A Framework for Adaptive Communication Modeling on Heterogeneous Hierarchical Clusters

Today, due to the wide variety of existing parallel systems consisting on collections of heterogeneous machines, it is very difficult for a user to solve a target problem by using a single algorithm or to write portable programs that perform well on multiple computational supports. The inherent heterogeneity and the diversity of networks of such environments represent a great challenge to model the communications for high performance computing applications. Our objective within this work is to propose a generic framework based on communication models and adaptive techniques for dealing with prediction of communication performances on based-clusters hierarchical platforms. Toward this goal, we introduce the concept of Poly-model of communications that corresponds to techniques to better model the communications in terms of the characteristics of the hardware resources of the target parallel system. We apply this methodology on collective communication operations and show that the framework provides significant performances while determining the best combination model-algorithm depending on the problem and architecture parameters