Efficient Utilization of GPU Cluster Resource for Stencil Computation

• Item type: Symposium(1)
• 公開日: 2015-05-12
• タイトル: Efficient Utilization of GPU Cluster Resource for Stencil Computation
• タイトル / 言語: en
• タイトル: Efficient Utilization of GPU Cluster Resource for Stencil Computation
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: ポスターセッション：システム・アーキテクチャ・ツール
• 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: Tokyo Institute of Technology／JST-CREST
• 著者所属: Tokyo Institute of Technology／JST-CREST
• 著者所属(英): en
• 著者所属(英): Tokyo Institute of Technology / JST-CREST
• 著者所属(英): en
• 著者所属(英): Tokyo Institute of Technology / JST-CREST
• 著者名: Guanghao, Jin; Toshio, Endo
• 著者名(英): Guanghao, Jin; Toshio, Endo
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: In common way case, the domain size of the stencil computation is limited by the memory capacity GPUs in GPU cluster. To efficiently use the resource of GPU cluster, this paper proposes and evaluates parallel optimization method for stencil computation to utilize GPU memory, CPU memory and SSD of the multiple nodes while maintaining high performance. Furthermore, our new method utilizes multiple GPUs in each node to achieve higher performance. Also, it uses the CPU memory and SSD to enable bigger domain computation in each node. Then, it proposes new decomposition method among the nodes to achieve scalability. Evaluation of stencil simulation on 3D domains show that our new method for 7-point achieves good scalability while achieving 2.14 times higher performance than other methods on average.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: In common way case, the domain size of the stencil computation is limited by the memory capacity GPUs in GPU cluster. To efficiently use the resource of GPU cluster, this paper proposes and evaluates parallel optimization method for stencil computation to utilize GPU memory, CPU memory and SSD of the multiple nodes while maintaining high performance. Furthermore, our new method utilizes multiple GPUs in each node to achieve higher performance. Also, it uses the CPU memory and SSD to enable bigger domain computation in each node. Then, it proposes new decomposition method among the nodes to achieve scalability. Evaluation of stencil simulation on 3D domains show that our new method for 7-point achieves good scalability while achieving 2.14 times higher performance than other methods on average.
• 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. 98-98, 発行日 2015-05-12
• 出版者 / 言語: ja
• 出版者: 情報処理学会