Critical Path Analysis for Characterizing Parallel Runtime Systems

• Item type: SIG Technical Reports(1)
• 公開日: 2016-08-01
• タイトル: Critical Path Analysis for Characterizing Parallel Runtime Systems
• タイトル / 言語: en
• タイトル: Critical Path Analysis for Characterizing Parallel Runtime Systems
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: 並列処理
• 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: University of Tokyo
• 著者所属: University of Tokyo
• 著者所属(英): en
• 著者所属(英): University of Tokyo
• 著者所属(英): en
• 著者所属(英): University of Tokyo
• 著者名: An, Huynh; Kenjiro, Taura
• 著者名(英): An, Huynh; Kenjiro, Taura
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Parallel programming models are increasingly relying on runtime systems to map logical parallelism onto available physical cores at runtime. In order to characterize runtime systems in doing this dynamic scheduling job, we analyze how they progress computation along the critical path of serially dependent tasks along which there is always a task ready or running. Hence, the entire execution time can be classified into three parts: work, during which a task was running; busy delay, during which no task on the path was running but all cores were busy; and scheduler delay, during which no task on the path was running and there was at least one available (idle) core. A large scheduler delay is likely to be an artifact of an implementation of the runtime system. Furthermore, we classify scheduler delays by the types of dependencies (e.g., end of a task, creation of a task, continuation of parent tasks) which can highlight differences in the design and implementation of the runtime systems.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Parallel programming models are increasingly relying on runtime systems to map logical parallelism onto available physical cores at runtime. In order to characterize runtime systems in doing this dynamic scheduling job, we analyze how they progress computation along the critical path of serially dependent tasks along which there is always a task ready or running. Hence, the entire execution time can be classified into three parts: work, during which a task was running; busy delay, during which no task on the path was running but all cores were busy; and scheduler delay, during which no task on the path was running and there was at least one available (idle) core. A large scheduler delay is likely to be an artifact of an implementation of the runtime system. Furthermore, we classify scheduler delays by the types of dependencies (e.g., end of a task, creation of a task, continuation of parent tasks) which can highlight differences in the design and implementation of the runtime systems.
• 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 2016-HPC-155, 号 23, p. 1-8, 発行日 2016-08-01
• 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