@techreport{oai:ipsj.ixsq.nii.ac.jp:00203235, author = {平澤, 将一 and 鯉渕, 道紘 and Shoichi, Hirasawa and Michihiro, Koibuchi}, issue = {1}, month = {Feb}, note = {一定のビット誤りを許容する Approximate ネットワークは,既存のネットワークと比較して広帯域かつ低レイテンシであるため, アプリケーションの高速化に有効である.しかしながら,Approximate ネットワークを活用した上で正しい計算結果を得るためには,アプリケーション側の対応が必要である. さらに,アプリケーション中で許容できる誤りの大きさと Approximate ネットワークのパラメータに依存して実行性能は大きく変化する上に,試行ごとに実行性能の揺らぎが発生する.そこで本研究では,Approximate ネットワークを用いてアプリケーションを実行する環境において,実行性能の揺らぎを考慮した上で高速に実行できるパラメータを動的に発見する最適化基盤を提案し,評価を行う., To resolve the next generation interconnection networks latency problem, our prior work has presented error-prone high-bandwidth low-latency networks that do not use the FEC operation. They enable not only high-bandwidth approximate data transfer but also low-bandwidth perfect data transfer, for supporting various kinds of parallel applications that are subject to different levels of bit flips. As the number of approximate data transfers increases, the parallel application can obtain a significant performance improvement at the expense of the degraded quality of results (QoRs). It is difficult for users to identify whether each communication should be approximate or not, so as to obtain the shortest execution time with enough QoRs for a given parallel application. In this study, we apply an optimization framework to approximate interconnection networks with performance fluctuation; it automatically execute applications multiple times to identify whether each communication should be approximate or accurate data transfer.}, title = {Approximateネットワークに対する動的最適化基盤}, year = {2020} }