CSAR Seminar

SPEAKER: Kenn K. Q. Zhang, University of Illinois at Chicago

TITLE: A Discrete Operator Splitting Superposition-Based Parallel Finite Element Method for Incompressible Flows

DATE: Wednesday, March 26, 2008
TIME: 12:00 Noon
PLACE: 2240 DCL
1304 W. Springfield Ave., Urbana, IL

ABSTRACT

The popular juxtaposition-based domain decomposition is intuitive, but it involves complicated pre-processing and sometimes communications of excessive quantities, is less adapted for continuous numerical methods and restricted to field problems, and requires the inter-subdomain boundary treatments which is different from the inter-element treatments, and is method dependent and sometimes problem dependent. In this presentation a superposition-based domain decomposition is proposed, which employs element-by-element construction, processor-level assembling, and condensed random data structure. Superposition-based parallelization shows great flexibility in partitioning the computational domains, communicates more consolidated data, and can be applied beyond field problems. Moreover, superposition-based parallelization can, as an option, follow the same numerical process as its serial counterpart to produce digit-by-digit the same result, which makes code development and debugging much easier. Solving large scale indefinite systems continues to pose as a challenging issue for incompressible flows and may lead to a way to solve compressible and incompressible flows in a unified approach. In this presentation, the discrete operator splitting (DOS) technique is proposed to break the original ill-natured large indefinite system into two smaller well-natured definite systems coupled through source terms. The underpinning idea of the technique is to seamlessly combine the splitting and iterations together. These two techniques (the parallelization and the DOS) are implemented with a continuous finite element method and applied to several incompressible Navier-Stokes benchmark flows. The algorithm for the logical parallelization, an upgrade from the superposition-based parallelization, will also be detailed.