Automatic Parallelization For Seismic Data Processing Programs On Grid Environment

Parallel compiling has been carried out with the development of the parallel computer architecture. And how to use parallel system efficiently to solve the computation-oriented problems is a difficulty in computer science. It is an all-important key to develop high performance parallel software.Grid is an innovative model of distributed memory, focusing on computing resources sharing for data processing and collaborative applications of Virtual Organizations. Grid is expected to integrate the scattered resources and processing ability into an organic whole, which can provide higher computing performance than any single high performance multi-processors. Thus, it is significant to research the technology of automatic parallelization on Grid.As a hot research topic in parallel computing area, automatic parallelization is attracting more and more researchers. Much progress in exploiting coarse-grain parallelism has been made in recent years, but application results are still disappointing, since many programs achieved little or no speedup while executed in parallel. In this paper, we study the Seismic data processing programs and mainly analyze the 3D Pre-stack Depth Migration, for the sake of applying automatic parallelization to grid computing so as to achieve data parallelism with large scale and high efficiency. The main contributions of this paper can be listed as follows:1. By reading a mass of source codes about 3D Pre-stack Depth Migration, we point out that, in pre-stack depth migration based on kirchhoff, the computation of travel time is fit for parallelization according to shot, the computation of imaging result is fit for parallelization according to offset; and in the split-step pre-stack depth migration, it is fit for parallelization according to single-shot record.2. We present APPSDMM (automatic parallelization of 3D Pre-stack Depth Migration Model) based on the analysis of 3D Pre-stack Depth Migration sequential programs and its introduction of key techniques.3. In order to resolve the problems of data and loops distribution, we partition it into two parts according as whether the processors communicate with others after distribution. When it does not relate to communications, we put forward three algorithms, which respectively resolve the problems of alignment between arrays and loops, arrays with the same name, and arrays which with different names. As for the other situation, we implement the problems through extracting APDG (Automatic Parallel Distribution Graph) from the multilayer nesting loops, which is subject to the restriction that the edges connected between different subsets after distribution are least.4. An algorithm based on improved ant algorithm was presented to solve the problem of task scheduling after the partition of APDG.