Accelerating Federated Learning at Programmable User Plane Function via In-Network Aggregation
Chanbin Bae, Hochan Lee, Sangheon Pack, Youngmin Ji · 2024
Recently, 5G mobile networks are evolving with emerging real-time interaction applications such as AR/VR, which require high throughput and low latency. To meet this demand, user plane function (UPF) should support high-speed data plane and protocol extensions with continuously evolving specifications. Therefore, UPF can be offloaded to a programmable data plan (PDP), which supports flexible packet processing and protocol extension. Meanwhile, as machine learning (ML) models have grown in size and privacy concerns have increased, federated learning (FL) was proposed as a distributed manner solution in mobile networks. To improve the performance of FL, PDP can be used to enhance communication efficiency and decrease learning delay by utilizing in-network aggregation (INA). In this context, solutions for accelerating FL at UPF can be implemented on PDP. In this paper, we present AccelFL that is designed to accelerate FL at UPF by aggregating local gradients in networks via INA. Our experimental results demonstrate that AccelFL can reduce job completion time (JCT) and communication overhead by 30% and 36.9%, respectively.