The open archive for STFC research publications

Full Record Details

Persistent URL http://purl.org/net/epubs/work/12417475
Record Status Checked
Record Id 12417475
Title Sparse Communication Avoiding Pivoting and GPUs
Abstract Modern GPUs provide a window on future hardware platforms that we need to design for. In this talk we describe our experience in implementing direct methods for sparse symmetric indefinite matrices in CUDA, a topic that presents both mathematical and technical challenges. Whilst a number of communication-avoiding methods have been developed for dense linear algebra, their applicability to sparse linear algebra is limited as they ignore sparsity in selecting pivots, generating significant fill. We will describe heuristics that attempt to manage this complexity through a series of fall-back strategies that in the best case have similar communication properties to the unpivoted Cholesky factorization, but for numerically challenging problems deliver stability on a par with the best existing sparse symmetric indefinite direct solvers.
Organisation STFC , SCI-COMP , SCI-COMP-CM
Funding Information
Related Research Object(s):
Licence Information:
Language English (EN)
Type Details URI(s) Local file(s) Year
Presentation Presented at SIAM Computational Science and Engineering 2015 (CSE15), Salt Lake City, Utah, USA, 13-18 Mar 2015. sps_ca_piv.pdf 2015