Toward petaflop numerical simulation on parallel hybrid architectures

june, 6-10, 2011
INRIA
Sophia Antipolis, France

Objectives

Benefiting from the  ever  increasing  performances  of  processing facilities,  scientific computing  finds nowadays  numerous industrial applications  as well as  applications of interest to well being in our daily society (environmental and health related applications). well.   In order to tackle the complexity of the problems  that scientific  research and  engineering design communities are facing nowadays, parallel computing has became a mandatory path for taking full benefit of these computing resources.  Modern petascale (and future exascale) computing systems are massively hybrid (heterogeneous) parallel architectures. These systems are based on Non-Uniform  Memory Architecture  (NUMA) nodes interconnected by high speed networks, each node being organized around  an  heterogeneous set  of  processing units:  multiple-core chips, Graphical Processing Units (GPU)  or other types of accelerator cards. These processing  units are fed  by complex multi-stage  memory units, and for the full systems  to be productive (i.e.  to deliver sustained application performance  in an acceptable  relation to the  peak) they require computer programs that can efficiently manipulate this kind of memory  hierarchy.  In this context, this school aims at presenting some   recent   developments   concerning   programming   models   and environments, as  well as  on generic numerical  algorithms, targeting modern high  performance computing (HPC) systems.  The program of the school is organized around the following topics:

• Introduction to modern HPC systems based on multiple-core, GPU, FPGA, etc.
• Low level programming models and environments (in particular, CUDA, OpenCL and HMPP) and their hybridization with MPI or OpenMP – Virtualization – Communication operations in heterogeneous systems
• High level programming models and environments (such as PyCUDA, CLyther, etc.)
• Numerical algorithms adapted to modern HPC systems: numerical linear algebra for dense and sparse linear systems, algebraic resolution algorithms (domain decomposition and hierarchical methods), FFT
• Methods for performance evaluation and optimization in relation to the hardware characteristics of modern HPC systems
• OpenSource environments for scientific computing on modern HPC systems

Moreover, the schedule of the school will also include presentations of   practical   experiences with the   development of numerical methodologies for solving computational physics problems.

Organizers

• Luc Giraud, Equipe‐projet HiePACS, INRIA Bordeaux ‐ Sud‐Ouest
• Loïc Gouarin, LAGA, Laboratoire de Mathématiques - Université de Paris - Sud 11
• Stéphane Lanteri, Equipe‐projet Nachos, INRIA Sophia Antipolis ‐ Méditerranée
• Violaine Louvet, Institut Camille Jordan, Lyon
• Jean Roman, LABRi et Equipe‐projet HiePACS, INRIA Bordeaux ‐ Sud‐Ouest

Registration fees

The registration fees include school attendance, lunches and coffe breaks,  from June 6-10, 2011  and a welcome cocktail.

• Students: 100€
• Academics: 300€
• Full rate: 600€