Modeling BBR v3 Congestion Control Behavior Using Supervised ML Techniques

Abstract

BBR v3 has recently emerged as the most sophisticated model-based congestion control. It measures different network parameters such as bottleneck bandwidth, round-trip time (RTT), packet loss rate, and explicit congestion notification (ECN) to get the true picture of the available network bandwidth and then sets its pacing rate. The goal is to operate near Kleinrock’s optimal operating point to prevent excessive queue formation in case of large buffers and to prevent overreacting in case of shallow buffers. However, there are still limitations that exist in properly setting up the pacing rate that matches the delivery rate at the receiver’s side. BBR v3 generally sets the pacing rate relatively high, as the pacing gain values are generally fixed in its probing for the bandwidth phase. In this paper, we have evaluated this issue using machine learning regression algorithms such as XGBoost, Random Forest, Neural Networks, Linear Regression, Support Vector Regression, Gradient Boosting, and Decision Tree by training these models and predicting its pacing rate. The machine learning (ML) regression models are then evaluated using various metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R² score. Our results show that Linear Regression and XGBoost provide the best results in terms of lowest error and superior predictability of BBR v3’s pacing rate.