学際大規模情報基盤共同利用・共同研究拠点

採択課題 【詳細】

jh260086 Study on Efficient File IO for GPU
課題代表者 塙 敏博(東京大学情報基盤センター)
Toshihiro Hanawa (Information Technology Center, The University of Tokyo)
概要 To use GPUs in practical applications, data transfer between host and GPU memory often becomes a bottleneck. GPUDirect Storage (GDS) aims to reduce this overhead by enabling direct RDMA transfer of GPU memory. This research focuses on optimizing file IO for GPUs, particularly in the Miyabi system with NVIDIA GH200 Grace-Hopper Superchip. The goal is to establish methods for handling direct file IO on the GPU and overlapping data transfer with computation, improving efficiency in various applications. Additionally, the research aims to provide libraries for optimal file IO across all JHPCN GPU systems.
関連Webページ
報告書等 研究紹介ポスター / 最終報告書
無断転載禁止