dBBR: Enhance BBR Performance With Delay Information

Abstract

BBR, a model-based congestion control algorithm developed by Google, is known for its high tolerance to random loss and low latency. However, it exhibits limited competitiveness in deep buffers and suffers severe retransmissions in shallow buffers. To address these issues, we propose dBBR, which adjusts the congestion window and pacing rate based on delay information. We comprehensively analyze the problems and identify that the crux lies in the underutilization of delay information. Based on the analysis, we adjust the window size according to the delay gradient, use the latest delay sample to identify congestion loss, and correct the bandwidth overestimation when the delay is abnormally high. We implement dBBR in the Linux kernel and compare it with other BBR variants through extensive evaluations in controlled environments. The results suggest that dBBR achieves strong competitiveness with Cubic, low retransmissions in small buffers, and high tolerance to random loss—capabilities unmet simultaneously by other state-of-the-art BBR variants. Finally, real-world deployment demonstrates that dBBR achieves consistently high performance across intra-continent and inter-continent scenarios.