Due to the unpredictability in job arrival times in clusters and widely varying resource requirements, dynamic scheduling of parallel computing resources is necessary to increase system throughput. Dynamically resizable applications provide the flexibility needed for dynamic scheduling. These applications can expand to take advantage of additional free processors, or to meet a Quality of Service (QoS) deadline, or can shrink to accommodate a high priority application, without getting suspended. This thesis is part of a larger effort to define a framework for dynamically resizable parallel applications. This framework includes a scheduler that supports resizing applications, an API to enable applications to interact with the scheduler, and libraries that make resizing viable. This thesis focuses on libraries for efficient resizing of parallel applications—efficient in terms of minimizing the cost of data redistribution, choosing and allocating the right set of additional processors, and focusing on the performance of the application after resizing. We explore the tradeoffs between these goals on both homogeneous and heterogeneous clusters. We focus on structured applications that have 2D data arrays distributed across a 2D processor grid. Our library includes algorithms for processor selection and processor mapping. For homogeneous clusters, processor selection involves selecting the number of processors that needs to be added and processor mapping decides the placement of the new processors in the context of the given topology such that it minimizes the amount of data that is to be redistributed. For heterogeneous clusters, since the processing powers of the processors vary, there is also an additional problem of choosing the right set of processors that needs to be added. We also present results that demonstrate the effectiveness of our approach. / Master of Science

Les infrastructures de cloud fournissent une grande variété de ressources de calcul à la demande avec différents compromis coût-performance. Cela donne aux utilisateurs des nombreuses opportunités pour exécuter leurs applications ayant des besoins complexes en ressources, à partir d’un grand nombre de serveurs avec des interconnexions à faible latence jusqu’à des dispositifs spécialisés comme des GPUs et des FPGAs. Les besoins des utilisateurs concernant l’exécution de leurs applications peuvent varier entre une exécution la plus rapide possible, la plus chère ou un compromis entre les deux. Cependant, le choix du nombre et du type des ressources à utiliser pour obtenir le compromis coût-performance que les utilisateurs exigent constitue un défi majeur. Cette thèse propose trois contributions avec l’objectif de fournir des bons compromis coût-performance pour l’exécution des applications sur des plates-formes hétérogènes. Elles suivent deux directions : un bon usage des ressources et un bon choix des ressources. Nous proposons comme première contribution une méthode de partage pour des accélérateurs de type FPGA dans l’objectif de maximiser leur utilisation. Dans une seconde contribution, nous proposons des méthodes de profilage pour la modélisation de la demande en ressources des applications. Enfin, nous démontrons comment ces technologies peuvent être intégrées dans une plate-forme de cloud hétérogène. / Cloud infrastructures provide on-demand access to a large variety of computing devices with different performance and cost. This creates many opportunities for cloud users to run applications having complex resource requirements, starting from large numbers of servers with low-latency interconnects, to specialized devices such as GPUs and FPGAs. User expectations regarding the execution of applications may vary between the fastest possible execution, the cheapest execution or any trade-off between the two extremes. However, enabling cloud users to easily make performance-cost trade-offs is not a trivial exercise and choosing the right amount and type of resources to run applications accordingto user expectations is very difficult. This thesis proposes three contributions to enable performance-cost trade-offs for application execution in heterogeneous clouds by following two directions: make good use of resources and make good choice of resources. We propose as a first contribution a method to share FPGA-based accelerators in cloud infrastructures having the objective to improve their utilization. As a second contribution we propose profiling methods to automate the selection of heterogeneous resources for executing applications under user objectives. Finally, we demonstrate how these technologies can be implemented and exploited in heterogeneous cloud platforms.