BBR's Sharing Behavior with CUBIC and Reno

TL;DR

This paper evaluates theoretical models of TCP BBR's behavior with loss-based flows across different versions, revealing their strengths and limitations, especially for the latest BBRv3, through extensive experiments on the FABRIC platform.

Contribution

It systematically compares steady-state and fluid models of BBR across multiple versions, highlighting their applicability and limitations in various network scenarios.

Findings

Steady-state model accurately predicts BBRv1 behavior against single loss-based flows.

Fluid model captures key behaviors of BBRv1 and BBRv2 but struggles with deep buffers and many flows.

Existing models do not accurately predict BBRv3 behavior, indicating the need for new models.

Abstract

TCP BBR's behavior has been explained by various theoretical models, and in particular those that describe how it co-exists with other types of flows. However, as new versions of the BBR protocol have emerged, it remains unclear to what extent the high-level behaviors described by these models apply to the newer versions. In this paper, we systematically evaluate the most influential steady-state and fluid models describing BBR's coexistence with loss-based flows over shared bottleneck links. Our experiments, conducted on a new experimental platform (FABRIC), extend previous evaluations to additional network scenarios, enabling comparisons between the two models and include the recently introduced BBRv3. Our findings confirm that the steady-state model accurately captures BBRv1 behavior, especially against single loss-based flows. The fluid model successfully captures several key behaviors of BBRv1 and BBRv2 but shows limitations, in scenarios involving deep buffers, large numbers of flows, or intra-flow fairness. Importantly, we observe clear discrepancies between existing model predictions and BBRv3 behavior, suggesting the need for an updated or entirely new modeling approach for this latest version. We hope these results validate and strengthen the research community's confidence in these models and identify scenarios where they do not apply.