Title: Architecture-Aware Algorithms and Combinatorial Techniques for High-Performance Computing

Speaker: Michael M. Wolf, Sandia National Laboratories

Date/Time: Tuesday, December 7, 2010, 11:00 am mountain time        

Location: CSRI Building/Room 90 (Sandia NM)

Brief Abstract: I give a broad overview of my recent work in the areas of combinatorial and high performance scientific computing. As part of this overview, I discuss into research on sparse matrix partitioning and the software developed to facilitate this partitioning.

I spend more substantial time discussing my work on architecture-aware algorithms for extreme-scale computing. In particular, I describe efforts to improve the scalability of linear systems, using hybrid MPI/threaded algorithms to solve these systems while exploiting the underlying system hierarchy. For this approach to yield scalable linear solvers, we need efficient threaded triangular solvers (important for preconditioning) to run on the multicore nodes. I briefly describe such a threaded triangular solver and issues involved, and present numerical results. For the integration of these hybrid MPI/threaded linear solvers into existing large-scale scientific simulations to be painless, we advocate using MPI methods for shared memory allocation on the multicore node. I discuss two relevant programming models and give an example of using MPI shared memory allocation in a mini-application. I finish by describing potential future research directions for the discussed combinatorial and high performance scientific computing research.

CSRI POC: Scott Collis, 1442, 284-1123