 |
(tel) 617.253.8027 (fax) 617.253.6385 |
|
(tel) 617.253.8027 (fax) 617.253.6385 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Very Large Scale Finite Difference Modeling of Seismic Waves
Giuseppe A. Sena
Submitted
to the Department of Earth, Atmospheric, and Planetary Sciences on September
21, 1994 in partial fulfillment of the requirements for the degree of Doctor
of Philosophy
Abstract
In this thesis we develop a method for solving very large scale seismic
wave propagation problems using an out of core finite difference approach.
We implement our method on a parallel computer using special memory management
techniques based on the concepts of pipelining, asynchronous I/O, and dynamic
memory allocation.
We successfully apply our method for solving a 2-D Acoustic Wave equation
in order to show its utility and note that it can be easily extended to solve
3-D Acoustic or Elastic Wave equation problems. We use second order finite
differencing operators to approximate the 2-D Acoustic Wave equation. The
system is implemented using a distributed-memory/message-passing approach
on an nCUBE 2 parallel computer at MIT’s Earth Resources Laboratory.
We use two test cases, a small (256 X 256 grid) constant velocity model and
the Marmousi velocity model (751 X 2301 grid). We conduct several trials-
with varying memory sizes and number of nodes- to fully evaluate the performance
of the approach. In analyzing the results we conclude that the performance
is directly related to the number of nodes, memory size, and bandwidth of
the I/O subsystem.
We demonstrate that is feasible and practical to solve very large scale seismic
wave propagation problems with current computer technologies by using advanced
computational techniques. We compare two versions of the system, one using
asynchronous I/O and the other using synchronous I/O, to show that better
results can be obtained with the asynchronous version with pipelining and
overlapping of I/O with computations.
