Section: New Results

BOAST: A Meta-Programming Framework to Produce Portable and Efficient Computing Kernels for HPC Applications

Participants : Brice Videau, Kevin Pouget, Luigi Genovese [CEA Inac, Grenoble] , Thierry Deutsch [CEA Inac, Grenoble] , Dimitri Komatitsch [LMA, CNRS, Marseille] , Frédéric Desprez, Jean-Francois Mehaut.

Application portability is an important issue that should be solved efficiently, especially given the large number of different processors now available for today’s supercomputers. The work needed to get performance portability is a tedious task, even for experienced programmers. The availability of semi-automatic tools is therefore mandatory for the development of large simulation applications. Computing kernels’ identification and optimization has to be carefully performed as they usually consume most of the computing resources.

BOAST is a framework (DSL and run-time) that aims at describing kernels in a high-level language and it allows the comparison of the performance of different versions of the code in a simple and seamless way. We described its application to three use cases from the Mont-Blanc project. Results are encouraging as BOAST proved to be a powerful and flexible tool that allowed gains in performance compared to hand-tuned codes. Performance portability of those codes is also improved.

This work was mainly developed by Brice Videau in the context of the Mont-Blanc FP7 European projects 8.3.1.1. It lead to a publication in the International Journal of High Performance Computing Applications (IJHPCA) [11] . A paper will also be published in 2018 describing the BOAST usage for the Gysela Application (see http://gyseladoc.gforge.inria.fr/).