Parallel/Distributed Software for the Petroleum Industry

Many applications in the petroleum industry can benefit tremendously from parallel computation. Examples include both seismic data processing and reservoir simulation.

One of the most popular software tools for parallel processing is Linda. Linda is a high-level language for developing portable parallel applications. Using Linda, it is possible to rapidly develop efficient versions of FORTRAN and C programs to run on a variety of platforms, including parallel MIMD computers or across networks of high performance workstations.

SCIENTIFIC has worked closely with the petroleum industry to develop parallel applications. In a joint project with Weidlinger Associates, SCIENTIFIC helped to develop parallel raytracing software for seismic modeling.

INTEGRA 3-D Seismic Modeling Program

Computing acoustic wavefields is extremely difficult for complex 3-D geophysical models. One successful technique for this task involves the use of ray-tracing, in which the software computes the travel paths of rays orthogonal to the acoustic wavefronts. The INTEGRA seismic modeling program, developed by Weidlinger Associates, is a commercial implementation of the ray-tracing approach.

To speed up the modeling process, Weidlinger Associates reimplemented INTEGRA for execution on parallel computes and on networks of workstations using Linda. Linda was chosen because of its efficiency, portability, ease of use, and commercial support from SCIENTIFIC.

To parallelize INTEGRA, the well-known master-worker paradigm was used. In this approach, the master process sets up the problem, creates a number of worker processes and generates a collection of ray-tracing tasks. Each worker process then repeatedly grabs a task, computes the required ray paths and records the results. Meanwhile, the master process collects the results and carries out post-processing such as visualization. The computation ends when all ray tracing tasks have been completed. So long as there are sufficiently many tasks, the total computation time may be decreased simply by creating a large number of workers.

Linda is ideal for this application since it permits dynamic load balancing among the worker processes in a particularly simple and natural way. The code was easy to develop and debug, and as shown in the accompanying chart, it performs well. Notice that the code displays essentially linear speedup--that is, the computational speed increases in direct proportion to the number of workers. Thus the program makes particularly effective use of system resources.

Looking for Oil with Linda

The search for oil and gas makes heavy use of reflection seismology to promote an understanding of subsurface geology. The first step is the collection of raw data by seismic surveys (experiments measuring the strength of signals reflected from pulses induced at the earth's surface). Typical surveys use multiple source and receiver locations and yield large quantities of data for analysis. Since the trend is toward the generation of large three-dimensional datasets to increase the accuracy of the subsurface information, the problem size can be expected to grow rapidly.

To build up an image of the subsurface geological structure for analysis by a trained geophysicist, the raw data must be correlated and corrected for distortion caused by such things as variations in signal strength and subsurface material densities.

A technique know as prestack depth migration gives the most focused and correctly positioned images. However, besides number-crunching power, it demands much in terms of storage and data movement. So most of the 3-D prestack migration work to date has been conducted on large vector or parallel supercomputers, and using such machines for such enormous computations has proved too costly for most exploration programs.

At IBM's Parallel and Geoscience Computing Group in Dallas, TX, however, James Black, Chen Bin Su, and their colleagues have called on Linda to process seismic data on networks of IBM RS/6000 workstations. They used Network Linda to run 3-D prestack migration on a dataset of 900 traces of 500 samples each. (The output image was a box containing 64 by 64 by 256 points) The computation was carried out on a cluster of 32 workstations, each with a 62.5-MHz clock and 64 MB of memory, connected by a token ring. For the basic migration computation (that is, for the parallelizable compute-intensive part of the processing), nearly perfect parallel speedups (based on elapsed times) were obtained with four to 31 workstations.

However, the speedups ignore the additional time required to integrate the partial images on the master workstation for display or other purposes. The time for image collection depends on the interconnection network, and for relatively slow networks based on Ethernet or token ring technology, it can be a significant fraction of the total application time. For this example, the limiting effects of the token ring network have little impact for cluster sizes of 20 nodes or so, but markedly degrade parallel performance past that point.

~ Leigh D. Cagan and Andrew H. Sherman