Transcript Draft.pptx

GPU Random Walk Method for Hyperbolic Problems
Tara
1,2
Aida ,
Dr. Sorin
1
Mitran ,
Mathematics Department at UNC-Chapel
Introduction
• Focus on solving wave equation: utt + Ñ u = f
• Make Fourier guess to get Helmholtz equation: Ñ2 v + k 2 v = f
• Solve Helmholtz equation (waves at single frequency)
• Superimpose these solutions with time dependent coefficients to
find solutions to the wave equation
2
• Traditionally, hyperbolic equations are solved numerically with gridbased methods, e.g. finite volume, differences methods
• However, these exhibit limitations [1]:
• Becomes computationally heavy if k2 has large range
• Domain discretization is time consuming
• Field solution must be produced
• Random Walk Method (RWM) is an attractive alternative that can
overcome these difficulties
Random Walk Method
Michael
1
Hill
and Harvard
Objectives
1. Finite Element Method (FEM)
• Gain familiarity with traditional techniques
• Solve Helmholtz equation with FEM, varying domain and k2 value
• Finer meshes lead to lower error, but require more computational
power
• Conduct convergence study: analyze the effect of k2 value on the
relation between error and meshn fineness
Figure 3. Plot of relative error versus measure of mesh fineness, heff = AW / n,
for FEM solutions of the Helmholtz equation, with linear fits for k2 varying
from 1 to 105, for domain shown.
Conclusion
• Through RWM, we hope to find solutions to the wave
equation, within some error, and with relatively low
computational power, even over a wide range of wave
frequencies
Benefits of RWM:
• Allows one to solve PDE at specific points, rather than over entire
domain
• Easily parallelized with low communication between each walk
 lower computational load
Figure 2. Sample FEM solution for Helmholtz equation on custom
domain, with k 2 = y / (y - x) .
2. CUDA Program for RWM
•
•
•
Write parallelized code to solve the wave equation and the
Helmholtz equation in inhomogeneous media
Use GO language and CUDA platform on NVIDIA GPUs
Conduct convergence study to compare with FEM study
3. Apply RWM to Acoustics
•
•
www.PosterPresentations.com
1Derived
from theory of diffusion processes, stochastic calculus, Itô formula
Results
Methods
• Applies to second-order partial differential equations [1].
• To find the solution v(x) at a point x in the domain:
1. Conduct random walks, starting at x, ending at the boundary
2. Calculate path integrals along walks
3. Solve for v(x) using formula1 relating it to boundary conditions,
and expectation value of path integrals.
RESEARCH POSTER PRESENTATION DESIGN © 2012
2
University
1. Investigate possibility of combing RWM with grid-based
methods to solve wave equation and the Helmholtz
equation.
2. Compare the computational load of RWM with
traditional methods for these solutions.
3. Apply RWM to a specific problem in acoustics.
Theory:
Figure 1. Two random walks produced using Mathematica random
number generators.
1
Malahe
Apply RWM to solve hyperbolic PDEs in specific acoustics problem.
Compare RWM to previous numerical solutions.
• Since hyperbolic equations apply to many areas in
physics, a faster more flexible method of solving these
equations could benefit a wide range of computational
research.
References
[1] M.K. Chati et al. “Random walk method for the two- and threedimensional Laplace, Poisson and Helmholtz’s equations.” 2001.
Acknowledgements
This research made possible by NSF Award OCI-1156614. I’d also like
to thank Dr. Mitran and Michael Malahe for their guidance as my main
mentors, as well as Dr. Kannappan for her work in coordinating the
CAP REU.