Improving per-flow fairness by ML-based estimation of competing flows’ congestion control algorithm
Keito Maeta, Gen Kitagata, Go Hasegawa · 2022 Thirteenth International Conference on Ubiquitous and Future Networks (ICUFN) · 2022
A variety of Internet congestion control algorithms have been developed to overcome the ever-increasing diversity of the Internet. As a result, more than ten different congestion control algorithms co-exist on the current Internet. Therefore, flows with different congestion control algorithms must compete with the bottleneck link, causing unfair bandwidth share. In this paper, we aim to improve per-flow fairness in such situations. In the proposed method, a flow estimates congestion control algorithms of competing flows by a machine learning-based estimation and majority vote algorithm. It then changes its congestion control algorithm based on the estimation results. We evaluated the performance of the proposed method by extensive experiments and found that the estimation accuracy of the proposed method was significantly larger than the chance level and that the per-flow fairness was improved by at most 77.9 [%].