ABSTRACT

Cartesian structured adaptive mesh refinement methods have become an increasingly popular modelling approach to solving partial differential equations in complex or irregular geometries. This chapter describes the primary framework optimizations that enable scalable simulations. The effects of local geometry caching for a high-fidelity incompressible viscous flow computation in pore-scale geometry of packed spheres are shown. While the algorithm is automatic and robust, it is a nontrivial amount of computation and is too large to be replicated across processing elements. Hence, EBIndexSpace generation must be computed in parallel and stored in a distributed fashion. High-order EB stencils are produced from a least-squares algorithm at runtime based on the geometric moments and current mesh hierarchy. For geometrically dense problem definitions, the sparse file format has little savings in terms of final storage and it incurs the collective communication burden of creating the prefix sum.