一句话：将对象存储客户端卸载到DPU并实现DPU到GPU直连，降低DL训练CPU开销与数据搬运。
对象存储因成本与扩展性被用于深度学习，但会加重DL集群CPU负担并引入多次数据搬运。DPU卸载看似可行，但直接迁移现有对象存储客户端性能严重下降；仅卸载客户端仍需DPU→主机→GPU的冗余搬运。本文提出OS2G架构，将对象存储客户端卸载至DPU，并设计GPUDirect DPU（GDD）使DPU加速存储系统与GPU间直接传输，绕过主机。OS2G客户端在DPU上利用异步、预读与并发策略提供高性能对象存储服务。实验表明，相比S3FS和S3Connector，OS2G将ResNet18执行时间分别降低34.3%和50.4%，CPU消耗分别降低61.9%和57.7%。
Object storage is increasingly attractive for deep learning (DL) applications due to its cost-effectiveness and high scalability. However, it exacerbates CPU burdens in DL clusters due to intensive object storage processing and multiple data movements. Data processing unit (DPU) offloading is a promising solution, but naively offloading the existing object storage client leads to severe performance degradation. Besides, only offloading the object storage client still involves redundant data movements, as data must first transfer from the DPU to the host and then from the host to the GPU, which continues to consume valuable host resources. In this paper, we propose OS2G, a high-performance offloading architecture designed to free up valuable CPU resources while providing high-performance storage services for DL applications. The key idea of OS2G is to offload the object storage client to a DPU and enable direct data transfer between the DPU and GPU. Specifically, we design a high-performance OS2G Client running on the DPU, utilizing asynchrony, pre-reading, and concurrency strategies to provide high-performance object storage services. Additionally, we propose the GPUDirect DPU (GDD) technique for OS2G to optimize the data path, allowing direct data transfer between the DPU-accelerated storage system and the GPU computing system, fully bypassing the host. Results demonstrate that compared to S3FS and S3Connector, OS2G reduces the execution time of the ResNet18 model by 34.3% and 50.4%, and also decreases CPU consumption by 61.9% and 57.7%, respectively.
@inproceedings{0008CLWFZZXLLZS25,
title = {OS2G: A High-Performance DPU Offloading Architecture for GPU-based Deep Learning with Object Storage},
author = {Zhen Jin and Yiquan Chen and Mingxu Liang and Yijing Wang and Guoju Fang and Ao Zhou and Keyao Zhang and Jiexiong Xu and Wenhai Lin and Yiquan Lin and Shushu Zhao and Wenkai Shi and Zhenhua He and Shishun Cai and Wenzhi Chen},
booktitle = {ASPLOS},
year = {2025},
doi = {10.1145/3676641.3716265}
}