next up previous
Next: About this document ...

Hormozd Gahvari
Tying Iterative Solvers to Performance Models: Experiences with Algebraic Multigrid

University of Illinois at Urbana-Champaign
Department of Computer Science
201 N Goodwin Avenue
Urbana
IL 61801
gahvari@illinois.edu
William Gropp
Kirk E. Jordan
Martin Schulz
Ulrike Meier Yang

Emerging machines are becoming more massively parallel and exhibiting an increasing diversity of architectural features. Performance modeling becomes key to understanding how solvers perform on these machines, and where the opportunities are to improve performance. In this talk, we discuss our experiences with using performance modeling to drive improvements to algebraic multigrid (AMG). We use a performance model to guide a data redistribution algorithm that results in reduced communication and improved performance. We also use this model to inform the selection of thread/task mixtures on multicore nodes when using a hybrid programming model. In the future, we envision solvers like AMG making extensive use of performance models to ensure continued good performance and scalability.





Copper Mountain 2014-02-23