RT Conference Proceedings
T1 Adaptive Partition Strategies for Loop Parallelism in Heterogeneous Architectures
A1 Vilches Reina, Antonio
A1 Asenjo-Plaza, Rafael
A1 Corbera-Peña, Francisco Javier
A1 González-Navarro, María Ángeles
K1 Computación heterogénea
K1 Procesos en paralelo (Informática)
AB This paper explores the possibility of efficiently using multicoresin conjunction with multiple GPU accelerators under a parallel taskprogramming paradigm. In particular, we address the challenge ofextending a parallel_for template to allow itsexploitation on heterogeneous systems. The extension is based on atwo-stages pipeline engine which is responsible for partitioning andscheduling the chunks into the computational resources. Under thisengine, we propose a dynamic scheduling strategy coupled with anadaptive partitioning heuristic that resizes chunks to preventunderutilization and load unbalance of CPUs and GPUs. In this paperwe introduce the adaptivepartitioning heuristic which is derived from an analytical model thatminimizes the load unbalance while maximizes the throughput in thesystem. Using two benchmarks we evaluate theoverhead introduced by our template extensions finding that it isnegligible. We also evaluate the efficiency of our adaptivepartitioning strategies and compared them with related work.
YR 2014
FD 2014-07-30
LK http://hdl.handle.net/10630/7956
UL http://hdl.handle.net/10630/7956
LA eng
NO Este trabajo describe nuestra contribución para la ejecución de bucles paralelos en arquitecturas multi-core/multi-GPU de forma que la carga computacional se distribuya de forma balanceada entre todas las unidades de computación.
NO Universidad de Málaga. Campus de Excelencia Internacional Andalucía Tech. TIN2010-16144, P08-TIC-3500, P11-TIC-08144
DS RIUMA. Repositorio Institucional de la Universidad de Málaga
RD 20 ene 2026