Date: 18th September 2012
Time: 9.00am - 5.00pm
Venue: Charles Babbage Room (Level 17, Connexis South, Fusionopolis)
Instructor: Prof. Serge G. Petiton, Université de Lille, Laboratoire d’Informatique Fondamentale de Lille (LIFL/CNRS)
About the Workshop:
The goal of this day-long series of talks is to introduce advanced high performance scientific computing including parallel and distributed algorithms and method designs, regular or irregular data structure adaptations, programming paradigms and methodologies. The course will cover all the main existing high performance execution and programming paradigms: flux parallelism (pipelined vector computing), data parallelism, control-flow parallelism, SIMD, SPMD, MSPMD,…along with linear algebra examples. Emerging multicore and many-core programming will be considered, and algorithmic optimizations to existing GPU and processor will be considered. We will also survey iterative Krylov-based methods for linear algebra problems and discuss how they are well-adapted for future post-petascale hypercomputers.
The course begins with an overview, following the history of supercomputing, of parallel and distributed architectures and programming paradigms, including associated concepts and terminology. Then, classical programming paradigms such as pipelined-vector and data parallel computing will be explored and illustrated with linear algebra examples from basic matrix-vector operations to hybrid Krylov methods. Large granularity parallel and distributed programming will be considered using the same examples. The last part of the course will focus on future challenge scientists would have to solve during this decade to be able to achieve adequate computational efficiency on hypercomputers. We will discuss unsolved problems on the road to exascale programming and computing.
A key aim of the course is for the participant to acquire state-of-art knowledge on high performance scientific computing by understanding how some classical linear algebra methods are to be programmed on such environments. Participants will be acquainted with the challenges at the forefront of post-petascale computing.
Part I (18 September, am):
A Brief history and survey of supercomputing; from vector to GRID and Cloud Computing; toward exascale computing.
o Main HPC architectures and programming paradigms, from pipelined vector computing to GRID and Cloud computing
o Vector and data parallel computing; dense and sparse linear algebra data structures and algorithms
o Vector and data parallel programming, with examples as y=A(Ax+x)+x, Gauss, Gauss-Jordan, the Conjugate Gradient with Polynomial preconditioning
Part II (18 September, pm):
Krylov methods; from basic parallel methods to smart hybrid auto-tuned methods
o The parallel Arnoldi’s (ERAM, IRAM) and GMRES restarted subspace iterative methods
o Vector orthogonalisation methods
o The hybrid MERAM and GMRES-ERAM/LS methods
o Auto-tuned (subspace size, orthogonalisation, sparse patterns) hybrid Krylov methods for large sparse no-symmetric linear algebra problems on cluster of GPU.
Part III (18 September, pm):
Toward exascale programming and computing
o Large Granularity parallel task computing; from SPMD, MSPMD to YML/XMP programming. Same example that previously, adapted for SPMD and MSPMD programming. Block-version of classical Linear Algebra methods. Large granularity hybrid asynchronous methods.
o Toward post-petascale programming, on the road to the exascale frontier. How the studied programming paradigms would have to be hybridized to obtained efficient post-petascale applications. Different problems to address (sparse matrices, GPU programming, energy consumption, asynchronously processes and other issues). Dynamic multi-parameter auto-tuning.
About the speaker:
Prof. Serge G. Petiton received the B.S. degree in mathematics, in 1982, the M.S. degree in Applied Mathematics, in 1984, the M.E. degree in Operating System, in 1985, the Ph.D. degree in computer science, 1988, and the “Habilitation à diriger des recherches”, 1993, from Pierre and Marie Curie University, PARIS 6. He was post-doctoral student at Yale University, 1989-1990. He has been researcher at the “Site Experimental en Hyperparallelisme” (supported by CNRS, CEA, and the French DoD) from 1991 to 1994. He also was affiliate research scientist at Yale and visiting research fellow in several US laboratories, especially in NASA-ICASE and the AHPCRC during the period 1991-1994.
Since 1994, Serge G. Petiton is tenured Professor at the Scientific and Technical University of Lille and leads the “Methodology and Algorithmic Parallel Programming” group of the CNRS “Laboratoire d’Informatique Fondamentale de Lille”. He participates in several projects of the INRIA Laboratory in Saclay and of the CNRS Japanese-French Laboratory on Informatics (JFLI) in Tokyo. He is senior lecturer at University of Paris 6, University of Versailles, and the University of Tsukuba in Japan. He is director of the board of the ORAP association (launched in 1994 by CNRS, INRIA and CEA) to promote HPC and he participates in several French and International HPC committees.
He has been scientific supervisor of more than 20 Ph.D. candidates and has authored more than 100 articles in international journals and conferences. His main current research interests are “Parallel and Distributed Computing”, “Post-Petascale Smart-tuned Dense and Sparse Linear Algebra”, and “Language and Programming Paradigm for Extreme Modern Scientific Computing”.
Presentation File Part 1a
Presentation File Part 1b
Presentation File Part 2
Presentation File Part 3
No of Participants: 33