Sie sind auf Seite 1von 15

This article was downloaded by: [The University of Manchester Library]

On: 29 October 2014, At: 05:15


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number:
1072954 Registered office: Mortimer House, 37-41 Mortimer Street,
London W1T 3JH, UK

International Journal of
Production Research
Publication details, including instructions for
authors and subscription information:
http://www.tandfonline.com/loi/tprs20

Evaluating scheduling
heuristics for non-identical
parallel processors
S. U. Randhawa & C.-H. Kuo
Published online: 14 Nov 2010.

To cite this article: S. U. Randhawa & C.-H. Kuo (1997) Evaluating scheduling
heuristics for non-identical parallel processors, International Journal of
Production Research, 35:4, 969-981, DOI: 10.1080/002075497195489

To link to this article: http://dx.doi.org/10.1080/002075497195489

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all
the information (the “Content”) contained in the publications on our
platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy,
completeness, or suitability for any purpose of the Content. Any
opinions and views expressed in this publication are the opinions and
views of the authors, and are not the views of or endorsed by Taylor
& Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information.
Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities
whatsoever or howsoever caused arising directly or indirectly in
connection with, in relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access
and use can be found at http://www.tandfonline.com/page/terms-and-
conditions
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014
IN T. J. PROD . RES., 1997, VOL . 35, N O. 4, 969±981

Evaluating scheduling heuristics for non-identical parallel processors


S. U. RANDH AWA² ³ and C.-H . K UO²

Job allocation and job sequencing decisions are combined to develop scheduling
heuristics for non-identical parallel processor systems. Several factors aŒecting the
system are examined and the relative performance of the heuristics are evaluated in
terms of ¯ ow time, tardiness and proportion of tardy jobs. An understanding of
the relationship between variables in a scheduling system leads to decision rules
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

that provide feasible and eŒective production schedules.

1. Introduction
A challenge facing many manufacturing and service industries is that of assigning
jobs to non-identical, parallel processors (machines or workers). A parallel processor
is the situation where a task can be done by more than one processor but only one
processor can actually work on the task. Non-identical processors are processors that
do not have the same capacities and/or capabilities. The occurrence of parallel, non-
identical processors is quite common in both manufacturing and service industries.
An example would be a typing pool where any typist could type a document, but only
one typist can be assigned the task. Other examples include an airline assigning a type
of airplane to service a route and a textile plant assigning jobs to looms. Also,
machines or other physical resources are typically acquired over time, so that
although they may be of nominally similar type, their production rates may be
diŒerent. The focus of this study is scheduling of multiple, non-identical, parallel
processors.

2. Background
Scheduling problems arise whenever a common set of resources (e.g. machines,
operators, transporters) are required to perform a set of tasks or jobs. Each task may
be accomplished in more than one way. Given a limited set of resources, the
scheduling problem is to assign tasks to resources according to some process plan
such that the use of shared resources and other performance measures (e.g. through-
put, lateness, inventory) is optimized and production constraints (e.g. resource
capacities, precedence of operations, production requirements) are satis® ed.

2.1. T he general job shop scheduling problem


The general job shop scheduling problem is extremely complex. Although the
production scheduling problem has been studied extensively as a mathemat ical
programming formulation, the resulting models are NP-hard and can only be
solved for small and trivial cases (Fox 1983, Randhawa and McDowell 1990).
Besides mathemat ical programming, the other main approach to scheduling has
been the use of heuristic or priority rules. A large number of heuristics have been
R evision received M arch 1996.
² D epartment of Industrial and M anufacturing Engineering, Oregon State U niversity,
Corvallis, Oregon 97331, U SA.
³ To whom correspondence should be addressed.
0020 ±7543/97 $12.00 € 1997 Taylor & Francis Ltd.
970 S. U. R andhawa and C.-H . Kuo

developed for the job shop scheduling environment and studied using simulation
(Conway et al. 1967, Panwalker and Iskander 1977, D ar-El and Wysk 1982, Baker
1984, Russell et al. 1987, Randhawa and Zeng 1995). Not surprisingly, no single
scheduling rule has been identi® ed that performs best for every measure of perfor-
mance. This is partly due to the complexity of the scheduling problem and partly due
to the non-consistency of the performance measures themselves.

2.2. Non-identical parallel processors


As compared to the general job shop problem, research in scheduling of parallel
processors is limited. No general results are known for situations where a job is to be
processed by one of many non-identical machines (Conway et al. 1967).
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

There have been three major studies associated with non-identical, parallel pro-
cessors. Marsh (1973) was primarily concerned with evaluating optimum solutions to
scheduling non-identical, parallel processors to minimize total setup time. Several
optimization techniques were evaluated, but the focus was on the branch-and-bound
programming technique. The research showed that the computation time require-
ments made solving for an optimal solution prohibitive, but for the simplest systems.
In Guinet’s (1991) study of scheduling textile production systems, graph theory
algorithms were adapted to model the non-identical, parallel processor scheduling
problem. An attempt was made to minimize the mean ¯ ow time which would in turn
minimize the mean tardiness by employing the linear programming approach.
Guinet’s investigation, like Marsh’ s, included sequence-dependent setup times.
As pointed out by Smith (1993), both Marsh (1973) and Guinet (1991) studies
showed that an optimum solution for all the smallest systems was not practica l in
common, every day scheduling situations. Smith (1991) examined the factors aŒecting
scheduling a system of non-identical, parallel processors using a series of experi-
mental designs. Several factors including loading of jobs on processors, the range and
distribution of processor capacities, ranking of jobs for processor assignment, job size
distribution, and product demand distribution were examined. The results showed
that system loading and job setup times on processors play a major role in system
performance (de® ned primarily in terms of ¯ ow time and due date measures).
F urthermore, grouping jobs by product type was also found to minimize setup
times and hence reduce the mean ¯ ow time and tardiness but at the expense of
controlling individual processor usage.
Two aspects of scheduling non-identical parallel processors are investigated in this
research: developing scheduling heuristics and evaluatin g their performance for a
range of diŒerent operating scenarios. Simulation modelling was used to model and
obtain insight into the dynamics of scheduling in parallel processor operations. The
results obtained from simulation were analysed using the experimenta l design
methodology to evaluate the impact of process parameters and scheduling rules on
system performance.

3. Problem de® nition


As de® ned earlier, a non-identical parallel processor system consists of several
resources (machines, workers, etc.) that are all capable of processing a job but diŒer in
their capabilities (for example, diŒerent processing rates for the same job). A job, as
used here, is synonymous with an order and represents an individual, distinct demand
(number of units) for a product or service. Products are classi® cations of jobs. Each
product may have one, or more than one, individual job but a job can only belong to
Scheduling heuristics for non-identical parallel processors 971

one product type. F inally, tasks are sets of jobs grouped by product types. A job is
characterized by a due date that represents its promised delivery date.
Changing from one product line to another requires setup time. Setup time could
be processor-dependent, product-dependent, or both. Processor dependent setup
means that setup times depend only on the processor, regardless of the product type.
Product dependent setups imply that the setup times are a function of the production
sequence.
The system consists of several non-identical parallel processors or machines.
Processors may have same or diŒerent capacities. The system can produce a number
of products but not all product types can be produced on every processor. F urther-
more, jobs for a product have varying quantities and due dates. The scheduling
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

problem then is to:


M inimize Z subject to:
(1) Resource constraints,
(2) Technologica l constraints,
(3) Due-date constraints.
The objective function Z typically represents multi-criteria performance
measures; examples include throughput, tardiness, % tardy jobs, utilization s and
inventories. Resource constraints generally refer to processor capacities and limita-
tions. Technologica l constraints include precedence relationships (e.g. processing for
a job is completed before the next job is scheduled for production) and alternative
routeings. Due-date constraints represent delivery date commitment s for jobs.
The three performance measures considered in this study are ¯ ow time, tardiness,
and proportion of tardy jobs. Flow time is de® ned as the time a job spends in the
system from the time it is available (or ready) to be processed until it is completed.
The focus of this study is the `static’ scheduling environment, where the system is
assumed empty initially and there is a ® nite set of jobs to be scheduled. Therefore, the
¯ ow time for each job is equal to its completion time. Smaller ¯ ow time is desirable
since it indicates that a job ¯ ows through the system faster. It also means responding
to customers quickly and reducing work-in-process inventories.
T ardiness is de® ned as the positive diŒerence between completio n time of a job
and its due date. The proportion of tardy jobs measures the percentage of jobs which
are completed after their due dates. Obviously, small values of tardiness and tardy
jobs are preferred.
M athematically let:
F i = ¯ ow time for job i
T i = tardiness of job i
P T = proportion of tardy jobs
C i = completion time of job i
r i = ready time of job i
di = due date of job i
Then,
Fi = Ci - ri ,
T i = Max (0, C i - di ), and
P T = Number of tardy jobs/Total jobs
Another important measure for many production systems is the processor
utilization spread (PUS). The PUS is de® ned as the diŒerence between the heaviest
972 S. U. R andhawa and C.-H . Kuo

and least loaded processors and measures how evenly jobs are distributed among the
processors. Depending on product demand and organization goals, it may be
desirable to use all processors evenly or use a minimum number of processors and
shut down unneeded capacity. The focus in this study is on even usage where
processors are used on a regular basis such as in glass ® bre production where
the cost of cleaning processors whenever they are taken oŒproduction is extremely
high. Process utilizatio n spread is explicitly included in developing scheduling
heuristics.

4. Methodology
There are four main steps in this study: (1) Identify factors aŒecting performance
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

of parallel processor systems, (2) Develop heuristic decision rules for scheduling jobs
on parallel processors, (3) Develop a simulation model for the system, and (4) Execute
the simulation model under a range of operating scenarios and analyse the results.

4.1. Ex periment variables


Based on prior research (Smith 1993, Randhawa and Smith 1995), three groups
of factors were considered in this study: Product-related, Process-related and
Other (that depend on both product and process characteristics). The experimenta l
variables and their de® nitions are summarized in Table 1.
The product- and processor-related variables were included to investigate the
eŒect of these variables on system performance. Of speci® c interest were the number
of product groups and processors, and job size and processor capacity spreads.
Both loading and setup time depend on products and processors. Loading
represents a target utilizatio n level for the system. The actual processors’ utilization s
may vary from this target level due to uncontrollable factors present in any
environment and the di culty in achieving a perfect balance among multiple
processors.
Table 2 summarizes the values of experimenta l variables used in this study. The
factor levels in Table 2 represent 216 combinations of variables, a wide range of
operating scenarios. Statistical distributions were used to model variables, where
appropriate. A distribution’s spread or diŒerent distributions for a variable would
provide an assessment of the variable’s impact on system performance.

F actor D e® nition
Process-related
Number of processors (N MACH ) N umber of parallel processors in the system.
Processor capacity (CAPACITY) Production rate (e.g., units per hour) for a processor.
Product-related
Number of products (N PR OD ) N umber of diŒerent products to be scheduled.
Job size distribution (JOB-SIZE) Variation in demand (number of units) for a job.
Other
Loading (LOAD ) Percent of capacity scheduled for usage.
Setup time (SETU P) Time to change from one product line to another;
includes both processor-dependent and product-
dependent setups.

Table 1. Experimental variables.


Scheduling heuristics for non-identical parallel processors 973

F actor Levels F actor Values


N M ACH 3 3 5 10
CAPACITY 2 U (80,120) U (50,150)
N PR OD 3 5 10 15
JOB-SIZE 2 U (800,1200) N (1000,300)
LOAD 2 75% 90%
SETU P 3 Low (10% ) M edium (20% ) H igh (30% )

U: Uniform distribution. N : N ormal distribution


Table 2. F actor levels for experimental variables.
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

The factor levels for loading and setup time, the two variables that depend on both
products and processors, represent realistic goals for manufacturing organizations.
Two levels of loading are evaluated : a `low’ level of 75% and a `high’ level of 90% .
The low level is based on the fact that organizations generally consider utilizatio n less
than 75% to be unacceptab le; utilization higher than 90% would likely result in most
jobs being tardy. Similarly, there are three levels of setups considered: 10% (low),
20% (medium), and 30% (high) of total capacity, where the total capacity is the total
available machine time during the scheduling period.

4.2. Scheduling heuristics


Smith (1993) showed that grouping jobs by product type minimized setup times
and hence reduced the mean ¯ ow time and tardiness but at the expense of controlling
individual processor usage. In order to control the individual processor usage, the
`processor utilizatio n spread’ (PUS) is used in developing the scheduling heuristics.
F or the study, a processor utilizatio n spread of 10% or less of the maximum loading
was considered to be `even’ loading. Sensitivity of results to processor utilization
spread criterion of 10% was also investigated and is described in the result section of
this paper.
In developing the heuristics, the idea of a two-phase approach (Baker 1974) was
used. The problem of scheduling multiple parallel processors contains both allocatio n
and sequencing dimensions. Allocation means assigning jobs on processors and
sequencing is simply the order in which the jobs are processed on the processors. A
sound heuristic procedure should address both the allocatio n problem and the
sequencing problem. Thus, the ® rst step is to allocat e (or assign) jobs on processors
and the second step is to determine the optimal sequence on each processor sepa-
rately. Using this two-phase method to schedule jobs on processors may not produce
an optimal schedule, but it will tend to provide a very good schedule (Baker 1974).
The basic heuristic developed in this research consists of four phases: group jobs
by product type; assign tasks (grouped jobs) to processors, evaluate the `processor
utilizatio n spread’ (PUS), and sequence individual jobs using a sequencing rule to
meet the PUS criterion.

4.2.1. Phase 1. Group jobs by product


As mentioned earlier, a job is associated with a product. The ® rst step in the
heuristic is to group jobs by products. These grouped jobs are referred to as tasks.
Processing similar jobs (i.e. tasks) together would tend to reduce setups between
products. A negative result may be violation of due dates or excessive tardiness of
some jobs; this concern is addressed by the sequencing component of the heuristic.
974 S. U. R andhawa and C.-H . Kuo

4.2.2. Phase 2. A ssign tasks to processors


Phase 2 of the heuristic consists of the following steps:
S tep 1. Assign products that could only be run on one processor to that processor.
S tep 2. Identify products that could be run on multiple processors, but not all
processors. Order these products by decreasing number of processors that
a product can use.
S tep 3. Identify the processor with the minimum loading.
S tep 4. Schedule the `minimum-machine’ product identi® ed in step 2 to the least
loaded processor. Calculate the processor loads.
S tep 5. Repeat steps 3 and 4 until all products identi® ed in step 2 are assigned. If
there are still some unassigned products with any processor restrictions,
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

assign them to the processor with the minimum loading.


S tep 6. Assign products with no restrictions (i.e. products that can be processed
on any processor) to the least scheduled processor, one at a time.
S tep 7. After assigning a product in step 6, calculat e the processor loads and
identify the minimum loading processor.
S tep 8. Repeat steps 6 and 7 until all products are scheduled.

4.2.3. Phase 3. Evaluate processor utilization spread ( PUS)


At this stage, all tasks have been assigned to processors. The processor utilization
spread criterion (CPUS) is de® ned as 10% of the maximum processor load. Phase 3 of
the heuristic consists of the following steps:
S tep 1. Calculate the processor utilizatio n spread (PUS), de® ned as
PUS = (M aximum processor load ) - (Minimum processor load )
S tep 2. If PUS is less than CPU S, then go to phase 4 of the heuristic; otherwise, go
to step 3.
S tep 3. Divide the last task assigned in Phase 2 into individual jobs and reassign
these jobs individually to the processor with the least loading. Else, select
the last product scheduled on the most busy processor that can also be
scheduled on the least busy processor and recalculat e the PUS. If PU S still
does not meet the CPUS criterion, break up the next to the last product
group. Repeat this sequence until no more groups can be selected or
broken.

4.2.4. Phase 4. S equence jobs on processors


Within a product group, individual jobs are processed in the order generated.
Since jobs are generated randomly, this rule represents a random processing order.
Some of the performance measures may be improved by considering criterion
other than random sequencing order in phase 4 of the heuristic. There are a number of
priority rules that have been developed in scheduling. The `shortest process time’
(SPT) rule is a commonly used rule in scheduling; it is shown to give better or
comparable results to more complex scheduling rules for job ¯ ow times (Bedworth
and Bailey 1987, Conway et al. 1967, Day and Hottenstein 1970, Randhawa and Zeng
1995). Scheduling jobs by earliest due date (EDD ) rule is shown to minimize the
maximum tardiness in single-stage scheduling problems (Jackson 1955, Baker 1974).
Another measure of urgency for a job is its slack time (SS) de® ned as the time
remaining before a job must be started if it is to be completed on time (i.e. its due date
Scheduling heuristics for non-identical parallel processors 975

minus its processing time). Since the job with the minimum slack oŒers the greatest
risk of being late, it is given precedence in the schedule.
Three extensions of the basic heuristic were evaluated. These diŒer from the basic
heuristic in phase 4, where the job sequence on processors is determined using the
SPT, ED D, or SS criterion (instead of a random sequencing order).

4.3. Simulation model and parameter speci® cation


A simulation model was developed to model the system of parallel, non-identical
processors. There are three main steps in the simulation model. The ® rst step was data
generation. The second step was modelling the scheduling heuristics. The last step
consisted of calculatin g the necessary statistics and generating the output report.
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

The simulation model is designed to be general enough to model a wide class of


parallel processor systems. A user interface aids speci® cation of values for the
experimental variable de® ned in Table 1.
In addition to speci® cation of experimental variables, procedures were developed
to specify a number of other parameters. The more important of these are sum-
marized below.
· Product type: A number between 1 and number of products is generated
randomly and assigned to each job as its product type.
· Processor requirement for each product type: A product may be processed on
one or more processors. A random number was used to determine the number
of processors on which a product can be processed (alternative routeing). A
second set of random numbers then identify the speci® c processors. To
illustrate, assume that the ® rst random number generated for a product is 2.
This means that this product can be processed on two processors. Now, two
new random numbers are generated, say, 1 and 4, implying that the product can
be processed on either processor 1 or processor 4.
· Process time for jobs: Since a job may be scheduled on more than one processor,
there may be multiple process times per job (one for each processor). The
process time used in scheduling depends on the processor to which the job is
assigned. A processing time for a job on processor i, P i , is given by
Quantity
Pi =
Capacity of processor i
F or example, a job consisting of 1000 units would require 8.33 hours on a
processor with a production rate of 120 units per hour (1000/120).
· Jobs due dates: Due date assignment for job i (di ) is speci® ed as a product of two
parameters: di = F * U (1,2), where F is a sample from the uniform distribution
over the range {maximum processing time, scheduling time span} and U (1,2) is
a uniform random variable between 1 and 2.
· L oading: To determine the number of jobs required to reach a speci® c loading
level, a loading level is speci® ed and jobs are accepted for scheduling until the
cumulative job quantity exceeds the desired loading level.
Once the ® nal schedule is obtained, statistics associated with the performance
measures are computed. Both the average value and variation for each measure are
reported.
The simulation model was implemented in F ORTR AN 77 (Kuo 1994). The
simulation was executed using an 80486-DX computer operating at 33 MHz. Based
976 S. U. R andhawa and C.-H . Kuo

on the factor settings in Table 1, there were 216 treatment s per run. The time needed
to run the simulation depends on the job size. With 3 processors and an average of 130
jobs for 90% loading, it took approximately 3 minutes to obtain a schedule. With 5
processors and 220 jobs for 90% loading, it took approximately 5 minutes. F or 10
processors, the execution time increased to around 15 minutes to complete 430 jobs.
Two runs per simulation were executed. The average of these two runs was used in the
statistica l analysis.

4.4. A nalysis
The `analysis of variances’ (AN OVA) technique is used to analyse the results.
Pairwise t-tests are used to compare the heuristics. All results are based on a 95% con-
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

® dence level. Statgraphics 5.0 (Statgraphics 1991), a commercia l software package, is


used to perform the ANOVA analysis. All interaction s greater than second order were
ignored because including all possible interaction s requires the use of all 2n degrees
of freedom which eliminates the possibility of correcting for experimenta l error.
Another problem with higher interactio n is di cultly in interpreting their meaning.

5. Discussion of results
The ANOVA results for the basic heuristic and the three extensions are
summarized in Table 3.

5.1. R esults for the basic heuristic


In the basic heuristic, within a product group individual jobs are processed in the
order generated. The three variables, number of products (N PROD), loading level
(LOAD) and setup time (SETUP), are identi® ed as being signi® cant for all three
performance measures.

F lowtime Tardiness Tardy Jobs


Variables ran spt edd ss ran spt edd ss ran spt edd ss
1 ´ ´ ´ ´ ´ ´ ´ ´ ´ ´
2
3 ´ ´ ´ ´ ´ ´ ´ ´ ´ ´ ´ ´
4 ´ ´ ´ ´ ´
5 ´ ´ ´ ´ ´ ´ ´ ´ ´ ´ ´ ´
6 ´ ´ ´ ´ ´ ´ ´ ´ ´ ´ ´ ´
1 *3 ´ ´ ´ ´ ´ ´ ´ ´ ´ ´
1 *4 ´ ´
1 *5 ´ ´ ´ ´ ´ ´ ´ ´
1 *6 ´ ´ ´ ´ ´ ´ ´ ´ ´
2 *4 ´ ´ ´ ´
2 *5 ´ ´ ´ ´ ´ ´ ´ ´
2 *6 ´ ´ ´ ´
1 N M ACH ´ indicates signi® cant eŒect
2 CAPACI TY ran R andom sequencing (refers to basic heuristic)
3 N PR OD spt Shortest Processing Time sequencing
4 JOB-SIZE edd Earliest D ue D ate sequencing
5 LOAD ss Shortest Slack sequencing
6 SETU P
Table 3. Summary of results.
Scheduling heuristics for non-identical parallel processors 977

Increasing the number of products increases the setup time because setput time
includes product dependent changeover time. This in turn changes the completion
time of the job, thus aŒecting all three performance measures.
The number of jobs needed to reach a desired loading level increases as the target
loading level increases. Thus the time to complete all jobs (referred to as makespan or
total production time) increases. Increased loading also translates into more jobs in
the system, with subsequent increase in tardiness and proportion of tardy jobs.
Similarly, increasing the setup time increases the job completion times and the
number of jobs waiting to be completed.
Besides the three variables mentioned above, NMACH (number of machines) was
signi® cant for ¯ ow time and tardiness. As the number of machines increase, the
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

number of jobs needed to reach a speci® c loading level increases. Since the completion
time equals the ¯ ow time (ready time for all jobs is zero), the mean ¯ ow time will
increase and tardiness will be aŒected. It is surprising that this variable is not
identi® ed as being signi® cant for the proportion of tardy jobs. This may be due to
other variables having such a dominant in¯ uence that the eŒect of NMACH cannot
be detected.
The other two variables, machine capacity (CAPACITY), and job size distribu-
tion (JOB-SIZE) are not signi® cant for any of the performance measure. F or the
processor related variable, CAPACITY, the mean for the diŒerent factor levels is the
same (refer to Table 2). Thus, the diŒerence tested for this variable is for the spread of
capacity between processors. As processor capacity spread increases, more of the
workload is shifted to the faster processors. Consequently, the schedule is not
in¯ uenced by the capacity spread distribution. The job size variable is also not
signi® cant, primarily due to the mean speci® cations for factor levels being the same.
F or the signi® cant variables, NMACH, NPROD, LOAD and SETUP, all two-
factor interactions were examined. There are no signi® cant interaction s for propor-
tion of tardy jobs. The interaction s between NM ACH and the other three variables, and
the interaction between LOAD and SETUP are signi® cant for ¯ ow time and tardiness.
The ANOVA results for ¯ ow time (Table 4) show the F -ratio for the two-way
interaction between N MACH and N PROD to be signi® cantly higher than the
corresponding values for other interaction s. There is a complex interactio n between

Sum of M ean
Source of variation squares df square F -ratio
M ain eŒects
N M ACH 41 826. 67 2 20 913 .34 151 .45
N PR OD 21 017 .88 2 10 508 .94 76.10
LOAD 87 824 .63 1 87 824 .63 636 .01
SETU P 67 582 .70 2 33 791 .35 244 .71
Interactions
N M ACH * N PR OD 151 608. 56 4 37 902 .14 274 .48
N M ACH * LOAD 1 822. 46 2 911 .23 6.60
N M ACH * SETU P 4 366. 17 4 1 091 .54 7.91
N PR OD * LOAD 1 940. 94 2 970 .47 7.03
Residual 27 065. 24 196 138 .09
Total 405 055. 24 215

Table 4. AN OVA results for ¯ ow time for the basic heuristic.


978 S. U. R andhawa and C.-H . Kuo

these two variables. With fewer machines, products need to wait longer to be
scheduled, resulting in longer completion times. As the number of machines increase
in relation to the number of products, more machines are available for processing;
thus jobs should ¯ ow through the system faster. However, recall that more machines
also mean more jobs to reach the same loading level. This results in higher setups.
F urthermore, job groups are split in the last phase of the heuristic to achieve a balance
in processors’ utilization . Groups can only be split to a certain level. Thus too few
products compared to the number of machines results in an imbalance in processors’
usage as utilization s of some processors may be relatively low while jobs wait to be
processed on other machines.
It is not surprising that the interaction s between NMACH and LOAD and
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

between NMACH and SETUP are signi® cant. As mentioned earlier, increased
machine utilizatio n can only be achieved at the expense of higher ¯ ow time, when
all other factors are held constant. Setup time has a signi® cant eŒect on ¯ ow time; the
¯ ow time increases when setup time increases.
The four interactio ns which are signi® cant for ¯ ow times are also signi® cant for
tardiness. For example, consider the interactio n between the number of machines and
loading. An increase in loading from 75 to 90% results in a 40% increase in tardiness
for all settings of the number of machines. Higher loading levels imply more jobs waiting
to be processed to meet target loading. Similarly, higher setup times will cause increase
in completion time of jobs. With ® xed due dates, this will result in increased tardiness.

5.2. Comparison of heuristics ( Table 3)


The SPT heuristic sequences jobs in Phase 4 of the heuristic by non-decreasing
process times. Since process time is determined by the ratio of the job quantity to a
processor’s capacity, sequencing jobs using SPT may be aŒected by processor
capacity or job size. Since the mean for all processor capacities is the same, this
eŒect is not signi® cant. H owever, the eŒect of job size is signi® cant for all the
performance measures.
The results for the EDD heuristic are similar to the basic heuristic. In EDD, jobs
within tasks are sequenced based on the earlier due date. Since job due dates are
assigned randomly over the interval {maximum processing time, scheduling time
span}, the job processing sequence approaches a randomly generated sequence.
F or the SS heuristic, jobs are sequenced by the slack time which is related to a job
due date and processing time. Therefore, job quantity is signi® cant, as with the SPT
heuristic. With both SPT and SS heuristics, interaction sinvolvingjob size aresigni® cant.
Pairwise t-tests are used to test the hypothesis whether there is any signi® cant
diŒerence in performance measures obtained from each heuristic based on a 95%
con® dence level. The null hypothesis tested is de® ned as (l i - l j ) = 0, where l i and l j
are the mean values of a performance measure using heuristics i and j, respectively.
The comparison results are summarized in Table 5. The only case where the null
hypothesis is not rejected is for the ¯ ow time comparison between the basic heuristic
and EDD. This shows that sequencing rules do aŒect the performance of the system.
The relative performance of heuristics results in the following rank order:
Flow time: SPT < ED D = RAN < SS
Tardiness: EDD < SPT < RAN < SS
Tardy jobs: EDD < SPT < RAN < SS
Scheduling heuristics for non-identical parallel processors 979

Hypothesis F low time Tardiness Proportion of tardy jobs


RAN -SPT 12.73 8.91 8.79
RAN -ED D 1.03* 16.44 14.50
RAN -SS - 14.76 - 9.42 3.74
SPT-ED D - 16.02 18.03 9.88
SPT-SS - 16.32 - 11.13 - 11.80
ED D -SS - 13.71 - 17.05 - 15.13

*Indicates not rejecting the null hypothesis.


Table 5. R esults from pairwise t-tests.
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

Heuristic CPU S % M ean ¯ ow time M ean tardiness Prop. of tardy jobs


RAN 0 284 .53 77.11 0. 4236
10 253 .53 65.82 0. 3798
20 253 .05 67.00 0. 3673
SPT 0 283 .11 76.94 0. 4188
10 249 .10 64.33 0. 3676
20 247 .62 62.40 0. 3678
EDD 0 296 .04 66.17 0. 4020
10 254 .71 51.87 0. 3151
20 253 .08 48.92 0. 3136
SS 0 292 .58 80.86 0. 4498
10 259 .27 69.56 0. 3827
20 258 .36 68.58 0. 3730

Table 6. Sensitivity of results to PU C criterion.


(NM ACH = 10, N PR OD = 25, LOAD = 90% , SETU P = HIG H).

The use of SPT results in minimum ¯ ow time. This is consistent with past
scheduling research where SPT is shown to be a superior sequencing rule in
minimizin g the mean ¯ ow time in the single machine system (Baker 1974), and in
¯ ow shop problems (Rowe 1958, Conway 1965, and Baker 1984).
The rankings for the two due-date criteria, tardiness and proportion of tardy jobs,
is identical. The EDD heuristic performing better than other heuristics for tardiness is
also consistent with prior scheduling research (see for example, Jackson 1955). The
number of tardy jobs is minimized by the Hodgson’s Algorithm (Baker 1974).
Basically, the algorithm is scheduling jobs using the EDD order. Sequencing jobs
by non-decreasing slack time maximizes the minimum tardiness (Conway et al. 1967).
Hence, the SS heuristic is not a good choice if the objective is to minimize tardiness or
proportion of tardy jobs.

5.3. Sensitivity of results to PUS criterion


The `processor utilizatio n spread’ (PUS) criterion used in the results described
above is 10% of the maximum processor load. In order to test the sensitivity of results
to this criterion, the system is evaluated using PUS criteria of 0% (breaking groups
into individual jobs) and 20% of the maximum processor load. The results from
simulation experiments for the scenario, NMACH = 10, NPROD = 25, LOAD =
90% and SETU P = HIGH, are summarized in Table 6.
980 S. U. R andhawa and C.-H . Kuo

M ean M ean Proportion of


Heuristic ¯ ow time tardiness tardy jobs
SPT 237. 93 60.23 0.3458
EDD 253. 35 47.22 0.3140
Hybrid (wSP T = wEDD = 0. 5) 243. 70 47.73 0.3136

Table 7. Comparison of SPT, ED D and hybrid heuristics.


(NM ACH = 10, N PR OD = 25, LOAD = 90% , SETU P = HIG H).

With no criterion (0% ), all processors are equally loaded. However, once groups
are broken into individual jobs, the proportion of setup time increases; thus ¯ ow time,
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

tardiness and proportion of tardy jobs increase. The computation time needed for
scheduling increased by approximately 10% compared to the 10% criterion. In
contrast, use of the higher criterion did not result in any signi® cant change in the three
performance measures.

5.4. A `H ybrid’ heuristic


Trying to `optimize’ performance measures simultaneously is generally not
feasible as some of the measures con¯ ict with others. F or example, high processor
utilizatio n can only be achieved at the expense of high ¯ ow times and more jobs
waiting in the system. The aim of a hybrid (or weighted) heuristic is to provide a
balance between ¯ ow time and due date measures.
In the hybrid heuristic, the rank order of processing of jobs is computed from the
following relationship:
Rank = wSP T * R SPT + wEDD * R EDD ,
where wSPT and wEDD represent the weights assigned to the two objectives, minimize
¯ ow times and minimize due dates (wSP T + wEDD ) = 1, and R SPT and R EDD are the
ranks of jobs when they are scheduled using SPT and EDD heuristics, respectively. If
SPT heuristic is used exclusively, wSPT = 1 and wEDD = 0. Similarly, use of EDD
heuristic implies wSPT = 0 and wEDD = 1. The hybrid heuristic schedules the
jobs within each product group by the rank determined by the weights, wSPT and
wEDD . A number of simulation experiments for diŒerent scenarios were executed. The
results for NMACH = 10, NPROD = 25, LOAD = 90% and SETUP = HIGH are
summarized in Table 7.
H igher weight on SPT results in better ¯ ow times whereas the due date results
improve with higher weight on EDD. Depending on managemen t objectives, assign-
ing appropriate weights to SPT and EDD criteria can provide a compromise as with
wSPT = wEDD = 0. 5 in Table 7.

6. Conclusions
Through this research, an extensive, systematic analysis of a parallel, non-
identical system is carried out. An understanding of the relationship between parallel
processors, products and scheduling system leads to the decision rules that can
provide feasible and eŒective production schedules. The schedules may not yield an
optimal solution, but solving the problem analytically to obtain an optimal solution is
diŒerent and impractical. The approach discussed in this paper uses a multi-stage
heuristic process for scheduling jobs for non-identical parallel systems. The use of
Scheduling heuristics for non-identical parallel processors 981

hybrid heuristics can aid in combining diŒerent managemen t objectives in the


decision making process.

References
BAK ER , K. R ., 1974, Introduction to Sequencing and S cheduling (New York: Wiley).
BAK ER , K . R ., 1984, Sequencing rules and due-dates assignments in a job shop. M anagement
Science, 30(9), 1093 ±1104.
BEDWOR TH , D . D., and BAILEY , J. E., 1987, Integrated Production Control S ystems:
M anagement, A nalysis and Design (N ew York: Wiley).
C ON WAY , R . W., 1965, Priority despatching and work-in-process inventory in a job shop. T he
Journal of Industrial Engineering, 16(2), 123±130.
C ON WAY , R. W., M AXWELL , W. L., and M ILLER , L. W., 1967, T heory of Scheduling (R eading,
Downloaded by [The University of Manchester Library] at 05:15 29 October 2014

M A: Addison-Wesley).
D AR -E L , E. M ., and WYSK , R . A., 1982, Job shop schedulingÐ a systematic approach. Journal
of M anufacturing Systems, 1(1), 77±88.
D AY , J. E., and H OTTENSTE IN , M . P., 1970, R eview of sequencing research. N aval R esearch
L ogistics Quarterly, 17(1), 118±146.
F OX , M . S., 1983, Constraint-directed search: a case study of job-shop scheduling. Technical
report, CM U -R I-TR-83-22/CM U -CS-83-161, Carnegie-M ellon U niversity.
G U INET , A., 1991, Textile production systems: a succession of non-identical parallel processor
systems. Journal of the Operational Research Society, 42(8), 655±671.
J ACKSON , J. R ., 1955, Scheduling a production line to minimize maximum tardiness. R esearch
R eport 43, M anagement Sciences R esearch Project, U niversity of California at Los
Angeles.
K U O , C.-H ., 1994, Evaluation of scheduling heuristics for non-identical parallel processors, M S
thesis, Oregon State U niversity, Corvallis, OR .
M AR SH , J. D ., 1973, Scheduling parallel processors. PhD thesis, G eorgia Institute of
Technology, Atlanta, G A.
P ANWALKE R , S. S., and I SK AN DER , W., 1977, A survey of scheduling rules. Operations Research,
25(1), 45±61.
R AN DH AWA , S. U., and M C D OWELL , E. D ., 1990, An investigation of the applicability of expert
systems to job shop scheduling. International Journal of M an-M achine Systems, 32(2),
203±213.
R AN DH AWA , S. U ., and SMITH , T. A., 1995, An experimental investigation of scheduling non-
identical parallel processors with sequence-dependent set-up times. International Journal
of Production R esearch, 33(1), 59±69.
R AN DH AWA , S. U ., and Z ENG , Y., 1995, Job shop scheduling: an experimental investigation of
the performance of alternate scheduling rules. Production Planning and Control, 7(1)
47 ±56.
R OWE , A. J., 1958, Sequential decision rules in production scheduling. PhD thesis, U niversity of
California Los Angeles, CA.
R USSELL , R . S., D AR -E L , E. M ., and T AYLOR , III, B. M ., 1987, A comparative analysis of the
COVER T sequencing rule using various shop performance measures. International
Journal of Production R esearch, 25(10), 1523 ±1540.
SM ITH , T. A., 1993, An experimental investigation of scheduling non-identical parallel
processors with sequence-dependent set-up times and due-dates, M S thesis, Oregon
State University, Corvallis, OR.
Statgraphics, 1991, Statgraphics User M anual, version 5, STSG Co.

Das könnte Ihnen auch gefallen