THROUGHPUT ANALYSIS OF TWO-STAGE MANUFACTURING SYSTEMS WITH MERGE AND BLOCKING

Abstract

Parallel lines are often used to increase production rate in many manufacturing systems where the main line splits into several lines in parallel, and after some operations, they merge into a main line again. Queueing networks with finite buffers have been widely used for modeling and analyzing manufacturing systems. This paper provides an approximation technique for multi-server two-stage networks with merge configuration and blocking which will be a building block for analysis of general manufacturing systems with parallel lines and merge configuration. The main idea of the method is to decompose the original system into subsystems that have two service stations with multiple servers, two buffers and external arrivals to the second stage are allowed. The subsystems are modeled by level dependent quasi-birth-and-death (LDQBD) process.

1. Introduction

A complex manufacturing system usually consists of one or more complex operations, such as assembly, disassembly, split, merge, rework, parallel, feedforward and scrap, etc. [10]. Parallel lines are often used to increase production rate in many manufacturing systems. In parallel systems, the main line splits into several lines in parallel, and after some operations, they merge into a main line again. For example, three types of lens module installed in smart phone are assembled in their own assembly line. After finishing assembly, all lens modules are sent to the resolution inspection machines which can inspect all types of lens modules automatically.

Queueing networks with finite buffers have been widely used for modeling and analyzing many manufacturing systems. However, exact analytical results for queueing networks with finite buffers are limited only to some special systems with one or two queues [4,6,13,16,19]. So many approximation methods are presented for analysis of queueig networks with finite buffers. Most of the approximation methods for serial lines or assembly systems are either based on decomposition or on aggregation approaches [3]. The principal procedure of decomposition methods is to decompose the long line into subsystems that are mathematically tractable, and derives a set of equations that determine the unknown parameters of each subsystem, and finally develops an algorithm to solve these equations. The basic idea of aggregation method is to replace a twonode system by a single equivalent node and reduce the number of nodes in the network. For a detailed list references for approximations and their application to manufacturing system, see review papers [3,11,14] and monographs [1,2,7,12,15]. The literature on the analysis of queueing networks with finite buffer and multiple servers is more scarce than the case of single server system, e.g. see [5,18] which are focused on the approximation of serial lines with reliable servers. Almost all decomposition method use subsystems with two nodes and single buffer between them. Recently, approximate analysis for multi-server tandem queues with exponential service times and finite buffers is presented in [17] where the subsystems with two nodes and two buffers are used to reflect the dependence between consecutive stages and improves the quality of approximation.

The literature on the approximation for the system with parallel lines with multiple servers and finite buffer is very limited. The purpose of this paper is to provide an approximation method for multi-server two-stage networks with merge configuration and blocking which will be a building block for analysis of multi-server-finite-buffer queueing networks with parallel lines and merge configuration. The main idea of the method is to decompose the original system into subsystems that have two service stations with multiple servers, two buffers and external arrivals to the second stage are allowed. The subsystems are modeled by level dependent quasi-birth-and-death (LDQBD) process.

This paper is organized as follows. In Section 2, the model is described in detail. Approximation procedure is presented in Section 3. Numerical results and concluding remarks are presented in Sections 4 and 5, respectively.

 

2. Model description

We model the manufacturing with parallel lines by the two-stage open queueing network in which the first stage consists of J multi-server queues with finite buffers and the second stage is one queue with multiple servers and finite buffer. Two stages are linked in merge configuration as depicted in Fig. 1. The queues (merging queues) at the first stage will be denoted by node i, i = 1, 2, · · · , J, and the queue (merged queue) at the second stage is denoted by node 0. The node i consists of a service station Mi with mi parallel identical servers and a buffer Bi of size bi and let Ii = mi + bi, i = 0, 1, 2, · · · , J. The index i can be viewed as the sequential number of the nodes at the first stage runs from 1 to J unless otherwise stated. The service time of each server at Mi is exponentially distributed with rate µi , i = 0, 1, 2, · · · , J. Customers arrive to node i from outside according to a Poisson process with rate λi, i = 1, 2, · · · , J. The arrivals to any node at the first stage are dispatched to some of its servers if there is one free, and if all servers are occupied the request waits in the corresponding buffer. When an arrival finds no free position in the buffer (buffer is full), it is lost. After completion of the service at node of the first stage the request proceeds to node 0, and is taken immediately into service if one of the servers is free (no queue at node 0), otherwise is placed into the buffer B0 and waits there until a server is available. If buffer B0 is full at the instant of service completion at node i, the customer is forced to stay at node i until a space becomes available at node 0. During this period, the server just completed its service is blocked. This type of blocking mechanism is called blocking-after-service (BAS). Since there are more than one node directly linked to node 0, the servers in more than one upstream node can be blocked simultaneously. In this case, we assume that the corresponding blocked customers enter node 0 on a random dispatching rule. Let αj (y1 , · · · , yJ) be the probability that a blocked customers at node j joins the node 0 upon a space in B0 is available when there are yi servers blocked at node i, i = 1, 2, . . . , J. A customer who has completed service at node 0 leaves the system.

FIGURE 1.A two-stage queueing network with merge configuration

Let Xi be the number of customers at node i which includes the customers being served at Mi and waiting requests at Bi but does not include the customers blocked to enter node 0 and Yi the number of blocked servers at Mi, i = 1, 2, · · · , J in stationary state. By X0 denote the number of customers at node 0 that include the customers blocked at node i, i = 1, 2, · · · , J. It can be seen that 0 ≤ X0 ≤ H, where H = l0 + m with Let X = (X1, · · · , XJ) and Y = (Y1, · · · , YJ). Then the state space of (X0, X , Y) is S = S0 ∪ S1, where and where x = (x1, · · · , xJ), y = (y1, · · · , yJ) and e is the column vector of appropriate size whose components are all 1. The state space S is so complex and its size increase drastically as the number J of nodes and/or mi, bi increase.

Main goal of this analysis is to provide an approximation method for the mean departure rate from node 0 that is the throughput of the system.

 

3. Approximation

In this section, we derive an algorithm for approximate analysis of throughput of the system. First we fix a node i. For an investigation of the behavior of node i and node 0, we consider the two-node queue denoted by Li where the arrivals from outside to the second stage are allowed as depicted in Fig.2. We denote the first node and the second node by respectively. The node is the same as node i of the original system and has a service station, that is, there are mi exponential servers with rate µi in parallel and a finite buffer of size bi. The BAS blocking rule is assumed in the system. The node consists of a service station M0 with m0 exponential servers with rate µ0 in parallel and a buffer of size b0. In order to reflect the effects of node j, j ≠ i, we assume that has J − 1 virtual buffers whose size is of mj, j = 1, 2, · · · , J, j ≠ i and type j customers arrive from outside to queue . If a type j customer finds that the queue is full upon arrival, the customer is blocked and stayed at the virtual buffer until the queue can accommodate the customer. We assume that the distribution of inter-arrival time of type j, j ≠ i customer is exponential with rate γj (h, y) that depends not only on the state h of but also on the state y = (y1, · · · , yJ) of the number of blocked customers Y. We assume that the arrivals of type j are blocked when Yj = mj, that is, γj (h, y) = 0 for yj = mj. If buffer is full upon type j arrival, we let Yj increase by 1. The state of Yj s decreased by 1 with probability αj (y) upon a service completion The arrival rate γj (h, y) will be approximated by iteration in the later.

FIGURE 2.Two-node tandem queue Li with virtual buffers and external arrivals to the second node

Let be the number of customers excluding the blocked one and the number of blocked customers at at time t. Denote by the number of type j, j ≠ i customers blocked and by the sum of the number of customers at and the number of blocked customers at time t. Then the stochastic process and is a Markov chain on the state space where h = {(h, n, 0), 0 ≤ n ≤ li}, h = 0, 1, · · · , l0 and for h = l0 + k, k = 1, 2, · · · , m, h = {(h, n, y), It can be easily seen that the Markov chain is a level dependent quasi-birth-and-death (LDQBD) process whose generator Qi is of the form

The block matrix is the square matrix of size |h|, the number of states of level h, h = 0, 1, 2, · · · , H and whose component are given as follows:

where I(A) is the indicator function of A, that is, if A is true, I(A) = 1, otherwise I(A) = 0. The components of are given as follows: for 0 ≤ h ≤ l0 − 1,

and for l0 ≤ h ≤ H − 1,

where ek is the J-dimensional vector whose ith component is 1 and others are all 0 and µi (n, yi) = min(n, mi − yi)µi is the service rate at node Si when The components of are given as follows: for 1 ≤ h ≤ l0,

and for l0 + 1 ≤ h ≤ H,

Let and Y* be the stationary version of and Y*(t), respectively and

The stationary distribution πi = (πi (0), πi (1), · · · , πi (H)) satisfies the following equations

where is obtained recursively as follows

The vector πi (0) is the unique solution of the equations

Noting that the number Ψi of busy servers at once πi is given, we can calculate the approximation formulae for arrival rate from node i to node 0 given as follows :

where and the arrival rate from node i to node 0 given

where Si (h, y) = {y : ye = max(h−l 0, 0)} and

We use the iterative algorithm to determine the parameters and calculate the throughput. be the system Li in the kth iteration and be the arrival rate to

Algorithm

Once γi (h, y) are obtained, calculate the stationary distribution πi and γi (h) using (2), i = 1, 2, · · · , J. Modeling node 0 as M/M/m0/H queue with state dependent arrival rate and service rate µ0 and calculate throughput (TP)

No proof of convergence or accuracy can be given for this algorithm, as for many similar algorithms for serial line decomposition. However extensive numerical experiments show the convergence of the algorithm.

 

4. Numerical results

In this section, some numerical results are presented for accuracy of approximations for the performance measures such as throughput, mean number E[X0] of customers at node 0 and mean number of blocked customers. The selection probability is used as

where y∗ = max1≤i≤J yi. Note that is the number of nodes in which the number of blocked customers is biggest among other nodes.

In the following tables, the vectors m = (m0, · · · , mJ) and b = (b0, · · · , bJ) denote the server allocation and buffer size allocation, respectively. The total service rate allocation is denoted by mµ = (m0µ0, · · · , mJµJ). For example, the vector (2, 3, 4) in the column corresponding to b means that b0 = 2, b1 = 3, b2 = 4 and the vector mµ = (2.0, 1.0, 1.0) means µ0 = 2.0/m0, µ1 = 1.0/m1, µ2 = 1.0/m2 in Table 1. The arrival rates from outside to node i are assumed to be the same and denote it by λi = λ, i = 1, · · · , J and the tolerance for the stopping criterion (3) is chosen as ϵ = 10−5.

TABLE 1.Approximation results for the system with J = 2, m = (4, 3, 2), mµ = (2.0, 1.0, 1.0)

Since the state space of the Markov chain for the system with J = 2 is small, we can have the exact results by solving the balance equations. In Table 1, approximation results (App) for the system with two merging nodes (J = 2) are compared with exact one (Exact). The relative percentage error is calculated by Err(%) = |App − Exact| × 100/Exact. The errors in Table 1 seem to be due to the approximation of arrival rates γi (h, y) by decomposition. In Tables 2-4, approximation results for the throughput, E[Y0] and E[Y] in the system with J = 3, m = (5, 4, 3, 2), mµ = (2.5, 1.0, 1.0, 1.0) are compared with those of simulation (Sim). Simulation models are developed with ARENA [9]. Simulation run time is set to 220,000 unit times including 20,000 unit times of warm-up period. Ten replications are conducted for each case and the average value and the half length of 95% confidence interval (c.i.) are obtained. The deviation (Dev) between simulation and approximation in Tables 2-4 is calculated by Dev=AppSim. The number of iterations of the algorithm to obtain each result in the tables is less than or equal to 4.

TABLE 2.The throughput for the system with J = 3, m = (5, 4, 3, 2), mµ = (2.5, 1.0, 1.0, 1.0)

TABLE 3.E[X0] for the system with J = 3, m = (5, 4, 3, 2), mµ = (2.5, 1.0, 1.0, 1.0)

TABLE 4.E[Y] for the system with J = 3, m = (5, 4, 3, 2), mµ = (2.5, 1.0, 1.0, 1.0)

Numerical results show the approximation works well. Almost all the approximations for throughput are contained in the 95% confidence intervals and the relative percentage errors are all less than 1%. The approximations for E[X0] and E[Y] also provide satisfactory results.

 

5. Conclusion

In this paper, we developed an approximation method for a multi-server two-stage queueing network with finite buffers and merge configuration. We decompose the system into subsystem. The subsystem is approximated by the tandem queue with two service stations and two buffer spaces and external arrivals at second stage. Approximate system is modeled with LDQBD process. An iterative algorithm is presented for calculating the parameters. The approximation method proposed can be used to analyze the more complex queueing networks with merge configuration.