Toward Computational and Data High Performance Intelligent Methods for Extreme Scale Distributed and Parallel Architectures

Future extreme-scale distributed and parallel hypercomputers are expected to have highly hierarchical architectures with nodes composed by lot-of-core processors and accelerators. The different programming levels will generate new difficult algorithm issues. New methods should be defined and evaluated with respect to modern state-of-the-art of applied mathematics and scientific methods. Adapted new programming paradigms would emerge and protocols allowing end-users to optimize codes have to be proposed.

In this talk we propose a survey of ours researches in high performance linear algebra and in programming paradigms. First we survey some auto/smart-tunning strategies we proposed and evaluated for some of the Krylov method parameters and we explain why “unite and conquer” methods we introduced on clusters of accelerators are well adapted for extreme scale architecture. We discuss some results obtained on several supercomputers.  Second, we present a multilevel programming paradigm for extreme scale computing based on graph of task-component programming, using YML. We present some results on several supercomputers, using some tasks developed with the PGAS language XMP.

As a conclusion, we discuss on the researches we are launching to mix computational and data oriented programming, using graphs of components and containers.