Federated Learning based Flow Aware AQM for 5G Networks and Beyond
Shreyanshu Agarwal, Vasanth Kanakaraj, Sukhdeep Singh, Jay Dilipbhai Rathod, Issaac Kommineni · 2023
Rapid proliferation of data intensive applications such as Augmented Reality (AR), Virtual Reality (VR), content streaming, online gaming etc., will create large (Elephant) flows in 5G network and beyond. This combined with mass deployment of Internet of Things (loT) devices can lead to bursty traffic conditions creating challenges for operator in traffic engineering. Traffic flow study in high bandwidth networks indicates that around 20% of flows (elephant) contribute to more than 80 % of data traffic in fast path. This negatively affects small (Mice) flow's Flow Completion Time (FCT). This paper explores a novel method to implement flow aware queue management in User Plane Function (UPF) of 5G Core. Flow classification (elephant or mice) is performed in User Equipment (UE) with help of Federated Learning (FL) approach. We demonstrate the effectiveness of our proposed method by comparison of the FCT, latency, bandwidth performance and queue utilization across the flow duration. The results from conducted simulations using proposed method indicate an improvement in the UPF capacity by 12%-18% by ensuring faster FCT and reduction in queueing delay by 28 % -38 % during buffer bloat.