LANS Informal Seminar
"Accelerating Algebraic Multigrid on GPUs"
DATE: December 1, 2011
TIME: 10:30:00 - 11:30:00 Description:
SPEAKER: Steven Dalton, PhD Student at UIUC
LOCATION: Building 240, 1404-1405, Argonne National Laboratory
Accelerating algebraic multigrid methods on massively parallel throughput-oriented processors, such as the GPU, demands algorithms with abundant fine-grained parallelism. Sparse matrix-vector multiplication operations dominate the performance of the cycling phase of algebraic multigrid and we use efficient GPU implementations to achieve notable speedup on a representative set of matrices. We also present novel sparse matrix operations required to construct the AMG hierarchy. The GPU sparse matrix-matrix and maximal independent operations avoid transfer operations and achieve an average of 2x speedup. Our algorithms are expressed as collections of data parallel primitives provided by the Thrust library and available as part of the Cusp sparse matrix library.
TIME: 10:30:00 - 11:30:00
Please send questions or suggestions to Krishna: snarayan at mcs.anl.gov.