In EDS ansehen

Loop Level Parallelization of a Seismic Inversion Code

Gespeichert in:
Bibliographische Detailangaben
Titel: Loop Level Parallelization of a Seismic Inversion Code
Autoren: M. Kern, W. W. Symes
Weitere Verfasser: The Pennsylvania State University CiteSeerX Archives
Quelle: ftp://ftp.caam.rice.edu/pub/people/trip/abstract/kern/kern1.ps.gz
Publikationsjahr: 1993
Bestand: CiteSeerX
Beschreibung: We present a parallel implementation of a seismic inversion code. Parallelism is exploited at the loop level within the finite difference modeling, as this is the most time consuming part of the code. We give details of our implementation, and present numerical results. We have reached a performance of 750 Mflops on one processor, and over 4 Gflops on 16 processors, on a Cray C90. 1 Introduction This paper presents experience with trying to exploit loop-level parallelism in DSO, a seismic inversion under development in the Rice Inversion Project. Seismic problems are characterized by the huge amount of data they generate. A typical seismic survey would generate several hundreds of Mbytes of data! Then, as will be explained below, the inversion procedure requires repeated solution of the wave equation (if we can be satisfied with acoustics !). It is thus of paramount importance that each of these solutions be done as fast as possible. Obviously, it is also desirable to solve as few o.
Publikationsart: text
Dateibeschreibung: application/postscript
Sprache: English
Relation: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.38.949
Verfügbarkeit: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.38.949
Rights: Metadata may be used without restrictions as long as the oai identifier remains attached to it.
Dokumentencode: edsbas.90317E2B
Datenbank: BASE
Beschreibung
Abstract:We present a parallel implementation of a seismic inversion code. Parallelism is exploited at the loop level within the finite difference modeling, as this is the most time consuming part of the code. We give details of our implementation, and present numerical results. We have reached a performance of 750 Mflops on one processor, and over 4 Gflops on 16 processors, on a Cray C90. 1 Introduction This paper presents experience with trying to exploit loop-level parallelism in DSO, a seismic inversion under development in the Rice Inversion Project. Seismic problems are characterized by the huge amount of data they generate. A typical seismic survey would generate several hundreds of Mbytes of data! Then, as will be explained below, the inversion procedure requires repeated solution of the wave equation (if we can be satisfied with acoustics !). It is thus of paramount importance that each of these solutions be done as fast as possible. Obviously, it is also desirable to solve as few o.