|
Multimethod Solvers: Algorithms, Applications, and
Software
Sanjukta Bhowmick
University of Illinois at Chicago
Hosted by Paul Hovland
10:30 AM, April 30, 2004
Building 221, Room A216
|
| Abstract |
My presentation is on developing ``multimethod'' solvers for faster and
more robust solution of large sparse linear systems, generated in course of solving nonlinear PDE-based simulations with Newton's method. For such
applications, the performance of different solvers can vary dramatically depending on the numerical properties of the system, the simulation
state, uniprocessor or multiprocessor execution, etc. Thus it is almost impossible to select a single
"best-performing'' solver.
We therefore, combine multiple solution methods to derive schemes such as dynamic method selection. In particular I will discuss a combinatorial
approach for constructing composite solvers which uses a sequence of methods on the same system to provide a robust solution. I will report on
the performance of serial and parallel multi method solvers and provide a brief discussion of their implementation as components. I will summarize
direction for future research including the extension of multi method and multilevel techniques for automatic differentiation.
|
