Study about parallel programming models design (Farouk Mansouri)

The Avalon working group will host a presentation by Farouk Mansouri on thursday 2nd of march, at 10h30 in amphi J. You are all welcome.

For ten years, the evolution of computing machines trend to parallel and heterogeneous architectures such as clusters. Composed of several nodes connected via the network and  including heterogeneous processing units, they allow a high performance computation. To program these architectures, the user must rely on programming models such as MPI, OpenMP or CUDA. These models have two objectives: to increase programmer productivity by providing an abstraction of the architectural specificities, and preserve the efficiency by exploiting its performances. However, these two goals are paradoxical, achieving at the same time is a real challenge for designers. Indeed, increasing the model’s level of abstraction leads to lose the control of hardware optimizations and reduce efficiency of implementations. In this presentation we focus on which design of parallel programming model could help to achieve both goals. We think that important ideas are: 1) Domain decomposition and application characterization. 2) Topology-aware hardware optimisations and 3) Smart execution engine as a middle level.