On the Use of Performance Models for Adaptive Algorithm Selection on Heterogeneous Clusters

Due to the increasing diversity and the continuous evolution of existing parallel systems, solving efficiently a target problem by using a single algorithm or writing efficient and portable programs is becoming a challenging task. In this paper, we present a generic framework that integrates performance models with adaptive techniques in order to design efficient parallel algorithms in heterogeneous computing environments. To illustrate our approach, we study the matrix multiplication problem, where we compare different parallel algorithms. Experiments demonstrate that accurate performance predictions obtained from analytical performance models allow us to select the most appropriate algorithm to use depending on the problem and the platform parameters.