Abstract

With the widespread application of deep learning in various domains, the demand for computational resources and training efficiency is growing exponentially. Cloud computing, with its robust computational power and flexible resource scheduling, has become a crucial platform for distributed deep learning training. However, existing distributed training methods still face challenges such as uneven task scheduling, excessive communication overhead, and low resource utilization. This paper proposes an optimization strategy for distributed deep learning training in cloud environments, encompassing dynamic task allocation, data partitioning and caching optimization, communication efficiency improvement, and heterogeneous resource scheduling methods. Experimental validation on typical cloud computing platforms with various deep learning tasks demonstrates that the proposed strategies significantly reduce training time, improve resource utilization, and effectively minimize communication overhead, providing strong support for the efficient execution of cloud-based deep learning tasks.