PowerPoint Slides (4 MB)

advertisement
Automation components
for
simulation-based engineering
October 2008
simulation-based engineering challenges
 Use of physics simulation as an
integral part of the design process
 Use of simulation early and often in
the design process
 Use of simulation to evaluate design
functional objectives
 Use of simulation to affect design
decisions
October 2008
simulation-based engineering challenges
 Increasing complexity
 Structural Analysis
 Thermal Analysis
 Computational Fluid
Dynamics (CFD)
 ElectroMagnetic Analysis
 Radiation Analysis
 Coupled Physics
 MEMS
 Application specific




 ...
October 2008
Turbo-machinery
Power generation
Combustion engines
Chemical Mixing
simulation-based engineering challenges

objectives of physics
simulation in the design
process are changing
 Drive reusable Analysis Data
Models from high level
requirements through detailed
analysis.
 Concept to detail design
phases
 Provide a means to support
robust design, systems
engineering, functional design
and design space exploration
for performance investigations.
 Provide Analysis data definition
to enable capture and reuse of
expertise.
October 2008
simulation-based engineering environment
Simulation Driver Frameworks
InstanceIndependent
Simulation
Definition
Simulation
Instance
Model
Simulation
Execution
Instance
Model
(Analysis
Abstract Model)
Solve for Frameworks
(Robust Design / Design Exploration / Systems Engineering)
Simulation
Results
Instance
Data
Data Management Frameworks (PLM/SDM)
Rules, Processes,
Templates
October 2008
Object Definition Instance
Data
(CAD, Mesh Based,
Concept, Non-Geometric)
Instance Based
Simulation Data
(Execution
Ready/Results Data)
Simulation Models
Simmetrix approach
 Provide software components to enable automation to
generate accurate simulation results applicable to design
decisions directly from the current Design Model instance on
a repetitive basis through the design process







October 2008
Abstract Model
Simulation Model
Direct geometry access
Automatic mesh generation
Automatic generation of run-ready data
Results management
Adaptive mesh modification
analysis abstract model
 Instance Independent Simulation Definition
Made up of Four Major Aspects
Physical Characteristics
Solution Specific Definitions
Conceptual Model
October 2008
Instantiation
Rules
and
Processes
analysis abstract model
 Instance Independent Simulation
Definition
 Physical characteristics
 Physical constraints and properties that are
instance independent
 Material definitions/libraries
 Physical constraints
 others
 Defined / utilized as attributes assigned to a
global / nothing Class in the Conceptual
Model
October 2008
analysis abstract model
 Instance Independent Simulation
Definition
 Solution Specific Definitions
 Definitions of the specific problem to be solved
 Defined as attributes assigned to data objects
(or global class) in the conceptual model
 Can be expanded to include definition of
derived results (Performance Requirements)
calculation
October 2008
analysis abstract model
 Abstract (Conceptual) Model
 Tag based approach
 Tags placed on our associated with Object Definition Data
 Typically string parameters or attributes
 Tags can be assigned to Assemblies, subassemblies, parts,
features, and explicit Faces
 Auxiliary file used for Discrete (mesh) data
 Benefits
 Independent of complexity of problem domain
 Can work with object definitions from multiple sources
 Limitations & Issues
 Requires creation and maintenance of persistent tag data
 Difficult to maintain tags at anything less than a feature
 (ie a-pillar flange )
 Applicable to a limited domain of problems
October 2008
analysis abstract model
 Abstract (Conceptual) Model
 Tag based approach
 Simple Heat Exchanger example
 CFD simulation defined once and alternate designs
instanced
 first design could be in CAD system A and second design in
CAD system B
 Boundary Layer meshing and wall boundary conditions applied
to all appropriate faces
October 2008
analysis abstract model
 Abstract (Conceptual) Model
 Abstract Reasoning based approach
 Abstract Geometry
 Component Functions & Filters
 Result in Component / Abstract Geometry
 Can be used to drive complex “rule” like structure (ie matching edge
loop pairs for pin or bolt hole locations)
 Operations on Components / Abstract Geometry
 Relations, Groups, Functions & Filters
 Result in Component / Abstract Geometry
 Can be nested to form complex abstract objects
 Benefits




Independent of object data structure (ie CAD features)
Removes need for tags
Can work with object definitions from multiple sources
Applicable to a broad domain of problems
 Limitations & Issues
 Complexity of problem domain determines complexity of Abstract
Reasoning
October 2008
analysis abstract model
 Abstract (Conceptual) Model
 Abstract Reasoning based approach
 Decklid example
 Decklid analysis starting with existing NASTRAN mesh
models
 Loads & boundary conditions defined abstractly and located
based on each instance
October 2008
analysis abstract model
 Abstract (Conceptual) Model
 Mixed approach
 Abstract Reasoning + Tags
 Best of both worlds approach
 Benefits
 Leverages knowledge that is available
 Minimizes need for tags to what is easy & appropriate to tag
 Reduces complexity of Abstract Reasoning by only using
Abstract Reasoning where tags are not appropriate
 Independent of object data structure (ie CAD features)
 Removes need for tags
 Can work with object definitions from multiple sources
 Applicable to complete domain of problems
 Limitations & Issues
 Requires planning when & what tags are appropriate
 Yet another layer of abstraction
October 2008
analysis abstract model
 Instance Independent Simulation
Definition
 Instantiation Rules and Processes
 Transformation mappings from various
object definition instance representation
types to valid simulation instance models
(implicit).
 Includes instantiation of relations data
 Includes instantiation of derived data
 inverse space for CFD/EM
 “sliver” feature removal
 Different for each object definition instance
representation type
 Included as part of GeomSim modules for
supported object definition representation
types
October 2008
simulation model
 The Object Definition Instance (“Design Model”) is
transformed into a non-manifold topology (Simulation
Model)
 Solids that touch share common faces and edges at the contact
interface
 Resulting interface faces may have material on both sides
 Allows for single or set of mesh entities at the interface
 Attributes may be used to create duplicate mesh entities at the
interface
 Attributes are recast from the Abstract Model to the
Simulation Model for the current Design Model instance
October 2008
simulation model
 A Unified Topology Model is created independent of
geometry source (also works with discrete models – stl/mesh models)
 Initially generated from Design Model instance
 Provides interrogations of geometry via direct access of
the Design Model
 Topological adjacencies, point classification, surface evaluation
(points, derivatives, normals, etc.), closest point queries, etc.
 Assembly modeling
 Represent assembly models as non-manifold model even if the
underlying modeling engine does not support this
 Allows for creation and modification of simulation related topology
 Suppression of “small” features
 Addition of bounding boxes
 Symmetry planes
 Recognition of void regions that are not explicitly defined in the
modeling source
 Simulation Model topology does not have to exactly
match the Design Model topology
October 2008
automatic mesh generation
 Mesh control attributes assigned to the Abstract Model
are mapped to the appropriate entities in the current
Simulation Model instance.
 Supports non-manifold topology models
 embedded vertices, edges, faces
 Maintains relationship of mesh entities to Simulation
Model topology
 Provides ability to put a full or partial mesh on a model
entity
October 2008
automatic mesh generation
 Fully automatic mesh
generation for surfaces
and solids
 Triangular, quadrilateral
and mixed surface meshes
 Tetrahedral volume
meshes
Courtesy Top Systems Ltd
Courtesy Ford Motor Company
October 2008
Courtesy Infolytica Corporation
automatic mesh generation
 Curved mesh generation
 Supports meshes of higher
order elements that
capture geometry
Courtesy Top Systems Ltd.
October 2008
automatic mesh generation
 Matched meshes for periodic boundary conditions
Courtesy Infolytica Corporation
October 2008
automatic mesh generation
 Boundary Layer meshing with
edge blends
October 2008
automatic mesh generation
 Extrusion meshing
 Extrude mesh
between two faces
with similar topology
 Supports
generalized
curvature between
faces
October 2008
automatic mesh generation
 Crack Tip meshing
 3D edge blends along crack tip edge
October 2008
automatic mesh generation
 Extensive mesh refinement control
 Specified refinement size
 Absolute value on model, model entity or
location in space
 Relative value on model or model entity
 Function based on location in space
 Boundary layer growth rate
 Curvature based mesh refinement
 User defined refinement
October 2008
 A unified representation of
simulation results data
that is independent of the
physics solver used
 Provide result feedback in
terms of design objectives
and criteria
 Provide improved data for
visualization software
 Provide high level access
and query functions
October 2008
results management
 Expressions
 Are based on operations on one or more fields
 Can be created from multiple fields
 Fields may be on the same or different meshes
 Can create new fields
 Can store what the field represents
 Can be evaluated over any part of the domain
 Can be evaluated over Components or Classes in the
Abstract Model
 Can be used to express results in terms of design
objectives (e.g. comfort index, bearing force, power
drop, …)
October 2008
results management
 Supports mapping solutions
between meshes




October 2008
Different physics
Same physics but different mesh
Adaptive mesh modification
Solution migration during
repartitioning
adaptive mesh modification
October 2008
adaptive mesh modification
 Provides refinement and
coarsening of existing
mesh
 Ensures new nodes on
boundary are placed
correctly on geometry and
mesh is valid
 Enables anisotropic target
mesh
October 2008
geometry based parallel
mesh generation & adaptivity
 Growing need to solve larger and larger problems





October 2008
Fluid domain applications (automotive & aerospace)
Biomedical applications
Environmental applications
Electromagnetic applications
Coupled applications
geometry based parallel
mesh generation & adaptivity
 Growing need to solve larger and larger problems
 Tens of millions of elements are becoming prevalent
 Hundreds of millions of elements are becoming common
 Applications requiring billions of elements are appearing
October 2008
geometry based parallel
mesh generation & adaptivity
 Solvers have made significant advances in parallelization
 Parallel CFD and EM solves are quite common
 Recent advances have shown excellent scaling results on large
paralleled clusters (ref. Ken Jansen)
 Meshing Technology has not kept up with solver technology
in the area of Parallel computing
 Some advances have been made in Parallel mesh adaptation
 Some work has been done in Parallel mesh generation
 A few applications have Distributed memory parallel meshing available
 A few more applications have Shared memory parallel meshing available
 Most applications start the parallel meshing with an initial facetted model
 Mesh generation for the large scale problems is clearly
becoming the bottleneck
October 2008
geometry based parallel
mesh generation & adaptivity
 Parallel mesh generation brings
with it its own set of problems /
issues
 For generalized meshing of arbitrary
geometry the problem is ill formed
for parallel computing
 An added complexity is that the
intent for CFD, EM and other farfield applications is to model the
space defining the field volume with
one or more complex volumes
 Just splitting the workload by parts in
an assembly is not appropriate
October 2008
geometry based parallel
mesh generation & adaptivity
 Accurate solutions require
accurate capture of geometry
 Curvature based refinement is
commonly used in serial mesh
generation applications
 Small details may be of interest
 Using a predefined facet model
may not be accurate enough for
the critical areas of interest
 Accurate capture of geometry
requires direct geometry access
as part of the parallel mesh
generation
 Creating a highly accurate facet
model is a serial operation and
would quickly become the new
bottleneck
October 2008
geometry based parallel
mesh generation & adaptivity
 Shared Memory .vs. Distributed Memory
 64 bit processors and multi-core systems have raised the
question of Shared Memory (multi-threaded) or Distributed
Memory (MPI) architectures
 The answer is basically related to the size of mesh required
 The limitation on Shared Memory Parallel has moved from the
addressable memory space to the amount of physical memory
available
 The main issue with Distributed Memory Parallel is to get enough
regions to be meshed on each processor to avoid a negative impact
from communications
 Three groupings of problems can be considered
 Moderately large – (millions to low tens of millions of elements) –
SMP
 Large – (low tens of millions to high tens of millions) – either
 Very Large – (> high tens of millions) - DMP
 Simmetrix has developed a set of toolkits for Geometry Based
Parallel Mesh generation for both SMP and DMP architectures
October 2008
geometry based parallel
mesh generation & adaptivity

A series of rooms with furniture and people (Parasolid model)
 (courtesy of Transpire , Inc.)

Walls, Furniture, people and space are meshed
 With BL for CFD type applications
 Without BL for EM type applications

Results shown for various configurations of Distributed Memory
Parallel (DMP) on a small cluster
October 2008
geometry based parallel
mesh generation & adaptivity
 Cluster configuration used for testing (low end cluster)




6 dual core Suns
2Ghz Opteron processors
2GB Ram per processor
Gigabit Ethernet connection
 Speedup Test
 3 different mesh sizes run with and without Boundary Layers
 ~ 6 Million mesh regions run on 2, 4, and 8 processors
 ~ 24 Million mesh regions run on 4 and 8 processors
 ~ 46 Million mesh regions run on 4, 6, 8,10 and 12 processors
 Normalized Speedup = ( t(b) * n(b) ) / ( t(n) * n )




October 2008
t(b) – meshing time for base (minimum number of processors run)
n(b) – number of base processors
t(n) – meshing time for n processors
n – number of processors used
geometry based parallel
mesh generation & adaptivity
 6 million mesh regions
 ~1.5 million mesh regions/minute on 2 processors
 ~2.3 million mesh regions/minute on 4 processors
 ~3.6 million mesh regions/minute on 8 processors
Meshing time (6 million regions)
300
250
200
w/out Boundary Layers
150
w/Boundary Layers
100
50
0
0
2
4
6
number of processors
October 2008
8
10
geometry based parallel
mesh generation & adaptivity
 6 Million mesh regions
Processors
2
4
8
w/out Boundary Layers
1.00
0.77
0.61
w/ Boundary Layers
1.00
0.80
0.62
Normalized Speedup (6 million regions)
1.2
speedup
1
0.8
w/out Boundary Layers
w/Boundary Layers
0.6
0.4
0.2
0
0
October 2008
2
4
6
number of processors
8
10
geometry based parallel
mesh generation & adaptivity

24 million mesh regions
 ~ 3 Million mesh regions/minute on 4 processors
 ~ 4.3 Million mesh regions/minute on 8 processors
Meshing time (24 million regions)
600
500
400
w/out Boundary Layers
300
w/Boundary Layers
200
100
0
0
2
4
6
number of processors
October 2008
8
10
geometry based parallel
mesh generation & adaptivity
 24 Million mesh regions
Processors
4
8
w/out Boundary Layers
1.00
0.72
w/ Boundary Layers
1.00
0.74
Normalized Speedup (24 million regions)
1.20
speedup
1.00
0.80
w/out Boundary Layers
w/Boundary Layers
0.60
0.40
0.20
0
October 2008
2
4
6
number of processors
8
10
geometry based parallel
mesh generation & adaptivity
 46 million mesh regions
 ~ 2.7 Million mesh regions/minute on 4 processors
 ~ 3 Million mesh regions/minute on 6 processors
 ~ 3.5 Million mesh regions/minute on 8 processors
 ~ 4.9 Million mesh regions/minute on 10 processors
 ~ 5.3 Million mesh regions/minute on 10 processors
Meshing Time (46 million regions)
1200
1000
800
w/out Boundary Layers
w/Boundary Layers
600
400
200
0
0
2
4
6
8
number of processors
October 2008
10
12
14
geometry based parallel
mesh generation & adaptivity
 6 Million mesh regions – normalized speedup
Processors
4
6
8
10
12
w/out Boundary
Layers
1.00
0.73
0.68
0.73
0.66
w/ Boundary Layers
1.00 (46 million
0.76
Speedup
regions)
0.62
0.69
0.65
1.20
speedup
1.00
0.80
w/out Boundary Layers
w/Boundary Layers
0.60
0.40
0.20
0.00
0
2
4
6
8
number of processors
October 2008
10
12
14
geometry based parallel
mesh generation & adaptivity
 Parallel Mesh
adaptivity
 Supports isotropic and
anisotropic mesh
adaptivity
 Supports refinement &
coarsening
 Adapted mesh
adheres to original
geometry
 Supports initial mesh
as partitioned or single
mesh
October 2008
Simmetrix technology is widely used
October 2008
Download