Department of Computational and Applied Mathematics
"Structure-Preserving High Performance Computational Methods for Transport in Porous Media"
Massively parallel subsurface simulators, such as PFLOTRAN, are widely used by application scientists for solving important subsurface problems. These simulators can of course still be improved, so this talk explores ways one can enhance the current predictive capabilities. However, it is intractable to determine whethe proposed modifications to these subsurface simulators would be fast, scalable, or efficient without a performance model of some sort. This talk is divided into three parts. First, I describe in detail a performance spectrum model which covers time-to-solution, arithmetic intensity, and equations solved per second. Second, I apply a PETSc-base optimization-based solver to enforce non-negative constraint for anisotropic diffusion and use a "perfect-cache" roofline model to justify its performance. Third, I extend the same PETSc framework to coupled flow and transport problems like miscible displacement, and discuss current and future work.