Data driven scheduling approach for the multi-node multi-GPU Cholesky decomposition
Tokyo Institute of Technology
19th Workshop on Job Scheduling Strategies for Parallel Processing (JSSPP), 2015
@article{tsujita2015data,
title={Data driven scheduling approach for the multi-node multi-GPU Cholesky decomposition},
author={Tsujita, Yuki and Endo, Toshio},
{year={2015}}
}
Recently large scale scientific computation on heterogeneous supercomputers equipped with accelerators is receiving attraction. However, traditional static job execution methods and memory management methods are insufficient in order to harness heterogeneous computing resources including memory efficiently, since they introduce larger data movement costs and lower resource usage. This paper takes the Cholesky decomposition computation, which is an important linear algebra kernel, as the target for optimization. And we describe a scalable data-driven scheduling method and a heterogenous memory management method in order to improve resource utilization and reduce amount of data movement. Through the performance evaluation on TSUBAME2.5, which is a heterogenous supercomputer with NVIDIA GPUs, we demonstrate the efficiency of the proposed task scheduling method and data replacement strategies considering data reusability.
March 25, 2015 by hgpu
