We are interested in this work by the combination of iterative solvers when solving linear systems of equations in an on-line setting. Our study targets users who may not be able to choose the best solvers for solving a set of linear systems while minimizing the total execution time. We propose a framework and algorithms in which the combination of solvers depends on informations gathered at runtime. The framework is assessed by extensive experiments using 5 SPARSKIT solvers over more than 70 matrices. The results show that the proposed approach is robust for solving linear sytems since we were able to solve more linear systems than each individual solver with an execution time nearly two times equal to those of the worst individual solver. Morever, we were able to predict a set of two solvers containing the best solver on more than 80% cases.