ML-Optimized BBR Congestion Control

• Machine Learning-Optimized BBR is a hybrid approach that combines ML models like GNNs and reinforcement learning with BBR to address inter-protocol unfairness.
• It dynamically tunes network parameters based on real-time data, leading to improved throughput, reduced latency, and enhanced convergence speed.
• Experimental results demonstrate significant gains in fairness and performance compared to traditional BBR, making it a promising solution for heterogeneous network conditions.

Machine Learning-Optimized BBR congestion control refers to approaches that integrate ML or data-driven algorithm generation into the core adaptation mechanisms of Google's Bottleneck Bandwidth and Round-trip Propagation Time (BBR) congestion control algorithm. These techniques address known limitations of BBR, most notably its inter-protocol unfairness against legacy TCP and its suboptimal parameter tuning within diverse network environments. Methods include incorporating classifiers such as Graph Neural Networks (GNNs), reinforcement learning for online parameter tuning, and recent frameworks using LLMs to automatically generate, evaluate, and specialize improved BBR code variants. Key advances have demonstrated substantial improvements in fairness, throughput, and convergence speed, as supported by experimental results across emulated and production-like network environments (Mhaske et al., 2023, Ketabi et al., 2023, He et al., 22 Aug 2025).

1. Background and Motivation

BBR congestion control operates by estimating the bottleneck bandwidth $B$ and the minimum observed round-trip propagation delay ($RT_{prop}$). Its core decisions are:

• $B = \frac{\text{delivered\_bytes}}{RTT}$
• $r_{pacing} = B \cdot G_{pacing}$
• $cwnd \approx B \cdot RT_{prop}$

Where $G_{pacing}$ is typically set to 1.0 in the steady state.

Standard BBR aggressively maintains a high bandwidth-delay product and constant pacing, leading to high throughput and low latency for isolated flows. However, when coexisting with CUBIC or Reno flows at a shared bottleneck, BBR tends to monopolize bandwidth, leading to pronounced inter-protocol unfairness. This is quantified using Jain's fairness index:

$$J = \frac{(\sum_{i=1}^n x_i)^2}{n \sum_{i=1}^n x_i^2}, \quad 0 \leq J \leq 1$$

where $x_i$ are the throughputs of the $n$ competing flows. Addressing this unfairness and achieving robust performance across heterogeneous network conditions motivates the integration of ML into BBR (Mhaske et al., 2023, He et al., 22 Aug 2025).

2. GNN-Based Inter-Protocol Fairness Enhancement

A representative approach applies a graph neural network (GNN) classifier to improve BBR's fairness in mixed-flow environments (Mhaske et al., 2023). The methodology comprises:

• Feature Engineering: At each congestion epoch, each flow collects:
  • $k$-window BtlBw and $RT_{prop}$0 histories
  • Pacing gain and $RT_{prop}$1 gain
  • Inflight bytes, recent loss rate, and/or ECN markings
• Graph Representation: Each flow is a node; undirected edges connect flows sharing a bottleneck (identified e.g., via traceroute).
• GNN Architecture: A 2-layer Graph Convolutional Network (GCN) processes the co-bottleneck subgraph. For layer $RT_{prop}$2, using node feature matrix $RT_{prop}$3, adjacency $RT_{prop}$4, degree $RT_{prop}$5, and weights $RT_{prop}$6:

$RT_{prop}$7

$RT_{prop}$8 is ReLU for the first layer. The output layer emits a softmax over "fair" and "unfair."

• Training: Datasets consist of multi-flow traces on controlled testbeds. Epochs are labeled as "unfair" if $RT_{prop}$9 or any flow exceeds double its equal share. Loss is standard cross-entropy.

• Integration with BBR: At end of each ProbeBW cycle:

  1. Collect features and assemble co-bottleneck subgraph.
  2. Run GCN classifier; if "unfair