"GPU Programming and Performance" Kevin R. Tubbs

1
"GPU Programming and Performance"Kevin R. Tubbs
IGERT Fellow Donald W. Clayton Graduate Program
in Engineering Science, Louisiana State
University, 3507 CEBA Building, Baton Rouge, LA
70803-6405 PH [phone] FAX (225)
578-6782, E-mail ktubbs2_at_lsu.edu

• IGERT Student Seminar Series

2
BackGround

• Computational Power

Source heise.de, tomshardware.com, wikipedia.org
3
Background

• CUDA Programming Model

Source nvidia.com
4
Overview CUDA Zone for CFD
http//www.nvidia.com/object/cuda_home.htmlstate
home
5
Applications

• Elegant Mathematics
• Iterative Solvers
• Linear System Solvers
• CFD EM Solvers
• Accelereyes
• Jacket GPU Engine
• Ports matlab code to GPU
• Supports functions friendly to GPU

6
Journal Papers

• Methods
• 3D Euler
• LBM (2D 3D)
• Discontinuous Galerkin (DG)

7
Presentations

• Smoothed Particle Hydrodynamics
• 23X Meshless/MeshFree
• Langrangian
• LES Navier Stokes Code
• 17X Cartesian
• Red-Black SOR Poisson Solver
• 3D Incompressible Navier Stokes Code
• 100X 4 GPU
• Cartesian

8
Applications Accelereyes

• Accelereyes
• Jacket GPU Engine
• Maximum Reported Speedup Using Jacket 50 X
• LBM Speedup 10X
• Graphics Toolbox
• OpenGL
• Basic Plot open source functions

9
Applications Elegant Mathematics

• EMGPUIter
• Avg. Peformance 100K unknowns
• 7 GFlop/s Single NVIDIA 260 (100 MFlop/s QC Xeon
  2.6Ghz)
• Block Versions 70 GFlop/s (3 GFlops )
• EMGPUHMatrix
• Linear System 163,840 unknowns
• Generated and Solved w/ 25X Speedup (QC Xeon)
• EMGPUPartBoltzmann
• Solves Boltzmann Eq. via Particle Method
• Collision 30X (QC Xeon)
• Stream 120X (QC Xeon)
• Problem Size
• 15 M particles on 1 GPU
• 1 B particles on GPU CPU w/ 64GB of Ram Using
  Out-of-Core

10
Method Performance LBM

• LBM
• Good Locality in Memory Acces
• 2D LBM for N-S
• D2Q9
• 10X
• 3D LBM
• D3Q15
• 28X 100X (4 GPU)

11
Method Performance 2D LBM

• 2D Shallow Water Mass Transport
• 10X
• Same Code Cast to GPU
• No Optimization

12
Method Performance DG

• Maxwells Eq.
• DG, 3D Unstructured Mesh
• DG
• Good Locality in Memory Access (element local)
• High order require fewer data points

A. Klocknera, T. Warburtonb, J. Bridgeb, J. S.
Hesthavena (2009) Nodal Discontinuous Galerkin
Methods on Graphics Processors
http//www.citebase.org/abstract?idoaiarXiv.org
0901.1024
13
Method Performance DG

• Maxwells Eq.
• DG, 3D Unstructured Mesh
• Single 400 GTX 280 GPU
• 40X to 60X relative to serial
• 200 GFlop/s

A. Klocknera, T. Warburtonb, J. Bridgeb, J. S.
Hesthavena (2009) Nodal Discontinuous Galerkin
Methods on Graphics Processors
http//www.citebase.org/abstract?idoaiarXiv.org
0901.1024
14
Method Performance 3D N-S Solver

• Lid-Driven Cavity
• Re 1000 Mesh 323 2563
• 2D flow physics 3D computations
• Numerical Method
• Incompressible flow Cartesian Method
• Projection Method Press/Vel Staggered Grid
• 1st order time 2nd order diff. Poisson Jacobi

Julien C. Thibault and Inanc Senocak, (2009)
CUDA Implementation of a Navier-Stokes Solver on
Multi-GPU Desktop Platforms for Incompressible
Flows 47th AIAA Aerospace Sciences Meeting,
Orlando FL, paper no AIAA-2009-758.
15
Method Performance 3D N-S Solver

• Workstation GPU Speedup
• Intel Core 2 Duo 3.0 GHz (Max 21X)
• AMD Opteron 2.4 GHz (Max 100X)

Julien C. Thibault and Inanc Senocak, (2009)
CUDA Implementation of a Navier-Stokes Solver on
Multi-GPU Desktop Platforms for Incompressible
Flows 47th AIAA Aerospace Sciences Meeting,
Orlando FL, paper no AIAA-2009-758.
16
Method Performance 3D N-S Solver

• Server GPU Speedup
• Scales at 75 Efficiency for Number of GPU

Julien C. Thibault and Inanc Senocak, (2009)
CUDA Implementation of a Navier-Stokes Solver on
Multi-GPU Desktop Platforms for Incompressible
Flows 47th AIAA Aerospace Sciences Meeting,
Orlando FL, paper no AIAA-2009-758.
17
IGERT Hardware Options

• Tesla
• 1300
• 240 Streaming Processor Cores
• Frequency of processor cores1.3GHz
• Single Precision floating point performance (peak
  ) 933
• Double Precision floating point performance
  (peak) 78
• Dedicated Memory 4GB GDDR3
• GTX 280
• 350
• 240 Processor Cores
• Processor Clock (MHz)1.3 GHz  
• Dedicated Memory 1GB GDDR3

18
Questions

• Thank You