PLP: Towards a realistic and accurate model for communication performances on hierarchical cluster-based systems

Abstract

Today, due to many reasons, such as the inherent heterogeneity, the diversity, and the continuous evolving of actual computational supports, writing efficient parallel applications on such systems represents a great challenge. One way to answer this problem is to optimize communications of such applications. Our objective within this work is to design a realistic model able to accurately predict the cost of communication operations on execution environments characterized by both heterogeneity and hierarchical structure. We principally aim to guarantee a good quality of prediction with a neglected additional overhead. The proposed model was applied on point-to-point and collective communication operations and showed by achieving experiments on a hierarchical cluster-based system with heterogeneous resources that the predicted performances are close to measured ones.