Performance analysis of a deep learning framework on a high-performance distributed file system

• Item type: SIG Technical Reports(1)
• 公開日: 2017-07-19
• タイトル: Performance analysis of a deep learning framework on a high-performance distributed file system
• タイトル / 言語: en
• タイトル: Performance analysis of a deep learning framework on a high-performance distributed file system
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: 機械学習
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Fujitsu Laboratories Limited
• 著者所属: Fujitsu Laboratories Limited
• 著者所属: Fujitsu Laboratories Limited
• 著者所属: Fujitsu Laboratories Limited
• 著者所属(英): en
• 著者所属(英): Fujitsu Laboratories Limited
• 著者所属(英): en
• 著者所属(英): Fujitsu Laboratories Limited
• 著者所属(英): en
• 著者所属(英): Fujitsu Laboratories Limited
• 著者所属(英): en
• 著者所属(英): Fujitsu Laboratories Limited
• 著者名: Saso, Stanovnik; Amir, Haderbache; Masahiro, Miwa; Kohta, Nakashima
• 著者名(英): Saso, Stanovnik; Amir, Haderbache; Masahiro, Miwa; Kohta, Nakashima
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Current deep learning framework data access patterns are not adapted to traditional HPC distributed file systems such as FEFS, as they do not take into account the data access latency and overhead. This causes performance degradation in high-scale environments, especially with large dataset sizes. We analyse data access patterns in an existing framework with both external and introspective analysis, then propose and implement a new pattern in order to achieve 5 to 6 times better performance when training on and by adapting to the capabilities of a distributed file system.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Current deep learning framework data access patterns are not adapted to traditional HPC distributed file systems such as FEFS, as they do not take into account the data access latency and overhead. This causes performance degradation in high-scale environments, especially with large dataset sizes. We analyse data access patterns in an existing framework with both external and introspective analysis, then propose and implement a new pattern in order to achieve 5 to 6 times better performance when training on and by adapting to the capabilities of a distributed file system.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN10463942
• 書誌情報: 研究報告ハイパフォーマンスコンピューティング（HPC）; 巻 2017-HPC-160, 号 39, p. 1-6, 発行日 2017-07-19
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8841
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会