| Title |
GPUDirect Storage Performance Analysis and DSGDS Implementation |
| Authors |
김정명(Jungmyoung Kim) ; 박정민(Jeongmin Park) ; 박문우(Piao Wenyu) ; 노재춘(Jaechun No) ; 박성순(Sungsoon Park) |
| DOI |
https://doi.org/10.5573/ieie.2023.60.9.27 |
| Keywords |
I/O; Distributed storage; GPU; CUDA; GPUDirect storage |
| Abstract |
GPU has mainly been used in graphics, but with the development of GPU technology, applications having leveraged CPU commonly utilize GPUs these days, due to their potential of high computing power, such as deep learning or parallel computing with CUDA. In addition, GPUs are used in various fields, including computer finance, data science and analysis, medical imaging, media, and entertainment. However, the traditional CUDA-based method is not effective because it first works reading data from the CPU and copying it to GPU memory. To overcome such a drawback, the proposed technology is GPUDirect Storage, which creates a direct data path between local NVMe or remote storage and GPU memory. This feature is activated through either a network adapter or Direct-Memory Access (DMA), to transfers data to GPU memory, while avoiding the use of CPU bounce buffer. In this paper, we first analyzed the performance of GPUDirect Storage, and then developed the DSGDS(GPUDirect Store based on Distributed Storage Environment) on top of GlusterFS, We compared our method with GPU I/O utilizing the traditional CUDA-based method. |