Presentation | 2021-03-25 Optimizing Data Transfer between CPU and GPU in Model Parallel Training with Mesh TensorFlow Hironori Yokote, Shinobu Miwa, Hayato Yamaki, Hiroki Honda, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | Since deep learning requires an enormous amount of computation time, it is often executed on multiple GPUs. Mesh TensorFlow has been proposed as a language for model parallelization, which is one of the parallelization methods for deep learning. In this paper, we optimize data transfer between CPU and GPU in model parallelization using Mesh TensorFlow. Specifically, our optimization enables training data to be transferred from the CPU to each GPU directly in parallel, though it is originally transferred via a specific GPU in the sample code of Mesh TensorFlow. Our experimental results show that our optimization can both reduce the time of the data transfer and improve the efficiency of GPU-memory utilization. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Mesh TensorFlow / Model Parallel / GPU |
Paper # | CPSY2020-56,DC2020-86 |
Date of Issue | 2021-03-18 (CPSY, DC) |
Conference Information | |
Committee | CPSY / DC / IPSJ-SLDM / IPSJ-EMB / IPSJ-ARC |
---|---|
Conference Date | 2021/3/25(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | Online |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | ETNET2021 |
Chair | Hidetsugu Irie(Univ. of Tokyo) / Hiroshi Takahashi(Ehime Univ.) / Yuichi Nakamura(NEC) / / Hiroshi Inoue(Kyushu Univ.) |
Vice Chair | Michihiro Koibuchi(NII) / Kota Nakajima(Fujitsu Lab.) / Tatsuhiro Tsuchiya(Osaka Univ.) |
Secretary | Michihiro Koibuchi(Univ. of Tokyo) / Kota Nakajima(Nagoya Inst. of Tech.) / Tatsuhiro Tsuchiya(Nihon Univ.) / (Chiba Univ.) / (Tokyo City Univ.) / (Kochi Univ. of Tech.) |
Assistant | Shugo Ogawa(Hitachi) / Eiji Arima(Univ. of Tokyo) |
Paper Information | |
Registration To | Technical Committee on Computer Systems / Technical Committee on Dependable Computing / Special Interest Group on System and LSI Design Methodology / Special Interest Group on Embedded Systems / Special Interest Group on System Architecture |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Optimizing Data Transfer between CPU and GPU in Model Parallel Training with Mesh TensorFlow |
Sub Title (in English) | |
Keyword(1) | Mesh TensorFlow |
Keyword(2) | Model Parallel |
Keyword(3) | GPU |
1st Author's Name | Hironori Yokote |
1st Author's Affiliation | The University of Electro-Communications(UEC) |
2nd Author's Name | Shinobu Miwa |
2nd Author's Affiliation | The University of Electro-Communications(UEC) |
3rd Author's Name | Hayato Yamaki |
3rd Author's Affiliation | The University of Electro-Communications(UEC) |
4th Author's Name | Hiroki Honda |
4th Author's Affiliation | The University of Electro-Communications(UEC) |
Date | 2021-03-25 |
Paper # | CPSY2020-56,DC2020-86 |
Volume (vol) | vol.120 |
Number (no) | CPSY-435,DC-436 |
Page | pp.pp.37-42(CPSY), pp.37-42(DC), |
#Pages | 6 |
Date of Issue | 2021-03-18 (CPSY, DC) |