Numerical modelling of seismic wave propagation is central to seismic imaging and inversion. Modelling of 3D heterogeneous anisotropic media can, however, simultaneously be both costly and of low fidelity. To address some aspects of the fidelity / cost problem, we present our implementation of seismic modelling within the Rayleigh- Sommerfeld algorithmic framework and implemented in a parallel computing environment. We demonstrate through example, that knowledge of relationships among hardware architecture, support software systems, and resource usage of our algorithm, is of crucial importance to the achievement of worthwhile performance using parallel computing. From an initial, simplistic expression of our parallel algorithm in the MATLAB system, exhibiting `negative improvement', we incrementally adapt the parallel control framework without losing the clarity or correctness of the original model expression, to bypass successive bottlenecks and achieve substantial performance gains.
View full article as PDF (0.22 Mb)