Improving Scientist Productivity, Architecture Portability, and Performance in ParFlow

Legacy scientific applications represent significant investments by universities, engineers, and researchers and contain valuable implementations of key scientific computations. Over time hardware architectures have changed. Adapting existing code to new architectures is time consuming, expensive, and increases code complexity. The increase in complexity negatively affects the scientific impact of the applications. There is an immediate need to reduce complexity. We propose using abstractions to manage and reduce code complexity, improving scientific impact of applications. This thesis presents a set of abstractions targeting boundary conditions in iterative solvers. Many scientific applications represent physical phenomena as a set of partial differential equations (PDEs). PDEs are structured around steady state and boundary condition equations, starting from initial conditions. The proposed abstractions separate architecture specific implementation details from the primary computation. We use ParFlow to demonstrate the effectiveness of the abstractions. ParFlow is a hydrologic and geoscience application that simulates surface and subsurface water flow. The abstractions have enabled ParFlow developers to successfully add new boundary conditions for the first time in 15 years, and have enabled an experimental OpenMP version of ParFlow that is transparent to computational scientists. This is achieved without requiring expensive rewrites of key computations or major codebase changes; improving developer productivity, enabling hardware portability, and allowing transparent performance optimizations