【摘要】 针对在异构环境中，计算节点之间的资源配置存在显著差异，产生的掉队节点会增加参数同步的时间开销，降低模型的整体训练效率问题，提出了一种基于强化学习的参数同步优化策略。融合各节点性能状态并构建智能调度网络对节点批次大小与通信压缩比进行自适应调整，以此平衡节点的计算与通信任务负载，使各节点的整体迭代时间趋于一致，从而消除参数同步过程中的长尾效应。实验结果表明，在CIFAR100和CIFAR10数据集上的训练中，该策略在实现相同准确率的情况下，将整体训练时间缩短了20%以上，相较于主流同步策略(如ADSP和LOCALSGD)，表现出更高的效率和优越的泛化能力。更多还原

【Abstract】 In heterogeneous environments,there are significant differences in resource configurations among computing nodes. The resulting straggler nodes increase the time overhead of parameter synchronization and reduce the overall training efficiency of the model. To address this issue,this paper proposes a parameter synchronization optimization strategy based on reinforcement learning. The strategy integrates the performance states of various nodes and constructs an intelligent scheduling network to adaptively adjust the batch size of nodes and communication compression ratio. This balances the computational and communication task loads of the nodes,aligns the overall iteration time of each node,and thereby eliminates the long-tail effect in the parameter synchronization process. Experimental results show that during training on the CIFAR100 and CIFAR10 datasets,the proposed strategy shortens the overall training time by more than 20% while achieving the same accuracy. Compared with mainstream synchronization strategies( such as ADSP and LocalSGD),it exhibits higher efficiency and superior generalization ability.更多还原