next up previous
Next: About this document ...

Andrey Prokopenko
Improving parallel performance of coarse grids in an algebraic multigrid

PO Box 5800
MS 1318
Albuquerque
NM 87185
aprokop@sandia.gov
Jonathan Hu
Sivasankaran Rajamanickam

Algebraic multigrid (AMG) solvers are an essential component in many large-scale scientific applications. However, increasing amounts of communication and data movement on coarser levels present significant challenges. In this talk, we discuss several methods targeted at improving the performance, such as reducing the number of coarse levels through aggressive coarsening, improving locality of communicating processors through repartitioning and careful mapping of subdomains to processors, reducing data movement through gathering and storing data redundantly. The results demonstrate substantial speedups on a variety of test problems.





Copper Mountain 2014-02-23