An Adaptive Distributed Simulation Framework for a Server Fulfillment Supply Chain

Yan Chen, John Fowler, Teresa Wu Thomas CallarmanEamonn Ambrose and Vincent

HargadenDept. of Industrial Engineering China Europe International Quinn Business School

Arizona State University Business School University Collage DublinTempe, AZ, United States Shanghai, China Dublin, Ireland{jennychen, john. fowler,

teresa.wu ,thomas.callarman}@asu.edu tecallarman@ceibs.edu {eamonn.ambrose,

vincent.hargaden}@ucd.ie

.Abstract – Supply chains that produce and distribute computer servers are globally dispersed and have a high degree of uncertainty. To excel at servicing customers, a supplier must be highly skilled in matching the assets in the system with customer demand. Discrete event simulation has been proven valuable for system state estimation of supply chains. However, irregularities and disruptions occurring at any site along the system and the resulting bullwhip effects can lead to significant departures of simulation-based estimation from the performance of the real system. These departures reduce the ability of the model to assist in making correct decisions. In this paper, we propose an adaptive distributed simulation framework for a server fulfillment supply chain, and a Kalman filter to improve our estimates of job completion times.

Index Terms – Simulation, Server fulfillment supply chain

I. INTRODUCTION

Computer server fulfillment supply chain studied in this work is illustrated in Figure 1. It consists of 6 main elements: a wafer fabrication facility; an assembly and test facility; a server fulfillment center; a peripheral warehouse; end-customers for Integrated Circuits (ICs); and end-customers for servers. Two types of products are produced in this system: integrated circuits (ICs) and configured servers.

The core material transformation flow in this supply chain is wafers to ICs to servers. The wafers go through an elaborate process in a wafer fabrication facility in which thousands of circuits are fabricated on the wafers. Once the wafers are completed, they are then cut, packaged and tested to create integrated circuits in an assembly and test facility. After that, ICs are shipped to either IC customers or the server fulfillment center depending on demand. In the server fulfillment center, the ICs go through a series of panel assemblies and system tests with other peripherals from a warehouse to configure ordered servers. After

.

configuration, the servers wait on the docks to be transported to customers.

Fig 1. Scope of the Server Fulfillment Supply Chain

For end-customers, good service means on-time receipt of parts ordered and the required quantities [1]. In this server fulfillment supply chain, each manufacturing facility faces a big challenge to meet the quota at the end of the each quarter. The cost of unfulfilled orders can run into millions of dollars. Two things in this server fulfillment supply chain make the task above challenging. The first is the scope of the problem, which includes complex manufacturing flow and long cycle time in wafer fabrication, globally dispersed locations of the wafer fabrication, assembly and test facilities, and the server fulfillment center. The second is irregularities and disruptions occurring at any point in the system without warning due to the dynamic nature of a supply chain.

Discrete event simulation has been proven valuable as a practical tool for representing complex interdependencies, evaluating alternative designs and policies, and analysing performance tradeoffs for supply chain systems [2, 3, 4, 5, 6, 7]. Jain et al. [8] describe a simulation study on the supply chain for a large logistics operation. The results indicate that improvement in forecast accuracy can provide larger savings than process automation changes. Applicability of distributed simulation for decision-making in semiconductor manufacturing has been demonstrated by Lendermann et al. [9]. Its popularity is also reflected in industry applications. IBM developed a supply chain

1

simulator, which has a mix of simulation and optimization functions, to model and analyse its own supply chain issues [10]. IBM also used its own simulation-based supply

chain analyzer to visualize and quantify the effects of making changes on a hypothetical supply chain, and the impact of the changes on system performance [11]. Further, the need for executing supply chain simulations based on a full-detailed model has also been pointed out: Jain et al. [12] compared two models with different levels of detail for semiconductor manufacturing supply chains. The result shows that simulations incorporating detailed models are required when attempting to determine the correct inventory levels for maintaining desired customer responsiveness. In these cases, abstracted models can give inaccurate results that may subsequently lead to erroneous decisions. Venkateswaran et al. [5] drew similar conclusions in their paper.

In this paper, we describe how a distributed, detailed simulation model is built as a prototype for the server fulfillment supply chain above to be studied. Meanwhile, we notice that irregularities and disruptions occurring at any site along the system and their resulting bullwhip effects can make significant departures of simulation-based estimates about the system state from the real situation, which subsequently impairs its functionality in making correct decisions. To address this issue, we propose a Kalman filter based approach to calibrate the estimates for entry and exit times at the wafer fabrication, the assembly and test facility, the server fulfillment center and the peripheral warehouse. Some preliminary work has been done for the server fulfillment center only [13].

II. DISTRIBUTED SIMULATION MODEL OF SERVER FULFILLMENT SUPPLY CHAIN

The distributed simulation test bed used in our study is an HLA-based discrete event simulation system that originated from a semiconductor supply chain simulator developed in C++ under a joint project between Singapore Institute of Manufacturing Technology and the School of Computer Engineering at Nanyang Technological University, Singapore [12]. The test bed is implemented using the Run Time Infrastructure (RTI), which is an implementation of the HLA Interface Specification [14].

The High Level Architecture (HLA), which is a framework developed by the Defense Modeling and Simulation Office (DMSO), provides the necessary infrastructure for large-

scale distributed simulation. In HLA, a federate can be viewed as a component simulation model that is taking part in a large simulation [14]. A federation consists of a set of federates. For instance, in the case of supply chain simulation, federates can be embodied factories or suppliers and the federation is then the entire supply chain itself.

In our distributed simulation framework, the federation includes 6 federates.

1. IC demand generator2. Server demand generator3. Wafer fabrication facility4. Assembly & Test facility5. Server fulfillment center6. Peripheral warehouseEach federate is a sub-model which executes on

separate process in workstations and can be geographically distributed. It improves the simulation execution speed, supports reusability of existing simulation models and interoperability between different simulation packages.

The simulation model is composed of the basic elements of a supply chain [12]. These elements include manufacturing, transportation, business processes and customer orders as depicted in Figure 2.

Three successive stages of material transformation, wafers to ICs to servers, are modeled. The transportation among wafer fabrication facility, assembly and test facility, server fulfillment center and peripheral warehouse are modelled. So is the shipment to end-customers. Forecasting, production and inventory planning that are related to business processes are incorporated in the model. Customer orders are generated with the actual rate allowed to be different from the forecasted rate so as to simulate real life situations. The major components in this distributed simulation framework and their interaction are summarized in Figure 2. The federates interact with each other through information and material flow. The information flow is represented using dashed lines while the material flow using solid lines. Each of the federates is described below.

2

Fig 2. Interaction between Federates in Distributed Simulation System

A. IC Demand Generator (D/G)

The IC D/G generates orders for integrated circuits daily based on a predefined demand profile. These orders are then fed to the assembly and test facility. The volume for the orders of ICs can be varied each week by altering the demand profile. Each order is randomly assigned a customer weight and due dates are randomly assigned to each customer order, based on a uniform distribution.

B. Server Demand Generator (D/G)

The server D/G generator works similarly to the IC D/G. Server orders are produced daily based on another predefined demand profile. The volume can be varied each week by altering the demand profile. Each order is randomly assigned a customer weight and due dates are randomly assigned to each customer order, based on a uniform distribution.

C. Wafer Fabrication

The production of wafers in the wafer fabrication facility is based on a make-to-stock strategy driven by forecast, whereas the production of ICs in the assembly and test facility is based on a make-to-order strategy. The W/F facility releases wafer lots into production based on the product’s work-in-progress level in the factory and in transit, the inventory level of the wafer product in the warehouse of the A&T facility, the desired safety stock level and the forecasted demand of the product. For further details, refer to the paper by Chong et al. [15]. On a daily basis, wafer fabrication ships completed wafers to the A&T warehouse with a shipment delay of one day.

The wafer fabrication plant data is based on factory data from Sematech dataset 1, which is available through the Internet [16]. It produces two wafer products, which go through 210 and 245 process steps respectively. There are 32 operator groups in the dataset. The primary dispatching rules for machines are FIFO and Setup Avoidance (only for the medium and high current implantation machines).

D. Assembly and Test (A&T)

The orders from IC D/G and the demand from the server fulfillment center are fed to the assembly and test facility. Based on the availability of factory capacity and wafer inventory, the A&T facility assigns lots to orders and releases the lots into the facility.

The data for the assembly and test facility is based on previous projects the authors have had with the semiconductor industry. The data, particularly volume release and factory capacities, has been adapted to ensure that the production quantities and the utilization of facilities are consistent with what is typically found in the industry. Approximately 25 process steps exist in the A&T facility and the major dispatching rule for machines is FIFO.

E. Server Fulfillment Center

The manufacturing process in the server fulfillment center is shown in Figure 3. The chips provided by the A&T facility along with other peripherals are put on the boards in panel assembly. Then the assembly is tested at this stage. Next, the tested assembly is put together with additional peripherals to form a basic untested server system. This basic untested server system then goes for system test. After the test, the server system is disassembled (Dekit) and the resulting tested components are put into storage to fulfill a future customer order. Once a customer order is issued by the server demand generator, servers are configured depending on the actual requirements. The configured customer servers are then tested and sent for packing and shipping.

The production of servers before the point of fulfillment assembly is based on a make-to-stock strategy driven by forecast, whereas the production starting after the fulfillment assembly is based on a make-to-order strategy.

The server fulfillment center releases multiple chip modules and other peripherals into production based on the product’s work-in-progress level in the factory and in transit, the inventory level in storage, the desired safety stock level and the forecasted demand for the server. At this moment, only one type of server is considered in this study.

Once the orders from the server D/G are issued, based on the availability of factory capacity and inventory, the server fulfillment center assigns orders to fulfillment assembly. Meanwhile, it re-evaluates its stock level and generates IC demand for Assembly and Test and peripheral demand for warehouse, if necessary.

Fig 3. Manufacturing Process in Server Fulfillment Center

F. Peripheral Warehouse

The peripheral warehouse supplies peripherals based on the demand issued by the server fulfillment center. A lead time consistent with industrial experience is randomly generated using a normal distribution.

While the distributed simulation testbed works well to represent the operations of the real world supply chain, the

3

uncertainties inherent to supply chains, (such as unpredictable environmental changes, unpredictable failures, and emergent behaviour) impose significant challenges on supply chain management and on supply chain simulation, which have been designed, planned, and controlled based on a static, deterministic paradigm of data acquisition and decision-making, and an (essentially) open-loop mode of control. Often, decisions on SC models cannot be changed until the gap between the planned system state and actual system state has become significant, making correction expensive. Therefore, there is a need for an integrated framework combines actual system state from the real operation and predicted system state from simulation results to narrow the gap and provide better overall system estimation.

Note hardware and software technologies are developing which could enable a new paradigm of real-time decision making in supply chains. In particular, technologies like RFID, GPS, grid computing, and universal access to the World Wide Web promise instant availability and communication of state change data.

In this study, we propose a framework integrating Kalman filter and discrete event simulation. The basics of the Kalman filter are introduced in next section.

III. KALMAN FILTER

Kalman filters have traditionally been used for stochastic estimation and control. Recently, the Kalman filter has been applied in a variety of applications as well including Inertial Navigation and Guidance [17], Global Positioning Systems [18], Target Tracking [19], Finance [20, 21], etc. In the supply chain domain, Aviv [22] proposes an adaptive inventory replenishment policy that utilizes the Kalman filtering technique. Wu and O’Grady [23] develop an integrated approach that uses Kalman filtering and a Petri Net model to obtain a better state estimation of a supply chain system. Vensim [http://www.vensim.com/], an optimizer tool provided by Ventana Systems, uses Kalman filtering to track the actual inventory. Ventana Systems claim that Kalman filtering tracks the inventory much better than either simple simulation alone or the measured inventory alone. In this work, a Kalman filter approach is proposed to take into account the measurement error to obtain a better state estimate, namely, the estimated start and end processing times of jobs at each major supply chain component.

A Kalman filter [24, 25] is a set of mathematical equations that supports estimations of past, present, and even future states. The power comes from the fact that it can do these estimations even when the precise nature of the modeled system is unknown [26].

Mathematically, a Kalman filter is a set of recursive equations used to estimate the state x Rn of a discrete-time controlled process such as a manufacturing process that is governed by a transition equation and a measurement equation.

Transition Equation: xk = G*xk-1 + k-1, x R n

Measurement Equation: zk = H* xk + k, z R m

where Gnxn is the system state matrix that relates the state at the previous time step k-1 to the present step k, and Hmxn

relates the system states to the measurements. The random variables k and k represent the process and measurement noise respectively. They are assumed to be white noise with normal distributions: p (k) N (0, Qk) and p (k) N (0, Rk). The equations for the Kalman filter fall into two groups: time based equations (Equation 1 and 2), applied to obtain the current system state, and measurement based equations (Equation 3-5), used to adjust the system state from the measurements.

(1)

(2)

(3)

(4)

(5)where Kk is the Kalman gain, Pk is the error covariance,

is the estimation of xk before the measurement, and is

the estimation of xk given measurement zk. The Kalman filter assembles the two groups of equations to give the best estimate of the system state. The system of measurement and transition equations can be combined into an iterative process to determine the state of the system x.

IV. INTEGRATED FRAMEWORK

In this section, we describe how to integrate the distributed simulation system with the Kalman filter to get an estimate of the entry and exit time for each job at the wafer fabrication facility, the assembly and test facility, the server fulfillment center and the peripheral warehouse in the server fulfillment supply chain.

The proposed framework is shown in Figure 4, where the estimated state for each job is obtained by running the distributed simulation model. When a job is in a process, an emulation module provides real progress update. The Kalman filter module will calibrate the results and provide a more realistic estimate.

First, an estimate of when an entity will pass through every major process is obtained. This estimate is obtained using simulation. The entire process is simulated for 30 replications. For each replication, the arrival time of the entity to the system, is kept the same. An estimate of the queue times for all entities and an estimate of the processing times for all entities at every process are calculated by taking the average from the 30 replications. The standard deviation in these estimates is also calculated.

To emulate the real world, an additional simulation is run, using the same distribution and using the same parameters for the processing times as used in the simulation with 30 replications. We call this emulation. In

4

Fig 4. Proposed Framework

other words, measurements obtained from emulation are assumed to represent measurements obtained from the real world. At this point, information from the two different sources is available and both these sources have variability

–simulation has variance in estimation and emulation has variance in measurement (one source of measurement variance might be human recording errors). The Kalman filter can then be applied to generate estimates by

optimally combining simulated predictions with measurement using equations (6) – (14). The variables in these equations are defined in Table 1.

(6)

(7)

(8)

(9)

(10)

(11)

(12)

(13)

(14)

Equation (6) estimates the start time of the process at the wafer fabrication facility and equation (7) estimates the variation in the estimate of the start time. Equation (8) calculates the Kalman correction required to account for the measurement error in the state. Equation (9) applies the Kalman correction and obtains the corrected estimates of the state. Equation (10) estimates the variance in the corrected estimate of the state and equation (11) uses the corrected estimate for intermediate states to estimate when an entity will complete the last process (process at server

fulfillment center). Equation (12) estimates the corrected variance in the estimate of the state when an entity will complete the last process (process at server fulfillment center). Equation (13) estimates the next entity state from the corrected current state. Finally, equation (14) estimates the variance in the estimate of the next entity state.

TABLE IVARIABLE DEFINITION

Li

Time when entity starts or completes a process. For example, L1 refers to the start of the first process, L2 refers to the end of the first process, L3 refers to the start of the second process, and so on.

EAT Entity arrival time

TB(EAT L1) Estimate of time between entity arrival and start time of the first process

5

Fig 4. Proposed Framework

TB(Li Li+1)

Estimate of time between the instant when entity is at the start of process i and the end of the process. This accounts for the queue time between the end and beginning of a process and the processing time at a process.

TB(Li Ln) Estimate of time between the start of one process to the end of the overall process

Estimated time for entity to be at the start or the end of the process

Measured time for entity to be at the start or the end of the process

Corrected time for entity to be at the start or the end of the process

Kalman correction

Estimate of variance in estimated time of Li

Estimate of variance in measurment of Li (assumed to be 36)

Estimate of variance in corrected time of Li

Estimate of variance in estimated time between entity arrival time and the start time of the first process

Estimate of variance in estimated time between the instance when entity is at the start and the end of the process

Estimate of variance in estimated time between the start of one process and the end of the overall process

Some preliminary work has been done with a focus on the manufacturing portion of the server fulfillment center in our previous research [13]. The initial experimentation results show that using a Kalman filter can help in getting a more realistic estimate of when an entity is likely to come out of the server fulfillment center. The improved estimates are then used to sense whether an entity is on course to meet customer delivery expectations. In the future, we will explore it for the entire server fulfillment supply chain.

V. SUMMARY

This paper describes a computer server fulfillment supply chain and its modular structure in a prototype distributed simulation model, and it explores the dynamics and interaction among the components across the system. In addition, a Kalman filter approach is proposed to calibrate the system state estimate to get more realistic job completion time forecasts.

REFERENCES

[1] K. Fordyce, “New supply chain management applications provide better customer service: serious gets exciting,” IBM Microelectronics, Second Quarter, 2001.

[2] M. Hennessee, “Challenges facing global supply chains in the 21st century,” Proceedings of the 1998 Winter Simulation Conference, ed. D.J. Medeiros, E.F. Watson, J.S. Carson and M.S. Manivannan, 3-4. 1998.

[3] L. Chwif, M.R.P. Barretto, E. Saliby. “Supply chain analysis: Spreadsheet or simulation?” Proceedings of the 2002 Winter Simulation Conference, ed. E. Yücesan, C.-H. Chen, J.L. Snowdon, and J.M. Charnes, 59-66. 2002.

[4] S. Jain, N.F. Choong, and W. Lee. “Modeling computer assembly operations for supply chain integration,” Proceedings of the 2002 Winter Simulation Conference, ed. E. Yücesan, C.-H. Chen, J.L. Snowdon, and J.M. Charnes, 1165-1173. 2002.

[5] J. Venkateswaran, Y.-J. Son, and B. Kulvatunyou. “Investigation of influence of modeling fidelities on supply chain dynamics.” Proceedings of the 2002 Winter Simulation Conference, ed. E. Yücesan, C.-H. Chen, J.L. Snowdon, and J.M. Charnes, 1183-1191. 2002.

[6] S.T. Enns, and P. Suwanruji. “A simulation test bed for production and supply chain modeling,” Proceedings of the 2003 Winter Simulation Conference, ed. S. Chick, P.J. Sánchez, D. Ferrin, and D.J. Morrice, 1174-1182. IEEE. 2003.

[7] B.P. Gan, L. Liu, S. Jain, S.J. Turner, W. Cai, and W.J. Hsu. “Distributed supply chain simulation across enterprise boundaries.” Proceedings of the 2000 Winter Simulation Conference, ed. J.A. Joines, R.R. Barton, K. Kang and P.A. Fishwick, 1245-1251. 2000.

[8] S. Jain, E.C. Ervin, A.P. Lathrop, R.W. Workman, L.M. Collins. “Analyzing the supply chain for a large logistics operation using simulation.” Proceedings of the 2001 Winter Simulation Conference, ed. B.A. Peters, J.S. Smith, D.J. Medeiros, and M.W. Rohrer, 1123-1128. 2001.

[9] P. Lendermann, N. Julka, B.P. Gan, D. Chen, L.F. McGinnis, and J.P. McGinnis. “Distributed supply chain simulation as a decision-support tool for the semiconductor industry,” Simulation, 79(3): 126-138. 2003.

[10]S. Bagchi, S.J. Buckley, M. Ettl, and G.Y. Lin. “Experience using the IBM supply chain simulator,” Proceedings of the 1998 Winter Simulation Conference, ed. D.J. Medeiros, E.F. Watson, J.S. Carson and M.S. Manivannan, 1387-1394. 1998.

[11]G. Archibald, N. Karabakal, and P. Karlsson. “Supply chain vs. supply chain: Using simulation to compete beyond the four walls,” Proceedings of the 1999 Winter Simulation Conference, ed. Farrington, H.B. Nembhard, D.T. Sturrock, and G.W. Evans, 888-896. 1999.

[12]S. Jain, C.C. Lim, B.P. Gan, and Y.H. Low. “Criticality of detailed modeling in semiconductor supply chain simulation,” Proceedings of the 1999 Winter Simulation Conference, ed. P.A. Farrington, H.B. Nembhard, D.T. Sturrock, and G.W. Evans, 888-896. 1999.

[13]D. Parmar, T. Wu, J. Fowler, T. Callarman and V.Hargaden. “An intergrated framework for responsive supply chain management,” The 16th International Conference on Flexible Automation and Intelligent Manufacturing, Limerick, Ireland, June, 2006.

[14]S.J. Turner, W. Cai, and B.P. Gan. “Adapting a supply chain simulation for HLA,” Proceedings of IEEE 4th International Workshop on Distributed Simulation and Real Time Applications, 71-78, San Francisco, USA. 2000.

[15]C.S. Chong, P. Lendermann, B.P. Gan, B.M. Duarte, J.W. Fowler, and T.E. Callarman. “Analysis of a customer demand driven semiconductor supply chain in a distributed simulation test bed,” Proceedings of the 2004 Winter Simulation Conference, ed. R.G. Ingalls, M.D. Rossetti, J.S. Smith, and B.A. Peters.1902-1909, 2004.

[16]MASMLab Test-Bed datasets [Online], maintained by Modeling and Analysis of Semiconductor Manufacturing Laboratory, Industrial Engineering department, Arizona State University, USA. http://www.eas.asu.edu/~masmlab/ftp.htm, 2004.

[17]P. Zarrchan., "Tactical and strategic missile guidance," AIAA, Inc., Washington, DC, 1994.

[18]R.P. Denaro and P.V.W. Loomis., "GPS Navigation Processing and Kalman filtering," AGARD, NO. 161, pp. 11.1-11.9, 1989.

[19]C.K. Chui and G. Chen, Kalman filtering with real-time applications, Springer-Verlag, New York, 1987.

[20]C. Wells, The Kalman filter in finance, Kluwer Academic Publishers, Dordrecht, 1995.

[21]P.J. Bolland and J.T. Connor, "A Constrained neural network Kalman filter for price estimation in high frequency financial data,"

6

International Journal of Neural Systems, Vol. 8., No. 4, August, 1997.

[22]Y. Aviv: “A Time series framework for supply chain inventory management,” Operations Research, Vol. 51, No. 2, pp. 210-227, 2003.

[23]T. Wu and P. O’Grady: “A Methodology for improving the design of a supply chain,” International Journal of Computer Integrated Manufacturing, Vol. 17, No. 4, pp. 281-293, 2004.

[24]R.E. Kalman, and R.S. Bucy., "New results in linear filtering and prediction theory," Transactions of the ASME series D: Journal of Basic Engineering, 83 (3): 95 - 108, 1961.

[25]R.E. Kalman, "A New approach to linear filtering and prediction problems," Transactions of the ASME series D: Journal of Basic Engineering, 82 (1): 35 - 45, 1960.

[26]G. Welch and G. Bishop: “An introduction to the Kalman filter,” Technical Report No. TR 95-041, Department of Computer Science, University of North Carolina, 1995.

[27]P.S. Maybeck. Stochastic Models, Estimation, and Control, Academic Press, Inc., New York, Vol 1, 1979.

7