Implementing Large-scale Parallel Geophysical Algorithms Using the Java Programming Language a Feasibility Study

Memory and time requirements of seismic calculations demand parallel implementations. One of the most widespread languages applied to this area is Fortran. It includes optimized functionality for scienti c computing. Unfortunately in Fortran, parallel program code is often error prone and lacks maintainability, reusabililty, and portability if Message Passing approaches are used. On the other hand Java contains parallel language constructs, provides portability, and demands strict object-orientation. These standard features of Java implicate superiority to Fortran in parallel computing. Unfortunately Java has reputation of poor performance. This study examines applicability of Java towards seismic computations. Two basic seismic operators, Velocity Analysis and Kirchho Migration, are implemented with Java in parallel. On an IBM SP/2 { one of the fastest emerging platforms for seismic processing { the maximal slowdown factor measured is always below 3. In some cases performance measurements even reach Fortran performance. These results show Java's applicability for seismic computations. Furthermore future Java releases exploiting more sophisticated compilation techniques focus only on further performance improvements. Thus they could eliminate Fortran's last superiority to Java almost completely. Acknowledgments This thesis has been supported by a diploma grant of the Prof. Dr.{Ing. Erich M uller Stiftung, Essen, Germany. I would like to thank Mobil Oil Corp. for supplying the marine AVO dataset which was used in this study. Furthermore, I want to express gratitude to the JavaParty group, especially Matthias Zenger, for their support. Matthias Schwab of the Stanford Exploration Project designed and provided the geophysical framework JAG. I would like to thank Maui High Performance Computing Center and the Computing Center of Karlsruhe University for the granting access on the IBM SP/2. 1