Marc Tchiboukdjian - Cache-efficient parallel algorithms based on divide and conquer

It is foreseen that data movements will be the main cause of power consumption in future architecture. Enhancing data locality is thus essential to achieve energy efficiency. In this talk, I will present several examples of how one can adapt simultaneously computation order and data layout to produce algorithms that exposes lots of parallelism use caches efficiently. These algorithms are based on a recursive formulation of the algorithm and the underlying data structure to expose locality and parallelism at all levels of granularity. I will also give an overview of related projects I worked on during my post-doc.