Comet: Un modèle de haute-performance pour la composition à grain fin
Julien Bigot, Thierry Gautier, Christian Pérez, Jérôme Richard · HAL (Le Centre pour la Communication Scientifique Directe) · 2017
This paper deals with the efficient combination of software components and task-based models for HPC. Task-based models are known to greatly enhance performance and performance portability while component models ease the separation of concerns and thus improves modularity and adaptability. The paper describe the C OMET programming model, a component model for HPC extended with task concepts. We demonstrate its prototype implementation built on top of the task model of OpenMP and the low level component model L2C. We evaluate the approach on five synthetic use-cases representative of common patterns from HPC applications. Experimental results show that the approach is very efficient on the use-cases. On one hand, independent software codes can be easily composed. On the other hand, fine-grained composition supports very good performance. It sometimes even outperforms classical hand-written OpenMP implementations thank to better task interleaving.