Paul E. Plassmann and Ivana Veljkovic
Department of Computer Science and Engineering
Penn State University
University Park, PA 16802, USA
Abstract:
With the increasing availability of high-performance parallel computers simulations that include complex multiscale models, such the the detailed chemistry in a reacting flow, are now possible. The central problem with such simulations is the computational expense of the chemistry calculations--a direct implementation of the required function evaluations often comprise more than 99of the overall time of computation. However, by archiving previously computed function values in an on-line database and approximating subsequent queries we have been able to dramatically speed up such calculations. In this talk we discuss the extension of this approach from a sequential to a parallel implementation. A number of interesting issues including load-balancing and the optimal manner to distribute the database search tree are considered. Finally, we present experimental results of the use of these algorithms in the parallel solution of a reacting flow with complex chemistry obtained from a high-order, Direct Numerical Simulation (DNS) simulation.