Offloading Techniques to Improve Performance on MPI Applications in NoC-Based MPSoCs

• Autores: Eduardo Fernández Alonso
• Directores de la Tesis: Jaime Joven Murillo (dir. tes.), J. Carrabina (dir. tes.)
• Lectura: En la Universitat Autònoma de Barcelona ( España ) en 2014
• Idioma: inglés
• Tribunal Calificador de la Tesis: Emilio Luque Fadón (presid.), Eduard Ayguadé Parra (secret.), Lionel Torres (voc.)
• Materias:
  • Matemáticas
    • Ciencia de los ordenadores
      • Logicales de ordenadores
    • Investigación operativa
      • Programación no lineal
• Enlaces
  • Tesis en acceso abierto en: DDD
• Resumen

  • Future embedded System-on-Chip (SoC) will probably be made up of tens or hundreds of heterogeneous Intellectual Properties (IP) cores, which will execute one parallel application or even several applications running in parallel. These systems could be possible due to the constant evolution in technology that follows the Moore¿s law, which will lead us to integrate more transistors on a single dice, or the same number of transistors in a smaller dice. In embedded MPSoC systems, NoCs can provide a flexible communication infrastructure, in which several components such as microprocessor cores, MCU, DSP, GPU, memories and other IP components can be interconnected. In this thesis, firstly, we present a complete development process created for developing MPSoCs on reconfigurable clusters by complementing the current SoC development process with additional steps to support parallel programming and software optimization. This work explains systematically problems and solutions to achieve a FPGA-based MPSoC following our systematic ¿ow and offering tools and techniques to develop parallel applications for such systems. Additionally, we show several programming models for embedded MPSoCs and propose the adoption of MPI for such systems and show some implementations created in this thesis over shared and distributed memory architectures. Finally, the focus will be set on the overhead produced by MPI library and on trying to find solutions to minimize this overhead and then be able to accelerate the execution of the application, offloading some parts of the software stack to the Network Interface Controller.