联邦边缘学习的低功耗带宽分配与用户调度

Low-power bandwidth allocation and user scheduling for federated edge learning

为了降低联邦边缘学习(federated edge learning, FEL)能耗,提出了应用于多服务器边缘蜂窝网络的联合资源优化分配策略,包含低功耗带宽分配(bandwidth allocation, BA)策略以及智能边缘用户调度机制。低功耗BA理论推导结果表明,为了实现在约束时间内的能耗最小化,应为计算能力较弱、信道条件较差的设备分配更大带宽。进一步,在本地边缘设备数据量差异较大和数据量近似两种情景下,模拟了智能边缘用户调度机制,并提出了时间平均筛选和时间峰值筛选两种优化策略。仿真结果表明,与参考算法相比,采用低功耗BA策略,传输功率最大可以降低43.5%,传输能耗最大降低17.7%;采用优化的用户调度策略可以进一步提升系统性能,其中,设备传输功率最大降低了56.5%,设备传输能耗最大降低了22%。

To reduce the energy consumption of federated edge learning(FEL),a joint resource optimization allocation strategy for multi-server edge cellular networks was proposed, including low-power bandwidth allocation(BA) strategy and intelligent edge user scheduling mechanism.The derivation of the low-power BA theory shows that in order to minimize the energy consumption within the constraint time, more bandwidth should be allocated to devices with weaker computing power and poor channel conditions.Furthermore, the intelligent edge user scheduling mechanism was simulated under two scenarios where the data volume of the local edge device is very different and the data volume is similar, and two optimization strategies of time average filtering and time peak filtering were proposed.The simulation results show that compared with the baseline strategy, the maximum transmission power can be reduced by 43.5% and the maximum transmission energy consumption can be reduced by 17.7% using the low-power BA strategy;the optimized user scheduling policy can further improve the system performance that the maximum transmission power is reduced by 56.5% and the maximum transmission energy consumption is reduced by 22%.