A coupled Newton-Krylov solver for improved CHAD cache utilization and performance.
Thomas Canfield, Tsai-Kan Chien, Henry M. Domanus, Adrian M. Tentner, Constantine P. Tzanos, Richard A. Valentin, David P. Weber · University of North Texas Digital Library (University of North Texas) · 2000
CHAD (Computational Hydrodynamics for Advanced Design) is a computer program that has been developed to analyze flows in automotive and defense applications. Extensive performance analysis of the CHAD computer program indicated a need to address cache memory use to increase computational performance. Several strategies have been adopted to achieve this goal: simultaneous solution of the coupled Navier-Stokes equations, data clustering, and data ordering. A coupled Newton-Krylov solver has been incorporated into a version of the CHAD program, resulting in consistent improvement in run times that varies from 50% to 200%. Further work will be required to tune the solver for optimal performance. In addition, experiments with data cluster and reordering indicate a potential for performance improvement.