A Comparison of Soft-Fault Error Models in the Parallel Preconditioned Flexible GMRES

Evan Coleman (1) , Aygul Jamal (2) , Marc Baboulin (2) , Amal Khabou (2) , Masha Sosonkina (1)

1 ODU - Old Dominion University [Norfolk]
2 ParSys - LRI - Systèmes parallèles (LRI)

The effect of two soft fault error models on the convergence of the parallel flexible GMRES (FGMRES) iterative method solving an elliptical PDE problem on a regular grid is evaluated. We consider two types of preconditioners: an incomplete LU factorization with dual threshold (ILUT), and an algebraic recursive multilevel solver (ARMS) combined with random butterfly transformation (RBT). The experiments quantify the difference between two soft fault error models considered in this study and compare their potential impact on the convergence.

Mots clés: fault tolerance; soft fault models; parallel iterative linear solvers; preconditioners; ARMS; ILUT; RBT randomization; FGMRES

Domaines: Analyse numérique [cs.NA]; Calcul parallèle, distribué et partagé [cs.DC]

Evan Coleman, Aygul Jamal, Marc Baboulin, Amal Khabou, Masha Sosonkina. A Comparison of Soft-Fault Error Models in the Parallel Preconditioned Flexible GMRES. International Conference on Parallel Processing and Applied Mathematics, Sep 2017, Lublin, Poland. pp.36-46, ⟨10.1007/978-3-319-78024-5_4⟩. ⟨hal-01768773⟩