Hierarchical Communication and Computation Scheduling for Efficient Federated Learning in Cloud-Edge Collaborative Intelligent Systems

Abstract

As federated learning and cloud-edge collaborative training are increasingly applied in distributed intelligent computing, issues such as high communication overhead, strong heterogeneity of computing resources, and insufficient efficiency of multi-layer node collaboration have gradually become key factors restricting system performance improvement. To address these issues, this paper studies a joint scheduling optimization method for communication and computation, addressing the distributed training needs in cloud computing scenarios, and constructs a unified modeling framework suitable for collaborative training across terminals, edge nodes, and the cloud. In terms of method design, firstly, starting from the multi-layer collaborative structure, a systematic description of client access relationships, edge node activation states, and cloud coordination processes is provided, incorporating communication latency, local computation overhead, and regional forwarding costs during training into a unified latency model. Secondly, to improve the effectiveness of node participation and the rationality of resource allocation, a utility evaluation mechanism integrating data quality, training status, and resource consumption is introduced to achieve dynamic selection of participating nodes under constraints of limited bandwidth and computational budget. Subsequently, combining the characteristics of cloud-edge layered training, a joint optimization objective for resource-constrained environments is further constructed, considering latency control, training quality, and energy consumption costs collaboratively, and enhancing the organizational capacity and execution feasibility among multiple nodes through hierarchical scheduling constraints. Finally, an aggregation mechanism combining latency and data contribution is designed during the model update phase to mitigate the adverse effects of heterogeneous node differences on the global model evolution. Experimental results show that the proposed method performs well in terms of accuracy, overall performance, and system coordination capabilities, effectively improving resource utilization and overall training efficiency in federated learning and cloud-edge collaborative training scenarios.