Section: Dissemination

Promoting Scientific Activities

Tomofumi Yuki was co-organizer of IMPACT'16 (International Workshop on Polyhedral Compilation Techniques, http://impact.gforge.inria.fr/impact2016/) with Michelle Strout (University of Arizona).

Alain Darte (with the help of Tomofumi Yuki for the program) co-organized with Violaine Louvet (Institute Camille Jordan in Lyon, now lead of UMS Gricad in Grenoble) a second polyhedral spring school, May 9-13 2016, targeting both the polyhedral community and HPC users from numerical analysis. This spring school has been labelled (and funded) as a CNRS interdisciplinary spring school (https://mathsinfohpc.sciencesconf.org/), with a total budget of roughly 39 Keuros, including funding from Labex MILYON, CNRS, GDR Calcul, ENS, LIP, and registrations fees, which were kept low to keep the spirit of the first spring school on polyhedral code analysis and optimizations.

This second spring school was motivated by the need for a more global approach for HPC applications, that combines the design of numerical methods with extensive hardware considerations, in interaction with languages and compilers, so as to take into account both the complexity of architectures and the needs of their non-expert users. Research communities in computer science (architecture, compilation) and applied mathematcs (numerical simulation) are not always aware of this need; at least their work do not always spread enough across the other discipline to lead to mutual influence. Automatic code optimizations and tools also require a better evaluation of their applicability. The goal of this research school – or meeting place of two communities – was to make the link between some of the most recent advances on automatic program optimizations (in particular polyhedral techniques and tools) and applied mathematics (schemes for numerical simulation), in relation with application needs. This school was therefore interdisciplinary, with a strong will to bring communities together on the common theme of supercomputing.

We finally opted for a single track instead of parallel sessions, which helped federate the two communities. The school included courses on architectures (M. Haefele, Maison de la simulation), on numerical schemes in connection with stencils (T. Dumont, ICJ), on simulation methods (discontinuous Galerkin) in particular for GPU (P. Helluy, Strasbourg), on polyhedral techniques and tiling (A. Darte, Compsys), on some polyhedral compilers such as Pluto (U. Bondhugula, Bangalore) and PPCG (S. Verdoolaege, ENS), on the roofline model for performance analysis (M. Püschel, ETH Zürich), on stencils and tensors optimizations (Ramanujam, Baton Rouge), on numerical precision (C. Rubio-Gonzalez, UC Davis), plus some additional talks on reproducibility, applications, the ECM model, etc. The school was a success, with 71 participants, roughly half from each community, with 29 coming from abroad (Italy, Algeria, USA, India, Canada, Germany, Croatia, Switzerland, Austria, Belgium), and a majority (37) being PhD students.

The future will tell if our objectives have been reached, i.e., if the two communities will interact more on the long term and rethink their work with an interdisciplinary look, to invent new computing schemes and compilers more suitable for the constraints of today's architectures, in particular their memory hierarchy and locality needs. In Compsys at least, one can already see some moves in this direction, with the interdisciplinary internship of Julien Versaci co-advised by Tomofumi Yuki, the participation of Alain Darte as a referee to the PhD jury of T. Gasc (CEA, Maison de la Simulation, ENS Cachan), a planned seminar by Alain Darte at Maison de la Simulation in early 2017, starting exchanges with the LMGC lab (Montpellier) on their applications, and a planned mini-symposium, following the line of this spring school, at SMAI 2017.