|
Efficient Parallel Algorithms for Computational Fluid Dynamics
Vivek Sarin
Purdue University
Hosted by Jorge More
10:30AM, Tuesday, February 22, 1999
Building 221, A-216
|
| Abstract |
This talk addresses a number of challenging issues in developing parallel
algorithms for large-scale scientific \ simulations. We will present recent results from
the simulation of particles in fluids, a problem of great interest to numerous industries,
e.g., the petroleum industry. As part of a Grand Challenge project, we have developed two
state-of-the-art packages called Particle Movers based on distinct modeling approaches,
and have implemented them on several parallel platforms. The most time-consuming component
of these simulations is the solution of nonsymmetric, indefinite, sparse linear systems at
each time step. This poses significant challenges in developing preconditioners for
iterative solvers that are robust and effective in addition to being parallelizable. We
have developed a number of innovative techniques that include a novel multilevel algorithm
to represent discrete divergence-free space for incompressible fluids and optimal
matrix-free parallel preconditioners for enforcing the rigid-body motion of the particles.
Our approach combines robust and effective preconditioning with excellent parallel
performance, and is superior to methods based on domain decomposition, multigrid,
approximate inverses, and incomplete factorizations for fluid problems. Using our
techniques, realistic three-dimensional flows with thousands of particles have been
simulated for the first time on high performance architectures such as the SGI Origin2000.
Ongoing work supported by a KDI award extends this work to various types of fluids.
Relevant publications and additional details are available on the Web. |
