Description
This tutorial will present different libraries made by the Innovative Computing Laboratory at University of Tennessee to address the problem of solving large dense linear problems on today's heterogeneous architectures. The course will first give reminders on dense linear algebra algorithms and the PLASMA library that targets shared memory multi-core architectures, will be presented as a way to introduce to dataflow programming models. Then, as of today the main concern is on accelerators, the focus will be on the MAGMA library and the development of GPU kernels. Finally, we will extend those to libraries to distributed heterogeneous architectures thanks to the PaRSEC runtime.
Mathieu Faverge is an Associate Professor at IPB - ENSEIRB-Matmeca at Bordeaux, France and is part of the Inria HiePACS team. Before that, he received a PhD degree in Computer Science from the University of Bordeaux 1, France and has been a Post Doctoral Research Associate at the University of Tennessee Knoxville's Innovative Computing Laboratory for three years. His main research interests are numerical linear algebra algorithms for sparse and dense problems on massively parallel architectures, and especially DAG algorithms relying on dynamic schedulers. He has experience with hierarchical shared memory, heterogeneous and distributed systems, and his contributions to the scientific community include efficient linear algebra algorithms for those systems.
Primary author
Dr
Mathieu Faverge
(Institut Polytechnique de Bordeaux, Inria Bordeaux)
