We have described our approach to modeling the earthquake-induced ground motion in large, heterogeneous basins on parallel computers. By paying careful attention to the impact on parallel execution of all components of the code, we are able to obtain excellent performance on highly unstructured mesh problems. In particular, through the use of (i) space- and time-localized absorbing boundaries; (ii) seismic input in the form of effective boundary or interior forces applied at the element level; (iii) explicit numerical techniques for the wave propagation problem; (iv) strict control of mesh resolution and aspect ratio; and (v) an asymptotically optimal mesh partitioner, we obtain excellent scalability of the parallel code. The Archimedes toolset integrates the basic components necessary for solving general PDE problems involving static unstructured meshes on parallel distributed memory systems. These components include meshing, partitioning, and parallel code generation.
We currently solve the meshing, partitioning, and parceling problems sequentially on a large shared-memory machine. Our ultimate target problem--the Greater Los Angeles Basin with an excitation of 2 Hz and with soil deposits having shear wave velocities as low as 200 m/s--will require meshes on the order of hundreds of millions of elements. Despite the fact that our sequential meshing and partitioning codes are fast, we may have to parallelize these steps in order to solve the target problem, primarily for memory reasons. The scalability of the parallel portion of our code suggests that our target problem is within reach.