The open archive for STFC research publications

Full Record Details

Persistent URL http://purl.org/net/epubs/work/25845401
Record Status Checked
Record Id 25845401
Title Numerically-aware nested dissection ordering
Abstract For numerically difficult sparse symmetric systems such as those often arising in interior point methods, sometimes the application of a scaling alone in the preprocessing step is insufficient to make the problem numerically tractable for a direct solver to solve without significant performance-inhibiting pivoting. In such cases the enforcement of specific 2x2 pivots through the use of a restricted ordering is often effective, such as the method based on MC64 suggested by Duff and Pralet. However such restrictions are normally crude and result in significantly more fill that an unrestricted ordering. In this talk we will describe a modified nested dissection method that takes into account numerical values of the matrix and avoids separators likely to result in delayed pivots. We will present results demonstrating that it can keep the number of delayed pivots in a subsequent LDL^T factorization low, being almost as numerically effective as existing methods, but with significantly less additional fill.
Organisation CSE-NAG , STFC , SCI-COMP
Funding Information EPSRC (EP/M025179/1)
Related Research Object(s): 24428477 , 35371260
Licence Information:
Language English (EN)
Type Details URI(s) Local file(s) Year
Presentation Presented at Sparse Days at Centre Européen de Recherche et de Formation Avancée en Calcul Scientifique 2016 (CERFACS 2016), Toulouse, France, 30 Jun 2016 - 1 Jul 2016. num_aware_order.pdf 2016