Title: HPC Seminar Series: “The Implementation of an All-In-One Numerical Procedure Within the Runtime StarPU”
Speaker: Dr. Vincent Perrier, Visiting Researcher, NIA
Date: Tuesday, February 28, 2017
Location: NASA/LaRC, Bldg 2102, Room 112-113
Sponsor: A. Mazaheri, Nasa/LaRC
Abstract: Over the last years, overcoming the MPI limitations on very large clusters has been mostly done with coarse grain OpenMP. If the exact architecture of the future computing clusters is not known yet, the current trend include more memory access heterogeneities (NUMA effects), and co-processors (GPU, Xeon-Phi). Performing efficient computations will thus be more and more difficult as the heterogeneities of computing platforms will increase. An alternative to static scheduling is given by runtime schedulers of task based implementations. In this talk, we will describe the implementation of an all-in-one numerical procedure within the runtime StarPU. In order to limit the complexity of the method, and for the sake of clarity, we will limit the numerics to first order in space and time. We will present our implementation, our approach for IO, and the results obtained on different architectures includinghybrid CPU/GPU and Intel KNL architectures. We will finish our talk by some results obtained for more complex systems and with higher order schemes.
The results of this talk were jointly obtained with Mohamed Essadki, Adam Larat and Milan Pelletier from the Energetics and Combustion EM2C department, and Jonathan Jung from Pau University and INRIA Cagire team. The talk will also include results obtained by Raphael Blanchard, Emeric Martin and Florent Renac, from ONERA – the french Aerospace lab.
Experiments presented in this talk were carried out using the PLAFRIM experimental testbed, being developed under the Inria PlaFRIM development with support from Bordeaux INP, LABRI and IMB and other entities: Conseil Régional d’Aquitaine, Université de Bordeaux and CNRS and ANR in accordance to the programme d’investissements d’Avenir. (see http://www.plafrim.fr/)
Speaker Bio: Dr. Vincent Perrier is a research scientist in France, at INRIA. He earned Ph.D in Applied Mathematics and Scientific Computing at Bordeaux University, France in 2007. He was assistant professor at Lyon University between 2007 and 2008, and then was hired at INRIA in Pau, France in 2008. His area of expertise is high order methods, multiphase compressible flows, low Mach compressible flows and high performance computing.