Networked Parallel Seismic Computing

Abstract

Abstract Workstations with high floating-point performance offer new ways of doing 3-D seismic computations: executing on a loosely-coupled set of workstations connected by a network. We have converted a standard one-pass 3-D finite-difference migration application to thk form and have achieved a speedup of 3 using 4 "worker" workstations connected via a relatively low-performance local-area network. We have also analyzed the factors (e.g. granularity y, communication bandwidth, memory per workstation) that determine the performance we have observed. 1 INTRODUCTION There is currently strong interest in modifying existing seismic processing algorithms to take advantage of parallel processing architectures. For example, at the 1991 Annual Convention of the Society of Exploration Geophysicists, there was one entire technical session [1] devoted to parallel programming of various depth-migration techniques, which are the most accurate (and most expensive) of the imaging techniques for seismic data. The current paper is part of that same general effort, but with a slightly different twist. Here we are interested in understanding the issues involved in running seismic algorithms across a loosely-coupled cluster of high-performance workstations rather than on a massivelyparallel computer. Clusters of R.ISC-based workstations offer good raw CPU performance and very attractive cost-performance ratios when compared with traditional supercomputers. In fact their cost-performance (roughly $300 per "Toward-Peak- Performance" [2] megaflop/s) rivals that of most massively parallel computers available today. However, doubts are often expressed as to their role in 3-D seismic computations, known to be compute-intensive, communication-intensive, and 1/0 intensive. The purpose of this paper is to analyze some aspects of these issues and to present some performance measurements we have made. One-pass 3-D finite-difference depth migration is a standard technique [3] in widespread use today as a post-stack imaging tool. Thk technique (and similar ones) are of interest because it currently takes several days to complete a typical run on a supercomputer that can sustain 100 megaflop/s. It is also a particularly good test case for parallel computing because it involves handling very large (several Gbytes)datasets that cannot be assumed to reside entirely in memory on any one of the computers in the network. Thus it forces us to address the issues of large-scale data management and data communication among the processors in the network. The type of parallel methodology applicable to a workstation cluster is MIMD (multiple instruction multiple data) rather than SIMD (Single instruction multiple data). More specifically we use the SPMD (single program multiple data) paradigm, where each "worker" workstation executes the same program on different parts of the data. Generallyavailable Fortran or C compiler technology is currently incapable of analyzing and exploiting the large-grain parallelism appropriate for the SPMD method. Fortunately there exist a number of tool sets (e.g. pre-compilersl subroutine libraries, and special daernons) which greatly simplify the program- ming effort, freeing the programmer of much of the drudgery of communication and synchronization tasks.