Parallel implementation of domain-based preconditioning of Newton-Krylov methods for a natural convection model problem
- Idaho National Engineering Lab., Idaho Falls, ID (United States)
Results are presented showing parallel implementations of domain based preconditioners used in conjunction with a Newton-Krylov solver for calculating natural convection in a square cavity. Newton-Krylov techniques are based on the use of Newton`s method to linearize the discrete equations and a Krylov projection method to solve the resulting linear systems. The calculations are based on a finite volume discretization of the incompressible Navier-Stokes equations and an energy equation in primitive variable form on a staggered grid. Viability of the Newton-Krylov technique often depends on the effectiveness of the preconditioner. Consequently, effective preconditioning can be the most CPU and memory intensive operation within the solution algorithm. For these reasons, domain decomposition based preconditioners are used because of their inherent parallelism. Results are presented for strip-wise, domain-based preconditioners on two different computational architectures: single CPU and distributed computing cluster. These parallel results are compared and contrasted to the use of global, Incomplete Lower-Upper (ILU) factorization type preconditioners in a serial implementation.
- DOE Contract Number:
- AC07-94ID13223
- OSTI ID:
- 500852
- Report Number(s):
- CONF-950828--; ISBN 0-7918-1710-5
- Country of Publication:
- United States
- Language:
- English
Similar Records
Investigation of Newton-Krylov algorithms for low Mach number compressible flow
Preconditioning a Newton-Krylov solver for all-speed melt pool flow physics