Distributed Deep Learning Framework Based on Cloud Oss for Efficient Model Synchronization

College of information engineering, Shanghai Maritime University, Shanghai, China

202210310195@stu.shmtu.edu.cn

Abstract

With the rapid development of deep learning technology, training deep neural networks (DNN) under limited computational resources has become a hot topic, especially in fields like computer vision and natural language processing. Traditional distributed training often relies on centralized parameter servers, which can lead to communication bottlenecks, limiting system performance. To address the issue, this paper proposes a distributed training framework based on cloud Object Storage Service (OSS), utilizing OSS for parameter synchronization and avoiding the bottlenecks associated with traditional parameter server models. Experimental results show that the OSS-based framework achieves faster convergence speed, higher training efficiency, and shorter training time compared to traditional distributed training frameworks. Specifically, the OSS framework reduces training time by approximately 30% and demonstrates better flexibility and scalability across different worker configurations, effectively enhancing performance in large-scale distributed training tasks. The significance of this study lies in providing an efficient and scalable solution for training deep neural networks under limited computational resources.