response time reduction in make-to-order and assemble-to-order supply chain design
DESCRIPTION
management productionTRANSCRIPT
-
IIE Transactions (2009) 41, 448466Copyright C IIEISSN: 0740-817X print / 1545-8830 onlineDOI: 10.1080/07408170802382741
Response time reduction in make-to-orderand assemble-to-order supply chain design
NAVNEET VIDYARTHI1,, SAMIR ELHEDHLI2 and ELIZABETH JEWKES2
1Department of Decision Sciences and MIS, John Molson School of Business, Concordia University, Montreal, Quebec, Canada,H3G 1M8E-mail: [email protected] of Management Sciences, University of Waterloo, 200 University Avenue West, Waterloo, Ontario, Canada, N2L 3G1,E-mail: {elhedhli, emjewkes}@uwaterloo.ca
Received December 2006 and accepted June 2008.
Make-to-order and assemble-to-order systems are successful business strategies in managing responsive supply chains, characterizedby high product variety, highly variable customer demand and short product life cycles. These systems usually spell long customerresponse times due to congestion. Motivated by the strategic importance of response time reduction, this paper presents modelsfor designing make-to-order and assemble-to-order supply chains under Poisson customer demand arrivals and general service timedistributions. The make-to-order supply chain design model seeks to simultaneously determine the location and the capacity ofdistribution centers (DCs) and allocate stochastic customer demand to DCs by minimizing response time in addition to the xedcost of opening DCs and equipping them with sufcient assembly capacity and the variable cost of serving customers. The problemis setup as a network of spatially distributed M/G/1 queues, modeled as a non-linear mixed-integer program, and linearized usinga simple transformation and a piecewise linear approximation. An exact solution approach is presented that is based on the cuttingplane method. Then, the problem of designing a two-echelon assemble-to-order supply chain comprising of plants and DCs servinga set of customers is considered. A Lagrangean heuristic is proposed that exploits the echelon structure of the problem and uses thesolution methodology for the make-to-order problem. Computational results and managerial insights are provided. It is empiricallyshown that substantial reduction in response times can be achieved with minimal increase in total costs in the design of responsivesupply chains. Furthermore, a supply chain conguration that considers congestion is proposed and its effect on the response timecan be very different from the traditional conguration that ignores congestion.
Keywords: Response time, make to order, assemble to order, supply chain design, cutting plane method, Lagrangean relaxation
1. Introduction
Make-to-Order (MTO) systems are successful businessstrategies to manage responsive supply chains that are char-acterized by high product variety, highly variable customerdemand and short product life cycles. Because of mass cus-tomization and competition on product variety, many rmsadopt an MTO strategy to offer a variety of products anddeal with product proliferation. Dells manufacturing anddistribution of Personal Computers (PCs) is an excellentexample of an MTO supply chain (Margretta, 1998; Dell,2000). Dell typically offers several lines of product, witheach allowing at least dozens of features from which cus-tomers can select when placing an orderdifferent combi-nations of CPU, hard drive, memory and other peripherals.In Dells supply chain, multiple components are procured
Corresponding author
and kept in inventory at various assembly facilities, fromwhich they are assembled into a wide variety of nishedproducts in response to customer orders. Whereas each ofthese components takes a substantial lead time to man-ufacture, the time to assemble all these components intoa PC is low, provided there is sufcient assembly capac-ity and the components are available. In traditional Make-To-Stock (MTS) supply chains, the customer orders aremet from stocks of an inventory of nished products thatare kept at various points of the network. This is doneto reduce the delay in fullling customer orders, increasesales and avoid stockouts. However, the problems asso-ciated with holding inventory of nished products mayoutweigh the benets, especially when those products be-come obsolete as technology advances or fashion changes.While an MTO strategy eliminates nished goods inven-tories and reduces a rms exposure to the risk of ob-solescence, it usually spells long customer response time(Gupta and Benjaafar, 2004).
0740-817X C 2009 IIE
-
Response time reduction in supply chain design 449
In order to reconcile the dual needs of a quick responsetime and high product variety, many rms such as GeneralElectric, American Standard, Compaq, IBM, BMW andNational Bicycle use a hybrid strategy (i.e., mix of MTOand MTS) called the Assemble-To-Order (ATO) strategy,in which a subassembly, or a number of common subassem-blies used in several products, are assembled and placed ininventory until an order is received for the nished prod-uct (Song and Zipkin, 2003). This allows the rm to cus-tomize the orders by having the product ready using theMTO strategy, while taking advantage of the economiesof scale using the MTS strategy. Also, investment in thesemi-nished product inventory is smaller compared to theoption of maintaining a similar amount of nished goodsinventory. Furthermore, demand pooling benets can berealized. Although, maintaining a semi-nished productinventory in ATO systems lowers the customer responsetime as compared to a pure MTO system, it can be furtherreduced by minimizing congestion at the point of differen-tiation. Naturally, the response time to deliver the productis critical and forms the basis for competition. Consumerswillingness to pay a premium for a shorter response timeprovides further incentives for rms to reduce response timein MTO and ATO supply chains.
Although various integrated models of supply chain de-sign have been proposed in recent years to support leadtime reduction, these models have continued to be largelyguided by more traditional concerns of efciency and cost inMTS settings, where the primary focus is on minimizing thexed cost of facility location and the variable transporta-tion cost under fairly stable and deterministic customer de-mand settings. This approach is personied by the work ofDogan and Goetschalckx (1999), Vidal and Goetschalckx(2000), Teo and Shu (2004), Shen (2005), Eskigun et al.(2005) and Elhedhli and Gzara (2008). For example, Vidaland Goetschalckx (2000) present a model that captures theeffect of change in transportation lead time and demand onthe optimal conguration of the global supply chain net-work, assuming that the demand is deterministic. Eskigunet al. (2005) incorporate delivery lead time and the choice oftransportation mode in the design of a supply chain under adeterministic demand setting. These models tend to ignorecongestion at the facilities and its effect on response time.Their solutions prescribe locating facilities whose capacityutilization is very high, resulting in an excessively long re-sponse time when subjected to variability in service timesand randomness in customer orders. Reviews by Vidal andGoetschalckx (1997), Erenguc et al. (1999) and Sarmientoand Nagi (1999) also point out that most of the existingsupply chain design models do not consider measures ofcustomer service such as response time in making loca-tion/allocation decisions. Also, refer to the recent reviewby Klose and Drexl (2005). This is not surprising given thecomplexity of the model and the interplay of locational andqueueing aspects of the problem. To the best of our knowl-edge, Huang et al. (2005) is one of the rst to model the effect
of congestion in the design of distribution networks. Theymodel capacity using the mean and variance of the Dis-tribution Centers (DCs) as continuous variables, whereasour model considers capacity as a set of discrete optionswith known means and variances. They propose solutionprocedures based on outer approximation and Lagrangeanrelaxation, and tested on small instances of the problem.
Another growing body of literature that is related to ourwork and accounts for congestion and its effect on responsetime in strategic planning is models for facility locationwith immobile servers, stochastic demand and congestion(such as location of emergency medical facilities, re sta-tions, telecommunication network design, automated tellermachines or internet mirror site location). For an exten-sive review, refer to Berman and Krass (2002). Due to thecomplexity of the underlying problem, most papers in thisarea make very strong assumptions: (i) either the numberor capacity of the facilities (or both) are assumed to bexed; (ii) the facilities are assumed to be identical; (iii) thedemand arrival process is assumed to be Poisson; and (iv)the service process is usually assumed to be exponential(see, Amiri (1997), Marianov and Serra (2002), Wang et al.(2003) and Elhedhli (2006) and references therein). Despitethat, most of the techniques proposed to date to solve theseproblems, with the exception of Elhedhli (2006), are eitherapproximate or heuristic based. Our work is also similarin spirit to models for capacity planning with congestioneffects, for which only heuristic solution procedures havebeen reported; see Rajagopalan and Yu (2001) and refer-ences therein.
The objective of this paper is to model the effect ofcongestion on the response time and analyze the trade-off among response time costs, facility location and capac-ity acquisition costs, and outbound transportation costs inthe design of supply chain networks. More specically, wepresent a model to determine the conguration of an MTOsupply chain, where the emphasis is on minimizing the cus-tomer response time through the acquisition of sufcientassembly capacity and the optimal allocation of workloadto the assembly facilities (DCs) under stochastic customerdemand settings. The DCs are modeled as spatially dis-tributed queues with Poission arrivals and general servicetimes to capture the dynamics of the response time. Themodel is formulated as a non-linear Mixed-Integer Pro-gramming (MIP) problem and is linearized using piecewiselinear functions. We present a cutting plane algorithm thatprovides the optimal solution to the problem. Furthermore,we present a Lagrangean relaxation heuristic procedurefor solving large-scale instances of such integrated models.Then, we present a model for the two-echelon ATO sup-ply chain design problem, where a set of plants and DCsare to be established to distribute various nished prod-ucts to a set of customers with stochastic demand. DCs actas assembly facilities, where semi-nished products, pro-cured from plants are held in inventories, from which theyare assembled into a wide variety of nished products in
-
450 Vidyarthi et al.
response to customer demands. We propose a Lagrangeanrelaxation heuristic that exploits the echelon structure ofthe problem and uses the solution methodology proposedabove for the MTO problem. Explicit consideration of con-gestion effects and their impact on response time in makinglocation, capacity and allocation decisions in supply chainsdistinguishes this work from most other supply chain designmodels.
The rest of the paper is organized as follows. Section 2provides a non-linear MIP formulation of the MTO sup-ply chain design problem, a piecewise linearization andan exact solution approach based on the cutting planemethod. The simplications resulting from assuming ex-ponentially distributed service times (M/M/1 case) anddeterministic service times (M/D/1 case) are also explic-itly described. In Section 3, we present the formulation ofthe two-echelon ATO supply chain design problem and aLagrangean heuristic. Computational results and manage-rial insights are reported in Section 4. Finally, Section 5concludes with some directions for future research.
2. MTO supply chain design
Consider the problem of designing an MTO supply chain,where a set of DCs are to be established and equipped withsufcient capacity to serve a set of customers. Sufcientcapacity here implies being able to obtain service withoutwaiting for an excessively long time after the order is placed.The DCs maintain inventory of multiple components andfacilitate the assembly and shipment of a wide variety of n-ished products in a timely fashion without carrying expen-sive nished-goods inventory and incurring a long responsetime. Response time refers to the interval between the plac-ing of an order and receipt of the ordered product. In MTOsupply chains, because a customer order triggers the assem-bly of nished product from components, the response timeconsists of the assembly lead time and the delivery lead time.The delivery time between individual DCs and customersis relatively constant compared to the order fullment timeat DCs in such settings. Moreover, it can further be reduced(using alternative transportation modes or expedited deliv-ery services) to respond quickly to customer orders on ashort-term basis. However, the assembly lead time is highlydependent on the DC capacity and the allocated workloadand is difcult to change (on a short-term basis) once theDC is established.
2.1. Model formulation
We consider the setting depicted in Fig. 1. We assume thatthe demand for each product from each customer is inde-pendent and occurs according to a Poisson process. Oncethe demand for a product is realized at the customers end,the order is placed at the DCs. DCs will act as assembly facil-
Fig. 1. An MTO supply chain network.
ities and the customers orders arriving at the DCs are meton a First-Come First-Serve (FCFS) basis. We assume thateach DC operates as a single exible-capacity server withinnite buffers to accommodate customer orders waitingfor service.
Under these assumptions, the MTO supply chain is mod-eled as a network of independent M/G/1 queues in whichthe DCs are treated as servers with service rates propor-tional to their capacity levels, where the capacity levels arediscrete. We also assume that there is an unlimited supply ofcomponents and their inventory holding costs at the DCsare insignicant. Hence, the model formulated below si-multaneously determines the location and capacity of DCsand the assignment of customer to DCs by minimizing theresponse time costs in addition to the xed location andcapacity acquisition costs, the assembly and transportationcosts from DCs to customers. Besides capacity restrictions(steady-state conditions) at the DCs, and the demand re-quirements, there are constraints which ensure that at mostone capacity level is selected at the DCs. To model thisproblem, we dene the following notation.
Indices and parameters:
i = index for customers, i = 1, 2, . . . , I ;j = index for potential DCs, j = 1, 2, . . . , J;k = index for potential capacity level at DCs, k =
1, 2, . . . , K;fjk = xed cost of opening DC j and acquiring capacity
level k ($/period);cij = unit cost of serving customer i from DC j ($/unit);t = mean response time cost per unit time per customer
($/period/customer);i = mean demand rate for the product from customer
i (units/period);jk = mean service rate at DC j, if it is allocated capacity
level k (units/period); 2jk = variance of service times at DC j, if it is allocated
capacity level k.
-
Response time reduction in supply chain design 451
Decision variables:xij = fraction of customer is demand served by DC j (0
xij 1);
yjk =1, if DC j is opened and capacity level kis acquired
0, otherwise.
Let the demand for the product at customer location ibe an independent random variable that follows a Poissonprocess with mean i. If xij is the fraction of customer isdemand served by DC j, then the aggregate demand ar-rival rate at DC j is also a random variable that follows aPoisson process with mean j =
Ii=1 ixij, due to the su-
perposition of Poisson processes. If the service times at eachDC follow a general distribution and each DC is modeledas an M/G/1 queue, then the mean service rate of DC j, ifit is allocated capacity level k, is given by j =
Kk=1 jkyjk
and the variance in service times is 2j =K
k=1 2jkyjk. This
service rate reects the server capacity or essentially thenumber of MTO products a DC can assemble and ship in agiven time period. Let j represent the mean service time atDC j (j = 1/j), j be the utilization of DC j (j = j/j)and CV2j be the squared coefcient of variation of ser-vice times (CV2j = 2j / 2j ). Under steady-state conditions(j < j) and FCFS queuing discipline, the expected aver-age waiting time (including the service time) at DC j is givenby the PollaczekKhintchine (PK) formula:
E[Wj(M/G/1)] =(
1 + CV2j2
)jj
1 j + j
=(
1 + CV2j2
)j
j(j j) +1j
j,and the expected total waiting time for DC j is obtainedby multiplying the waiting time at DC j by the expecteddemand as(
1 + CV2j2
)2j
j(j j) +j
jj.
The expected total waiting time for the entire system isgiven by
E[W (M/G/1)] =J
j=1
[(1 + CV2j
2
)2j
j(j j) +j
j
],
and can be written as:
12
Jj=1
[(1 + CV2j
) jj j +
(1 CV2j
) jj
].
This is equivalent to
12
Jj=1
[(1 +
Kk=1
CV2jkyjk
) Ii=1 ixijK
k=1 jkyjk I
i=1 ixij
+(
1 K
k=1CV2jkyjk
) Ii=1 ixijK
k=1 jkyjk
]. (1)
The resulting non-linear MIP formulation is
(PN) : minJ
j=1
Kk=1
fjkyjk+I
i=1
Jj=1
cijixij+tE[W (M/G/1)],(2)
subject to
Ii=1
ixij K
k=1jkyjk j, (3)
Kk=1
yjk 1 j, (4)J
j=1xij = 1 i, (5)
0 xij 1, yjk {0, 1} i, j, k. (6)The rst term in the objective function (2) represents the
xed cost (amortized over the planning period) of locatingDCs and equipping them with adequate assembly capacity.The second term accounts for the variable cost of assem-bly and shipment of products from DCs to customers. Thethird term is the expected total response time cost or the lostsales due to excessive response times between DCs and cus-tomers. The expected total response time cost is expressedas a product of average response time cost per unit time thatone of its customers spends in the system and expected totalwaiting time in the system. In this paper, for simplicity andtractability reasons, we assume that the cost of the responsetime is linearly proportional to the waiting time. However,the model can be extended to other cost functions such asa piecewise linear function or a cost function proportionalto max{0, E[W (M/G/1)] W0}, where W0 is the maximumtolerated waiting time for a customer. Moreover, the averageresponse time cost t may vary from customer to customer(tij) if desired, but we assume for simplicity that it is thesame across customers. It can be interpreted as a penaltyfunction that reects the true cost of not fullling customerorders in the committed lead time. In practice, determiningthe values of average response time cost can be challenging,however, one can rely on techniques outlined in Rao et al.(2000). To accurately reect lost sales due to unacceptableresponse time, Rao et al. (2000) surveyed Caterpillar dealersto determine the percent of customers who would renege ifa product was not available immediately, after 2 weeks, andafter 4 weeks. A lower bound on the response time cost canbe provided by the inventory holding costs (Eskigun et al.,2005). Furthermore, the problems can be solved iterativelywith different values of t to obtain a trade-off curve fromwhich decision makers may choose a solution based on theirpreference between location and capacity acquisition cost,transportation cost, and response time costs.
Constraints (3) ensure that the steady-state conditions(j j) at the DCs are met. Constraint set (4) ensuresthat at most one capacity level is selected at a DC, whereasconstraint set (5) ensures that the total demand is met.
-
452 Vidyarthi et al.
Constraints (6) are non-negativity and binary restrictions.The formulation can easily handle single-sourcing require-ments by imposing binary restrictions on xij. This wouldrestrict the assignment of customer is demand to one andonly one DC j.
The non-linearity in (PN) arises due to the expressionof the total waiting time at the DCs, E[W (M/G/1)]. Itcan be shown that E[W (M/G/1)] is convex in aggregatearrival rate j, for a xed value of j and convex in ser-vice rate j, for a xed value of j, where j =
Ii=1 ixij
and j =K
k=1 jkyjk. Intuitively, one would expect thatthe waiting time increases with increasing marginal returnsas the arrival rate increases and decreases with decreasingmarginal returns as the service rate increases. In the nextsection, we deal with the non-linearity due to the expres-sion of the total average waiting time using a linearizationbased on a simple transformation and a piecewise linear ap-proximation. We also present an exact solution procedurebased on the cutting plane algorithm. Our cutting planealgorithm is similar to the outer-approximation algorithm(Duran and Grossman, 1986).
2.2. Linearization and cutting plane method
In order to linearize Equation (1), let us dene non-negativeauxiliary variables Rj, such that:
Rj = jj j =
Ii=1 ixijK
k=1 jkyjk I
i=1 ixijj
which implies thatI
i=1ixij = Rj1 + Rj
Kk=1
jkyjk
= jK
k=1jkyjk =
Kk=1
jkzjk (7)
where
zjk ={
0 if yjk = 0j if yjk = 1 j, k (8)
and j = Rj/(1 + Rj) is the server (DC) utilization.Since there is at most one k with yjk = 1 while yjk = 0 for
all other k = k, the expression zjk = jyjk can be ensuredby adding the following constraints:
zjk yjk j, k,K
k=1zjk = j j.
The function j = Rj/(1 + Rj) is concave. Given a set ofpoints Rhj indexed by H, j can be approximated by aninnite set of piecewise linear functions that are tangent toj at points Rhj (as shown in Fig. 2) that is
j = minhH
{1(
1 + Rhj)2 Rj +
(Rhj
)2(1 + Rhj
)2}
j
or
j 1(1 + Rhj
)2 Rj +(Rhj
)2(1 + Rhj
)2 j, h H.
Fig. 2. A piecewise linear approximation of Rj/(1 + Rj).
-
Response time reduction in supply chain design 453
The expression for E[W (M/G/1)] reduces to
E[W (M/G/1)] = 12
Jj=1
{(1 +
Kk=1
CV2jkyjk
)Rj
+(
1 K
k=1CV2jkyjk
)j
}
= 12
Jj=1
(Rj +
Kk=1
CV2jkwjk + j K
k=1CV2jkzjk
)
where
wjk ={
0 if yjk = 0Rj if yjk = 1 and zjk =
{0 if yjk = 0j if yjk = 1 j, k.
Similarly, because there exists at most one k with yjk = 1while yjk = 0 for all other k = k, the expression wjk = Rjyjkcan be ensured by adding the following constraints:
wjk Myjk j, k,K
k=1wjk = Rj j,
where M is the usual Big-M.The resulting linear MIP formulation is
(PL(H)) : minJ
j=1
Kk=1
fjkyjk +I
i=1
Jj=1
cijixij
+ t2
Jj=1
{Rj + j +
Kk=1
CV2jk(wjk zjk)}
, (9)
subject to
Ii=1
ixij K
k=1jkzjk = 0 j, (10)
Kk=1
yjk 1 j, (11)J
j=1xij = 1 i, (12)
zjk yjk 0 j, k, (13)
j 1(1 + Rhj
)2 Rj (Rhj
)2(1 + Rhj
)2 j, h H, (14)j
Kk=1
zjk = 0 j, (15)
wjk Myjk 0 j, k, (16)K
k=1wjk Rj = 0 j, (17)
yjk {0, 1}; 0 xij, zjk 1; j, Rj, wjk 0; i, j, k.(18)
The steady-state conditions (j < j) translate into ca-pacity constraints, and are enforced by the constraints (10)and (13) and forced to
-
454 Vidyarthi et al.
and varying batch sizes. For exponentially distributed ser-vice times at the DCs, the total expected waiting time forthe entire system is given by
E[W (M/M/1)] =J
j=1
j
j j
=J
j=1
( Ii=1 ixijK
k=1 jkyjk I
i=1 ixij
)=
Jj=1
Rj.
The resulting linear MIP model is
(PM/M/1L(H)
): min
Jj=1
Kk=1
fjkyjk+I
i=1
Jj=1
cijixij + tJ
j=1Rj
(19)
subject to Equations (10) to (15),
yjk {0, 1}; 0 xij, zjk 1; j, Rj 0 i, j, k.The model is structurally identical to the service sys-
tem design model presented in Amiri (1997) and Elhedhli(2006).
2.3.2. Systems with deterministic service times(M/D/1 case)
In many cases, the processing/assembly of nished prod-ucts at the DCs often involves repeated steps without muchvariation (Kim and Tang, 1997). This is particularly truefor MTO products with limited options and batch size ofone such as Dells PCs. For deterministic service times, theexpected waiting time for the entire system is given by
E[W (M/D/1)] = 12
Jj=1
(j
j j +j
j
)= 1
2
Jj=1
(Rj + j).
The resulting linear MIP model is as follows:
(PM/D/1L(H)
): min
Jj=1
Kk=1
fjkyjk
+I
i=1
Jj=1
cijixij + t2J
j=1(Rj + j), (20)
subject to Equations (10) to (15),
yjk {0, 1}; 0 xij, zjk 1; j, Rj 0 i, j, k.The cutting plane algorithm described above can be used tosolve these models to optimality. Alternatively, one can relyon the Lagrangean heuristics. Some computational resultsare provided in Section 4.
3. ATO supply chain design
In this section, we consider the design of an ATO supplychain, where we seek to locate a set of plants and DCs todistribute a product with a non-trivial bill of materials to
Fig. 3. An ATO supply chain network.
a set of customers with stochastic demand. The DCs willact as intermediate facilities between the plants and thecustomers and facilitate the shipment of products betweenthe two echelons, as shown in Fig. 3.
The semi-nished products are produced at the plantsand shipped to the DCs. Once the demand is realized at thecustomers end, the order is placed to the DC and the nalproduct is assembled and the demand is met. Hence, thesupply chain network is a combination of the MTS echelon(plant-DC echelon) and the MTO echelon (DC-customerechelon). The problem environment is characterized by astochastic customer demand that has to be satised froma set of DCs and where sufcient capacity has to be ac-quired in order to avoid long response times. To model thisproblem, we dene the following additional notation.
m = index for potential plants, m = 1, 2, . . . , M;n = index for semi-nished products, n = 1, 2, . . . , N;gm = xed cost of opening a plant at location m ($/pe-
riod);Pm = maximum available capacity of plant m (units);cjmn = unit production and transportation cost for semi-
nished product n from plant m to DC j;n = number of units of semi-nished product n required
to make one unit of nished product;um = decision variable that equals one, if plant m is
opened; zero, otherwise;vjmn = number of units of semi-nished product n pro-
duced from plant m and shipped to DC j.
Under the assumption that the demand at customer i isan independent random variable that follows a Poisson pro-cess with mean i and the service time at each DC followsa general distribution, each DC is modeled as an M/G/1queue, whose mean service rate, if it is allocated capacitylevel k, is given by j =
Kk=1 jkyjk and the variance in
service times is given by 2j =K
k=1 2jkyjk. Under steady-
state conditions (j < j) and FCFS queuing discipline, the
-
Response time reduction in supply chain design 455
total average waiting time for the entire system (service plusqueuing time) is given by Equation (1). The resulting non-linear MIP formulation that simultaneously determines thelocation and capacity of plants and DCs, the shipment lev-els from plants to DCs, and allocation of customers to DCsby minimizing response time costs in addition to xed facil-ity location and capacity acquisition costs, production andtransportation costs between echelons is as follows:
(PATO) : minM
m=1gmum +
Jj=1
Kk=1
fjkyjk +J
j=1
Mm=1
Nn=1
cjmnvjmn
+I
i=1
Jj=1
cijixij + t2J
j=1
(1 +
Kk=1
CV2jkyjk
)
I
i=1 ixijKk=1 jkyjk
Ii=1 ixij
+ t2
Jj=1
(1
Kk=1
CV2jkyjk
) Ii=1 ixijK
k=1 jkyjk, (21)
subject to
Jj=1
Nn=1
vjmn Pmum m, (22)
Mm=1
vjmn =I
i=1nixij j, n, (23)
Ii=1
ixij K
k=1jkyjk j, (24)
Kk=1
yjk 1 j, (25)J
j=1xij = 1 i, (26)
0 xij 1, yjk, um {0, 1}, vjmn 0i, j, k, m, n. (27)
The objective function (21) consists of the xed cost ofopening plants, the xed cost of locating DCs and equip-ping DCs with the required capacity level, the variable costof producing and procuring semi-nished product, the vari-able cost of serving customers from DCs and the total wait-ing time costs at the DCs. Constraints (22) are capacityrestrictions on the opened plants and permit the use ofopened plants only. Constraints (23) are commodity owconservation equations at the DCs. Constraints (24) ensurethat the steady-state conditions at the DCs are met. Con-straints (25) ensure that at most one capacity level is selectedat a DC whereas constraints (26) ensure that the total de-mand is met. Constraints (27) are non-negativity and binaryconstraints.
3.1. Lagrangean relaxation
There are number of ways in which the model can be re-laxed in Lagrangean fashion (see Klose and Drexl (2005)and reference therein). In this paper, we exploit the eche-lon structure of the ATO supply chain using Lagrangeanrelaxation to decompose the model into two subproblems.Note that in (PATO), constraints (22) relate to the MTSechelon, and constraints (24) to (26) relate to the MTOechelon, whereas constraints (23) are the ow conservationconstraints that link the two echelons. Upon relaxing theow conservation constraints (23) with dual multipliers jn,the problem decomposes into two subproblems:
(SPMTS) : minM
m=1gmum +
Jj=1
Mm=1
Nn=1
(cjmn jn)vjmn, (28)
subject to
Jj=1
Nn=1
vjmn Pmum m, (29)
um {0, 1}, vjmn 0 j, m, n. (30)
(SPMTO) : minJ
j=1
Kk=1
fjkyjk +I
i=1
Jj=1
Nn=1
(cij + njn)ixij
+ t2
Jj=1
(1 +
Kk=1
CV2jkyjk
) Ii=1 ixijK
k=1 jkyjk I
i=1 ixij
+ t2
Jj=1
(1
Kk=1
CV2jkyjk
) Ii=1 ixijK
k=1 jkyjk, (31)
subject to
Ii=1
ixij K
k=1jkyjk, (32)
Jj=1
xij = 1 i, (33)
Kk=1
yjk 1 j, (34)
0 xij 1, yjk {0, 1} i, j, k. (35)
Subproblem (SPMTS) is a linear MIP model that deter-mines the location of plants and the ow of semi-nishedproducts into the DCs, whereas the subproblem (SPMTO) isa non-linear MIP model that provides the location and ca-pacity level of DCs and the allocation of customers to DCs.Note that the subproblem (SPMTO) is the MTO supply chaindesign model presented in the previous section and hencewe use the proposed cutting plane algorithm to solve it.
From model (PATO), we can derive some valid con-straints. For example, consider the following set of
-
456 Vidyarthi et al.
constraints:
Mm=1
Pmum I
i=1i, (36)
Mm=1
vjmn n(
maxk
jk) j, n, (37)
Jj=1
Mm=1
vjmn nI
i=1i n. (38)
Constraints (36) are aggregate capacity constraints forthe MTS echelon. Constraints (37) are derived from Equa-tions (23) and (24) and constraints (38) follow from Equa-tions (23) and (26). Constraints (37) imply that the totalow of semi-nished products through a DC should notexceed the DCs maximum throughput capacity, whereasconstraints (38) ensure that the ow of every semi-nishedproduct from plants to DC is at least equal to the bill ofmaterials multiplied by the demand of that product fromall the customers. These constraints are redundant in theoriginal MIP formulation, but they improve the quality ofthe subproblem solutions in terms of the feasibility to theoriginal problem upon relaxing the ow conservation con-straints. This results in better heuristic solutions. Therefore,we add these set of constraints to (SPMTS) as follows:
(SPMTS) : minM
m=1gmum +
Jj=1
Mm=1
Nn=1
(cjmn jn)vjmn,
(39)subject to
Jj=1
Nn=1
vjmn Pmum m, (40)
Mm=1
Pmum I
i=1i, (41)
Mm=1
vjmn n(
maxk
jk) j, n, (42)
Jj=1
Mm=1
vjmn nI
i=1i n, (43)
um {0, 1}, vjmn 0 j, m, n. (44)
3.1.1. The lower boundThe Lagrangean lower bound is given by the solution of theLagrangean dual problem, max[v(SPMTS) + v(SPMTO)]which is equivalent to
max
{minhIu,v
Mm=1
gmuhm +J
j=1
Mm=1
Nn=1
(cjmn jn)vhjmn
+ minhIx,y
Jj=1
Kk=1
fjkyhjk +I
i=1
Jj=1
Nn=1
(cij + njn)ixhij
+ t2
Jj=1
(1 +
Kk=1
CV2jkyhjk
) Ii=1 ix
hijK
k=1 jkyhjk
Ii=1 ix
hij
+ t2
Jj=1
(1
Kk=1
CV2jkyhjk
) Ii=1 ixijK
k=1 jkyjk
}.
Thus can be explicitly written as
(MP) : max
1 + 2,
subject to
1 +J
j=1
Mm=1
Nn=1
vhjmnjn M
m=1gmuhm
+J
j=1
Mm=1
Nn=1
cjmnvhjmn h Iu,v,
2 I
i=1
Jj=1
Nn=1
(nixhij
)jn
Jj=1
Kk=1
fjkyhjk
+I
i=1
Jj=1
cijixhij
+ t2
Jj=1
(1 +
Kk=1
CV2jkyhjk
) Ii=1 ix
hijK
k=1 jkyhjk
Ii=1 ix
hij
+ t2
Jj=1
(1
Kk=1
CV2jkyhjk
) Ii=1 ixijK
k=1 jkyjkh Ix,y,
where Iu,v is the index set of feasible points of the set:{(um, vjmn): Equations (40) to (43); um {0, 1}; vjmn
0, j, m, n},and Ix,y is the index set of feasible points of the set:
{(xij, yjk): Equations (32) to (34); xij 0; yjk {0, 1}, i, j, k}.
We use Kelleys cutting plane method, in which the point is the solution of the relaxed master problem (RMP), de-ned on subsets Iu,v Iu,v and Ix,y Ix,y (Kelley, 1960).This from (RMP) is used to solve the subproblems(SPMTS) and (SPMTO), and generate two cuts of the form:
1 +J
j=1
Mm=1
Nn=1
vijmnjn M
m=1gmuim +
Jj=1
Mm=1
Nn=1
cjmnvijmn,
(45)
2 I
i=1
Jj=1
Nn=1
(nixi
ij
)jn
Jj=1
Kk=1
fjkyi
jk
+I
i=1
Jj=1
cijixi
ij +t2
Jj=1
(1 +
Kk=1
CV2jkyijk
)
I
i=1 ixiijK
k=1 jkyijk
Ii=1 ix
iij
+ t2
Jj=1
(1
Kk=1
CV2jkyijk
)
I
i=1 ixiijK
k=1 jkyijk
. (46)
-
Response time reduction in supply chain design 457
The index sets Iu,v and Ix,y are updated as Iu,v {i} andIx,y {i}, respectively, as the algorithm proceeds throughthe iterations.
3.1.2. The heuristic: nding a feasible solutionThe rst subproblem (SPMTS) provides the location ofplants (um) and the ow of semi-nished products into theDCs (vjmn), whereas the second subproblem (SPMTO) pro-vides the location and the capacity decisions of the DCs(yjk), the assignment of customers to DCs (xij). Note thatthe link between the two subproblems is the ow balance ofproducts in and out of DCs. Hence, a feasible solution toproblem (PATO) can be constructed by solving (SPMTS) withthe additional set of constraints
Mm=1 vjmn =
Ii=1 nixij,
where xij is obtained from the solution of (SPMTO). Theoverall procedure is shown in Fig. 4. Computational resultsare provided next.
4. Computational results and insights
In this section, we report our computational experienceswith the proposed solution methodologies and presentsome insights. All the proposed solution procedures were
coded in C and the MIP problems were solved using ILOGCPLEX 10.1 (using the Callable Library) on a Sun Blade2500 workstation with 1.6-GHz UltraSPARC IIIi proces-sors. In the implementation of the iterative cutting planealgorithm and after the solution of the relaxed MIP, we usethe procedure CPXaddrows() to append the cuts generatedand exploit warm starting.
4.1. Test problems
The test problems are derived from the 2000 census dataconsisting of 150 largest cities in the continental UnitedStates (see Daskin (2004)). We generate nine sets of testproblems by setting the number of customers (I) to the 50,100 and 150 largest cities, and the potential DC locations(J) to the ve, ten and 20 most populated cities. The meancustomer demand rates i are obtained by dividing the pop-ulation of those cities by 103. The unit transportation costscij are obtained by dividing the great-circle distance betweenthe customer i and the potential DC location j by 100. Theservice rate of DC j equipped with capacity level k, is set tojk = k
i i (where k = 0.15, 0.20, 0.45 for I = 50; k =
0.10, 0.20, 0.30 for I = 100; k = 0.10, 0.15, 0.20, 0.30, 0.45for I = 150). The xed costs, fjk are set to 100 jk to
Fig. 4. Solution procedure for two-echelon ATO supply chain design.
-
458 Vidyarthi et al.
reect economies of scale. For the M/M/1 case, the vari-ance of service times 2jk are set to the mean service rates,jk whereas for the M/G/1, the 2jk is obtained by settingcoefcient of variation (CV ) to 1.5. The average responsetime cost t is set to (i j(icij)/(I J)), where isthe response time cost multiplier and icij denotes the totalproduction and transportation cost associated with the or-der from the ith customer served by the jth DC. In order toexplore the sensitivity of the solution to different levels ofresponse time costs, the multiplier is tested for 0.1, 1, 5, 10,50, 100 and 200 (higher value of models the situation inwhich losing a customer order due to high expected waitingtime is extremely costly). For the ATO supply chains, theinstances are generated by setting the capacities of plantsto: Pm = U [0.1, 0.5]
Ii=1 i whereas their xed cost are
set to: gm = U [1000, 2000]
(Pm). In order to comparethe performance of the cutting plane method and the La-grangean heuristic for different values of the ratio of plantcapacities to total demand (r ), the instances are generatedby setting the capacities of plants to: Pm = (r
Ii=1 i)/M
and the xed costs to: gm = U [100, 110]
(Pm). The pro-duction coefcients (bill of materials) n were randomlygenerated in the range U [1, 5] and rounded up to the near-est integer value.
4.2. An illustrative case study
Using the rst set of test problems (where I = 50, J = 5 andK = 3), we illustrate that the MTO supply chain congu-ration that considers congestion and its effect on responsetime can be different from the conguration that ignorescongestion. Furthermore, we show empirically that sub-stantial reduction in response times can be achieved withminimal increase in total costs in the design of responsivesupply chains.
The rst set of test problems consists of 50 customers,ve potential DC locations (j = 1, New York, NY; j = 2,Los Angeles, CA; j = 3, Chicago, IL; j = 4, Houston, TX;j = 5, Philadelphia, PA) that can be equipped with threecapacity levels (k = 1, small; k = 2, medium; k = 3, large).The problem is solved to optimality (with a gap of 106) us-ing the cutting plane algorithm. Table 1 summarizes the re-sults for different values of the response time cost by setting to 0, 0.1, 1, 10, 100 and 1000 under four different cases:M/G/1 case with CV = 1.5, M/M/1 case, M/G/1 casewith CV = 0.5, and M/D/1 case. The table shows the totalobjective function value (TC), xed cost (FC), variable cost(VC), total response time cost (RC), total expected wait-ing time (E(W )), average DC utilization (), DCs openedand their capacity levels (Open DCs (yjk)), expected wait-ing time at the DCs (Wj), DC utilization (j), the numberof cuts generated (CUT), the number of iterations required(ITR) and the CPU time in seconds (CPU(s)) under dif-ferent scenarios. The supply chain network congurationfor two extreme values of response time cost ( = 0 versus
= 1000) are shown in Fig. 5. Figure 6(a) shows the effectof changing response time cost on the total expected waitingtime (E(W )), and Fig. 6(b) shows the effect of changing thetotal expected waiting time E(W ) on the sum of the xedand transportation costs. The observations are as follows.
1. Figure 5 shows that the supply chain conguration thatignores congestion opens four DCs (medium size DCsin New York, Los Angeles and Chicago and a smallDC in Houston) whereas the conguration that consid-ers congestion opens ve DCs, all with capacity level 3.Also, there is substantial reallocation of customer de-mand among the DCs in order to balance the work-load to reduce the DC utilization and overall responsetime in the system. Hence, the supply chain congurationthat considers congestion and its effect on response timecan be different from the traditional conguration thatignores congestion. However, we observe that at veryhigh values of , the congurations (location, capacityand demand allocations) are not signicantly differentamong the M/G/1, M/M/1 and M/D/1 cases.
2. As we see from Table 1, even with very small valuesof average response time cost, t = 0.1 meani,j(icij)or t = 1 meani,j(icij), substantial improvement in thetotal expected waiting time (E[W ]) can be achievedover t = 0. For example, for the M/G/1 case (CV =1.5), E[W ] decreases from 77.41 units to 48.37 units for = 0.1 and 1 respectively. This is due to the even distri-bution of demand among DCs. Figure 6(a) also showsthat the substantial reduction in response time can beachieved with a small value of the response time costs.This is because, as we increase the magnitude of the re-sponse time cost, DCs with higher capacity are usedand/or the number of DCs opened increases, averageDC utilization decreases, thereby reducing congestionand improving average response time. From Fig. 6(b),we see that the left portion of the curves are quite at,indicating that substantial improvement (decrease) in re-sponse time can be achieved with a small increase in xedand transportation costs.
3. As the response time cost becomes dominant comparedto other cost components, DCs with higher capacity lev-els are opened, and average DC utilization decreases,thereby improving (decreasing) the response time. Forexample, in Table 1, in the M/G/1 case (CV = 1.5),for = 1, four medium size DCs are opened, whereasfor = 1000, ve large size DCs are opened. The aver-age DC utilization decreases from 0.83 to 0.44 and theexpected waiting time decreases from 48.37 to 5.44.
4. As we increase the magnitude of the response time cost,the transportation cost decreases initially and then in-creases. For example, in Table 1, in the M/G/1 case (CV= 1.5), for = 0, TC = 103, 577; for = 1, TC = 101,849; and for = 1000, TC = 106, 029. This is due tothe reallocation of customer demand among DCs in anattempt to reduce the total expected waiting time.
-
Response time reduction in supply chain design 459
Table 1. Comparison of the MTO supply chain network congurations for M/G/1, M/M/1 and M/D/1 cases: an illustrativeexample
M/G/1 (CV = 1.5) M/M/1 (CV = 1) M/G/1 (CV = 0.5) M/D/1 (CV = 0) = 0 TC 146, 965 146, 965 146, 965 146, 965
FC 43, 388 43, 388 43, 388 43, 388VC 103, 577 103, 577 103, 577 103, 577RC 0 0 0 0E(W ) 0.96 0.96 0.96 0.96Open DCs 1(2), 2(2), 3(2), 4(1) 1(2), 2(2), 3(2), 4(1) 1(2), 2(2), 3(2), 4(1) 1(2), 2(2), 3(2), 4(1)Wj [138.79, , 5.27, ] [138.79, , 5.27, ] [138.79, , 5.27, ] [138.79, , 5.27, ]j [0.99, 1.00, 0.84, 1.00] [0.99, 1.00, 0.84, 1.00] [0.99, 1.00, 0.84, 1.00] [0.99, 1.00, 0.84, 1.00]CUT 0 0 0 0ITR 1 1 1 1CPU(s) 0.14 0.14 0.14 0.14
= 0.1 TC 148, 567 148, 333 148, 042 146, 724FC 46, 816 43, 388 43, 388 43, 388VC 101, 494 104, 235 104, 093 104, 033RC 257 710 560 503E(W ) 77.41 214.05 169.00 152.40 0.83 0.96 0.96 0.96Open DCs 1(2), 2(2), 3(2), 4(2) 1(2), 2(2), 3(2), 4(1) 1(2), 2(2), 3(2), 4(1) 1(2), 2(2), 3(2), 4(1)Wj [10.09, 33.51, 2.49, 2.83] [33.93, 98.34, 7.24, 74.55] [42.56, 124.89, 6.66, 94.06] [48.12, 139.37, 6.43, 107.05]j [0.99, 0.97, 0.71, 0.74] [0.97, 0.99, 0.88, 0.99] [0.98, 0.99, 0.87, 0.99] [0.98, 0.99, 0.87, 0.99]CUT 4 20 20 20ITR 2 6 6 6CPU(s) 0.35 0.88 1.69 0.9
= 1 TC 150, 269 149, 683 149, 286 149, 139FC 46, 816 46, 816 46, 816 46, 816VC 101, 849 101, 744 101, 649 101, 609RC 1604 1,124 822 714E(W ) 48.37 33.9 24.79 21.55 0.83 0.83 0.83 0.83Open DCs 1(2), 2(2), 3(2), 4(2) 1(2), 2(2), 3(2), 4(2) 1(2), 2(2), 3(2), 4(2) 1(2), 2(2), 3(2), 4(2)Wj [10.09, 15.08, 2.49, 3.39] [10.09, 18.11, 2.49, 3.21] [10.09, 22.02, 2.49, 3.06] [10.09, 24.19, 2.49, 3.00]j [0.91, 0.94, 0.71, 0.77] [0.91, 0.95, 0.71, 0.76] [0.91, 0.96, 0.71, 0.75] [0.91, 0.96, 0.71, 0.75]CUT 4 8 16 12ITR 2 3 5 4CPU(s) 0.33 0.41 0.91 0.45
= 10 TC 157, 274 155, 674 154, 324 153, 874FC 52, 078 49, 447 49, 447 49, 447VC 101, 407 101, 640 101, 638 101, 616RC 3789 4,587 3,240 2811E(W ) 11.43 13.84 9.77 8.48 0.67 0.75 0.75 0.75Open DCs 1(3), 2(3), 3(2), 4(2) 1(2), 2(3), 3(2), 4(2) 1(2), 2(3), 3(2), 4(2) 1(2), 2(3), 3(2), 4(2)Wj [1.75, 2.27, 2.10, 1.94] [6.63, 2.27, 2.54, 2.39] [6.63, 2.27, 2.58, 2.36] [6.78, 2.27, 2.65, 2.28]j [0.64, 0.69, 0.68, 0.66] [0.87, 0.69, 0.72, 0.71] [0.87, 0.69, 0.72, 0.70] [0.87, 0.69, 0.73, 0.69]CUT 12 12 8 12ITR 4 4 3 4CPU(s) 0.65 0.62 0.49 0.60
= 100 TC 182, 451 176, 255 172, 486 170, 940FC 57, 340 57, 340 57, 340 54, 709VC 101, 987 101, 494 101, 494 101, 557RC 23, 124 17, 421 13, 650 14, 675E(W ) 6.98 5.26 4.12 4.43 0.56 0.56 0.56 0.61Open DCs 1(3), 2(3), 3(3), 4(3) 1(3), 2(3), 3(3), 4(3) 1(3), 2(3), 3(3), 4(3) 1(3), 2(3), 3(3), 4(2)Wj [1.45, 1.68, 0.96, 1.05] [1.54, 1.84, 0.91, 0.97] [1.54, 1.84, 0.91, 0.97] [1.75, 2.15, 1.00, 1.54]j [0.59, 0.63, 0.49, 0.51] [0.61, 0.65, 0.48, 0.49] [0.61, 0.65, 0.48, 0.49] [0.64, 0.68, 0.50, 0.61]CUT 16 4 4 16ITR 5 2 2 5CPU(s) 0.77 0.26 0.35 0.93
(Continued on next page)
-
460 Vidyarthi et al.
Table 1. Comparison of the MTO supply chain network congurations for M/G/1, M/M/1 and M/D/1 cases: an illustrativeexample (Continued)
M/G/1 (CV = 1.5) M/M/1 (CV = 1) M/G/1 (CV = 0.5) M/D/1 (CV = 0) = 1000 TC 358, 156 316, 316 289, 863 280, 873
FC 71, 675 71, 675 71, 675 71, 675VC 106, 029 102, 974 99, 683 99, 460RC 180, 453 141, 666 118, 506 109, 738E(W ) 5.44 4.27 3.57 3.31 0.44 0.44 0.44 0.44Open DCs 1(3), 2(3), 3(3), 4(3), 5(3) 1(3), 2(3), 3(3), 4(3), 5(3) 1(3), 2(3), 3(3), 4(3), 5(3) 1(3), 2(3), 3(3), 4(3), 5(3)Wj [0.64, 1.39, 0.78, 0.84, 0.55] [0.66, 1.51, 0.77, 0.83, 0.50] [0.68, 1.67, 0.76, 0.84, 0.43] [0.72, 1.67, 0.76, 0.85, 0.41]j [0.39, 0.58, 0.44, 0.46, 0.35] [0.40, 0.60, 0.44, 0.45, 0.34] [0.40, 0.63, 0.43, 0.46, 0.30] [0.42, 0.63, 0.43, 0.46, 0.29]CUT 35 35 25 10ITR 8 8 6 3CPU(s) 1.66 1.45 1.2 0.39
4.3. Performance of the cutting plane method for MTOsupply chain design
Table 2 displays the performance of the cutting planemethod for MTO supply chain design problems underM/G/1 (CV = 1.5), M/M/1 (CV = 1.0) and M/D/1
(CV = 0) cases for varying problem sizes. The columnsmarked FC, VC and RC represent the xed costs, the vari-able production and transportation, and the response timecosts respectively, expressed as a percentage of total costs(TC). The columns marked E(W ) are the total average wait-ing time in the system, is the average DC utilization and
Fig. 5. Effect of changing response time cost on the supply chain network conguration: (a) = 0; (b) the M/D/1 case, = 1000; (c)the M/M/1 case, = 1000; and (d) the M/G/1 case (CV = 1.5), = 1000.
-
Response time reduction in supply chain design 461
Fig. 6. (a) Effect of changing response time cost (t) on the total expected waiting time E[W ]; and (b) effect of changing expectedwaiting time E[W ] on the xed costs plus the transportation costs.
DC represents the number of DCs opened. The table alsodisplays the number of constraints generated (CUT), thenumber of iterations of the algorithm (ITR) and the to-tal CPU time in seconds required to obtain the optimalsolution.
The results show that the average CPU time for M/G/1,M/M/1 and M/D/1 cases are 48, 21 and 9 seconds re-spectively, whereas the average number of cuts requiredare 26, 25 and 25 respectively. Also, the maximum CPUtime for M/G/1, M/M/1 and M/D/1 cases are 1147,410 and 107 seconds respectively, whereas the maximumnumber of cuts required are 66, 60 and 50 respectively.The computation times reveal the stability and the ef-ciency of the cutting plane algorithm for different percent-ages of xed, variable and response time costs, whereasthe number of iterations imply that only a fraction of theconstraints in (PL(H)) is required. As the magnitude ofresponse time cost (t) increases, the percentage of responsetime cost becomes more signicant with respect to othercost components and the algorithm seems to require moreCPU time and iterations as large number of cuts are gen-erated. Furthermore, in almost all of the instances, theM/G/1 case requires more cuts, and hence more CPUtime to solve than the M/M/1 and M/D/1 cases. Thisis attributed to the non-linearity in the expression of ex-pected waiting time for M/G/1 queues. It is also worth-while noting that the computational times for the secondset of instances (I = 50, J = 20 and K = 3) are compara-tively higher than others because the optimal solution hashighly congested DCs. In the model, this corresponds to thevalue of R/(1 + R) approaching one. At the at portionof R/(1 + R), a higher number of cuts is needed to closethe gap.
4.4. Performance of the Lagrangean heuristic for ATOsupply chain design
The computational performance of the Lagrangeanheuristic for the two-echelon ATO supply chain modelfor the M/G/1, M/M/1 and M/D/1 cases is shownin Table 3. The second subproblem pertaining to theMTO echelon was solved to optimality using the cuttingplane approach. In all these test problems, the heuristicis activated at the nal iteration of the Lagrangeanprocedure. The Lagrangean bound (LAG-H) is ex-pressed as the percentage of heuristic solution and thequality of the heuristic solution (GAP) is expressed as:100 (Heuristic Solution LAG-H)/LAG-H. The tableshows the computational time of the subproblems (SP),the master problem (MP) and the heuristic (H) expressedas a percentage of the total computational time (CPU)for various single (L = 1) and multiple (L = 3 and 5)product instances. From these results, it is evident that theproposed heuristic succeeds in nding feasible solutionsthat are within an average of 2.81, 2.58 and 2.99% of theLagrangean bound in reasonable computational time: 427,390 and 345 seconds for the M/G/1, M/M/1 and M/D/1cases respectively. The total computational time can be ashigh as 1038 seconds in some cases. In terms of the size ofthe test problems, the heuristic is able to solve problems withup to 35 plants, 20 DCs and 150 customers, ve products,ve capacity levels and ve semi-nished products within amaximum of a 6% gap from the optimal solution. Table 4shows that the solution of subproblem 2 accounts for mostof the computational time, 89.08, 89.50 and 89.93% forM/G/1, M/M/1 and M/D/1 cases respectively. The mas-ter problem accounts for 9.87, 9.36 and 9.03%, whereas the
-
Tab
le2.
Com
puta
tion
alpe
rfor
man
ceof
the
cutt
ing
plan
em
etho
d:M
TO
supp
lych
ain
desi
gn
M/G
/1ca
se(C
V=
1.5)
M/M
/1ca
seM
/D/1
case
No.
IJ
K
FC
(%)
VC
(%)
RC
(%)
TC
E(W
)
DC
CU
TIT
RC
PU
FC
(%)
VC
(%)
RC
(%)
TC
E(W
)
DC
CU
TIT
RC
PU
FC
(%)
VC
(%)
RC
(%)
TC
E(W
)
DC
CU
TIT
RC
PU
150
103
0.1
4060
013
3,71
322
10.
936
245
240
600
133,
594
152
0.93
618
41
4060
013
3,48
495
0.92
618
41
139
592
136,
202
198
0.93
630
62
3959
113
5,18
112
60.
936
245
140
601
134,
356
660.
936
245
15
3958
314
0,63
560
0.85
624
52
4058
213
9,14
841
0.85
618
42
4059
113
7,86
226
0.86
630
63
1040
563
143,
888
350.
776
245
339
574
141,
829
370.
856
306
340
582
139,
498
230.
856
184
250
4252
615
8,72
314
0.61
66
22
4053
815
3,35
517
0.70
618
43
3954
614
8,62
814
0.77
618
42
100
3949
1216
8,46
714
0.61
618
45
4051
916
2,91
710
0.61
66
22
3952
915
6,21
511
0.7
618
43
250
203
0.1
5941
012
2,77
541
90.
9211
445
1059
410
122,
625
296
0.92
1144
58
5941
012
2,48
318
80.
9211
445
71
5643
112
4,84
319
80.
9110
405
1056
431
124,
153
139
0.91
1040
57
5941
112
3,48
299
0.92
1144
57
557
394
129,
842
117
0.86
1155
624
5839
312
7,93
780
0.87
1155
612
5542
212
6,04
563
0.91
1040
58
1057
376
134,
288
890.
8311
556
4457
385
131,
372
740.
8611
556
2258
393
128,
366
450.
8711
445
1050
5137
1215
4,12
341
0.71
1050
654
153
3710
147,
618
330.
7410
506
281
5438
814
0,91
524
0.76
1040
550
100
5233
1617
0,68
929
0.64
1166
711
4749
3515
160,
786
270.
7110
607
410
5237
1215
0,93
220
0.74
1050
610
73
100
53
127
730
196,
048
510.
834
124
127
730
195,
869
370.
834
83
127
730
195,
713
210.
834
124
15
2772
119
7,59
441
0.83
44
20
2773
119
6,93
226
0.83
44
20
2773
019
6,38
916
0.83
412
41
1026
722
199,
377
370.
834
165
127
721
198,
142
260.
834
124
127
731
197,
085
150.
834
42
050
2869
320
6,96
412
0.67
416
51
2770
320
4,85
813
0.75
412
41
2671
320
2,21
113
0.83
48
31
100
2867
521
2,39
912
0.67
416
51
2868
420
9,16
88
0.67
412
41
2769
420
6,23
58
0.75
412
41
200
2965
622
0,24
97
0.56
416
51
2866
521
6,43
06
0.61
48
31
2868
521
1,63
76
0.67
412
41
410
010
31
4752
016
5,40
213
40.
748
325
834
661
178,
665
284
0.93
630
65
3466
017
8,13
515
30.
936
245
35
5347
016
6,56
432
0.65
918
36
3762
118
0,10
457
0.86
721
43
3862
017
9,57
638
0.86
721
43
1052
471
167,
152
310.
659
274
637
621
181,
077
490.
857
285
337
621
180,
216
320.
867
214
250
5146
317
1,45
328
0.65
936
59
3860
318
6,48
026
0.77
728
54
3761
218
4,15
525
0.85
728
53
100
5044
617
6,74
928
0.65
927
412
3758
519
1,08
424
0.77
728
55
3859
318
7,49
415
0.77
728
54
200
4846
618
4,57
214
0.55
88
212
3559
619
8,18
216
0.70
618
44
3461
519
2,88
213
0.77
624
54
510
020
31
4752
016
5,21
815
90.
748
325
947
520
165,
061
119
0.74
832
56
4852
016
4,90
379
0.74
824
45
547
521
166,
275
860.
748
325
947
520
165,
867
650.
748
244
547
520
165,
458
440.
748
325
610
5347
016
6,75
132
0.65
918
35
5347
016
6,50
323
0.65
918
34
4752
116
5,93
936
0.74
824
45
5052
462
169,
629
290.
659
365
752
471
168,
556
200.
659
274
552
471
167,
667
130.
659
274
510
051
454
173,
133
280.
659
365
1051
463
170,
995
200.
659
274
552
462
169,
281
130.
659
274
520
046
477
179,
780
250.
658
325
2150
456
175,
861
200.
659
365
951
464
172,
442
130.
659
274
56
150
55
124
760
225,
407
149
0.91
416
52
2476
022
5,12
411
40.
914
165
224
760
224,
830
810.
924
165
25
2475
122
7,40
096
0.91
48
31
2476
122
6,56
364
0.91
412
41
2476
022
5,80
742
0.91
416
52
1025
741
229,
437
380.
824
83
124
751
227,
963
590.
914
165
224
761
226,
646
340.
914
124
150
2573
223
4,23
619
0.75
416
52
2673
123
2,79
413
0.75
44
21
2574
123
0,86
813
0.82
416
52
100
2572
323
8,46
218
0.75
412
41
2572
223
5,75
813
0.75
416
52
2673
223
3,47
78
0.75
44
21
200
2670
424
3,90
212
0.67
412
42
2671
324
0,68
98
0.67
48
31
2572
323
7,12
38
0.75
416
52
715
010
51
3169
020
7,16
429
40.
926
245
631
690
206,
933
224
0.93
624
55
3169
020
6,70
714
40.
936
245
65
3268
020
8,30
787
0.88
66
23
3466
020
7,97
588
0.89
728
55
3169
020
7,47
480
0.92
624
55
1032
681
209,
108
810.
886
184
532
681
208,
534
550.
886
62
234
660
208,
034
470.
897
285
650
3464
221
3,57
154
0.83
728
510
3464
221
1,75
937
0.84
728
56
3564
121
0,22
923
0.84
721
44
100
3165
321
7,91
737
0.8
624
513
3463
321
5,17
035
0.83
721
47
3464
221
2,31
422
0.84
721
45
200
3264
422
3,72
225
0.73
618
421
3165
422
0,22
024
0.8
624
516
3463
421
6,27
121
0.83
721
47
815
020
51
4852
017
7,44
433
20.
910
405
1148
520
177,
277
245
0.9
1040
58
4852
017
7,11
316
20.
910
405
85
4554
117
8,54
316
90.
899
274
1145
540
178,
141
115
0.89
927
47
4852
017
7,69
399
0.9
1040
59
1045
541
179,
538
157
0.89
936
516
4554
117
8,81
810
30.
899
274
845
540
178,
190
610.
899
274
750
4652
218
4,59
362
0.81
927
438
4553
218
3,14
764
0.85
936
527
4553
218
0,89
351
0.88
936
514
100
4551
418
8,33
559
0.81
927
454
4552
318
5,82
741
0.81
927
427
4552
218
3,62
036
0.85
936
524
200
4947
519
4,47
637
0.71
1030
421
144
515
190,
699
390.
819
274
6845
523
186,
730
240.
819
274
25m
in24
330
122,
775
70.
554
42
024
350
122,
625
60.
614
42
024
370
122,
483
60.
654
42
0m
ax59
7616
243,
902
419
0.93
1166
711
4759
7615
240,
689
296
0.93
1160
741
059
7612
237,
123
188
0.93
1150
610
7m
ean
4156
318
3,02
282
0.77
726
448
3958
318
2,99
567
0.80
725
421
3959
218
6,51
646
0.82
725
49
462
-
Tab
le3.
Com
puta
tion
alpe
rfor
man
ceof
the
Lag
rang
ean
heur
isti
c:A
TO
supp
lych
ain
desi
gn
M/G
/1ca
se(C
V=
1.5)
M/M
/1ca
seM
/D/1
case
No.
IJ
KL
MN
LA
G-H
(%)
GA
P(%
)S
P(%
)M
P(%
)H (%
)C
PU
(s)
LA
G-H
(%)
GA
P(%
)S
P(%
)M
P(%
)H (%
)C
PU
(s)
LA
G-H
(%)
GA
P(%
)S
P(%
)M
P(%
)H (%
)C
PU
(s)
150
103
110
30.
199
.78
0.22
84.3
114
.36
1.33
108
96.8
33.
1786
.06
12.7
51.
1985
98.3
31.
6792
.96
6.29
0.75
561
95.8
44.
1684
.55
13.5
11.
9411
199
.59
0.41
85.2
112
.96
1.83
8594
.58
5.42
91.3
48.
060.
6062
594
.77
5.23
87.1
11.5
11.
3911
899
.97
0.03
86.1
312
.19
1.68
8698
.84
1.16
92.0
55.
942.
0169
1098
.33
1.67
8314
.98
2.02
121
95.7
74.
2393
.14
5.53
1.33
8896
.94
3.06
88.4
310
.90
0.67
7150
98.9
11.
0991
.52
6.84
1.64
122
94.1
25.
8889
.78.
821.
4810
199
.86
0.14
88.5
89.
591.
8373
100
97.2
02.
8088
.78
11.1
30.
0912
795
.40
4.6
91.7
37.
011.
2610
498
.65
1.35
90.9
17.
931.
1675
250
203
110
30.
197
.07
2.93
87.8
511
.17
0.98
191
99.6
60.
3485
.99
12.4
11.
617
698
.23
1.77
93.5
95.
440.
9713
11
98.2
41.
7685
.03
13.1
51.
8219
296
.36
3.64
91.1
78.
290.
5417
994
.79
5.21
91.8
76.
361.
7714
75
99.3
20.
6885
.22
13.0
41.
7419
999
.07
0.93
92.5
66.
141.
318
195
.92
4.08
94.0
95.
390.
5214
810
97.8
92.
1188
.68
10.7
30.
5920
999
.37
0.63
90.5
58.
590.
8618
599
.73
0.27
86.7
711
.58
1.65
155
5094
.06
5.94
94.6
65.
170.
1721
199
.02
0.98
87.1
812
.46
0.36
186
97.7
12.
2991
.02
6.97
2.01
158
100
94.6
55.
3589
.75
9.67
0.58
213
95.0
94.
9192
.35
7.59
0.06
189
97.0
52.
9593
.33
6.03
0.64
162
310
05
33
205
199
.97
0.03
93.7
85.
680.
5418
298
.35
1.65
89.1
810
.14
0.68
168
98.5
61.
4486
.09
12.0
91.
8215
05
96.5
83.
4285
.45
14.3
90.
1618
599
.75
0.25
92.9
6.7
0.4
169
95.3
34.
6788
.18
11.2
70.
5515
910
97.2
32.
7786
.79
12.6
40.
5718
794
.26
5.74
87.1
410
.91.
9617
195
.07
4.93
86.8
911
.50
1.61
159
5097
.41
2.59
93.2
25.
830.
9519
096
.71
3.29
92.6
65.
431.
9117
699
.89
0.11
86.5
12.7
20.
7816
110
099
.68
0.32
93.1
16.
640.
2519
499
.22
0.78
92.2
56.
421.
3318
098
.85
1.15
88.0
211
.86
0.12
165
200
95.6
44.
3693
.61
5.06
1.33
196
99.5
50.
4588
.63
10.3
51.
0217
594
.97
5.03
85.3
812
.94
1.68
165
410
010
33
205
194
.47
5.53
91.7
38.
090.
1823
894
.01
5.99
87.3
511
.95
0.7
217
95.4
04.
6091
.29
8.25
0.46
197
594
.51
5.49
90.3
18.
261.
4326
895
.93
4.07
92.1
7.29
0.61
218
94.2
15.
7986
.72
11.7
21.
5620
110
98.8
31.
1790
.63
9.29
0.08
277
98.7
71.
2387
.36
11.1
81.
4622
297
.99
2.01
92.5
15.
881.
6120
250
97.2
22.
7886
.47
11.8
71.
6628
495
.41
4.59
86.5
112
.94
0.55
232
99.5
80.
4287
.82
11.8
40.
3420
410
097
.40
2.60
91.1
77.
411.
4229
698
.45
1.55
90.2
19.
10.
6923
496
.62
3.38
85.7
112
.99
1.30
210
200
95.9
44.
0685
.513
.22
1.28
298
97.4
52.
5590
.95
7.79
1.26
234
94.8
05.
2091
.81
6.52
1.67
215
510
020
33
205
198
.97
1.03
85.8
112
.14
2.05
297
98.0
81.
9287
.23
10.9
21.
8528
395
.40
4.60
92.2
25.
791.
9925
15
97.6
12.
3992
.95.
911.
1929
894
.37
5.63
92.1
57.
570.
2828
598
.65
1.35
93.1
15.
431.
4626
810
95.8
24.
1887
.46
11.1
31.
4130
097
.12
2.88
87.5
410
.71
1.75
285
99.6
10.
3988
.22
11.1
00.
6826
950
95.5
54.
4591
.68
7.47
0.85
309
97.0
32.
9789
.47
9.11
1.42
286
94.9
25.
0886
.04
12.1
31.
8327
110
099
.12
0.88
85.5
912
.88
1.53
313
98.6
61.
3487
.43
11.7
90.
7828
994
.12
5.88
86.8
112
.00
1.19
277
200
95.4
64.
5489
.19.
641.
2634
896
.44
3.56
92.2
27.
550.
2329
396
.12
3.88
91.1
68.
600.
2427
86
150
55
535
51
98.6
71.
3389
.13
9.93
0.94
551
99.9
40.
0686
.61
12.2
61.
1348
695
.08
4.92
92.2
16.
980.
8142
35
98.3
51.
6585
.54
13.9
60.
5056
394
.75
5.25
86.2
11.9
11.
8949
995
.08
4.92
93.1
95.
551.
2642
710
94.7
35.
2793
.29
6.17
0.54
576
96.0
53.
9590
.43
9.19
0.38
513
98.6
91.
3187
.81
10.8
61.
3344
450
99.8
30.
1789
.76
9.06
1.18
593
98.6
61.
3486
.77
11.6
81.
5553
699
.16
0.84
90.3
39.
110.
5645
110
099
.01
0.99
85.8
113
.66
0.53
599
96.9
03.
192
.06
7.38
0.56
544
97.2
22.
7891
.59
7.71
0.70
484
200
99.8
70.
1387
.42
12.0
70.
5161
799
.44
0.56
92.7
35.
441.
8354
794
.26
5.74
88.2
910
.55
1.16
485
715
010
55
355
199
.23
0.77
90.3
99.
290.
3271
599
.91
0.09
89.1
89.
531.
2964
594
.90
5.10
93.5
46.
090.
3758
85
97.2
32.
7791
.57.
630.
8771
799
.00
189
92
690
96.4
73.
5390
.24
9.10
0.66
598
1095
.34
4.66
94.5
55.
320.
1372
698
.73
1.27
89.7
18.
641.
6569
199
.17
0.83
89.6
9.16
1.24
615
5097
.13
2.87
90.7
47.
411.
8574
198
.15
1.85
91.6
7.61
0.79
697
95.2
74.
7388
.82
10.4
70.
7162
410
095
.27
4.73
91.7
46.
661.
6074
398
.50
1.5
87.9
111
.34
0.75
698
99.9
40.
0687
.06
12.8
80.
0662
720
098
.12
1.88
91.4
47.
151.
4174
796
.97
3.03
91.4
57.
141.
4170
697
.41
2.59
92.4
96.
960.
5563
28
150
205
535
51
98.3
11.
6991
.96.
981.
1298
195
.47
4.53
87.5
811
.25
1.17
922
95.0
74.
9394
.15.
130.
7776
55
96.6
83.
3289
.86
8.58
1.56
1000
99.8
10.
1991
.22
7.21
1.57
935
95.9
04.
1085
.95
12.4
01.
6580
010
95.1
24.
8893
.25.
870.
9310
0496
.02
3.98
87.9
610
.64
1.4
945
98.3
51.
6592
.89
6.46
0.65
810
5095
.71
4.29
84.3
214
.16
1.52
1007
99.5
00.
591
.95
7.83
0.22
955
94.3
55.
6586
.913
.03
0.07
871
100
98.6
41.
3685
.82
12.3
51.
8310
1994
.38
5.62
89.9
69.
180.
8697
299
.94
0.06
90.3
29.
600.
0889
020
094
.38
5.62
84.7
214
.67
0.61
1038
94.0
95.
9188
.84
9.49
1.67
976
99.7
20.
2892
.09
6.17
1.74
904
min
94.0
60.
0383
.00
5.06
0.08
108
99.9
70.
0385
.21
5.43
0.06
8594
.12
0.06
85.3
85.
130.
0656
max
99.9
75.
9494
.66
14.9
82.
0510
3894
.01
5.99
93.1
412
.96
2.00
976
99.9
45.
8894
.10
13.0
32.
0190
4m
ean
97.1
92.
8189
.08
9.87
1.05
427
97.4
22.
5889
.50
9.36
1.14
390
97.0
12.
9989
.93
9.03
1.04
345
463
-
Tab
le4.
Com
pari
son
betw
een
the
cutt
ing
plan
em
etho
dan
dth
eL
agra
ngea
nhe
uris
tic
for
AT
Osu
pply
chai
nde
sign
for
I=
100,
J=
10,K
=3,
M=
20an
dN
=1
Cut
ting
plan
em
etho
dL
agra
ngea
nhe
uris
tic
FC
(%)
VC
(%)
RC
(%)
E(W
)
DC
sC
UT
ITR
CP
U(s
)F
C(%
)V
C(%
)R
C(%
)E
(W)
D
Cs
LR
(%)
GA
P(%
)C
PU
(s)
Tig
htca
paci
ties
,r=
30.
150
500
689.
30.
965
103
714
4159
068
9.3
0.96
595
.88
4.12
251
149
501
209.
20.
965
103
1029
4159
121
8.2
0.96
596
.78
3.22
268
548
502
169.
30.
955
103
1585
4357
119
5.9
0.94
595
.04
4.96
269
1051
491
32.5
0.69
510
321
8042
562
51.8
40.
765
95.8
14.
1927
150
5147
214
.51
0.6
55
210
0244
552
17.4
70.
625
96.7
23.
2817
710
051
472
8.39
0.54
55
281
043
543
11.4
70.
65
94.9
95.
0127
850
050
464
8.39
0.54
55
296
443
498
10.8
20.
655
96.4
3.6
423
1000
4944
76.
210.
445
52
1100
4046
149.
820.
445
96.8
83.
1242
720
0046
4113
5.99
0.44
55
211
4839
4121
4.75
0.43
697
.11
2.89
444
5000
3832
303.
950.
327
285
1699
3333
343.
980.
327
97.0
12.
9945
1M
oder
ate
capa
citi
es,r
=5
0.1
4654
026
1.9
0.84
612
315
0039
610
332.
30.
854
98.1
11.
8920
51
4654
098
.12
0.83
66
296
439
610
122.
70.
844
96.8
33.
1740
75
4158
134
.16
0.83
44
210
1938
611
40.7
20.
844
96.6
93.
3132
510
4158
134
.16
0.83
44
289
838
611
40.7
20.
844
95.5
24.
4828
050
4156
421
.84
0.76
44
214
6739
583
27.5
60.
794
97.2
72.
7332
210
043
552
6.81
0.56
44
214
9540
572
7.74
0.58
497
.11
2.89
333
500
4254
46.
810.
564
42
1122
3753
107.
740.
584
96.8
83.
1236
510
0040
5010
6.81
0.56
44
213
7034
4818
7.74
0.58
494
.01
5.99
291
2000
4144
155.
520.
445
103
2524
4038
226.
630.
476
95.6
4.4
589
5000
3029
414.
550.
376
123
1981
3029
406.
530.
446
97.6
72.
3323
2L
oose
capa
citi
es,r
=10
0.1
4060
081
.01
0.83
44
211
0340
600
81.6
60.
834
96.5
73.
4354
61
4060
062
.75
0.83
44
295
539
600
63.8
50.
834
96.6
53.
3543
55
4060
140
.27
0.83
44
290
639
601
42.2
30.
834
98.9
11.
0948
710
3959
140
.27
0.83
44
287
539
601
41.7
80.
834
94.4
15.
5932
750
4058
211
.71
0.67
44
297
940
582
12.9
70.
674
96.1
93.
8135
410
040
564
11.7
10.
674
42
1140
4057
311
.75
0.61
496
.99
3.01
654
500
4155
57.
020.
564
42
1080
3852
118.
960.
564
97.2
72.
7376
510
0038
5210
6.94
0.56
48
317
6534
4719
7.84
0.56
498
.11.
964
220
0034
4719
6.85
0.56
48
318
8441
3821
7.54
0.37
695
.59
4.41
822
5000
3229
394.
430.
376
123
2148
3129
394.
430.
376
96.3
73.
6398
4M
in30
290
3.95
0.32
44
271
430
290
3.98
0.32
494
.01
1.09
177
Max
5160
4168
9.3
0.96
728
525
2444
6140
689.
30.
967
98.9
15.
9998
4M
ean
4350
763
.38
0.66
57
213
1439
529
640.
645
97.5
13.
4942
1
464
-
Response time reduction in supply chain design 465
heuristic accounts for 1.05, 1.14 and 1.04%, on average forM/G/1, M/M/1 and M/D/1 cases respectively.
In Table 4, we compare the performance of the cuttingplane algorithm and the Lagrangean heuristic and reportthe results for one problem set (I = 100, J = 10, K = 3,M = 20 and N = 1) for different values of the ratio of plantcapacities to total demand (r = m Pm/ i i). The resultsshow that the Lagrangean heuristic outperforms the cuttingplane method in terms of computational time. On average,the Lagrangean heuristic takes 421 seconds whereas thecutting plane method takes 1344 seconds.
5. Conclusions
In this paper, we modeled and analyzed the effect of re-sponse time consideration on the design of MTO and ATOsupply chain networks. We presented an MTO supply chaindesign model that captures the trade-off among responsetime, the xed cost of opening DCs and equipping themwith sufcient capacity, and the transportation cost asso-ciated with serving customers. Under the assumption thatthe customer demand follows a Poisson process and servicetimes follow general distribution, the DCs were modeled asa network of single-server queues, whose capacity levels andlocations are decision variables. We presented a non-linearMIP formulation, a linearization procedure and an exactsolution approach based on a cutting plane method. Ourcomputational results indicate that the cutting plane algo-rithm provides optimal solution for moderate instances ofthe problem in few iterations and reasonable computationtimes.
We also presented a model for designing two-echelonATO supply chain networks, that consists of plants andDCs serving a set of customers. Lagrangean relaxation wasapplied to decompose the problem by echelonone for theMTS echelon and the other for the MTO echelon. While La-grangean relaxation provides a lower bound, a heuristic isproposed that uses the solution of the subproblems to con-struct an overall feasible solution. Computational resultsreveal that the heuristic solution is on average within 6% ofits optimal. We used the models to demonstrate empiricallythat substantial improvement (decrease) in response timecan be achieved with a minimal increase in total cost asso-ciated with designing supply chains. Also, we showed thatthe supply chain conguration (DC location and capacity,and allocation of customers to DCs) obtained using themodel that considers congestion can be very different fromthose obtained using the traditional models that ignoresresponse time.
The focus of our ongoing research is to develop modelsfor the design of MTO and ATO supply chains under moregeneral settings: general demand arrivals, general servicetime distributions and multiple servers. Due to the lack ofan exact expression for the waiting time in these settings, weplan to explore two approaches to model the problem. Therst approach relies on approximations of expected waiting
times whereas the second approach integrates simulationwithin an MIP framework.
Acknowledgement
The authors would like to thank two anonymous refereesand the Associate Editor for helpful comments.
References
Amiri, A. (1997) Solution procedures for the service system design prob-lem. Computers and Operations Research, 24, 4960.
Berman, O. and Krass, D. (2002) Facility location problems with stochas-tic demands and congestion, in Facility Location: Applications andTheory, Drezner, Z. and Hamacher, H.W. (eds.), Springer-Verlag, NewYork, pp. 329369.
Daskin, M.S. (2004) SITATIONfacility location software. Depart-ment of Industrial Engineering and Management Sciences, North-western University, Evanston, IL. Available at http://users.iems.nwu.edu/mdaskin.
Dell, M. (2000) Direct from Dell: Stratagies that Revolutionized an Indus-try, Harper Collins, New York, NY.
Dogan, K. and Goetschalckx, M. (1999) A primal decomposition methodfor the integrated design of multi-period production-distribution sys-tems. IIE Transactions, 31(11), 10271036.
Duran, M.A. and Grossman, I.E. (1986) An outer approximation algo-rithm for a class of mixed-integer nonlinear programs. MathematicalProgramming, 36, 307339.
Elhedhli, S. (2006) Service system design with immobile servers, stochas-tic demand and congestion. Manufacturing and Service OperationsManagement, 8(1), 9297.
Elhedhli, S. and Gzara, F. (2008) Integrated design of supply chain net-works with three echelons, multiple commodities, and technology se-lection. IIE Transactions, 40(1), 3144.
Erenguc, S.S., Simpson, N.C. and Vakharia, A.J. (1999) Integrated pro-duction/distribution planning in supply chains: an invited review.European Journal of Operational Research, 115, 219236.
Eskigun, E., Uzsoy, R., Preckel, P.V., Beaujon, G., Krishnan, S. and Tew,J.D. (2005) Outbound supply chain network design with mode selec-tion, lead times, and capacitated vehicle distribution centers. EuropeanJournal of Operational Research, 165(1), 182206.
Gupta, D. and Benjaafar, S. (2004) Make-to-order, make-to-stock, ordelay product differentiation? A common framework for modellingand analysis. IIE Transactions, 36, 529546.
Huang, S., Batta, R. and Nagi, R. (2005) Distribution network design: se-lection and sizing of congested connections. Naval Research Logistics,52, 701712.
Kelley, J. E. (1960) The cutting plane method for solving convex programs.Journal of the SIAM, 8, 703712.
Kim, I. and Tang, C.S. (1997) Lead time and response time in a pullproduction control system. European Journal of Operational Research,101, 474485.
Klose, A. and Drexl, A. (2005) Facility location models for distributionsystem design. European Journal of Operational Research, 162, 429.
Margretta, J. (1998) The power of virtual integration: an interview withDell computers Michael Dell. Harvard Business Review, 76(2), 7384.
Marianov, V. and Serra, D. (2002) Location-allocation of multiple-serverservice centers with constrained queues or waiting times, Annals ofOperations Resear