High Performance Fortran Gives Boost to Seismic Migration Project

Calgiare, 20-10 -1996 Although High Performance Fortran (HPF) has been following the tradition set by Fortran 77 and Fortran 90 in that the market has been slow to adopt it, today there are many users experiencing great success with HPF. CRS4, a Sardinian Research Center in Italy, working with a commercial oil exploration and production company is one of these successes. HPF is currently being used on the CRS4 seismic migration project known as the Geophysical Computing (GeoComp) Project and has helped to increase hardware efficiency and data accuracy.

The primary commercial application of the GeoComp project is the geophysical prospecting for oil. An established numerical technique known as "migration" is used to image the earth's interior starting from the echo of acoustic signals produced and detected on the earth's surface, going down for as many as several miles. Migration generates huge quantities of data that must be reduced, transformed, visualized and interpreted in order to obtain meaningful answers. For the GeoComp project specifically, 3D depth migration is being used, a more complex imaging technique that typically generates gigabytes of data.

The challenge of 3D depth migration is the accurate imaging of seismic data in the presence of complex geological structures. Given this challenge, we made it a requirement upfront that a standard language be adopted for the entire GeoComp project; we wanted the enhanced portability across different high-performance computers and the minimized maintenance effort a standard language would offer.

CRS4 first began migrating codes from CMFortran to the HPF language two years ago when HPF emerged as a standard in data parallel computing. With the GeoComp project, we started using HPF in the early stages with a public domain compiler. As the project progressed, we found we needed a maintained, robust and strongly-supported compiler to which we could code industrial applications. We chose The Portland Group, Inc.'s (PGI) HPF compiler.

Our first concern when choosing an HPF compiler was portability. With the PGI compiler, we were able to port two different migration kernels (about six thousand lines of code) developed on the IBM SP2 at CRS4 to an SGI PowerChallenge installed at the oil company in Milan in just a few days. Performance of the HPF codes has been comparable to similar codes which have been parallelized for the PowerChallenge by using SGI proprietary directives. To note, the HPF porting took significantly less time.

A major benefit HPF offers our project is the single address space across the physical processors. With HPF, seismic data are seen as global objects; in actuality they are distributed and aligned inside local processor memories by means of HPF directives. This single address space across processors makes the parallel platform appear to be a shared-memory machine; this is much more difficult to achieve with explicit message-passing programming.

Another key benefit HPF has brought to the GeoComp project is that data mapped in a single computational node are accessible through "extrinsic local" procedures written in Fortran 77, Fortran 90 or C. Those procedures can be compiled and linked separately to optimized scientific libraries (ESSL, NAG, SGIMATH, etc.) and/or message-passing libraries (MPI, PVM, IBM MPL, etc.). This means that with HPF it is possible to program MIMD machines in a SPMD style. Communications are possible through the previous mechanism or by using the array syntax, FORALL statement, and intrinsic and library procedures (SUM, "scatter" and prefix operations). Therefore, one can take the best of the two worlds by considering explicit message-passing as the "low-level language" of parallel programming.

The adoption of PGI's HPF compiler has enabled easier access to the parallel computing hardware used in the GeoComp project. The processing of enormous quantities of data that used to take weeks now takes only days and, by reducing our parallel software development effort, we have been able to validate more than one algorithm. The hardware efficiency experienced with this project has enabled us to recognize buried structures as potential oil reservoirs with a much higher degree of accuracy. This accuracy drastically reduces some of the risks involved in the exploration activities.

Ernesto Bonomi
Carlo Nardone