The Department of Biomedical Informatics
presents
Dr. Serge G. Petiton
Laboratoire d’Informatique Fondamentale de Lille,
University of Lille, Sciences and Technologies and
Maison de la Simulation, CNRS, Saclay
“YML-XMP as an example of Multi-Level Programming Paradigms for Graphs of PGAS-written tasks on hypercomputers”
Thursday, April 30, 2015 10:00AM
Computer Science Building Rm 2311
Hosted by Dr. Joel Saltz, Cherith Professor and Founding Chair, BMI Department
Abstract
Exascale hypercomputers are expected to have highly hierarchical architectures with nodes composed by lot-of-core processors and accelerators. Methods have to be redesigned and new ones introduced or rehabilitated in terms of communication optimizations and data distribution.The different programming levels (from clusters of processors loosely connected to tightly connected lot-of-core processors and/or accelerators) will generate new difficult algorithm issues. New language and framework should be defined and evaluated with respect to modern state-of-the-art of scientific methods. We propose a framework, called YML (yml.prism.uvsq.fr), associated with a multilevel programming paradigm, to explore extreme computing and avoid costly global communications and reductions.YML with its high level language permits to automate and delegate the managements of dependencies between loosely coupled clusters of processors to a specialized tool which controls the execution of the application. Besides, the tightly coupled processors inside each cluster could be programmed through a PGAS language such as XMP. Thanks to the component-oriented software architecture of YML, it is relatively easy to integrate new components such as numerical libraries, encapsulated XMP programs for lower level of the computer architecture, etc. Each of the components may also use runtime system or tools to use accelerators.In this talk, we present this multilevel programming paradigm for exascale computing and propose our approach based on YML. We discuss orchestration and scheduling strategies to develop in order to minimize communications and I/O. We present the Block Gauss-Jordan method to invert dense matrices, and the Multiple Explicitly Restarted Arnoldi Method (MERAM) to compute eigenvalues of sparse matrices as study cases. We also propose experiments using components implemented in XMP and discuss projects using StarPU to address accelerators.Experimental results are obtained on Japanese “K” and “T2K” supercomputers, on the French Grid5000 platform, and on the “Hooper” supercomputer in LBNL. We conclude, first, on the correctness of this approach and we point out, next, the performances of these methods on the targeted multi-level parallel architectures in the context of the YML/XMP multi languages integrated framework. On the K computer, we obtained much better results using YML and XMP than only XMP, illustrating the interest of our approach, even if supercomputers scheduler are not yet smarter enough to exploit YML graph analysis.
Contact the Department of Biomedical Informatics at (631) 444-8459 with any questions regarding this event.