FRUGAL: Reducing GPU Memory Requirement of HPC Applications

• Item type: SIG Technical Reports(1)
• 公開日: 2024-08-01
• タイトル: FRUGAL: Reducing GPU Memory Requirement of HPC Applications
• タイトル / 言語: en
• タイトル: FRUGAL: Reducing GPU Memory Requirement of HPC Applications
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: 並列計算
• 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: Tokyo Institute of Technology
• 著者所属: RIKEN Center for Computational Science
• 著者所属: Tokyo Institute of Technology／RIKEN Center for Computational Science
• 著者所属: National Institute of Advanced Industrial Science and Technology (AIST)／RIKEN Center for Computational Science
• 著者所属: Tokyo Institute of Technology
• 著者所属: RIKEN Center for Computational Science
• 著者所属(英): en
• 著者所属(英): Tokyo Institute of Technology
• 著者所属(英): en
• 著者所属(英): RIKEN Center for Computational Science
• 著者所属(英): en
• 著者所属(英): Tokyo Institute of Technology / RIKEN Center for Computational Science
• 著者所属(英): en
• 著者所属(英): National Institute of Advanced Industrial Science and Technology (AIST) / RIKEN Center for Computational Science
• 著者所属(英): en
• 著者所属(英): Tokyo Institute of Technology
• 著者所属(英): en
• 著者所属(英): RIKEN Center for Computational Science
• 著者名: Tengfei, Wang; Lingqi, Zhang; Ivan, R. Ivanov; Peng, Chen; Toshio, Endo; Mohamed, Wahib
• 著者名(英): Tengfei, Wang; Lingqi, Zhang; Ivan, R. Ivanov; Peng, Chen; Toshio, Endo; Mohamed, Wahib
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Modern GPUs use High Bandwidth Memory (HBM) to mitigate the memory wall effect. Both vendors and consumers often pursue maximal HBM configurations, despite its notoriously high price. Reducing the memory requirement of High-Performance Computing (HPC) and AI applications could allow HPC centers and clouds to use cheaper GPUs with lower memory capacity while maintaining performance, or use the saved memory for solving larger problems or running other applications in parallel. In this paper, we present FRUGAL, a framework designed for reducing the memory requirement of a given workload. FUGAL finds the optimal data migration plan between high bandwidth near memory and slower far memory, subject to the memory and time constraints set by the user. It is based on empirical profiling and analysis of the workload, and does not require modifications to underlying algorithm. We demonstrate the effectiveness and generality of FRUGAL by optimizing several HPC and AI applications. Our evaluation shows that FRUGAL can reduce the memory requirement of Tiled Cholesky Decomposition by 49.3% with 3.9% overhead in time, and Tiny-CUDA-NN by 64.8% with 16.0% overhead in time.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Modern GPUs use High Bandwidth Memory (HBM) to mitigate the memory wall effect. Both vendors and consumers often pursue maximal HBM configurations, despite its notoriously high price. Reducing the memory requirement of High-Performance Computing (HPC) and AI applications could allow HPC centers and clouds to use cheaper GPUs with lower memory capacity while maintaining performance, or use the saved memory for solving larger problems or running other applications in parallel. In this paper, we present FRUGAL, a framework designed for reducing the memory requirement of a given workload. FUGAL finds the optimal data migration plan between high bandwidth near memory and slower far memory, subject to the memory and time constraints set by the user. It is based on empirical profiling and analysis of the workload, and does not require modifications to underlying algorithm. We demonstrate the effectiveness and generality of FRUGAL by optimizing several HPC and AI applications. Our evaluation shows that FRUGAL can reduce the memory requirement of Tiled Cholesky Decomposition by 49.3% with 3.9% overhead in time, and Tiny-CUDA-NN by 64.8% with 16.0% overhead in time.
• 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 2024-HPC-195, 号 27, p. 1-8, 発行日 2024-08-01
• 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