面向联合学习的D2D计算任务卸载

D2D computation task offloading for efficient federated learning

联合学习是一种分布式机器学习,边缘节点的计算和通信资源受限等因素是限制其性能优化的瓶颈。当边缘节点的计算和通信能力异构时,需要对通信和计算进行联合优化。提出了一种面向联合学习的D2D计算任务卸载方案,不同边缘节点通过D2D通信交换数据样本,平衡节点的处理能力和任务负载,以最小化联合学习模型训练过程的总时延。仿真结果表明,所提出的D2D计算任务卸载方案能显著提高联合学习的模型训练速度和效率。

Federated learning is a kind of distributed machine learning technique.The factor of communication and computation resource constraints at the edge node is becoming the performance bottleneck.In particular,when different edge node has distinct computation and communication capabilities,the model training performance may degrade severely,thus necessitating the joint communication and computation optimization.To tackle this challenge,a computational task offloading scheme enabled by device-to-device(D2D)communications was proposed,in which different edge node exchanged data samples via D2D communication links to balance the processing capability and task load,in order to minimize the total time delay for machine learning model training.Simulation results show that compared to the benchmark scheme without such D2D task offloading the training speed and efficiency of federated learning has be improved significantly.