Accurate Analytic Models to Estimate Execution Time on GPU Applications
Institute of Informatics, Federal University of Rio Grande do Sul (UFRGS) – Porto Alegre, RS – Brazil
11th Workshop on Parallel and Distributed Processing (WSPPD), 2013
@article{velho2013accurate,
title={Accurate Analytic Models to Estimate Execution Time on GPU Applications},
author={Velho, Pedro and de Oliveira, Daniel AG and Padoin, Edson L and Navaux, Philippe OA},
year={2013}
}
Today top ranked HPC systems feature several GPUs which present high processing speed at low power budget with various parallel applications. Many scientific applications still claim for even more computing speed than the available today. A general approach to provide more processing speed is to scale the system. However, aspects such as interference, the amount of resources, heterogeneity of resources and failure probabilities hinder the research towards the future HPC systems. For that reason, many research use models to understand and estimate the behavior of the systems. GPUs are present and future of HPC systems, however the community lack of models to fast estimate the behavior of several GPUs with the desired accuracy and scale. The goal of this paper is to characterize the GPU resources to infer models that can be used in large scale simulators. To create this models with the two main concerns in mind, speed and accuracy, we conduct linear regression to estimate the time. Our test show that when the workload is within the range of realistic parameters the error is less than 11 %.
August 21, 2013 by hgpu
