OpenMP performance analysis for many-core platforms with non-uniform memory access
TEISA, University of Cantabria Santander, Cantabria 39005, Spain
IJCSI International Journal of Computer Science Issues, Vol. 10, Issue 2, No 2, 2013
@article{de2013openmp,
title={OpenMP performance analysis for many OpenMP performance analysis for many-core platforms with core platforms with non-uniform memory access uniform memory access uniform memory access},
author={de Aledo Marug{‘a}n, Pablo Gonz{‘a}lez and Bay{‘o}n, Javier Gonz{‘a}lez and Espeso, Pablo S{‘a}nchez and Mart{‘i}n, Juan Casal},
year={2013}
}
One of the first steps in embedded-system design flow is to choose the most efficient implementation of the embedded software application. However, this is difficult to do at the earliest design stages because particular details of the final manycore HW platform are usually unknown and many possible mappings of the software tasks/threads have to be evaluated. This paper presents a complete framework for early performance estimation of parallel programs in many-core platforms. The proposed framework is based on a specific native-simulation approach oriented to many-core platforms, which enables fast simulation and profiling. The software parallelism is specified in OpenMP, a commonly used application software interface (API) for shared-memory parallel programming. In order to support Non-Uniform Memory Access (NUMA) architectures (which are dominant in high-performance many-core platforms), the paper proposes some OpenMP extensions. These extensions improve performance analysis and facilitate the automatic translation from OpenMP to OpenCL (a low-level API for heterogeneous computing), which are commonly used for NUMA programming). Results show that the proposed OpenMP extension and specific parallel modeling techniques provide reliable results even for NUMA architectures.
May 2, 2013 by hgpu
