|
We are concerned with the issues involved in designing, running, verifying, and validating large-scale simulations of nature, in particular the ground motion due to large earthquakes in Southern California. The process requires (1) parallel simulations running on thousands of processors, each of which can generate many terabytes of output datasets, followed by (2) analysis and visualization of the output datasets, verification against against other codes, and validation with observed data. Our approach for (1) introduces a new capability called end-to-end in-situ simulation, where mesh generation, partitioning, solving, and volume vizualization are performed in place and in parallel on the processors of a parallel machine. In this approach, a sophisticated distributed octree mesh data structure replaces cumbersome intermediate files, eliminating the "heroic" nature of typical large-scale simulation runs and enabling highly scalable general wave-adaptive meshing. Our approach for (2) introduces the idea of a computational database system, where all meshes, input, and output datasets are stored in compressed and indexed disk structures with an identical query interface. Queries to such datasets produce other indexed datasets and include operators such as transpose, transform, filter, resample, interpolate, and volume rendering, which are themselves complex HPC applications. |