Locality Optimization on a NUMA Architecture for Hybrid LU Factorization

Adrien Rémy (1, 2) , Marc Baboulin (1, 2) , Masha Sosonkina (3) , Brigitte Rozoy (2)

1 POSTALE - Performance Optimization by Software Transformation and Algorithms & Librairies Enhancement
2 ParSys - LRI - Systèmes parallèles (LRI)
3 ODU - Old Dominion University [Norfolk]

We study the impact of non-uniform memory accesses (NUMA) on the solution of dense general linear systems using an LU factorization algorithm. In particular we illustrate how an appropriate placement of the threads and memory on a NUMA architecture can improve the performance of the panel factorization and consequently accelerate the global LU factorization. We apply these placement strategies and present performance results for a hybrid multicore/GPU LU algorithm as it is implemented in the public domain library MAGMA.

Mots clés: LU factorization; thread placement; ccNUMA; dense linear systems; MAGMA library

Domaines: Analyse numérique [cs.NA]

Adrien Rémy, Marc Baboulin, Masha Sosonkina, Brigitte Rozoy. Locality Optimization on a NUMA Architecture for Hybrid LU Factorization. Parallel Computing: Accelerating Computational Science and Engineering, 25, pp.153-162, 2014, Advances in Parallel Computing, ⟨10.3233/978-1-61499-381-0-153⟩. ⟨hal-00987284⟩