Comparison of Parallel STL with C/C++ GPU Programming Models

• Item type: SIG Technical Reports(1)
• 公開日: 2023-07-27
• タイトル: Comparison of Parallel STL with C/C++ GPU Programming Models
• タイトル / 言語: en
• タイトル: Comparison of Parallel STL with C/C++ GPU Programming Models
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: アクセラレータ
• 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: Department of Mechanical and Aerospace Engineering, School of Engineering, Tohoku University
• 著者所属: Cyberscience Center, Tohoku University／Graduate School of Information Sciences, Tohoku University
• 著者所属: Cyberscience Center, Tohoku University
• 著者所属: Cyberscience Center, Tohoku University／Graduate School of Information Sciences, Tohoku University
• 著者所属(英): en
• 著者所属(英): Department of Mechanical and Aerospace Engineering, School of Engineering, Tohoku University
• 著者所属(英): en
• 著者所属(英): Cyberscience Center, Tohoku University / Graduate School of Information Sciences, Tohoku University
• 著者所属(英): en
• 著者所属(英): Cyberscience Center, Tohoku University
• 著者所属(英): en
• 著者所属(英): Cyberscience Center, Tohoku University / Graduate School of Information Sciences, Tohoku University
• 著者名: Joanna, Imada; Keichi, Takahashi; Yoichi, Shimomura; Hiroyuki, Takizawa
• 著者名(英): Joanna, Imada; Keichi, Takahashi; Yoichi, Shimomura; Hiroyuki, Takizawa
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: The C++ 17 standard introduced a set of parallel algorithms, referred to as Parallel STL, that is designed to be programmer-friendly and portable across CPU and GPU. Several studies compared the performance between GPU programming models including Parallel STL. However, the reasons behind the performance differences are not well discussed yet. This study thus investigates what causes the performance differences among GPU programming models: CUDA, Kokkos, OpenACC, OpenMP, and Parallel STL. Three benchmarks are selected to compare the models: BabelStream, Himeno benchmark, and CloverLeaf. In BabelStream, Parallel STL achieves similar performance to other models. In the Himeno benchmark, it achieves 12% higher performance than CUDA for the large problem size. However, for the largest problem size, it performs 23% worse than CUDA. Profiling reveals that Parallel STL has a low cache hit ratio compared to other models in the larger problem sizes.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: The C++ 17 standard introduced a set of parallel algorithms, referred to as Parallel STL, that is designed to be programmer-friendly and portable across CPU and GPU. Several studies compared the performance between GPU programming models including Parallel STL. However, the reasons behind the performance differences are not well discussed yet. This study thus investigates what causes the performance differences among GPU programming models: CUDA, Kokkos, OpenACC, OpenMP, and Parallel STL. Three benchmarks are selected to compare the models: BabelStream, Himeno benchmark, and CloverLeaf. In BabelStream, Parallel STL achieves similar performance to other models. In the Himeno benchmark, it achieves 12% higher performance than CUDA for the large problem size. However, for the largest problem size, it performs 23% worse than CUDA. Profiling reveals that Parallel STL has a low cache hit ratio compared to other models in the larger problem sizes.
• 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 2023-HPC-190, 号 2, p. 1-7, 発行日 2023-07-27
• 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