Improving the efficiency of multicore systems through software and hardware cooperation

• Autores: Víctor Javier Jiménez Pérez
• Directores de la Tesis: Francisco J. Cazorla (dir. tes.), Mateo Valero Cortés (codir. tes.)
• Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2016
• Idioma: español
• Tribunal Calificador de la Tesis: Ramón Beivide Palacio (presid.), Xavier Martorell Bofill (secret.), Gábor János Dózsa (voc.)
• Programa de doctorado: Programa de Doctorado en Arquitectura de Computadores por la Universidad Politécnica de Catalunya
• Materias:
  • Ciencias tecnológicas
    • Tecnología de los ordenadores
      • Arquitectura de ordenadores
• Enlaces
  • Tesis en acceso abierto en:  TDX  UPCommons 
• Resumen

  • Increasing processors' clock frequency has traditionally been one of the largest drivers of performance improvements for computing systems. In the first half of the 2000s, however, it became clear that continuing to increase frequency was not a viable solution anymore. Power consumption and power density became prohibitively costly, and processor manufacturers moved to multicore designs. This new paradigm introduced multiple challenges not present in single-threaded processors. Applications running on multicore systems share different resources such as the cache hierarchy and the memory bus. Resource sharing occurs at much finer degree when cores support multithreading as well. In this case, applications share the processor¿s pipeline too. Running multiple applications on the same processor allows for better utilization of its resources¿which otherwise may just lie idle if an application does not use them. But sharing resources may create interferences between applications running on the system. While the degree of these interferences depends on the nature of the applications, it is typically desirable to reduce them in order to improve efficiency.

    Most currently available processors expose a set of sensors and actuators that software can use to monitor and control resource sharing among the applications running on a system. But it is typically up to end users to analyze their workloads of interest and to manually use the actuators provided by the processor. Because of this, in many cases the different mechanisms for controlling resource sharing are simply left unused.

    In this thesis we present different techniques that rely on software/hardware interaction to monitor and improve application interference¿and thus improve system efficiency. First we conduct a quantitative study showing the benefits of hardware/software cooperation on system efficiency. Then we narrow our focus on a given hardware knob: data prefetching. Specifically we develop and evaluate several adaptive solutions for improving the efficiency of hardware data prefetching on multicore systems. The impact of the solutions presented in this thesis, however, goes beyond the particular case of data prefetching. They serve as illustrative examples for developing software/hardware cooperation schemes that enable the efficient sharing of resources in multicore systems.