Performance evaluation of interconnection networks using simulationtools and case studies

Résumé

This dissertation focuses on the performance evaluation of interconnection networks, It briefly introduces supercomputing and discusses the most common performance evaluation methodologies: analytical, simulation-based and empirical. Subsequently, it describes a simulation environment developed within the author's research group and all the related tools. Remarkable contributions to this environment are the trace-driven engine and the application-kernels that allow the evaluation of interconnection networks using realistic loads. This environment is used to perform four case studies. The first is the evaluation of the twisted torus topology, in which a pitfall of the derivation of the theoretical throughput from the bisection bandwidth is shown. The second case study evaluates the thin-tree topology, an alternative to the over-dimensioned k-ary n-tree topology which offers better performance/cost figures. The third case study evaluates the interconnection network of SpiNNaker, a large-scale system-on-chip-based architecture with severe restrictions in terms of power consumption and chip area; this evaluation is mainly focused on stability and fault-tolerance. The last case study measures the influence that job and task allocation policies have on the execution time of parallel applications. These evaluations have been carried out using mainly simulation, although some results have also been mathematically derived.