Multicore architectures have been one of the primary driving forces in the recent rapid growth in high-end computing systems, contributing to its growing scales and capabilities. With significant enhancements in high-speed networking technologies and protocol stacks which support these high-end systems, a growing need to understand the interaction between them closely is realized. Since these two components have been designed mostly independently, there tend to have often serious and surprising interactions that result in heavy asymmetry in the effective capability of the different cores, thereby degrading the performance for various applications. Similarly, depending on the communication pattern of the application and the layout of processes across nodes, these interactions could potentially introduce network scalability issues, which is also an important concern for system designers. In this thesis, we analyze these asymmetric interactions and propose and design a novel systems level management framework called SIMMer (Systems Interaction Mapping Manager) that automatically monitors these interactions and dynamically manages the mapping of processes on processor cores to transparently maximize application performance. Performance analysis of SIMMer shows that it can improve the communication performance of applications by more than twofold and the overall application performance by 18%. We further analyze the impact of contention in network and processor resources and relate it to the communication pattern of the application. Insights learnt from these analyses can lead to efficient runtime configurations for scientific applications on multicore architectures. / Master of Science

La tendance actuelle des constructeurs pour le calcul scientifique est à l'utilisation de grappes de machines dont les noeuds comportent un nombre de coeurs toujours plus grand. Le modèle basé uniquement sur MPI laisse peu à peu la place à des modèles mélangeant l'utilisation de threads et de MPI. Ce changement de modèle entraîne de nombreuses problématiques car les implémentations MPI n'ont pas été conçues pour supporter les applications multi-threadées. Dans cette thèse, afin de garantir le bon fonctionnement des communications, nous proposons un module logiciel faisant interagir l'ordonnanceur de threads et la bibliothèque de communication. Ce gestionnaire d'entrées/sorties générique prend en charge la détection des événements du réseau et exploite les multiples unités de calcul présentes sur la machine de manière transparente. Grâce à la collaboration étroite avec l'ordonnanceur de threads, le gestionnaire d'entrées/sorties que nous proposons assure un haut niveau de réactivité aux événements du réseau. Nous montrons qu'il est ainsi possible de faire progresser les communications réseau en arrière-plan et donc de recouvrir les communications par du calcul. La parallélisation de la bibliothèque de communication est également facilité par un mécanisme d'exportation de tâches capable d'exploiter les différentes unités de calcul disponible tout en prenant en compte la localité des données. Les gains obtenus sur des tests synthétiques et sur des applications montre que l'interaction entre la bibliothèque de communication et l'ordonnanceur de threads permet de réduire le coût des communications et donc d'améliorer les performances d'une application. / The current trend of constructors for scientific computation is to build clusters whose node include an increasing number of cores.The classical programming model that is only based on MPI is being replaced by hybrid approaches that mix communication and multi-threading. This evolution of the programming model leads to numerous problems since MPI implementations were not designed for multi-threaded applications. In this thesis, in order to guarantee a smooth behavior of communication, we propose a software module that interact with both the threads scheduler and the communication library. This module, by working closely with the thread scheduler, allows to make communication progress in the background and guarantees a high level of reactivity to network events, even when the node is overloaded. We show that this permits to make communication progress in the background and thus to overlap communication and computation. The parallelization of the communication library is also made easier thanks to a task onloading mechanism that is able to exploit the available cores while taking data locality into account. The results we obtain on synthetic application as well as real-life applications show that the interaction between the thread scheduler and the communication library allows to reduce the overhead of communication and thus to improve the application performance.