Tools for the realistic evaluation of parallel computing systems

• Autores: Francisco Javier Ridruejo Pérez
• Directores de la Tesis: Javier Navaridas Palma (dir. tes.), José Miguel Alonso (dir. tes.)
• Lectura: En la Universidad del País Vasco - Euskal Herriko Unibertsitatea ( España ) en 2013
• Idioma: español
• Tribunal Calificador de la Tesis: Clemente Rodríguez Lafuente (presid.), Alexander Mendiburu Alberro (secret.), Ramón Beivide Palacio (voc.), Maria del Carmen Carrion Espinosa (voc.), Fernando Vallejo Alonso (voc.)
• Materias:
  • Ciencias tecnológicas
    • Tecnología de los ordenadores
      • Arquitectura de ordenadores
• Texto completo no disponible (Saber más ...)
• Resumen

  • Top 500 supercomputers are very complex and expensive machines, but they are essential for scientific and technological advancement. For these reasons these distributed memory parallel computers with message passing through interconnection networks are the subject of extensive research work by groups around the world. This dissertation focuses on performance prediction and evaluation tools of those interconnection networks using simulation techniques. All methodologies and tools we have developed have been integrated into INSEE, an Interconnection Network Simulation and Evaluation Environment. It allows evaluating different multicomputer architectures and interconnection network topologies, with different levels of detail, aiming the highest levels of accuracy while keeping low resource consumption. The input workload used in the simulation is the main factor in determining the level of simulation fidelity. INSEE fills a gap in the simulation field as it allows the simulation of large supercomputers, and incorporates a comprehensive workload generation mechanism. For preliminary design phases INSEE provides fast statistical distributions, burst-based traffic generation and application micro-kernels. Trace-based traffic generation and full-system simulation provide the highest fidelity, necessary to fully understand all the mechanisms affecting the performance of a system. Traffic generation has been thoroughly studied and is the backbone of this dissertation, which focuses on explaining how the different traffic generation models were devised, developed, implemented and tested into INSEE. All these features have made INSEE to be used by other research groups around the world. All traffic generation models are designed to reflect better the way scientific applications exchange messages, incorporating causality among them, a feature that classical traffic generation based on statistical distributions lack. These traffic generation models were designed and implemented as they were needed during our research. INSEE and these traffic models have been used in the context of this dissertation to research on IN topologies and their performance, analyze network-level congestion control mechanisms and network-level policies, and to examine the implications of using realistic full-system simulation. Use cases and conclusions of this research using INSEE and its traffic generation models are provided as results of this dissertation.