Workshop: IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms
Authors: Lorenzo Migliari (Università della Svizzera italiana) and Dimosthenis Pasadakis and Olaf Schenk (Università della Svizzera italiana, Panua Technologies)
Abstract: The numerical factorization of sparse symmetric matrices is an irregular, memory-bound kernel. State-of-the-art solvers leave the placement of the factor implicit on multi-socket NUMA architectures, binding it to a domain by first touch. This work instead binds the factor during the analysis phase, partitioning it along the elimination tree so that each supernode is placed close to those updating it. Each thread then factors the supernodes bound to its own memory and steals from the nearest domain. Furthermore, no thread binding or memory policy has to be chosen since every decision is derived from the discovered compute node topology. The method is implemented in a production-level LDL solver and evaluated on Intel Ice Lake, AMD Genoa, and NVIDIA Grace over 34 SuiteSparse matrices. It outperforms MKL PARDISO, Panua PARDISO, CHOLMOD, and PaStiX, achieving geometric-mean speedups from 1.3 to 9.5 and, at flop parity, 2.7 less cross-socket traffic than Panua PARDISO.
Back to IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms Archive Listing
Back to Full Workshop Archive Listing