Вы находитесь на странице: 1из 4

Computers ind. Engng Vol. 31, No. 1/2, pp.

327 - 330,1996

Pergamon
S0360-8352(96) 00142-8

Copyright 1996 Elsevier Seimee Ltd


Printed in Great Brits~ All fights reserved
0360- 8352/96 $15.00 + 0.03

Parallel Discrete Event Simulation of Manufacturing Systems: A Technology Survey


Chong Peng and F. Frank Chert
Dept. of Industrial & Systems Engineering
Florida International University
University Park, ECS 442
Miami, FL 33199

Abstract
Parallel discrete event simulation (PDES) refers to a simulation platform initiated by Chandy and Misra
[1979]. The basic idea of PDES is that a simulation model is decomposed into submodels which are then run on
different processes coucurrently.The objective of this paper is to provide a comprehensive overview of PDES' use
for manufacturing systems. First, the PDES platform and basic analytical methods associated with this platform
are briefly introduced, the ditIiculty of PDES compared with conventional simulation method is discussed. A
stuvey of PDES' use in manufacturing systems is then provided, which covers: (1) a summary of different PDES
models for different kinds of manufacturing systems, and (2) a comparison of PDES characteristics in the context
of decomposition method and synchronization scheme. Based on the survey, it is concluded that using PDES for
manufacturing systems has not been successful and its success depends on both the availability of efficient model
decomposition method and the improvement of suitable hardware architecture.
_Key Words: Parallel Descrete Event Simulation, Flexible Manufacturing System

1. Introduction
Parallel discrete event simulation fPDES) refers to a simulation platform initiated by Chandy and Misra
[1979]. The basic idea of PDES is the decomposition of a simulation model into submodels and runs on different
processes coucurrently (note that processes, not processors referred here, because several processes could be run
on the same processor). Processes communicate to each other only through messages passing. Processes do not
have any shared variables and there is no central process for message routing or process scheduling. This is totally
different from conventional sequential discrete event simulation, which utilizes the state variables to describe the
state of the system, an event list to contain all pending events that have been scheduled but have not taken effect,
and a global clock variable to denote how far the simulation has progressed. In PDES:

The system being modeled (usually referred to as physical system) is viewed as being composed of a number
of physical processes (denoted by PP) that interact at various points in simulated time.
The simulator is constructed as a logical processes (denoted by LP) network. In the LP network, each LP
represents a PP, containing a portion of the state corresponding to the PP it models; the directed channels
between LPs represent possible interactions between PPs. The direction of each channel represents the
direction of possible message interaction in the physical system (when a channel is directed to a LP, it is
called the input channel of the LP; otherwise, it is called the output channel of the LP).
All actual interactions between PP at a certain time are modeled by time-stamped event message sent between
the corresponding LP through channels. The time-stamp of an event message correspond to the occurrence
time of the event in the physical system.
Each LP may execute sequential code and two special commands: send and receive. In sending, an LP names
an output channel and a message that is to be sent along that channel. In a receiving, an LP names one or
more input channels and receives a message from one of them.
Each LP has a clock value (called LP time or local time) associated with it, which denotes how far the
simulation for a PP has processed by the corresponding LP. LPs may have simulated their corresponding PPs
to different point in time.

The difficulty of PDES is present in that the internal parallelisms of the simulation must be exploited
while causality error should be avoided at the same time [Fujimoto 1990]. Speaking of the causality error, assume
327

328

19th International Conference on Computers and Industrial Engineering

that two events, say E1 and E2, are under consideration, and the occurrence time of E2 is larger than the
occurrence time of E1 in the physical system, then it is possible that El modifies the state variables used by E2 in
the simulation process. If the simulation is uncarefully arranged to make E2 modify the state variables used by El,
causality error may occur. Causality error in the simulation actually means the future could affect the past in the
physical system. This is absolutely unacceptable. It is easy to understand that causality error is not even a problem
in the sequential simulation method, because under this method, the event with the smallest occurrence time is
always selected from the event list as the one to be processed next at every simulation step. But, since the
occurrence time of event in the physical system is modeled by time stamp of event message, causality error
becomes a major problem in PDES.
Causality error avoidance scheme broadly falls into two categories: conservative scheme [Chandy and
Misra 1979; Misra 1986] and optimistic scheme (time warp) [Jefferson 1985]. Conservative method requires that
message from any LP to any other LP be transmitted in the chronological order according to their time stamps.
This strictly avoids the possibility of any causality error. On the other hand, optimistic scheme uses a detection
and recovery approach, in which causality errors are detected and a roll back mechanism is invoked to recover.
Speedup is defined to be the time it takes a single processor to perform a simulation divided by the time it
takes the parallel system to perform the same simulation [Righter and Walrand 1989]. It is an important index to
evaluate the performance of PDES systems and the objectives for many PDES researchers to pursue.

During the last 15 years, PDES has found its use in many areas [Fujimoto 1993] suc~ as digital logic
circuit, computer architecture, telecommunication network, transportation network, and manufacturing systems.
The motivation of applying PDES to the manufacturing systems mainly comes from requirements from the real
world to increase the computational performance. As observed by many researchers, simulation is a indispensable
tool to studying manufacturing systems since it closely duplicates the behavior of the real system and does not
require simplifying assumptions like the analytic method [Elmaragly 1991]. But as the complexity of modern
manufacturing systems increases, the conventional simulation method has become inadequate particularly in the
analysis and control of highly automated systems such as Flexible Manufacturing System (FMS). A typical
manufacturing system simulation effort requires the evaluation of multiple scenarios or control policies, the large
number of those scenarios and control polices, plus the complication of the structure of modern manufacturing
systems, making the computation burden of the simulation very heavy. It is often found that simulation of complex
manufacturing systems are excessively slow, a fatal shortfall making the simulation technology confined to mostly
off-line, not real-time applications.
From literature to date, the result of research works in PDES for manufacturing system is not quite
encouraging. This paper attempts to analyze why this have happened (Section 3) based on a brief literature review
(in Section 2).

2. PDES of Manufacturing Systems - A Literature Review


Comparing with the number of papers about PDES' applications in other areas (such as
telecommunication and computer architecture), the number of papers about its use in manufacturing system is
quite limited:
Robert and Shires [1984] presents a progress report on a project conducting the simulation of FMSs based
on the use of multiprocessing technique. Even PDES was not cited in the paper, the idea there i~ almost the same
as that of PDES. In this project, entities considered as parts of an FMS are limited to: buffers (storing and L/UL),
machines. The whole system is based on an Intel SYS 310 computer connecting with multiprocessing slave
computers through Intel multibus, the operating system is Intel RMX 86 (a real-time multi tasking executive). As
the time the paper was published, the work had resulted in a simple multi-tasking simulation system on one
single board computer. Five program modules (written in Pascal programming language) are used to simulate a
simple example FMS consisting of three buffers and two machine centers: one program module simulates the
machine centers (setting, cutting, tool changing, loading and unloading), the second module simulates all of the
buffers, the third simulates operation decisions, the fourth serves as animation purpose together with statistical
data, and the fifth detects the occurrence of event.

19th International Conference on Computers and Industrial Engineering

329

Nevison [1989] developed a conservative scheme to coordinate simulated time in PDES for
manufacturing systems based on the examination of special structure (closed loops) which occurs in some
manufacturing systems. In the paper, the simulation model of manufacturing system is composed of eight different
basic building blocks: linear node, match node, source node, sink node, join node, fork node, transfer node and
transfer pair. A series of rules is then developed to determine whether dcadiock will occur in a closed loop
containing certain types of nodes. The algorithm was implemented on a network of transputers.
Fujii [1989] proposed a PDES model for manufacturing system. Three components are considered in the
simulation model of manufacturing system: facility (fixed component, e.g., machine centers in FMS), transa~ion
(objects which move among facilities, e.g., work pieces) and cart (transportation devices which move by itself, e.g.
AGV). The connection among facilities is described as a network, a hierarchical structure called subsystem in the
network description is introduced. Each facility is allocated on a processor, a group of facilities (processors) form
a subsystem, each subsystem has a subsystem monitor (which is also allocated to a processor and connect to every
other processors in that group), the message passes between two facilities (processors) which belong to two
differem subsystems must go through subsystem monitor. Cart is allocated to another processor called cart
monitor. The hardware platform is fifty 16-bit microprocessors (IBM 5440) connected with a mainframe
computer (IBM 3083) by serial lines (IBM 3270 lines).
/

Hon and Ismail [1991] presented some preliminary result on a PDES model of manufacturing system
based on Occam-transputer platform. Each of the entities in the simulation model is represented by an Occam
process that is executed concurrently on a transputer. The entities in manufacturing system are described as either
material handling process or operation/queue process. In addition, each process is also connect to a central Occam
process called ~"monitor" which collects information on events and the reality changes in the entities. These
processes are lt~aked using software defined channels which convey both event and the process status information.
The system is t~ed to simulate a manufacturing cell contains an AGV, a robot, three identical machine centers
and four ImffeE~ocks with all machining and operation times are deterministic. The hardware system consists of
an INMOS B 4 ~ 1 transputer board mounted on an IBM PC/AT microcomputer.
Parga~!et al. [1990] described the design and implementation of a near-term simulator for apparel
manufacturing. !,':Thesimulator use conservative schema and composed of seventeen T-800 INMOS trnsputers with
a COMPAQ deskpro 386/25 serving as front-end processor.
Bhuskute and Mize [1993] proposed a PDES model of a manufacturing system by "taking advantage of
their deterministic routing structures" (fork, join, tandem, disjoin, and fork+join). The simulation model of the
manufacturing system described as "M" number of machines processing "N" number of parts, each part can
follow a distinct route through the system before being converted into a final product. The size of the model can be
changed by simply specifying new input values for "M" and "N" and by specifying the desired routings.
Simulation execution of each submodel running on a single processor has its own event calendar which simulates
the event within the boundary of the submodel, while the dependence of the events between the submodels is
reflected by the transfer of events across the submodels. The system is implemented on Intel's ips/2 concurrent
compute~ with 32 node (a processor and distributed memory combination along with a math co-processor), each
node rffns the NX/2 operating system.
Fujii et al. [994] conduct PDES in a factory-level simulation using time bucket (a variation of time warp)
synchronization schema. In the project, a CIM system (factory) is partitioned into some areas. Each area is
modeled on one processor and simulated with conventional sequential method. Transportation system connecting
all areas is mode~l on one processor transferring network. Since the dependence among areas is not so strong as
that of cell level-~ by assigning one area to one processor, simulation can be made independently in most of the
time. The whole"!system is composed of six Sun SPARC workstations connected with ethernet. A visual factory
with two processing areas, one assembly area, and one test area is used as a test example.

3: Why PDES for Manufacturing System Not Successful


The success of PDES for a special application depends on model decomposition and model
synchronization [Overstreet et al. 1991]. As was mentioned in Section 1, all model synchronization scheme in
PDES can be classified in two categories: conservative and optimistic. This is mainly the research topic of PDES
theorists and great achievement has been obtained through their efforts. For model decomposition, it means how

CAIE 31-1/2A4

330

19th International Conference on Computers and Industrial Engineering

to break a simulation model into submodels and assign them to different processes. Although a general model
decomposition criterium is that decomposing the model in the way in which submodels are as independent as
possible, it is quite application specified. No general decomposition method exists. Different applications have
different models for their simulation object and the decomposition approaches are always different.
Modern manufacturing systems are composed of machining equipment (e.g. NC machine centers) and
storage equipment (e.g. station shuttles), connected by a transportation system (like AGV network) and operated
by a complicated computer control system. Since the interconnection among entities is very complicate in a
manufacturing system and those connections depend a great deal on the system type, the model decomposition for
a manufacturing system could be very difficult. The decomposition method in most of literature is at the "cell
level". Because of the complicate relationship among submodels, it is possible that the benefit from the
partitioning of simulation model can not overcome the costs required for the communications between processes.
This is the main reason why PDES of manufacturing system has not been condidered satisfactory. PDES models of
manufactuirng systems are oversimplified to facilitate the decomposition. For example, in Bhusknte's PDES
model for manufacturing system [Bhuskute and Mize 1993], transportation system is not considered and the
process route for each kind of parts is considered deterministic. This is totally different from the situation in a
real-world manufacturing system. Oversimplified models simply cannot be used in the real application where
detailed simulation for the system is required. It seems that some researchers' effort effort [e.g. Fujii et al. 1994] to
extend the decomposition to the "factory level" is a promising method, but the speedup (the time it takes a single
processor to perform a simulation divided by the time it takes the parallel system to perform the same simulation)
has a limit.
Another reason why PDES of manufacturing system is not satisfactory is the availability of the hardware
architectures. Table 1 is the summary of sychronization method, hardware and software platform found in the
literature. From the table, we can conclude that PDES for manufacturing systems is based on either super
computer (e.g. networked work stations or IBM parallel computers) or transputer network. Both are too expensive
for applications at an ordinary manufacturing company. This makes it unrealistic for the PDES to be used in a real
manufacturing setting.
Table 1: Sychrouization method, hardware and software in the literature
Hardware platform
Paper
Sychronization
Robert, 1984
N/A
An Intel SYS 310 computer connecting with
multiprocessin6 slave computers throush intel multibns
Nevison, 1989
Conservative
Transpnter network
Fujii, 1989
Conservative Fifty 16-bit microprocessors (IBM 5440) connected with
a mainframe computer (IBM 3083) by serial line (IBM
3270 lines)
Transputer network
Paragas, 1990
Conservative
Conservative
Transputer network
Hon, 1991
N/A
Intel ips/2 concurrent computer with 32 nodes
Bhuskute, 1993
Opitimistic
Six SUN Sparc workstations connected with ethernet
Fujii, 1994

Software
Pascal
Occam
N/A

Occam
Occam
N/A
C

Acknowledgment: The work presented in this paper has been supported, in part, by a grant from National
Science Foundation (DMI-9496337).

Reference:
18 references. A complete list of references can be obtained from the authors on request.

Вам также может понравиться