1: \begin{abstract}

2:   The simulation of lattice QCD on massively parallel computers stimulated the 

3:   development of scalable algorithms for the solution of sparse linear systems.

4:   We tackle the problem of the Wilson-Dirac operator inversion by 

5:   combining a Schwarz alternating procedure (SAP) in multiplicative form 

6:   with a flexible variant of the GMRES-DR algorithm.   

7:   We show that restarted GMRES is not able to converge when the system 

8:   is poorly conditioned. By adding deflation in the form of the FGMRES-DR 

9:   algorithm, an important fraction of the information produced by the iterates

10:   is kept between successive restarts leading to convergence in cases in which 

11:   FGMRES stagnates.

12:   

13: \end{abstract}

14: