Research on next generation accelerator-devices and file IO


- 1.1 Data Processing Infrastructure (Cloud, Large-Scale Data Processing Systems, Machine Learning Frameworks, Databases, Data Structure)
- 2.7 Disaster Prevention, Reconstruction (Earthquake, Tsunami, Volcano, Meteorology, Flood, Natural Disaster)
- 2.9 Space Science (astronomical observation data, etc.)
Toshihiro Hanawa
Information Technology Center
Professor
Efficient execution of actual applications on GPU clusters requires deep consideration of data IO to GPUs; therefore, reducing IO latency, and overlapping data transfer and computation are promising techniques. This study aims to establish methods that facilitate GPUDirect Storage (GDS) utilization, which enables direct file IO on GPU, or the overlap of computation and file IO. The goal is to improve the IO performance between GPU and file in various applications which have different file IO characteristics.
Research collaborators
University of Tsukuba, Nagoya University
Related publications
Mizuki Tominaga, Toshihiro Hanawa, Yohei Miki: IPSJ SIG Technical Reports, 2023-HPC-190(6), 1-8 (2023-07-27) (In Japanese)
Contact
- Toshihiro Hanawa
- Email: hanawa[at]cc.u-tokyo.ac.jp
※[at]=@