benefits and drawbacks of simple models for complex
TRANSCRIPT
Oliver Rose
Dresden University of TechnologyDepartment of Computer ScienceChair for Modeling and Simulation
Benefits and Drawbacks of Simple Models
for Complex Production Systems
2
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Dresden University of Technology
Dresden located in the state Saxony, often called “Silicon Saxony” (factories from Infineon, Qimonda, AMD, and ZMD; lots of suppliers)Largest German University of TechnologyFocus on Computer Science, Electrical Engineering, and Mechanical EngineeringAbout 35,000 students, 3,000+ students in Computer Science, 600+ beginners in winter semester 2006/2007Faculty of Computer Science consists of 28 chairsInstitute of Applied Computer Science has 5 chairs, dealing with different aspects of factory planning, control, and automation
3
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Research Overview
Complex Production System
Model
Simulation
Dispatching Simulation-based
Scheduling
Me
asu
rem
en
t &
An
aly
sis
Au
tom
ati
on
of
Da
ta T
ran
sfe
rHigh-level Production Control
4
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Cooperations
IndustryInfineon Technologies AG, Munich and DresdenAMD Saxony LLC & Co. KG, DresdenKBA Planeta, Radebeul (Offset Printing Machines)Airbus Deutschland GmbH, Hamburg
AcademiaArizona State University, Tempe, AZ, USASingapore Institute of Manufacturing TechnologyGeorgia Institute of Technolgy, Atlanta, GA, USAFernUni Hagen, GermanyFraunhofer IPA, Stuttgart, Germany
5
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Semiconductor Manufacturing
Back End
Front End
Wafer Fab
TestAssembly
Probe
6
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Flow of material
Fotos: Fullman-Kinetics, Varian, Sematech International
7
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Characteristics of wafer fabrication facilities
Large number of processing steps, typically several
hundreds
Large number of tools of different types: photo
equipment, ovens, etching equipment, ion implanters, …
Wafer are build up in layers: reentrant flow of material,
jobshop type way of production
Machine breakdowns (typical availability: 70-90%)
Auxiliary resources, e.g., reticles (photo masks)
Batch tools with complex batching criteria
Sequence dependent setups
8
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Important performance measures
Cycle time: low
Output: high
Machine utilization: high
Inventory (work in process, WIP): low
Yield (percentage of good dies on a wafer): high
Cost per die: low
Conflicting goals!
9
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Motivation for simple models
Traditionally, only full detail models used for operational planning and control of semiconductor fabs
Consequences:Long run times of simulation experimentsLong run times of scheduling algorithmsToo complex to be included in enterprise models for SCM (Supply Chain Management)
Need for simple fab models
10
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Delay
Simple modeling approaches
RequirementsCorrect representation of characteristic curve (cycletime-over-utiliziation curve), i.e., typically 1/(1-utilization) shapeSame cycle time distributions as for real fabMimic typical behavior of fab over time
Very simple model: cycle time distributionDoes not depend on utilizationHas no capacity
0
5
10
15
0 0.2 0.4 0.6 0.8 1
11
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Simple modeling approaches
Simple queuing systemBehavior over time not appropriateIn general, shape of characteristic curve problematic
Delay DelayBuffer
Bottleneck
12
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Simple modeling approaches
Simple queuing system with loop(re-entrant flow of material)
Delay Delay
Bottleneck
Delay
?
13
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Mimic behavior of real fab
Goal of the study: Estimate the transient behavior of
a wafer fab after recovery from a catastrophic
failure
Motivation: fab engineers reported that large amounts of
WIP are present in the fab even weeks after end of repair
Problem of the analysis: estimation of the required
inventory curves not feasible with full model; hundreds of
replications of experiment needed
14
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Simple fab model
fixed number of cycles
bottleneckworkcenter
delay unit(rest of the fab)
finished?
yes
no
lot release
15
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Model details
Dispatching strategies:
FIFO
Critical Ratio
Slack Time
Delay time distributions of the delay unit
Lot release policy
Partial bottleneck breakdowns
16
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Modeling the Rest of the Fab
?
17
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Delay time distributions
Constant amount of time for sum of processing times
Distributions of different variability for sum of non-
processing times (waiting times, transport times, etc.):
Constant (coefficient of variation: 0.0)
Erlang-5 (coefficient of variation: 0.447)
Exponential (coefficient of variation: 1.0)
Independence of consecutive and parallel delays assumed
18
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Constant Delay
FIFOCritical RatioSlack Time
time
WIP
19
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Shifted Exponential Delay
FIFOCritical RatioSlack Time
time
WIP
20
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Shifted Erlang Delay
FIFOCritical RatioSlack Time
time
WIP
Dura
tion o
fbre
akdow
n
21
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Delay Distribution from Real Data
0
0.01
0.02
0.03
0.04
0.05
0.06
0 20 40 60 80 100 120 140hours
datagamma
22
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Characteristic curve
Utilization
Simple
Full detail
Flow
fac
tor
23
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Model improvement approach
Make delays load dependent! But how to measure load?
Delay Delay
Bottleneck
Delay
?
24
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Load measurement
Simply count lots in loop!
Delay Delay
Bottleneck
Delay
?
Lots in loop
25
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Generation of the load dependent data
Very long simulation run with stepwise (5% steps) increase of fab loadCollection of the delays for given range of lots in loop (interval width of 5 lots)
Simulated time [days]
Lots
in loop
26
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Load dependent “loop” delays
Lots in loop
Del
ay [
min
]
27
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Delay time percentiles
Lots in loop[100,124]
Lots in loop[25,49]
Probability
Del
ay [
min
]
In most cases, very close to exponential (correlation > 90%)
28
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Improvement of the characteristic curve
Utilization
Simple
Full detail
Flow
fac
tor
Load dependent
29
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Estimation of cycle time distributions
Detailedmodel
Simplemodel
30
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Estimation of cycle time distributions
0
0.005
0.01
0.015
0.02
0.025
0.03
0 100 200 300 400 500 600 700 800 900hoursHours
0
0.005
0.01
0.015
0.02
0.025
0.03
0 100 200 300 400 500 600 700 800 900hoursHours
0
0.002
0.004
0.006
0.008
0.01
0.012
700 750 800 850 900 950 1000 1050hours
fullsimple
FullSimple
Hours
FIFO
0
0.01
0.02
0.03
700 800 900 1000hours
fullsimple
FullSimple
Hours
CR
31
Dresden University of Technology
Department of Computer Science
Oliver Rose
Simple Models
Conclusions
Simple models useful for analyzing and understanding
complex production systems
Not a tool for beginners
Not appropriate for all problems
Pitfall of oversimplification
Simplification must not be the goal but only the method
to reach the goal
Keep the model as simple as possible but not simpler!