HR: 09:45h
AN: S31E-08 INVITED [Abstracts]
TI: Full waveform seismic inversion: Prospects for scaling to petaflops computers
AU: * Ghattas, O
EM: omar@ices.utexas.edu
AF: The University of Texas at Austin, Jackson School of Geosciences and Institute for
Computational Engineering and Sciences, 1 University Station, C0200, Austin, TX 78712, United States
AU: Akcelik, V
EM: volkan@slac.stanford.edu
AF: Stanford Linear Accelerator Center, Advanced Computations Department, SLAC, Menlo
Park, CA 94025, United States
AU: Epanomeritakis, I
EM: ioannis.epanomeritakis@gmail.com
AU: Burstedde, C
EM: carsten@ices.utexas.edu
AF: The University of Texas at Austin, Institute for Computational Engineering and Sciences, 1
University Station, C0200, Austin, TX 78712, United States
AU: Bielak, J
EM: jbielak@cmu.edu
AF: Carnegie Mellon University, Department of Civil and Environmental Engineering,
Pittsburgh, PA 15213, United States
AB:
The U.S., Japanese, and European governments are all pursuing
petaflops computing programs, and the first systems capable of a peak
petaflops performance are expected to appear in 2008. Problems in the
geosciences have been among the important drivers for the development
of such systems. One such problem is the seismic inverse problem of
determining the distribution of earth properties from surface
observations of earthquake-induced high-frequency ground motion in
large heterogeneous elastic regions. Scalability of inverse methods on
supercomputers requires both algorithmic scalability (scaling to large
problems sizes) and parallel scalability (scaling to large numbers of
processors). Here we focus on the deterministic inverse problem, and
we address both parallel efficiency and algorithmic efficiency of a
class of optimization methods for solution of full elastic waveform
seismic inverse problems as resolution limits are pushed. We discuss
algorithmic choices designed to assure scalability for various
components of the inverse method, including inexact Newton for
nonlinear iterations, conjugate gradients for linear iterations,
primal-dual active sets for treatment of parameter bounds, and
primal-dual treatment of total variation regularization. We conclude
with an assessment of the prospects of scalability of high-resolution
seismic inversion to upcoming petascale computing systems.
DE: 7270 Tomography (6982, 8180)
SC: Seismology [S]
MN: 2007 Fall Meeting