Department of Computer Science and Engineering
Abstract: Unstructured meshes are a powerful computational tool used in the numerical modeling of physical phenomena on complex, irregular domains. Instead of requiring a uniform distribution of grid points, unstructured meshes allow grid points to be strategically placed in the computational domain. Thus, these meshes are particularly effective in modeling irregular boundaries, multiscale geometries, and rapidly changing solutions. The use of these meshes on parallel machines has been limited because of the number of diverse tasks that must be accomplished in a scalable manner to use this approach effectively. These tasks include the generation, adaptive refinement, quality improvement, and partitioning of the unstructured mesh. In addition, many applications require the assembly and solution of large, sparse linear systems. In this talk, we describe new parallel algorithms for these tasks and discuss the vertical intergration of software for accomplishing these tasks as part of the SUMAA3d project. We present a P-RAM analysis showing the scalability of a strategy based on computation on independent sets. Experimental results obtained on a diverse set of parallel architectures, including the IBM SP and ATM-connected networks of workstations, are discussed. Finally, we compare the parallel performance of the two- and three-dimensional versions of these algorithms from the point of view of scalability and computational efficiency.