{"updated":"2025-01-21T20:17:57.799949+00:00","metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00079271","sets":["1164:1579:6260:6627"]},"path":["6627"],"owner":"10","recid":"79271","title":["MPI_Allreduceの「京」上での実装と評価"],"pubdate":{"attribute_name":"公開日","attribute_value":"2011-11-21"},"_buckets":{"deposit":"eab8aa81-0d72-4fca-a41d-49680b402b07"},"_deposit":{"id":"79271","pid":{"type":"depid","value":"79271","revision_id":0},"owners":[10],"status":"published","created_by":10},"item_title":"MPI_Allreduceの「京」上での実装と評価","author_link":["0","0"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"MPI_Allreduceの「京」上での実装と評価"},{"subitem_title":"Implementation and Evaluation of MPI_Allreduce on the K computer","subitem_title_language":"en"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"通信","subitem_subject_scheme":"Other"}]},"item_type_id":"4","publish_date":"2011-11-21","item_4_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"富士通株式会社"},{"subitem_text_value":"富士通株式会社"},{"subitem_text_value":"富士通株式会社"},{"subitem_text_value":"富士通株式会社"},{"subitem_text_value":"九州大学"},{"subitem_text_value":"九州大学"},{"subitem_text_value":"理化学研究所"},{"subitem_text_value":"理化学研究所"},{"subitem_text_value":"理化学研究所"},{"subitem_text_value":"理化学研究所"}]},"item_4_text_4":{"attribute_name":"著者所属(英)","attribute_value_mlt":[{"subitem_text_value":"Fujitsu Ltd.","subitem_text_language":"en"},{"subitem_text_value":"Fujitsu Ltd.","subitem_text_language":"en"},{"subitem_text_value":"Fujitsu Ltd.","subitem_text_language":"en"},{"subitem_text_value":"Fujitsu Ltd.","subitem_text_language":"en"},{"subitem_text_value":"Kyushu University","subitem_text_language":"en"},{"subitem_text_value":"Kyushu University","subitem_text_language":"en"},{"subitem_text_value":"RIKEN","subitem_text_language":"en"},{"subitem_text_value":"RIKEN","subitem_text_language":"en"},{"subitem_text_value":"RIKEN","subitem_text_language":"en"},{"subitem_text_value":"RIKEN","subitem_text_language":"en"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"item_publisher":{"attribute_name":"出版者","attribute_value_mlt":[{"subitem_publisher":"情報処理学会","subitem_publisher_language":"ja"}]},"publish_status":"0","weko_shared_id":-1,"item_file_price":{"attribute_name":"Billing file","attribute_type":"file","attribute_value_mlt":[{"url":{"url":"https://ipsj.ixsq.nii.ac.jp/record/79271/files/IPSJ-ARC11197006.pdf"},"date":[{"dateType":"Available","dateValue":"2013-11-21"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-ARC11197006.pdf","filesize":[{"value":"305.4 kB"}],"mimetype":"application/pdf","priceinfo":[{"tax":["include_tax"],"price":"660","billingrole":"5"},{"tax":["include_tax"],"price":"330","billingrole":"6"},{"tax":["include_tax"],"price":"0","billingrole":"16"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_date","version_id":"1702ed2d-6ddb-4a06-9be9-58780b4c6d05","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2011 by the Information Processing Society of Japan"}]},"item_4_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"松本, 幸"},{"creatorName":"安達, 知也"},{"creatorName":"田中, 稔"},{"creatorName":"住元, 真司"},{"creatorName":"曽我, 武史"},{"creatorName":"南里, 豪志"},{"creatorName":"宇野, 篤也"},{"creatorName":"黒川, 原佳"},{"creatorName":"庄司, 文由"},{"creatorName":"横川, 三津夫"}],"nameIdentifiers":[{}]}]},"item_4_creator_6":{"attribute_name":"著者名(英)","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Yuki, Matsumoto","creatorNameLang":"en"},{"creatorName":"Tomoya, Adachi","creatorNameLang":"en"},{"creatorName":"Minoru, Tanaka","creatorNameLang":"en"},{"creatorName":"Shinji, Sumimoto","creatorNameLang":"en"},{"creatorName":"Takeshi, Soga","creatorNameLang":"en"},{"creatorName":"Takeshi, Nanri","creatorNameLang":"en"},{"creatorName":"Atsuya, Uno","creatorNameLang":"en"},{"creatorName":"Motoyoshi, Kurokawa","creatorNameLang":"en"},{"creatorName":"Fumiyoshi, Shoji","creatorNameLang":"en"},{"creatorName":"Mitsuo, Yokokawa","creatorNameLang":"en"}],"nameIdentifiers":[{}]}]},"item_4_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AN10096105","subitem_source_identifier_type":"NCID"}]},"item_4_textarea_12":{"attribute_name":"Notice","attribute_value_mlt":[{"subitem_textarea_value":"SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc."}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_18gh","resourcetype":"technical report"}]},"item_4_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"本報告では，8 万台以上のノードを直接網で結合した 「京」 における MPI 集団通信の高速化について述べる．従来の MPI 集団通信アルゴリズムは，間接網向けのアルゴリズムが主体であり，これを直接網に適用してもメッセージの衝突のため効率的な通信ができない．このため，高い通信性能を得るためには直接網を意識した集団通信アルゴリズムが必須となる．そこで我々は，トーラス向け Allreduce アルゴリズム Trinaryx3 Allreduce を設計し，「京」 向けの MPI ライブラリに実装した．Trinaryx3 Allreduce は，「京」 の特長である複数 RDMA エンジンを同時に活用することができる．実装を評価した結果，既存の間接網向けアルゴリズムと比較して，5 倍程度バンド幅が向上することを確認した．","subitem_description_type":"Other"}]},"item_4_description_8":{"attribute_name":"論文抄録(英)","attribute_value_mlt":[{"subitem_description":"This paper reports a method of speeding up MPI collective communication on the K computer, that consists of more than 80 thousand computing nodes connected by direct network. Almost all existing MPI libraries only implement algorithms optimized for indirect network. However, such algorithms perform poor on direct network because of collisions of the messages. Thus, in order to achieve high performance on direct network, it is necessary to implement collective algorithms optimized for the network topology. In this paper, Trinaryx3 Allreduce algorithm is designed and implemented in the MPI library for the K computer. The algorithm is optimized for torus network and enables utilizing multiple RDMA engines, one of the strengths of the K computer. The evaluation result shows that the new implementation achieves five times higher bandwidth than existing one, optimized for indirect network.","subitem_description_type":"Other"}]},"item_4_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"10","bibliographic_titles":[{"bibliographic_title":"研究報告計算機アーキテクチャ（ARC）"}],"bibliographicPageStart":"1","bibliographicIssueDates":{"bibliographicIssueDate":"2011-11-21","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"6","bibliographicVolumeNumber":"2011-ARC-197"}]},"relation_version_is_last":true,"weko_creator_id":"10"},"created":"2025-01-18T23:34:02.877019+00:00","id":79271,"links":{}}