Rank Reordering and Data Padding for Optimizing Large-Scale Parallel Image Composition

• Item type: Symposium(1)
• 公開日: 2015-05-12
• タイトル: Rank Reordering and Data Padding for Optimizing Large-Scale Parallel Image Composition
• タイトル / 言語: en
• タイトル: Rank Reordering and Data Padding for Optimizing Large-Scale Parallel Image Composition
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: アプリケーション
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_5794
• 資源タイプ: conference paper
• 著者所属: RIKEN Advanced Institute for Computational Science
• 著者所属: RIKEN Advanced Institute for Computational Science／Light Transport Entertainment Inc.
• 著者所属: RIKEN Advanced Institute for Computational Science
• 著者所属(英): en
• 著者所属(英): RIKEN Advanced Institute for Computational Science
• 著者所属(英): en
• 著者所属(英): RIKEN Advanced Institute for Computational Science / Light Transport Entertainment Inc.
• 著者所属(英): en
• 著者所属(英): RIKEN Advanced Institute for Computational Science
• 著者名: Jorji, Nonaka; Masahiro, Fujita; Kenji, Ono
• 著者名(英): Jorji, Nonaka; Masahiro, Fujita; Kenji, Ono
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: In-situ visualization techniques have been widely recognized as the most rational approach for visual data analysis and mining of large data sets generated by modern leading-edge HPC systems. Among them, the Co-Processing method uses the same HPC systems for both numerical simulation and visualization processing, and in such case, Sort-last visualization method arises as the most feasible solution on massively parallel environments which can involve tens of thousands of nodes. The main drawback of the Sort-last method is the communication intensive final image composition. Popular image composition approaches for massively parallel rendering, such as Binary-Swap and Radix-k methods have reported low scalability when tens of thousands of rendering and composition nodes are involved. This is mostly due to the necessary use of MPI_Gatherv which suffers with scalability problem on massively parallel environments. In this paper, we present an appraoch that uses rank reordering, based on bit-reversal permutation, and data padding, to ensure uniform data subdivision, in order to enable the use of MPI_Gather in substitution of MPI_Gatherv. We implemented this proposed approach on Binary-Swap image composition method and evaluated on the K computer, installed at RIKEN AICS, using up to 4096 processing nodes acting as 32,768 composition nodes in Flat MPI fashion. By applying the proposed technique, the performance degradation of Binary-Swap when using large number of composition nodes was significantly reduced, thus confirming the effectiveness of the proposed approach for improving the scalability of parallel image composition on massively parallel environments.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: In-situ visualization techniques have been widely recognized as the most rational approach for visual data analysis and mining of large data sets generated by modern leading-edge HPC systems. Among them, the Co-Processing method uses the same HPC systems for both numerical simulation and visualization processing, and in such case, Sort-last visualization method arises as the most feasible solution on massively parallel environments which can involve tens of thousands of nodes. The main drawback of the Sort-last method is the communication intensive final image composition. Popular image composition approaches for massively parallel rendering, such as Binary-Swap and Radix-k methods have reported low scalability when tens of thousands of rendering and composition nodes are involved. This is mostly due to the necessary use of MPI_Gatherv which suffers with scalability problem on massively parallel environments. In this paper, we present an appraoch that uses rank reordering, based on bit-reversal permutation, and data padding, to ensure uniform data subdivision, in order to enable the use of MPI_Gather in substitution of MPI_Gatherv. We implemented this proposed approach on Binary-Swap image composition method and evaluated on the K computer, installed at RIKEN AICS, using up to 4096 processing nodes acting as 32,768 composition nodes in Flat MPI fashion. By applying the proposed technique, the performance degradation of Binary-Swap when using large number of composition nodes was significantly reduced, thus confirming the effectiveness of the proposed approach for improving the scalability of parallel image composition on massively parallel environments.
• 書誌情報: ハイパフォーマンスコンピューティングと計算科学シンポジウム論文集; 巻 2015, p. 64-72, 発行日 2015-05-12
• 出版者 / 言語: ja
• 出版者: 情報処理学会