Hub-based Simulation and Graphics Hardware Accelerated Visualization for Nanotechnology Applications

1
Hub-based Simulation and Graphics Hardware
Accelerated Visualization for Nanotechnology
Applications

• Wei Qiao qiaow_at_purdue.edu
• Michael McLennan mmclennan_at_purdue.edu
• Rick Kennell kennell_at_purdue.edu
• David S. Ebert ebertd_at_purdue.edu
• Gerhard Klimeck gekco_at_purdue.edu
• Purdue University

2
Our Goals

• Provide advanced interactive visualization of
  scientific simulations to users worldwide without
  the user needing special graphics capabilities
• Approach - integrate hardware-accelerated remote
  visualization into nanoHUB.org

3
nanoHUB Remote Simulation and Visualization
4
Outline

• nanoHUB.org
• Challenges and requirements
• Related work
• Our system design
• Performance and optimization
• Case studies
• Summary and future work

5
nanoHUB.org

• A nano-science gateway for nanotechnology
  education and research
• Created by the Network for Computational
  Nanotechnology (NCN)
• Educational material
• Animations
• Courses
• Seminars
• Simulation tools accessible from a web browser

6
User Community and Usage

• Nanoelectronics Community
• Researchers
• Educators
• Students
• Usage (last year)
• More than 10,000 users viewed online materials
• 1,800 users ran more than 54,000 simulation jobs
  consuming over 28,500 hours of CPU time

7
nanoHUB Simulation Architecture
Open Science Grid and NSF TeraGrid
Simulation Cluster
Web Server
Gig Net
Internet
Gig Net
Virtual Machine
8
DEMO!
9
System Requirements

• Transparency in service delivery
• Scalability to increased workload
• Responsiveness to user command
• Flexibility in handling simulation data
• Extensibility in software and hardware

10
Visualization Challenges

• Architecture
• Lack state of the art visualization systems
• Mismatch between CPU and GPU resources
• Users
• Predominantly remote
• Vast diversity of computing platforms and
  capabilities

11
Related Work

• Molecular Dynamics Visualization
• Surface rendering
• Structure rendering
• Volume visualization
• Electron potential fields
• Electronic wave function
• Electro-magnetic fields

12
Related Work (Cont.)

• Flow Visualization
• Texture synthesis
• CPU (Wijk 91 and Cabral and Leedom 93)
• GPU (Heidrich et al. 99, Jobard et al. 00,
  Weiskopf et al. 2003 and Telea and Wijk 03)
• Particle tracing
• CPU (Sadarjoen et al. 94)
• GPU (Kolb et al. 04 and Krüger et al. 05)
• Remote Visualization
• Data is too large to transfer over network
• Local workstation cannot handle the data
• Distance collaboration

13
Practical Obstacles to nanoHUB

• VNC session run on cluster nodes with no graphics
  hardware acceleration
• Cluster nodes are rack mounted machines with
  neither AGP nor PCI Express interfaces
• nanoHUBs virtual machine layer cannot directly
  access graphics hardware

14
Our System Design

• Client-server architecture
• nanoVIS render server
• Visualization engine library
• Vector flows
• Multivariate scalar fields
• Rappture GUI client
• User front end
• nanoSCALE service daemon
• Monitors render loads
• Track GPU memory usage
• Starts nanoVIS servers

15
Schematic View
Open Science Grid and NSF TeraGrid
Simulation Cluster
Web Server
Gig Net
Internet
Gig Net
Virtual Machine
16
Hardware

• Linux cluster render farm
• 1.6GHz Pentium 4
• 512MB of RAM
• nVIDIA Geforce 7800GT graphics hardware
• Advantages
• Extremely cost effective
• Flexible to upgrade and expand
• Integrates tightly into the nanoHUB architecture

17
Rappture Toolkit

• Rapid Application Infrastructure Toolkit
• Accelerate development of basic infrastructure
• Declare simulator input / output using XML
• Automatic generation of GUI

18
nanoVIS

• Fully accelerated by graphics hardware
• Visualize a variety of nanotechnology simulations
• Volumetric and multivariate scalar fields
• Texture-based volume visualization
• FFC volume (zinc-blende) Qiao et al. 2005
• Vector fields
• GPU particle tracing
• 2D texture synthesis
• Geometric drawing to illustrate simulation
  geometry
• GL primitive drawing

19
Vector Field Visualization (Cont.)

• Particle Implementation
• Kolb et al. 2004 Krüger et al. 2005
• Framebuffer Object (FBO)
• Vertex Buffer Object (VBO)
• Particles stay in GPU memory
• 2D texture synthesis
• Complement particles

Vertex Data
VBO
Particle Render
20
Client-Server Interaction

21
Performance and Optimization

• Work load consideration
• GPU heavy
• Rendering
• CPU light
• Network communication
• GPU-oriented optimization
• GPU load estimation scheme
• Node selection scheme based on estimated GPU load

22
GPU Load Estimation

• Fragment processing cost
• Number of rasterized fragments
• Computation per fragment
• Unified measurement for particle system and
  volume
• Hard to compare cost of particle rendering to
  advection
• Experimental data allows a unified measurement
• Render cost is factor of 0.2 to advection
• Estimation equation
• Primary cost of the shader execution is texture
  access

23
Performance

• Measure turn around time (from issue command to
  image received)
• 128 x 128 x 128 scalar field
• 512x512 render window
• Simulated user interaction
• Transfer function modification, rotation, zoom,
  cutting plane, etc.

24
Case Studies

• Successfully developed several nanotechnology
  tools
• SQUALID-2D
• Quantum Dot Lab
• BioMOCA
• nanoWire

25
2-D Electron Gas Simulator

• Goal
• Study the effects of impurity in a nanowire
• Device composition
• Electrodes are positioned on the top
• GaAs and AlGaAs semiconductor layers
• A narrow channel constraining the electrons in
  the middle
• Experiments
• Vary magnetic field
• Electron flows
• Electron potential fields

26
2-D Electron Gas Simulator
Flow and Electron Potential
Particle Tracing and LIC
27
BioMOCA

• Goal
• Study the flow of ions through a pore in a cell
  membrane
• Method
• Compute random walks of ions through a channel
  with a fixed geometry within a cell membrane.

Cell Wall
Cell Wall
28
Quantum Dot Lab

• Goal
• Study the wave functions (orbitals) of electrons
  trapped in a quantum dot device
• Method
• Configure incidental light source, shape and size
  of the quantum dot

s and p orbitals
p orbital
s orbital
29
Conclusions

• Hub-based remote visualization is a powerful,
  flexible solution
• Seamlessly delivers hardware-accelerated
  visualization to remote scientists with minimal
  requirements on their computing environments
• Intuitive interface and ease of use are key for
  wide-usage
• Enables rapid development and deployment of new
  simulation tools
• Tight integration into the simulation and
  interactive performance can speed scientific
  discovery and change science work flow
• nanoVis tools is huge success

30
Future Work

• Expand to generic scientific hub-based
  visualization engine
• Our system can be adopted to economically deliver
  accelerated graphics to other hub-based
  multi-user environments
• Expand to large data support
• GPGPU nano-electronics simulations and integrated
  visualization
• More accurate GPU load estimation using nVidia
  newly released NVPerfKit 2.1 for Linux

31
Acknowledgement

• Martin Kraus, Nikolai Svakhine, Ross Maciejewski,
  Xiaoyu Li
• Anonymous reviewers for many helpful discussions
  and comments
• nVIDIA
• National Science Foundation under Grant No.
  EEC-0228390

32
Vector Field Visualization

• GPU accelerated particle tracing
• Similar to Kolb et al. 2004 and Krüger et al.
  2005
• Particle position equation
• GPU Eulerian integration using fragment shader

33
Node Selection

• Selection criteria
• Sufficient GPU memory to fit the data
• Least amount of GPU workload
• Historical bias

34
Selection Details

• nanoVIS receives a render request
• nanoVIS estimates the GPU workload of the request
• nanoVIS sends estimate to local nanoSCALE using
  pipe
• nanoSCALE broadcasts cost sum to all peer render
  nodes
• Rappture contacts an initial host
• Initial host chooses a target host to start
  nanoVIS service with a redirection threshold
• Initial host updates its record of the target
  hosts load average
• Target host broadcasts its most recent workload
• Initial host update its record

35
Summary

• Developed / deployed a remote visualization
  hardware and client-server software architecture
  for nanoHUB.org
• Flexibly handle a variety of nanoscience
  simulation data
• GPU load estimation model and render node
  selection scheme
• Seamlessly deliver hardware-accelerated
  visualization to remote scientists with minimal
  requirements on their computing environments
• Enable rapid development and deployment of new
  simulation tools
• Our system can be adopted to economically deliver
  accelerated graphics to other hub-based
  multi-user environments