Algorithms for Scientific Computing and Informatics · SOS 10 March 6-9, 2006 Algorithms for...
Transcript of Algorithms for Scientific Computing and Informatics · SOS 10 March 6-9, 2006 Algorithms for...
SOS 10 March 6-9, 2006
Algorithms for Scientific Computing
and Informatics
David E. WombleSandia National Laboratories
SOS10 March 6-9, 2006
SAND2006-1407C
SOS 10 March 6-9, 2006
The Forward Simulation Problem
System Design
Geometric Modeling
Meshing
Physics
Model Equations
Discretization
Partitioning and Mapping
Nonlinear solve
Linear solve
Time integration
Informatics and Visualization
Adapt
SOS 10 March 6-9, 2006
Drivers
• Physics– Multiscale
• Scale– Billions of unknowns– Coupled physics
• Quality– Mesh– Discretization
• Implementation– Code and framework complexity– Machine architectures
SOS 10 March 6-9, 2006
Drivers
• Interactions between components– Solvers with– Discretization with– Meshing with– Load balancing with– Error estimation with– …
• Example– Z-pinch simulations
Liner implosion
SOS 10 March 6-9, 2006
What’s Missing?
System Design
Geometric Modeling
Meshing
Physics
Model Equations
Discretization
Partitioning and Mapping
Nonlinear solve
Linear solve
Time integration
Informatics and Visualization
AdaptV&V
Data
SOS 10 March 6-9, 2006
Verification and Validation
• Right now, we are using an ad-hoc, black box approach– Convergence studies and MMS– Point-wise comparisons to experimental data
• Need– Integration of V&V with code design, i.e., designed in.
» Element library with error estimators» Symbolic manipulation of equations and automatic differentiation» Designed to be part of a V&V framework (not a stand-alone code) and must
allow some “intrusion” into the code (e.g., sensitivity calculations)– Algorithms
» Error estimators, statistical methods, input distributions and stochastic equations (including solvers)
– Presentation of results
• Problem– Established infrastructure– Need knowledgeable users working with analysts and developers
SOS 10 March 6-9, 2006
V&V
But What Do We Really Want To Do
System Design
Geometric Modeling
Meshing
Physics
Model Equations
Discretization
Partitioning and Mapping
Nonlinear solve
Linear solve
Time integration
Informatics and Visualization
AdaptOptimization
and UQ
Improved design and understanding
Data
SOS 10 March 6-9, 2006
Visualization
• Geometric information (current)– Large scale– Desktop
• Non-geometric information (future)– Aggregate information (many
simulations)– Risk and uncertainties
• “Data-mining”– Need to think about how people really
use information– Tools to help extract information
(“features”)– Integration with simulation
SOS 10 March 6-9, 2006
Optimization and UQ
• Can include V&V in this loop, but adds design and “understanding”
– Ensemble of “single-point runs insufficient– Need to design for system performance, not just assess system
performance
• Need integrated tools– Many of the same tools as for V&V (e.g., error estimation, sensitivity
analysis)– Identify critical phenomena (turning points, bifurcations points)– Rigorous exploration of design space– Integrated design and solvers
• Need multifidelity– Surrogate methods– Hierarchy of resolutions and scale/physics couplings
• More integration of tools (designed in)– Including V&V, meshing and visualization
-3.0E+07
-2.0E+07
-1.0E+07
0.0E+00
Impl
osio
nV
elo c
it y(c
m/ s
)
0.001
0.002
0.003
Fuel Density (g/cm 3)0.10
0.120.14
Capsule Radius (cm
SOS 10 March 6-9, 2006
Dealing With Uncertainty
• Simply propagating distributions and quantifying errors is not enough
• We need to be able to incorporate experimental data to further reduce uncertainties
SOS 10 March 6-9, 2006
Model Real World
Data
What About The Informatics Problem?
System Design
Geometric Modeling
Meshing
Physics
Model Equations
Discretization
Informatics and Visualization
Partitioning and Mapping
Nonlinear solve
Linear solve
Time integrationAdaptOptimization
and UQ
Improved design and understanding
Processing ToolsValidation and UQ
SOS 10 March 6-9, 2006
Informatics Drivers
• Scale– Huge amounts of data– Visualization on multiple scales– Streaming data – “online” processing
• Presentation of Information– Making it meaningful
• Implementation– Code complexity– Architectures
• Quality– Representation of data– Dealing with model uncertainty– Dealing with missing data
SOS 10 March 6-9, 2006
Representing Data
• Challenge: Piecing the data together and extracting critical, relevant information in a timely manner
• Semantic Graphs or Attributed Relational Graphs (ARGs) are one way to integrate data from disparate sources
– Vertices represent people, places, locations, events, etc. – Edges represent the relationships between the vertices– ARG encodes web of relationships
• Graphs represent relationships and are the link between the model and processing
SOS 10 March 6-9, 2006
Motif Finding
Image Source:T. Coffman, S. Greenblatt, S. Marcus, Graph-based technologies for intelligence analysis, CACM, 47(3, March 2004): pp 45-47
SOS 10 March 6-9, 2006
One More Challenge
• Merging data with large-scale modeling and simulation– Informatics and modeling and simulation are not separate problems
• Example - Climate simulation– Climate predictions will drive economic decisions– Climate and economics will both drive behavior
Climatological feedback
SOS 10 March 6-9, 2006
Summary
• Scale will drive computing– Size of problem (in the traditional scientific computing model)– Amount of data
• Need to focus on how machines/simulation/information will be used– Presentation (visualization) of information– Support for design and decision making
• Software environments and complexity are limiting– Need a lightweight, simple environment that allows flexibility and linking of
components. (No problem)– Data representation is a key question. This will drive both software and
hardware design