Abstract: Bandwidth slicing is introduced to support federated learning in edge computing to assure low communication delay for training traffic. Results reveal that bandwidth slicing significantly improves training efficiency while achieving good learning accuracy.
1. Introduction
Federated Learning (FL) has emerged as a critical technique for collaborative machine learning at the network edge, allowing devices to train models locally without sharing raw data. However, the performance of FL is heavily dependent on the synchronization of model updates at a Centralized Parameter Server (CPS). In edge computing environments like Passive Optical Networks (PONs), FL traffic often competes with background broadband data, leading to unpredictable synchronization delays. These delays, caused by "straggler" nodes with slow hardware or congested network paths, can severely bottleneck the training process. This paper proposes a Bandwidth Slicing (BS) algorithm to provide dedicated network resources for FL traffic, ensuring deterministic and low-latency communication.
2. System Model and the BS Algorithm
The researchers focus on a TDM-PON architecture where Edge Computing (EC) nodes are associated with Optical Network Units (ONUs), and the CPS is co-located with the Optical Line Terminal (OLT).
As illustrated in Figure 1, a single training round involves distributing the global model ($T^{DL}$), performing local training at the ECs ($T^{UD}$), and uploading local updates ($T^{UL}$). The proposed BS algorithm calculates a specific bandwidth slice for these updates, defined by a starting time ($t_s$), ending time ($t_e$), and capacity ($B$). By integrating these slices into the PON's periodic polling cycles, the algorithm isolates FL traffic from background noise, effectively mitigating network-induced delays.
Unlike traditional methods that might drop slow nodes to save time—thereby sacrificing accuracy—the BS algorithm handles heterogeneity by using the varying computation and communication times of nodes as inputs to optimize the network slice schedule.
3. Experimental Setup
The approach was evaluated using the FEMNIST (Federated Extended MNIST) dataset with a Convolutional Neural Network (CNN) and the standard FedAvg aggregation algorithm. The simulation environment modeled a 10 Gbps TDM-PON with 128 ONUs and background traffic following a Poisson distribution. The BS algorithm was benchmarked against a standard First-Come-First-Served (FCFS) queuing protocol.
4. Key Results and Discussion
The experimental results highlight the dual benefits of the BS algorithm: improved training speed and maintained model accuracy.
As shown in Figure 4(a), the system achieves a peak accuracy of 82% with full client participation. More importantly, Figure 4(b) demonstrates that under high network load (0.8), the BS algorithm (dashed magenta line) saves 36% of training time compared to the FCFS benchmark to reach the target accuracy. While FCFS performance degrades as background traffic increases—with synchronization times spiking from ~6.7s to ~8s—the BS algorithm maintains stable, low communication delays regardless of the total network load.
5. Conclusion
Bandwidth slicing offers a robust solution for the synchronization bottleneck in edge-based Federated Learning. By providing Quality of Service (QoS) through network isolation, the BS algorithm significantly accelerates training without compromising the accuracy gains provided by full client participation. Future research will focus on validating these results on physical testbeds and incorporating energy consumption metrics into the optimization process.
Summary updated