University of Houston

In the past, most numerical modeling of seismic acquisition has been confined to either 3D scalar/acoustic models or 2D elastic models. However, the convergence of several factors has spurred a renewed interest in 3D elastic modeling of realistic geologic environments. Given this context, I have attempted to identify the difficulties involved in using the finite difference modeling method for highly heterogeneous 3D elastic models such as encountered in deepwater OBC experiments. The typical application results in input discretized model sizes on the order of tens of gigabytes such that domain decomposition and the communication methodologies between the processes becomes a crucial factor. Efficient domain decomposition can increase performance by reducing caches misses while better processor memory bandwidth and an efficient communication strategy may reduce the time computation to communication ratio. I evaluate and calibrate alternative domain decomposition and commun ication strategies through the simulation of moderate-sized 3D elastic model running on a 42-node Beowulf system.