Spectral Element Method and GPU Computing for Seismic Imaging

Post on 24-Feb-2016

27 views 0 download

Tags:

description

Spectral Element Method and GPU Computing for Seismic Imaging. Chaiwoot Boonyasiriwat May 1, 2010. Outline. Introduction to Seismic Imaging Spectral Element Method GPU Computing and Computational Steering Summary. 1. Seismic Imaging. Given a seismic data set. - PowerPoint PPT Presentation

transcript

Spectral Element Method andGPU Computing for Seismic Imaging

Chaiwoot BoonyasiriwatMay 1, 2010

Outline

1

• Introduction to Seismic Imaging• Spectral Element Method• GPU Computing and Computational Steering• Summary

Seismic Imaging

2

• Given a seismic data set )( trueobs A md

• Find a model that best describes the given data or minimizes the misfit function where

)(~ trialcalc A md 2

)()( mddm calcobsE

m?CSG

Common shot gather

Seismic Imaging

3

obsd

m?

• Use iterative gradient-based methods

Seismic Imaging in 3D

4

• Improve efficiency• Forward modeling: use better numerical methods• Improve convergence rate: preconditioning• Use many sources simultaneously

Forward Modeling

5

• Important component of seismic imaging methods)(md A

• Numerically solving a wave equation, e.g., acoustic wave equation

spvptt 22

• Numerical methods:• Finite difference• Spectral/pseudo-spectral• Finite element/spectral element

Outline

6

• Introduction to Seismic Imaging• Spectral Element Method• GPU Computing and Computational Steering• Summary

Finite Element Methods

7

8

FDM vs FEM

Finite Element Formulation

9

Acoustic wave equation),(),()(),( 22 tstpvtptt rrrr

rrr vsdpdvvvpdtt2 Weak form

n

iii tPtp

1

)()(),( rr Linear combination of basis functions

FKPPM tt

rM dM jiijij ][

rK dvK jiijij 2][

where

Matrix form

rdspvpv tt22 Weighted residual method

10

Geophysical Applications of FEM

• Global-scale• Exploration-scale

Outline

11

• Introduction to Seismic Imaging• Spectral Element Method• GPU Computing and Computational Steering• Summary

Seismic Imaging on GPUs

12

• High performance and large band-width• General-purpose computing on GPUs using CUDA• Advanced hardware (GPUs) + algorithm

(multisource imaging)• Computational steering: computation + visualization

13Courtesy of NVIDIA

1000

Peak

GFL

OP/

s 750

500

250

0

Performance of GPU vs CPU

Memory Bandwidth of GPU vs CPU

14Courtesy of NVIDIA

120

BandwidthGB/s

100

80

60

40

20

0

15

Conventional vs Multisource Method

Conventional RTM Image using 200 gathers

16

Multisource RTM Image using 20 supergathers

10x speedup

Conventional vs Multisource Method

Computational Steering

17

• Initial model• Line search method• Regularization• Iteration number• Perturb model

manually (steering)

User Interface

Iteration numberM

isfit

val

ue

True model

Tomogram 1

Tomogram 218

Summary

19

• Seismic imaging using spectral element methods• Global-scale waveform inversion using SEM• 3D seismic imaging on GPU cluster (CUDA+MPI)• Computational steering on GPU cluster