ALF Amdhal’s Law is Forever In PetaQCD
New technological context Where we come from ? CAPS (1994-2008) Architecture and compilation for high-end processors (GP and embedded/special purpose) Four successive positive evaluations Caps Entreprise start-up created in 2003 New technological context the multicore era
The ALF group A. Seznec, DR INRIA, microarchitecture P. Michaud, CR INRIA, microarchitecture I. Puaut, Pr Univ. Rennes I, worst case execution time F. Bodin, after sabbatical (exp. 2010), compilation E. Rohou, INRIA specialist engineer, compilation + 4 Ph.D students + 4 engineers
The advent of the multicore era Up to 2002, throwing transistors in uniprocessor was the most cost-effective path for « performance for the masses » Since 2002, progressively entering the multicore era is it « parallelism for the masses » ? Hardware: yes Applications: no
Context Hardware: 1000’s processors on a chip in 2015-2020 will be feasible. E.g. the 80-cores Intel Terascale prototype in 2006 Effective high performance on the chip is the issue Applications: every application will have to run on // hardware Architecture: defining the 2020’s manycore architecture Compilation: code generation/adaptation for manycores Performance predictability: predicting/guaranteeing performance on manycores A L F
Objectives within PetaQCD Understand issues on a Petaflops range applications Modelize future multi manycores particularly memory hierarchy to: explore dimensionning: Cores vs local memory/cache sizes Computing power vs bandwidth at all levels People: A. Seznec + PhD to be hired