學術產出-Theses
Article View/Open
Publication Export
-
題名 大數據分析於GPU平台之效能評估：以影像辨識為例
Evaluation of Big Data Analytical Performance on GPU Platforms: Computer Vision as an Example作者 曾豐源
Tseng, Feng-Yuan貢獻者 胡毓忠
Hu, Yuh-Jong
曾豐源
Tseng, Feng-Yuan關鍵詞 大數據分析
深度學習
ImageNet
NVIDIA
GPU
NVIDIA DGX A100
NVIDIA DGX Station
Big Data Analysis
Deep Learning
ImageNet
NVIDIA
GPU
NVIDIA DGX A100
NVIDIA DGX Station日期 2021 上傳時間 2-Sep-2021 18:17:33 (UTC+8) 摘要 本研究以ImageNet Large Scale Visual Recognition Challenge (ILSVRC)作為資料集，結合ResNet50深度學習模型，從企業角度為出發點，比較不同的GPU運算環境在AI 大數據分析流程中，探討硬體效能及性價比。本研究以政大電算中心私有雲NVIDIA DGX A100、NVIDIA DGX Station，以及Desktop Computer三種GPU運算環境進行效能測試，並且利用系統監控技術，取得各流程中硬體資源的使用情況，並分析總體效能。因此實驗結果顯示，NVIDIA DGX A100在訓練階段能夠減少模型訓練時間，而在上線階段Desktop Computer其性價比優於NVIDIA DGX A100和NVIDIA DGX Station。
This research adopts the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) as data set, combined with the ResNet50 deep learning model to compare the performance and cost-effectiveness of a hardware under different GPU computing environments applied throughout the AI big data analysis process from an enterprise’s perspective. Performance tests are conducted under three different GPU computing environments, including NVIDIA DGX A100 and NVIDIA DGX Station, hosted as two seperate private clouds owned by the NCCU Computer Center, and the typical desktop computer. We use the system monitoring technology to obtain the usage of hardware resources in each analysis process and to examine the overall performance. The results show that NVIDIA DGX A100 can reduce the time needed for model training during training phase, while Desktop Computer is more cost-effective than NVIDIA DGX A100 and NVIDIA DGX Station during the online phase.參考文獻 [1] Nvidia dali documentation. https://docs.nvidia.com/deeplearning/dali/user-guide/docs/. [Online; accessed 30May2021].[2] Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., and FeiFei, L. Imagenet: A largescale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition (2009), Ieee, pp. 248–255.[3] He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (2016), pp. 770–778.[4] Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012), 1097–1105.[5] Lawrence, J., Malmsten, J., Rybka, A., et al. Comparing tensorflow deep learning performance using cpus, gpus, local pcs and cloud.[6] Lin, C.Y., Pai, H.Y., and Chou, J. Comparison between baremetal, container and vm using tensorflow image classification benchmarks for deep learning cloud platform. In CLOSER (2018), pp. 376–383.[7] Peter Mattson, C. C., and Cody Coleman, e. a. Mlperf training benchmark, 2020.[8] Reddi, V. J., Cheng, C., and David Kanter, e. a. Mlperf inference benchmark, 2020.[9] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. Going deeper with convolutions, 2014.[10] Wikipedia contributors. Huang’s law — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Huang%27s_law&oldid=996423603, 2020. [Online; accessed 27January2021].[11] Wikipedia contributors. Imagenet — Wikipedia, the free encyclopedia, 2021. [Online; accessed 26May2021].[12] Wikipedia contributors. Kubernetes — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Kubernetes&oldid=1024839217, 2021. [Online; accessed 28May2021]. 描述 碩士
國立政治大學
資訊科學系碩士在職專班
107971025資料來源 http://thesis.lib.nccu.edu.tw/record/#G0107971025 資料類型 thesis dc.contributor.advisor 胡毓忠 zh_TW dc.contributor.advisor Hu, Yuh-Jong en_US dc.contributor.author (Authors) 曾豐源 zh_TW dc.contributor.author (Authors) Tseng, Feng-Yuan en_US dc.creator (作者) 曾豐源 zh_TW dc.creator (作者) Tseng, Feng-Yuan en_US dc.date (日期) 2021 en_US dc.date.accessioned 2-Sep-2021 18:17:33 (UTC+8) - dc.date.available 2-Sep-2021 18:17:33 (UTC+8) - dc.date.issued (上傳時間) 2-Sep-2021 18:17:33 (UTC+8) - dc.identifier (Other Identifiers) G0107971025 en_US dc.identifier.uri (URI) http://nccur.lib.nccu.edu.tw/handle/140.119/137165 - dc.description (描述) 碩士 zh_TW dc.description (描述) 國立政治大學 zh_TW dc.description (描述) 資訊科學系碩士在職專班 zh_TW dc.description (描述) 107971025 zh_TW dc.description.abstract (摘要) 本研究以ImageNet Large Scale Visual Recognition Challenge (ILSVRC)作為資料集，結合ResNet50深度學習模型，從企業角度為出發點，比較不同的GPU運算環境在AI 大數據分析流程中，探討硬體效能及性價比。本研究以政大電算中心私有雲NVIDIA DGX A100、NVIDIA DGX Station，以及Desktop Computer三種GPU運算環境進行效能測試，並且利用系統監控技術，取得各流程中硬體資源的使用情況，並分析總體效能。因此實驗結果顯示，NVIDIA DGX A100在訓練階段能夠減少模型訓練時間，而在上線階段Desktop Computer其性價比優於NVIDIA DGX A100和NVIDIA DGX Station。 zh_TW dc.description.abstract (摘要) This research adopts the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) as data set, combined with the ResNet50 deep learning model to compare the performance and cost-effectiveness of a hardware under different GPU computing environments applied throughout the AI big data analysis process from an enterprise’s perspective. Performance tests are conducted under three different GPU computing environments, including NVIDIA DGX A100 and NVIDIA DGX Station, hosted as two seperate private clouds owned by the NCCU Computer Center, and the typical desktop computer. We use the system monitoring technology to obtain the usage of hardware resources in each analysis process and to examine the overall performance. The results show that NVIDIA DGX A100 can reduce the time needed for model training during training phase, while Desktop Computer is more cost-effective than NVIDIA DGX A100 and NVIDIA DGX Station during the online phase. en_US dc.description.tableofcontents 誌謝 i摘要 iiAbstract iii目錄 iv圖目錄 vi表目錄 vii1 導論 11.1 研究動機 11.2 研究目的 21.3 論文架構 32 研究背景 42.1 AI大數據分析流程 42.2 ImageNet數據集 62.3 ResNet網路模型 62.4 政大NVIDIA GPU運算環境架構 73 相關研究 113.1 模型訓練基準評估研究案例 113.2 模型推理基準評估研究案例 124 研究構架與方法 144.1 實驗資料集 144.2 選擇網路模型 154.3 各GPU 運算環境之配置 164.4 訓練階段實驗方法 164.5 上線階段實驗方法 185 實驗結果 205.1 訓練階段之實驗結果 205.2 上線階段之實驗結果 265.3 總體效能與性價比分析 276 結論與未來展望 306.1 結論 306.2 未來展望 30參考文獻 31 zh_TW dc.format.extent 2284020 bytes - dc.format.mimetype application/pdf - dc.source.uri (資料來源) http://thesis.lib.nccu.edu.tw/record/#G0107971025 en_US dc.subject (關鍵詞) 大數據分析 zh_TW dc.subject (關鍵詞) 深度學習 zh_TW dc.subject (關鍵詞) ImageNet zh_TW dc.subject (關鍵詞) NVIDIA zh_TW dc.subject (關鍵詞) GPU zh_TW dc.subject (關鍵詞) NVIDIA DGX A100 zh_TW dc.subject (關鍵詞) NVIDIA DGX Station zh_TW dc.subject (關鍵詞) Big Data Analysis en_US dc.subject (關鍵詞) Deep Learning en_US dc.subject (關鍵詞) ImageNet en_US dc.subject (關鍵詞) NVIDIA en_US dc.subject (關鍵詞) GPU en_US dc.subject (關鍵詞) NVIDIA DGX A100 en_US dc.subject (關鍵詞) NVIDIA DGX Station en_US dc.title (題名) 大數據分析於GPU平台之效能評估：以影像辨識為例 zh_TW dc.title (題名) Evaluation of Big Data Analytical Performance on GPU Platforms: Computer Vision as an Example en_US dc.type (資料類型) thesis en_US dc.relation.reference (參考文獻) [1] Nvidia dali documentation. https://docs.nvidia.com/deeplearning/dali/user-guide/docs/. [Online; accessed 30May2021].[2] Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., and FeiFei, L. Imagenet: A largescale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition (2009), Ieee, pp. 248–255.[3] He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (2016), pp. 770–778.[4] Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012), 1097–1105.[5] Lawrence, J., Malmsten, J., Rybka, A., et al. Comparing tensorflow deep learning performance using cpus, gpus, local pcs and cloud.[6] Lin, C.Y., Pai, H.Y., and Chou, J. Comparison between baremetal, container and vm using tensorflow image classification benchmarks for deep learning cloud platform. In CLOSER (2018), pp. 376–383.[7] Peter Mattson, C. C., and Cody Coleman, e. a. Mlperf training benchmark, 2020.[8] Reddi, V. J., Cheng, C., and David Kanter, e. a. Mlperf inference benchmark, 2020.[9] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. Going deeper with convolutions, 2014.[10] Wikipedia contributors. Huang’s law — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Huang%27s_law&oldid=996423603, 2020. [Online; accessed 27January2021].[11] Wikipedia contributors. Imagenet — Wikipedia, the free encyclopedia, 2021. [Online; accessed 26May2021].[12] Wikipedia contributors. Kubernetes — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Kubernetes&oldid=1024839217, 2021. [Online; accessed 28May2021]. zh_TW dc.identifier.doi (DOI) 10.6814/NCCU202101203 en_US
