Visualization Techniques for Terascale Particle Accelerator Simulations

1
Visualization Techniques for Terascale Particle
Accelerator Simulations

• Kwan-Liu Ma Greg Schussman Brett Wilson
• University of California at Davis
• Kwok Ko
• Stanford Linear Accelerator Center
  
• Ji Qiang Robert Ryne
• Lawrence Berkeley National Laboratory

SC2002, Baltimore, MD, November 19
2
Outline

• Advanced Computing for 21st Century Accelerator
  Science Technology
• A hybrid rendering technique for visualizing
  particle beam dynamics data
• A scalable hardware-accelerated technique for
  visualizing electromagnetic fields and particle
  trajectories
• Further research

3
Advanced Computing for 21st Century Accelerator
Science Technology

• Particle accelerators helped enable some of the
  most remarkable discoveries of the 20th century
• Given the complexity and importance of
  accelerators, it is imperative that the most
  advanced HPC tools be brought to bear on their
  design, optimization, commissioning, and
  operation
• The objective of our SciDAC project is to
  establish a comprehensive terascale simulation
  environment needed to solve the most challenging
  problem in particle accelerator design
• http//scidac.nersc.gov/accelerator

4
Parallel Beam Dynamics Simulations

• Modeling a large number of charged particles as
  they move through the accelerator and respond to
  various forces
• Millions to billions of particles
• How to visualize and
  better understand the
  time-varying phenomena?

5
Data

• Each particle consists of spatial coordinates (x,
  y, z) and momenta (px, py, pz)
• 100-1000 million particles
• 5-48GB per time step
• High density beam and low density halo

(x, Px, y) (x, Px, z)
(Px, Py, Pz)
6
Previous Approach

• Convert the particles data into volume data
• Interactive visualization was made possible by
  using a graphics supercomputer
• The size of the volume that can be visualized
  efficiently is limited by the amount of available
  texture memory as well as the fill rate of the
  available hardware
• Uniform subsampling would leave out fine features
  in the very low density areas

7
512 mapping of 300 million particles
3
Pat McCormick, LANL
Spatial Coordinates (x, y, z)
Phase Space (x, Px, z)
8
Hybrid Rendering

• Texture-based volume rendering for regions of low
  interest/detail
• Point-based rendering for regions of high
  interest/detail
• Reduce storage requirements
• Faster data transport
• Better utilization of the
• video memory

9
High Performance Preprocessing

• Compute a hybrid representation of the data on
  the same parallel supercomputer at NERSC/LBL that
  the simulation ran on
• Partition step organizing the raw data into an
  octree
• Extraction step converting the octree data into
  the hybrid representation
• The preprocessing on average generates under
  100MB per distribution per time step for the 100
  million particle case

10
Viewing

• The more compact data can be interactively
  visualized on a graphics-enhanced desktop PC
• An image is created by classifying each octant as
  belonging to a volume rendered region or a
  point-rendered region, depending upon the
  transfer function for each region

Extraction threshold
point
volume
Point density
11
Varying Threshold Values
12
Selected Time Steps (x, y, z)
13
Misaligned Beam (x, y, z)
14
Misaligned Beam (x, Px, y)
15
Summary

• Volume rendering alone lacks the spatial
  resolution and the dynamic range to resolve
  regions with very low density
• Point rendering alone lacks the interactive speed
• The hybrid rendering approach allows interactive
  exploration of the region of high interest
• Parallel preprocessing and parallel rendering
  must be used for billion points or more cases
• Point-based feature enhancement
• Correlate particles over time

16
Hybrid rendernig for Volume Data
17
Electromagnetic Fields

• Parallel time domain electromagnetic field solver
  using unstructured hexahedral meshes
• Modeling of the reflection and transmission
  properties of open structures in an accelerator
  design
• To attain the needed accuracy a very large number
  of time steps are required and half million
  hexahedral elements are needed to model a 30-cell
  structure

18
Electromagnetic Fields

• Parallel time domain electromagnetic field solver
  using unstructured hexahedral meshes
• Modeling of the reflection and transmission
  properties of open structures in an accelerator
  design
• To attain the needed accuracy a large number of
  time steps are required and half million
  hexahedral elements are needed to model a 30-cell
  structure

19
Visualization Challenges

• Large time-varying data
• Dense collection of electric and magnetic field
  lines
• Clear spatial relationships
• Unambiguous global and local details
• Interacting with the field line visualization

20
Visualization Requirements

• A compact field line representation
• Perceptually effective
• -- illuminated, textured, semi-transparent,
  haloing
• Good seeding strategy and a hierarchical
• field line organization
• Interactive visualization

21
Visualization Requirements

• A compact field line representation
• Perceptually effective
• -- illuminated, textured, semi-transparent,
  haloing
• Good seeding strategy and a hierarchical
• field line organization
• Interactive visualization

22
Self-Orienting Surfaces (SOS)

• A triangle strip computed on the fly
• Low storage requirements
• Fast rendering
• Hardware accelerated bump mapping
• Haloing for free
• Perceptually correct depth cuing
• Flexible for adding other texturing options

23
Self-Orienting Surfaces (SOS)
24
Performance
150 lines 800 lines 10k lines
Polygonal tubes no display list 0.445 3.001 38.2
SOS, Finely tessellated (No HW Bump Map) 0.077 0.512 5.54
Polygonal tubes display list 0.027 0.173 1.82 (628 MB ! )
SOS with Hardware Bump Map 0.019 0.124 1.28 (0 MB ! )
25

26
(No Transcript)
27
(No Transcript)
28
Particle tracking (both primary and secondary
paths)
29

30
Summary

• A new representation for 3D vector field lines
• Scalable
• Excellent texture support
• Fast transfer and low memory requirement
• Perceptually effective visualizations
• An hierarchical field line representation makes
  possible interactive exploration
• Extreme dense field lines
• One hundred thousands to millions of electron
  paths to visualize
• Limited resolution of the display
• Abstraction of the field complexity
• User interfaces and interaction to reveal
  structures

31
Concluding Remarks

• HPC plays an important role in accelerator design
• Appropriate visualization technologies must be
  developed to address the challenges introduced by
  the large scale accelerator simulations
• Current effort must be expanded
• Interactive visualization techniques
• Time-varying data
• New interface technology
• Web-based collaborative visualization
• Parallel visualization

32
acknowledgments

• DOE SciDAC
• DOE LANL/LLNL
• NSF PECASE
• NSF LSSDSC
• NERSC
• The visualization group at Lawrence Berkeley
  National Laboratory
• Pat McCormick at Los Alamos National Laboratory

33
Kwan-Liu Ma

• CIPIC Department of Computer Science
• University of California at Davis
• ma_at_cs.ucdavis.edu
• http//www.cs.ucdavis.edu/ma

34
(No Transcript)
35
(No Transcript)