Runtime Techniques for Scientific Applications

1
Run-time Techniques for Scientific Applications

• Curtis Lisle, Ph.D.
• Systems Architect
• SGI Professional Services

2
Agenda

• Scientific Visualization Software Architectures
• Analytical and Rendering models
• Distribution of parallel datastructures
• Parallel performance
• Interactive Performance with Very Large Models
• Grid Computing

3
Components of a Visualization System

• Database Representation (save and reload)
• Method of traversing, and changing the dataset
  (scene graph)
• Architecture to render a view of the data
  (Performer)
• Software Packages for Problem Domains
• CFD (computational Fluid Dynamics)
• Exa Powerflow
• Fluent
• Finite Element Analysis
• Automotive and CAD applications

4
2 D Transient Flow (powerflow)
5
Turbulent Jet (powerflow)
6
Techniques for Handling Large Datasets

• SGI Graphics APIs
• OpenGL Optimizer decimation operations
• OpenGL Performer graphics rendering
• Volumizer - accelerated volumetric rendering
• Level of Detail
• Manual Insertion - time consuming and awkward in
  the modeling process
• Auto Level-of-detail insertion
• Third Party Packages
• Visualization Toolkit (VTK) and Applications
• www.kitware.com
• Rational Reducer

7
Auto Level of Detail

• Optional process for converting files to
  Performer native format for fast loading
• Auto-LOD algorithm rearranges the scene graph for
  higher rendering performance

(PF Loader)
PFB
Performer Converter
Performer Rendering Environment
Large Database
Auto-LOD algorithm
(Either path OK)
(IV or VRML)
8
Auto-LOD Algorithm

• Export Scene Graph to VTK Datatype
• Spatial Partitioning and decimation
• Recombine in modified scene graph

VTK
VTK
PF
PF
LOD
low
high
9
Dual Modeling Approach

• Dont try to put analytical values on a polygonal
  database
• Instead, regenerate the rendering data structures
  from the analytical ones as needed (when state
  changes appreciably enough)
• Dead Reckoning - technique for controlled
  approximations

Analytical Model
Rendered Model
10
Recursive, Object-Oriented Structures for
Molecular Modeling

• Project Goals
• Create a software environment for object-oriented
  molecular modeling
• Construct and debug models using sequential
  object-oriented design (in C)
• Provide a methodology to consistently extend the
  sequential hierarchy for parallel or distributed
  execution using standard message passing layers
  (i.e. BSP, MPI)
• Preserve a simple, sequential interface at the
  user level for non-computer scientists (hide the
  parallelism)

Problem-specific abstractions
Hidden parallelism
11
Recursive Model Definition

• A molecule is a set of smaller molecules.
• All applications use recursive molecule class
  hierarchy
• Some applications are sequential, some are
  distributed

12
Molecule Browser
13
Property Viewer
14
Parallel Speedup

• Peak speedup of 10 (using 12 to 14 processors)
• Then msg overhead / context switches begin to
  dominate

Speedups Random
Speedup
Number of Processors
15
Distribution Affects Performance

• Atoms and molecules were scattered across the
  processors
• More efficient messaging resulted from inteligent
  allocation

(localized)
(random)
16
Distributed Science Grid Computing

• The SGI vision of Visual Area Networks introduces
  the concept that the data is stored and processed
  in one placebut it can be interacted with by
  users using any client device, across any network.

17
Parallel Processing Techniques

• MPI - Message Passing Interface
• flexible message passing Interface
• Forces application to be cut up with explicit
  messaging
• OpenMP
• Shared memory computing
• Restricted number of CPUs
• Tightly integrated threading model
• Easier programming model
• greater speedup
• Multi-Level Parallelism
• OpenMP within Shared-memory computers
• MPI between computers in a cluster

18
Conclusions

• SGI scalable computation and scalable graphics
  are uniquely suited for interactive simulation
  applications
• A large set of commercial products and shareware
  research programs are available
• VTK
• Globus toolkit
• SGI is commited to continuing its leadership in
  graphics APIs (i.e. Performer, Volumizer, OGL
  Multipipe, etc.)