The open archive for STFC research publications

Full Record Details

Persistent URL http://purl.org/net/epubs/work/65757
Record Status Checked
Record Id 65757
Title Sparse Communication Avoiding Pivoting
Abstract On future machines communication will be the most expensive operation that algorithms seek to avoid as computation is spread across many cores. Recent developments in dense linear algebra focus on communication avoiding algorithms, particularly for pivoting. However the ideas and methods cannot be ported directly to sparse solvers as their design does not consider fill-in, which still needs to be avoided. We present a new technique called compressed pivoting that seeks to extend the traditional sparse threshold pivoting algorithms in a way that reduces the amount of communication required. Two variants are considered: (1) a pessimistic version that guarantees pivoting tests will always be met, and (2) an optimistic heuristic that works most of the time with a back-tracking mechanism for when it gets things wrong.
Organisation STFC , SCI-COMP , SCI-COMP-CM
Keywords SCI COM 2013-2014
Funding Information
Related Research Object(s):
Licence Information:
Language English (EN)
Type Details URI(s) Local file(s) Year
Presentation Presented at 25th Biennial Conference on Numerical Analysis, University of Strathclyde, Scotland, 25-28 Jun 2013. capiv.pdf 2013