Deflnition 1. The time to satisfy the jth demand, -rj, is given by
=

Consider three possible processing times, S(') < S(2)< V). To obtain convexity, we show that
, (2) _ (1) T (3) _
-

(1)

S -(2

1) S(3)

for any of three possible cases:


=O as well, so that both Tr(3)=-(1) = 0. Clearly, -0(2) sides of (3) are zero. 2. r(3)> 0, r(1) > 0. In this case, we can rewrite(3) as

1.

[T(1, j

mo)

Dj]+. T 2) - T()
S(2)
-

(Recall that the first term representsthe time the stock for the jth demand is available.) Deflnition 2. The expected time to satisfy demand after the nth demand is defined to be
I Un= n

S(1) *1

-<

T(3) -T()
S(1) -S1'

S3 S(3)

since T is convex in S.
3. r(1) = 0, _r3) > 0. There are two subcases.
T(2) = 0. In this case, the left-hand side of (3) is zero while the right is not. b. T(2) > 0. Rewriting (3) as

a.

j=l

E[iV].

Deflnition 3. Assuming the limit exists, define U= lim Un to be the average time to satisfy demand. Note that the measure Uncan increase without bound if there is insufficient capacity in the production line. We consider only cases in which U is defined. The following lemmas are useful in latter proofs. Lemma 1. T(i, j) is nonincreasing in mi. Proof. Since T(i, j) < T(i, j + 1) < ... then from definitions 1 and 2 we see that increasingmi will either decrease T(i, j) or will have no effect. Lemma 2. T(i, j) is increasing convex in Sik, k - 1,
2, . .. , j.

T D2

S(2) -S(1)

T3)

D-

-3)

S(1)

which is true since T is convex in S. 2.3. Stochastic Ordering in Kanban Systems In this section, we examine the effect of changing the characteristics of processing times on the customer service level of a kanban system. For our first result we use the notion of "stochastically less than" (see, e.g., Ross 1983, p. 251), i.e., for random variables X - F and Y - G that X <st Y if and only if F(x) 3 G(x) for all x.
Theorem 1. Consider two kanban systems K(S(k), m), k = 1, 2. Then S?M <St S(2) implies U(1) < U(2).

Proof. First note that for i = 0 the result is obvious. For i > 0 we expand (2) back to the T(O,j) - Soj terms and note that T(i, j - 1) > T(i, j - mi-,) for mi-, > 0 yielding,
T(i, j)
=

Proof. We prove the result using the notion of "coupling," as described in Ross. We first assume that S(') - F and S(2) - G and then generate successive service times for station i from
S(-)

= FT1(Uj) G (Uj),

max{Dj, T(O, j T(i,j


-

1), T(1, j

S(-2)=
-

1),..., + Sij,

1), T(i + 1,j -mi))


Sik,

where Uj, j = 1, 2, . . . represent iid uniformly distrib-

which is clearly increasing convex in * ,J

1, 2,

uted random variables. From the definition of stochastic ordering, S() < SV) i = 1, ..., N; j = 1, 2, .... Likewise from (1) note that T(')(i, j) T(2)(i, j). Since U, are increasing functions of T, U(1)s U(2). Corollary 1. Consider two systems having processing
timesSf,
-

Lemma 3. rj is increasing convex in Sij. Proof. First note that rj is increasing in T(1, j - mo) which is increasingin Si,j1,_ so that rj is a nondecreasing function of Si,j-m0 . For the remainderof the proof, we drop the subscript j and the station number 1.

6k) + 1j,k= 1, 2; i= 1, . ..,N;j=


ft(k)

1, 2,

where is a constant and fj are iid random variables with zero mean. Then 0(1) s a(2) implies
U( ) < U(2).

Servicein PullSystems / Customer


Proof. This follows immediately from Theorem 1. Thus, for systems with processing times whose means represent a location parameter (e.g., the normal), faster processing times imply better service. Of course, stochastic orderingis strongerthan this in that all the moments of one distribution will be less than or equal those of the other. That a system having greatercapacity also has better customer service is not surprising although we have shown this to be true under very general conditions. The following results deal with variabilityreduction in a kanban system. We first present a theorem that makes use of increasing convex ordering, i.e., X Bi, Y if and only if E[f(X)] : E[f(Y)] for any increasing convex functionf

953

Example 3. If S), k = 1, 2, are iid random variables with Weibull distributions with equal means and f30l)> X3(2), then U(')< U(2). shapeparameters, Note that the coefficient of variation of Weibull random variablesis a function of the shape parameter only with increasing values of 3 implying smaller coefficients of variation. 2.4. The Effect of Increasing Inventory Levels We now look at the effect of increasing the inventory level in a kanban system on customer service.

3. Consider Theorem twokanban systems, K(S, m(j)) j = 1, 2. Then m') >, m(2); i = 1, ..., N implies
U(1) S< U(2).

Theorem 2. Consider twokanban m systems, K(S i),


j = 1, 2. Then S(1)aicx S(2), i = 1, 2, ..., N implies T(')(O,j)3icx T(2)(0,j), j = 1, 2, . . . and U(1) 3 U(2). Proof. The first part is an immediate consequence of Lemma 2. The second part follows from the definition of U.

Proof. From Lemma 1 we see that increasing the initial WIP level at one or more stations can only make shipment times earlier, i.e.,

T(')(O, j) - T12)(0,j), j = 1, 2,.


Hence, the time to satisfy demand in system 1 will not be greaterthan that in system 2. This appears to be a classic tradeoff of inventory versus service (as in the safety stock setting problem). Also, at first glance, this appears to contradict the popular JIT maxim, "inventory is evil." A closer look shows that although extra inventory improves service it also reducesflexibility. This flexibility is important in the face of a changing product mix (as discussed earlier) and changing products (i.e., engineering changes). Also, Theorem 2 shows that one can improve service (or maintain service with less inventory) if one can reduce processing variability. Since major sources of variability are unreliable machines, rework, and the lack of standardization,excess inventory is evil because it hides problems (Monden), or rather, allows us to live with problems that could be eliminated. Indeed, Schonberger is correct in stating that "variabilityis a universal enemy." Theorem 3 also highlights a distinct difference in push and pull production environments. In a maketo-order push system with a fixed quoted lead time and constant throughput rate, increasing WIP levels will only degrade customer service. This is because the average flow time is given by Little's law, Average Flow Time
-

Corollary2. Considertwo kanban systems, K(S'(j) F and G m), j = 1, 2. If V) - F and S(2)- G, where have the same mean and whereF crossesG at most
once andfrom below, then U(') <
U(2).

Proof. From Stoyan (1983, p. 12) we see that the "singlecut criterion"describedin the corollaryimplies increasing convex ordering, i.e., S(2) 3icj V). The result is then obtained from Theorem 2. These results have immediate application to several common distribution functions (Stoyan). Example 1. If S(,k), k = 1, 2 are iid random variables having normal distributionswith means MI1) ,U(2) and variances " < C2), then U(') <2
-

A similar statement can be made for normally distributed random variables conditioned to be positive. A more general statement can be made about distribution functions of the class, P, with a common mean, p, and differing only by a scale parameter,aC(k), k= 1, 2, i.e.,
g(t; A, O-)= bo((t
t WU/o); t, p

0, a > 0.

Example 2. If S(Y)_ 4(.; ,A, (k)), k = 1, 2, and a(l) < < U(2). 0(J2) for i = 1, 2, . . .N,N, then U(') Similar statements can be made for a very different class of distribution functions.

AverageWIP Throghput Throughput

and that whenever flow times increase without a correspondingincrease in lead time, customer service will deteriorate. Of course, the main difference between

954 / SPEARMAN
these two systems is that the pull system is make-tostock, while the push system is make-to-order. 3. A COMPARISONOF KANBAN WITH A HYBRIDSYSTEM Earlier we discussed the problems of implementing kanban in a production environment with either a constantly changing product mix and/or an extremely large part number set. These conditions led us to consider what, exactly, does kanban offer in the way of improving production control and can these improvements be made using a more general construct than kanban. 3.1. CONWIP In an earlier study comparing push and pull systems, we concluded that: 1) controlling WIP is more robust than controlling throughput, and 2) the fact that WIP is bounded is more important than the practice of "pulling" everywhere (Spearman and Zazanis). This led us to consider a system that does not pull at every station and in which WIP was the control variable. This system maintains a constant amount of WIP on each production line and is called CONWIP. In this system, work is started at the first station in a line only when the WIP level for the line has fallen below a specified level. Otherwise,work is pushed within the line. Gong and Matsuo (1990) have proposed a similar system (constant net WIP) that considers yield losses. Unlike kanban, in a CONWIP line production quantities are measured in terms of standard parts, represented by the time a part takes at the slowest station in the line (the bottleneck) (Spearman, Woodruff and Hopp 1990). If part A requires one time unit at the bottleneck, while part B requirestwo, the completion of a job of 100 units of B would authorize the start of two jobs of 100 units of A. Also, unlike kanban the sequence of work is important. This is so a CONWIP system can be used when there are significant setup times (with setup times measured in terms of standard parts). The "CONWIP Backlog" is used by the operator at the first station in the line to determine what job to start next. This backlog is generated by attempting to group jobs sharinga common setup (at the bottleneck) while ensuring that all jobs finish on time (Woodruff and Spearman 1992). So while the pull in kanban occurs between stations and is to replenish the particular part that has just been used, the pull in CONWIP is over the entire line and for parts having the same routing (see Figure 4). Thus, under CONWIP the completion of part A can authorize the start of part B (if B is next to be pulled

Figure 4. A CONWIP production system with backlog. on the backlog). This feature helps CONWIP to accommodate both small and infrequent orders as well as a changing product mix. A CONWIP line is, therefore, quite similar to a closed queueing network. As such, the flow time and throughputrate tend to be less variablethan an equivalent open network having the same throughput (Whitt 1984). This allows for better predictability of job completion times that can be exploited in a higher level planning system (Spearman, Hopp and Woodruff 1989). While CONWIP can be implemented to include all operations from raw materials to finished goods, this does not have to be the case. We are in the process of implementing a version of CONWIP in a large circuit board plant that will be composed of five separate CONWIP lines in series. Buffers are maintained between the lines in much the same fashion as buffers occur between stations in a kanban line. CONWIP and kanban are equivalent when each CONWIP line consists of one process. 3.2. Comparison In this section, we compare the customer service of the kanban and CONWIP systems composed of the same servers. This extends the results of Spearman and Zazanis, where they showed that a CONWIP system will have greater limiting throughput than a kanban system when the serversare exponential. We first state the relation between completion times, service times, and demand times in the CONWIP system. T(O,j)
=

max$Dj, T(1,j-

mo), T(O,j- 1)} + Soj, j= 1, 2, ... (4)

T(i, j) = max{T(i + 1, j - mi), T(i, j - 1)J + S,, j= T(N, j)


=

1,2,...;i=1,2,...,N-1,
SNj,

(5)

max{T(O,j) - SOj, T(N, j - 1)J + j= 1, 2, ....

(6)

The first equation is similar to the one for the kanban system. The second, for interior stations, does not include the time of arrival of the kanban (since they are moved with the containers). The last equation

Customer Servicein PullSystems /


does not include the time of arrival of the next required standard container because we assume parts are always available. Note that using the above definitions, we can prove theorems similar to Theorems 1, 2, and 3 for CONWIP systems. We denote a given CONWIP system as C(S, m). The following lemma dealing with kanban systems is also useful.

955

Also, T(l, j - mo + 1) 3 T(l,jmo + 1) because the only change to the system since t occurred at station 0. Then from (1) and (4) and from the fact that T(O,j) = T(O,j), T(O,j + 1) 3,:T(O,j + 1). Case 2. (0 < k < N) Again, because the only change to the system since t occurred at station k, T(k + 1, j
-

Lemma4
T(i, j) fori=
-

mO +

1) > T(k + 1, j

mO +

1).

Sj : T(i

1, j)-S,1,2,....

Then from (2) and (5) and the fact that T(k, j) = T'(k,j), T(k, j + 1) >- T(k, j + 1).

1,2,...,N;j=

Proof. This follows directly from (1) and (2). A similar lemma can be shown for the CONWIP system. In Lemma 5, we compare the completion times in a kanban system to those in a CONWIP system on a common probability space. Lemma 5. Consider a kanban system K(S, m5)and a CONWIP system C(S, mi) such that the service times are on the same probability space. Then

Case 3. (k = N) The result follows from


T(N,j+
=

1)

maxl T(N- lj+ 1)

1) - SN-I,j+l, T(N,j)}

SN,j+l

T(N,j+
=

max{T(O,j + 1) - So,j+1, T(N, j)} +

SN,j+

and Lemma 4. The following theorem relates the customer service in a kanban system to that in a CONWIP system. Theorem 4. Consider a kanban system (K(S, m') and a CONWIP system C(S, miz).Then both systems having the same service time distributionsat each station implies U > U. Proof. The proof is an immediate consequence of Lemma 5 and the definitions of r and U.
3.3. Discussion and Simulation Results

T(i, j) -> T(i, j),


i= 0, 1, 2, ..., N; j1, 2 .... (7)

Proof. We prove the lemma by constructingthe sample paths of the two systems on the same probability space. To facilitate the proof we define several counting processes, Ai(t), as the number of completions at station i at time t, and
N

Nt =

Ai(t).

Given our initial starting conditions, the theorem is clearly true at t = 0. Assume that at t > 0, Nt = n,

T(i, j)

>

T(i, j) for all i, j. Now consider the

= n + 1. Note that if system at t + 6 such that N1+6 T(i, Ai(t)) > T(i, Ai(t)), i = 1, . . ., N at time t the condition will continue to hold trivially. However, suppose that at t, T(k, Ak(t)) = T(k, Ak(t)) and that Ak(t + 3) = Ak(t) + 1, i.e., the most recent completion occurred at station k. We show that (7) continues to hold under all conditions by considering three cases. Case 1. (k = 0) Note that T(0,j+
=

1)
-

max$Dj+1,T(1,j 1)

mo + 1), T(O,j)} + So,j+1

T(O,j+
=

max{Dj+1, T(1, j - mo + 1), T(0, j)} + So,j+.

This result is not that surprisingbecause the CONWIP system behaves (almost) like a closed queueing network, while the kanban system looks somewhat like a closed queueing network with blocking. However, the result is important given that it holds for any distribution of processingtimes and for any demand point process. We should note, however, that although the card counts are the same, the CONWIP system will tend to have a higher average WIP level than the kanban system since the latter system can have unattached cardsat interior stations. This is, no doubt, one reason why CONWIP has superior service. In spite of this, it appearsthat CONWIP can dominate kanban in terms of both service and average WIP. Unfortunately, because no exact analytic solutions for either system exist, such comparisons are difficult to perform. For this reason we have performed a systematic comparison of the two systems (see the Appendix for details).

956 / SPEARMAN
We compare kanban and CONWIP systems having three stations (the first station is the "staging"operation that has zero process time). In each case, we have 18 cards in both systems. We set card counts in the kanban system according to a procedureadvocated in several practitioner texts on kanban (e.g., Hall and Monden). Then we run the kanban system and observe the average WIP level and the average time to satisfy orders. The WIP level in the CONWIP system is then set at a value that is usually the smallest integer greater than the average WIP in the kanban system. In several cases, the CONWIP level was set to a value closer to the kanban system (e.g., for the kanban system having an average WIP level of 12.4, the CONWIP system used 12 cards). In all but two cases, the CONWIP system exhibited customer service that was statistically superior. In the other two cases, there was no significant difference. There are, also, more practical issues to consider. If the reason for the success of pull systems is that they limit WIPgrowth,then clearly both CONWIP and kanban accomplish this. Since it seems that the implementation of CONWIP is simpler than that of kanban (with one card count to set rather than N), then CONWIP may be preferable.Also, as discussed earlier,CONWIP appearsto be less demanding on the manufacturingenvironment than kanban, and hence, allows the benefits of JIT to be realized in more production systems. Finally, since CONWIP lines can be split into as many segments as needed, kanban is a subset of CONWIP and therefore cannot be superior to CONWIP. Thus, an important question is, why has kanban been used in the first place? To answer this, several authors cite the need for in the production environment and the interlearning actions of kanban facilitate this (see e.g., Suri and de Treville 1986). Another answer may be found in the teachings of Taiichi Ohno. Although his book gives few details, there is a figurethat looks remarkably similar to several CONWIP lines feeding an assembly line. There is no indication of pulling between stations on the lines and the backlog is obtained by looking up the assembly line to see what models are next in sequence. His own remarks yield another insight, as indicated by Myers (1990): the systemso powerful, Toyotaconsidered Ono [sic]claims, that the company deliberately coined difficultand even misleading wordsto describe it. 'If in the beginning the U.S. hadunderstood whatToyotawasdoing.'Onosays,'it would havebeenno good for us.' Could some of these words be "stocklessproduction," "zero inventories," or even "just-in-time"?Perhaps kanban was never designed to be used on short serial production lines, but instead between process centers composed of several stations and separatedby either distance or process differences. 4. CONCLUSIONS Our results can be summarized as: 1. Customer service in both kanban and CONWIP production systems is improved by: a. faster machines; b. extra WIP; c. less variable processingtimes. 2. CONWIP systems have better customer service than do pure kanban systems. While some of these results are ratherintuitive, the analysis requires very few assumptions. The only requirement for the processing times is that they are iid, while the demand process can be any point process such that the limits in (2) exist. The weakness of these assumptions strongly suggests that the above conclusions are true under a wide variety of situations. Finally, our results suggest that the use of pure kanban may not be justified in some cases. CONWIP solves many practical implementation problems of kanban and appears to offer better service with no increase in WIP. For these reasons, practitioners should consider CONWIP as a possible alternativeto kanban. APPENDIX Simulation of Kanban and CONWIPSystems For our comparisons, we consider a three-station system, where each station has iid process times and for which the demand is a Poisson process. Machine zero is the "pick" machine and has zero process time (so our measure is the average time between the arrival of a customer and the time the demand is filled). The other two machines have processing times equal to either unity or one-half. Machines one and two (numbering from right to left as in the paper) have process times of either unity or one-half. We test cases involving three factors: utilization, line balance, and variabilityof processtimes. For each factor, we establish two levels yielding a total of eight cases. We consider 75% and 90% utilization (interarrival times of 1.1111 and 1.3333 time units), coefficients of variation of one and one-half, and identical machines versus one machine that is twice the speed of the other. For the coefficient of variation equal to one, the service times are exponential; for the smaller coefficient of variation, the service times are Weibull

CustomerService in Pull Systems / Table I

957

Results of SimulationsComparingKanbanand CONWIP


Kanban Time Between Arrival 1.1111 1.1111 1.1111 1.1111 1.3333 1.3333 1.3333 1.3333 Wait Time Coef. Var. 1.0 1.0 0.5 0.5 1.0 1.0 0.5 0.5 Line Bal UnBal Bal UnBal Bal UnBal Bal UnBal WIP Avg. 9.9 12.4 15.2 14.7 16.0 16.5 17.6 17.5 Mean 52.79 3.198 2.411 0.9848 0.6053 0.1470 0.0554 0.0103 Std. Dev. 36.9 2.17 1.66 0.773 0.429 0.121 0.0419 0.0137 WIP Avg. 10 12 15 15 16 17 18 18 CONWIP Wait Time Mean 55.82 3.331 1.892 0.6500 0.2788 0.0406 0.0028 0.00016 Std. Dev. 36.4 2.10 1.40 0.629 0.239 0.0527 0.0063 0.00075 t Statistic -0.702 -0.529 2.867 4.033 7.983 9.674 14.897 8.868

with a shape parameter of 2.1013 and a location parameter of 1.1291 (for a mean of 1.0) and 0.5645 (for a mean of one-half). In each case, there are 18 kanbans. For the balanced kanban systems, these are allocated evenly between the two stations. For the unbalanced cases, 12 kanbans are placed with the slower machine while six are with the faster. These allocations are consistent with that advocated in Hall, pp. 235-236. At the start of the simulation all 18 kanbans are attached to standard containers so that there are no free kanbans in the system. Work begins immediately at all stations upon realization of the first demand. Work is considered "in process" immediately upon receipt of a kanban at station two. One hundred forty-fourobservations of the statistic were recorded for a total of 576,000 customers. U4,ooo The large batch size was needed because of the severe correlation of the observations. The results of the simulations are shown in Table I. ACKNOWLEDGMENT The author thanks Professors W. J. Hopp and M. A. Zazanis for their thoughtful comments regardingthis work. He is also grateful for the assistance provided by Mr. David Schwartzbardin performing the simulation runs. This work has been supported, in part, by grant DDM-8905638 from the National Science Foundation and by a grant from the International Business Machines Corporation. REFERENCES
ADAN, I., AND

BUZACOTT,

J. A. 1988. Kanban and MRP Controlled

Production Systems. Technical Report, Department of Management Sciences, University of Waterloo, Waterloo, Ontario, Canada. DELEERSNYDER, J. L., T. J. HODGSON,H. MULLER (MALEK) AND P. J. O'GRADY. 1989. Kanban Controlled Pull Systems: An Analytic Approach. Mgmt. Sci. 35, 1079-1091. DENARDO, E. V., AND C. S. TANG. 1992. LinearControl of a Markov Production System. Opns. Res. 40, 259-278. GONG,L., AND H. MATSUO.1990. Stabilizing Work-inProcess and Smoothing Production in a Production System With Random Yield. Working Paper, The Graduate School of Business, The University of Texas at Austin. GROENEVELT, H., AND U. S. KARMARKAR. 1988. A Dynamic Kanban System Case Study. Prod. Inv. Mgmt. J. 29, 46-5 1. HALL, R. W. 1983. Zero Inventories. Dow Jones-Irwin, Homewood, Ill. Hopp, W. J., AND J. T. SIMON.1989. Bounds and Heuristics for Assembly-Like Queues. Queue. Syst. 4, 137-156.
KARMARKAR,

U. S. 1986a. Kanban Systems. Working

J. VAN DER WAL. 1989. Monotonicityof the Throughputof a Closed Queueing Network in the Numberof Jobs. Opns.Res. 37, 953-957. BITRAN, G. R., AND L. CHANG. 1987. A Mathematical Programming Approachto a DeterministicKanban System.Mgmt. Sci. 33, 427-441.

Paper Series No. QM8612, Center for Manufacturing and Operations Management, The Graduate School of Management, University of Rochester, Rochester, New York. U. S. 1986b. Push, Pull and Hybrid ConKARMARKAR, trol Schemes. Working Paper Series No. QM8614, Graduate School of Management, University of Rochester, Rochester, New York. KIMBALL, G. E. 1988. General Principles of Inventory Control. J. Manuf Opns. Mgmt. 1, 119-130. KIMURA, O., ANDH. TERADA. 1981. Design and Analysis of Pull Systems, a Method of Multistage Production Control. Int. J. Prod. Res. 19, 241-253, KRAJEWSKI, L. J., B. E. KING, L. P. RITZMAN, AND D. S. WONG. 1987. Kanban, MRP, and Shaping the Manufacturing Environment. Mgmt. Sci. 33, 39-57. MONDEN, Y. 1983. Toyota Production System: Practical

958

SPEARMAN SHANTHIKUMAR, J. G., AND D.

Approach to Management. Industrial Engineering and Management Press, Norcross, Ga. MYERS, F. S. 1990. Japan's Henry Ford. Scient. Am. 262,

P. 98. OHNO,T. 1988. Toyota Production System: Beyond


Large Scale Production. Productivity Press, Cambridge, Mass. PHILIPOOM, P. R., L. P. REES,B. W. TAYLOR III AND P. Y. HUANG. 1987. An Investigation of the Factors Influencing the Number of Kanbans Required in the Implementation of the JIT Technique With Kanbans. Int. J. Prod. Res. 25, 457-472. Ross, S. 1983. Stochastic Processes. John Wiley, New York.

SCHONBERGER, R. J. 1986. WorldClass Manufacturing: The Lessons of SimplicityApplied.Free Press,New


York. SHANTHIKUMAR, J. G., ANDD. D. YAO. 1987. Stochastic Monotonicity of the Queue Lengths in Closed Queueing Networks. Opns. Res. 35, 583-588. SHANTHIKUMAR, J. G., AND D. D. YAO. 1988a. SecondOrder Properties of the Throughput of a Closed Queueing Network. Math OR 13, 524-534. SHANTHIKUMAR, J. G., ANDD. D. YAO. 1988b. Throughput Bounds for Closed Queueing Networks With Queue-Dependent Service Rates. Perf Eval. 9, 69-78. SHANTHIKUMAR, J. G., AND D. D. YAO. 1989a. Monotonicity and Concavity Properties in Cyclic Queueing Networks With Finite Buffers. In Queueing Networks With Blocking, H. G. Perros and T. Altiok (eds.). North-Holland, Amsterdam, 325-344.

D. YAO. 1989b. SecondOrder Stochastic Propertiesin Queueing Systems. Proc.IEEE 77, 162-170. SIMPSON,K. F. 1958. In-Process Inventories.Opns.Res. 6, 863-873. SPEARMAN, M. L., W. J. HoPP AND D. L. WOODRUFF. 1989. A Hierarchical Control Architecture for CONWIPProductionSystems.J. Mantufand Opns. Mgmt. 2, 147-171. SPEARMAN, M. L., D. L. WOODRUFF AND W. J. Hopp. 1990. CONWIP:A Pull Alternative to Kanban.Int. J. Prod.Res. 28, 879-894. SPEARMAN,M. L., AND M. A. ZAZANIS. 1992. Push and Pull ProductionSystems:Issues and Comparisons. Opns.Res. 40, 521-532. STOYAN, D. 1983. Comparison Methodsfor Queuesand Other Stochastic Models. John Wiley, Chichester, U.K. SUGIMORI,Y., K. KUSUNOKI, F. CHOAND S. UCHIKAWA. 1977. Toyota ProductionSystem and KanbanSystem: Materialization of Just-In-Timeand RespectFor-HumanSystem.Int. J. Prod.Res. 15, 553-564. SURI, R., AND S. DE TREVILLE.1986. GettingFrom 'Justin-Case'to 'Just-in-Time': InsightsFrom a Simple Model.J. Opns.Mgmt. 4, 295-304. WHITT, W. 1984. Open and ClosedModelsfor Networks of Queues.AT&TBell Lab. Tech.J. 63, 1911-1978. WIGHT, 0. 1970. Input/OutputControl a Real Handle on Lead Time. Prod.Inv. Mgmt. J. 11, 9-3 1. WOODRUFF, D. L., AND M. L. SPEARMAN. 1992. Bottleneck Sequencingfor Two Classesof Job With Deadlines and Setup Times. Prod. Opns. Mgmt. 1, 87-102.

Вам также может понравиться