Locality-Aware Work Stealing on Multi-CPU and Multi-GPU Architectures
Federal University of Rio Grande do Sul (UFRGS), Brazil
hal-00780890, 24 January 2013
@inproceedings{gautier:hal-00780890,
hal_id={hal-00780890},
url={http://hal.inria.fr/hal-00780890},
title={Locality-Aware Work Stealing on Multi-CPU and Multi-GPU Architectures},
author={Gautier, Thierry and Ferreira Lima, Joao Vicente and Maillard, Nicolas and Raffin, Bruno},
language={Anglais},
affiliation={MOAIS – INRIA Grenoble Rh{^o}ne-Alpes / LIG Laboratoire d’Informatique de Grenoble , Instituto de Inform{‘a}tica da UFRGS – UFRGS},
booktitle={6th Workshop on Programmability Issues for Heterogeneous Multicores (MULTIPROG)},
address={Berlin, Allemagne},
audience={internationale},
year={2013},
month={Jan},
pdf={http://hal.inria.fr/hal-00780890/PDF/joao-lima-multiprog2013.pdf}
}
Most recent HPC platforms have heterogeneous nodes composed of a combination of multi-core CPUs and accelerators, like GPUs. Scheduling on such architectures relies on a static partitioning and cost model. In this paper, we present a locality-aware work stealing scheduler for multi-CPU and multi-GPU architectures, which relies on the XKaapi runtime system. We show performance results on two dense linear algebra kernels, Cholesky (POTRF) and LU (GETRF) factorization, to evaluate our scheduler on a heterogeneous architecture composed of two hexa-core CPUs and eight NVIDIA Fermi GPUs. Our experiments show that an online locality-aware scheduling achieve performance results as good as static strategies, and in most cases outperform them.
January 25, 2013 by hgpu
