Feng Qiu, Ye Zhao, Zhe Fan, Xiaomin Wei, Haik Lorenz, Jianning Wang, Suzanne YoakumStover, Arie Kauf

1
GPU Accelerated Dispersion Simulation for Urban
Security
Feng Qiu, Ye Zhao, Zhe Fan, Xiaomin Wei, Haik
Lorenz, Jianning Wang, Suzanne Yoakum-Stover,
Arie Kaufman, Klaus Mueller Center for Visual
Computing and Department of Computer Science,
Stony Brook University
http//www.cs.sunysb.edu/vislab/projects/urbansec
urity

• Overview
• Simulating and visualizing the propagation of
  dispersive contaminants
• Open environments characterized by sky-scrapers
  and deep urban canyons
• Multiple Relaxation Time Lattice Boltzmann Model
  for flow simulation
• GPU accelerated computation and visualization

• Visualization
• Rendering of building
• Building textures from real images
• Reserve texture memory for LBM simulation
• Shader program adds weathering and cracks to
  textures
• Rendering of smoke
• Splatting of smoke particles
• Splats distorted for correct projection
• Half angle slicing for self-shadowing

• Single GPU Results
• West Village area of New York City (10 blocks)
• Lattice size 90x30x60 with grid unit 3.8m
• Speedup (GPU/CPU) 7

• Lattice Boltzmann Model (LBM)
• LBM models Boltzmann particle dynamics on a
  regular lattice
• Streaming and collision in discrete time steps
• 2nd order space-time accurate CFD method
• Advantages GPU accelerated, complex boundary,
  easy to implement, multi-resolution, sensor
  feedback

• GPU Acceleration
• Local operations in LBM are accelerated by GPU
• Data layout
• One state variable stored in one volume
• Each volume packed into a series of 2D textures
• Boundary link information packed into small 2D
  textures
• Computation
• LBM operations mapped to fragment program
• Results stored in pixel buffers and copied back
  to textures for next step
• Simulation and visualization on same GPU,
  reducing data transfer

Snapshots of simulation with smoke and flow
streamlines
Original façade
Façade variation

• Multi-GPU Results
• Time Square area of New York City (110 blocks)
• Lattice size 320x80x320 with grid unit 4.5m
• Speedup on 30 nodes (GPU cluster/CPU cluster) 4-5

• Acknowledgement
• NSF CCR-0306438
• Department of Homeland Security, Environment
  Measurement Laboratory

• Sensor Feedback
• Two methods
• Incorporate sensor data as body force
• Modify boundary nodes affecting sensor readings

Smoke in city model
Streamlines in Time Square