Algorithm Development for Large Scale Computing

We present a case study examining the creation of an efficient preconditioner for solution of incompressible flow problems. We trace algorithm development from theoretical considerations, through initial implementation and testing, to parallel performance for large-scale systems. We demonstrate that early coupling of the development process with existing code can reduce debugging time and ease the transition from toy problems to real applications. 1 Overview A crucial step in integrating state-of-the-art algorithmic developments into large-scale simulation code is the migration of prototype code to efficient, precompiled library code. Ideally, the prototype implementation of a new algorithm should be high-level, modular, portable, and run in an interpreter so that the entire state is readily available. On the other hand, the final library version should be highly efficient and optimized for a target architecture. Accomplishing this transition can be extremely difficult, and usua..