Contributions algorithmiques au calcul scientifique sur architectures haute performance
Pierre Fortin · HAL (Le Centre pour la Communication Scientifique Directe) · 2018
High performance architectures are constantly evolving in order to deliver ever greater compute powers, as well as ever greater energy efficiencies. This applies to multi-core CPUs (with higher core count and wider vector units) as well as to various many-core, possibly heterogeneous, architectures (GPUs, Xeon Phi processors . . . ). Considering performance, power efficiency or performance portability, and relying on new and relevant programming paradigms, we have focused on algorithmic changes allowing to adapt at best specific or key applications in scientific computing to such high performance architectures. Our research work has been structured according to three research directions: (i) designing algorithms for many-core architectures via massive parallelism, for multi-core architectures via task parallelism, or for both via hybrid algorithms; (ii) handling the vector divergence on high performance architectures; and (iii) taking advantage of new heterogeneousarchitectures for scientific applications. We present here our algorithmic contributions, their interdisciplinary context, and the close combination they require between application specificities, algorithmics, programming and architectural features.