Home // ICAS 2020, The Sixteenth International Conference on Autonomic and Autonomous Systems // View article

Deep Learning Workload Analysis for Efficient Resource Allocation

Authors:
Sayaka Takayama
Takashi Shiraishi
Shigeto Suziki
Masao Yamamoto
Yukihiro Watanabe
Masato Oguchi

Keywords: Workload analysis; MLPerf; Zabbix; Deep learning.

Abstract:
In recent years, with the prosperity of deep learning, Graphics Processing Units (GPUs) have become popular as hardware accelerators specialized for this purpose. However, compared to CPUs, which are general-purpose computing resources, GPUs are very scarce and valuable resources. Therefore, in this paper, we would like to consider some control that reduces GPU resource waste by determining GPU allocation based on the difference in application performance when using different GPUs. As a basic study, we evaluate the performance of 9 types of benchmarks executed on the framework using GPU and compare the performance when changing machine conditions. From this examination, it is judged whether the above control is possible. In addition, we estimate how much performance improvement can be expected by preferentially allocating GPUs with high performance to workloads that have a large impact on GPU performance using the data we collected. From this estimate, it is found that GPU priority control can reduce the total execution time by 8.24%.

Pages: 8 to 15

Copyright: Copyright (c) IARIA, 2020

Publication date: September 27, 2020

Published in: conference

ISSN: 2308-3913

ISBN: 978-1-61208-787-0

Location: Lisbon, Portugal

Dates: from September 27, 2020 to October 1, 2020