Research Outline

I have mainly interest in issues linked to high-performance computing (HPC) and distributed applications (Big Data or Cloud Computing for example). I focus in particular on scheduling aspects, more precisely on static scheduling (computed before execution) based on theoretical models. During my (short) career I deal with several specific problems.


Work in progress (If you can read it, French version is more complete).


To study parallel matrix multiplication, we use a square partitioning problem where load-balancing is given as an input. With this load-balancing, the goal is to minimise the sum of the half-perimeter of the covering rectangle of created zones. This is not a new problem but we provided some additional results. First we proposed two approximations algorithms (with proven ratio), one based on recursive strategy, the other on fractal curve. We also propose an alternative model based on cube partitioning and prove the NP-completeness of this model while providing two approximation algorithms. Finally we compare the efficiency of our approach against dynamic scheduling using simulations and practical implementation with the StarPU library. Our experiment trend to show that static approach with the addition of work-stealing provide slightly better makespan and significant decreasing of communication cost.


Adapted from a template by FreeHTML5.co