BBR Congestion Control Algorithms

Updated 16 January 2026

BBR congestion control is a model-based approach that optimizes throughput by explicitly measuring the bottleneck bandwidth and minimum RTT.
It employs techniques like pacing rate control, active probing cycles, and window adjustments to minimize queuing delays and maximize efficiency.
Recent variants such as BBRv2, BBRv3, and Delay-BBR integrate fairness mechanisms and adaptive controls, addressing RTT-unfairness and stability challenges.

The Bottleneck Bandwidth and Round-trip propagation time (BBR) congestion control algorithms represent a model-based approach to network congestion control, explicitly targeting the optimal operating point where throughput is maximized while queueing delays are minimized. BBR's architecture and numerous variants have fundamentally redefined sender-side congestion management, both in single-path and multipath networking scenarios. Key innovations include bandwidth-delay product tracking, explicit window and pacing rate control, active probing cycles, and more recent integrations of fairness and delay responsiveness mechanisms.

1. The Core Model: BBR’s Control Philosophy

BBR fundamentally departs from loss-based additive-increase multiplicative-decrease (AIMD) algorithms (e.g., Reno, Cubic) by ignoring packet loss as a congestion signal, instead explicitly modeling the available bottleneck bandwidth (BtlBw) and minimum round-trip propagation delay (RTprop) using sender-measured statistics. At each packet acknowledgment, BBR updates:

BtlBw as the maximum delivery rate over a sliding window of recent RTTs (typically 8–10 RTTs).
RTprop as the minimum RTT sample seen over a longer epoch (usually 10 s).

The core control variables are calculated as:

$\mathrm{pacing\_rate} = \mathit{pacing\_gain} \cdot \mathrm{BtlBw}$

$\mathrm{cwnd} = \mathit{cwnd\_gain} \cdot \mathrm{BtlBw} \cdot \mathrm{RTprop}$

BBR's state machine cycles through phases such as Startup (aggressive gain), Drain (below-available bandwidth to empty queues), ProbeBW (oscillatory, to actively probe available bandwidth with gain cycles), and ProbeRTT (minimal inflight to refresh RTprop). In classical BBR (v1), the ProbeBW phase uses an 8-RTT periodic gain sequence: {1.25, 0.75, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0} (Abreu et al., 25 Oct 2025).

2. BBR Derivatives and RTT Fairness Enhancements

2.1 RTT Unfairness in BBRv1

Deployment studies and analytical models revealed a severe intra-protocol unfairness: longer-RTT flows inject more excess data during probe phases (because their BDP is larger), dominating throughput to the detriment of shorter-RTT flows. Experimental data show, for instance, in a 10 ms vs 50 ms RTT pair, the short flow captured only ~6% of the link; this could lead to starvation at extreme RTT ratios (Ma et al., 2017). The root cause is that the excess queue contribution per probe is proportional to the flow’s RTT.

2.2 Fairness-Preserving Variants

Fair-BBR (BBQ) (Ma et al., 2017): Imposes a cap (α, typically 3 ms) on the probe-up slot duration, ensuring all flows contribute a RTT-independent volume of probe excess. When the bottleneck is full (detected via RTT inflation), the probe duration is capped; underutilized links revert to BBR’s default probe window.
FaiRTT (Abrol et al., 2024): Instead of global probe duration caps, it introduces a per-ACK BDP adjustment. If a flow’s instantaneous RTT exceeds the empirical fairness threshold, it scales its BDP down by $(\mathrm{minRTT} \cdot \gamma) / \mathrm{lastRTT}$ , reducing over-injection by long-RTT flows without sacrificing bottleneck utilization. In extensive simulations, FaiRTT achieves average throughput ratios near unity and fairness indices ≈0.98–0.99, outperforming BBRv2 (fairness ≈0.95).

3. Forward Evolution: BBRv2, BBRv3, Delay-BBR, and Beyond

3.1 BBRv2

BBRv2 introduces several mechanisms centered around robustness and fairness (Ivanov, 2020, Yang et al., 2021, Abreu et al., 25 Oct 2025):

Loss/ECN-sensitivity: Explicit inflight limitation on high loss/marking (e.g., inflight_hi, inflight_lo set adaptively).
Differential probing: ProbeBW uses gentler gain cycles, e.g., {1, 1+ε, 1-ε} instead of ±25% up/down. ε≈0.15 is typical. This reduces queue swings and coexists more equitably with AIMD flows.
Adjustable ProbeBW periods: The periodicity of probe windows is randomized to avoid phase alignment and improve fairness on multi-flow, multi-bottleneck scenarios.
Startup headroom and early exit: Startup exits early on plateaued bandwidth or excessive loss/ECN marks, preventing BW-hogging by early or long-standing flows.

BBRv2 is shown empirically to reduce median loss and standing queue sizes (down by ≈3× compared to v1), improve RTT fairness, and raise the fairness index versus Cubic from ~0.6 (v1) to ≥0.9. However, in the presence of very deep buffers, BBRv2 can still exhibit bufferbloat due to inflight_hi initialization artifacts (Scherrer et al., 2022).

3.2 BBRv3 and BBRv2+

BBRv3 further introduces ECN-driven slope control mimicking DCTCP, randomizes gain-cycle ordering, and shortens ProbeRTT exit for rapid convergence (Abreu et al., 25 Oct 2025). BBRv3 achieves or exceeds inter-protocol fairness to Cubic (Jain's index ≈0.95; near-perfect with shallow buffer + ECN), while maintaining high throughput. BBRv2+ (Yang et al., 2021) integrates MinRTT-delta-based (delay-based) probing to guide aggressiveness, multiplexes between BBRv2-mode and its enhanced logic depending on persistent queue state, and adds explicit jitter-driven BDP compensation. Resulting throughput in high-mobility traces exceeds BBRv2 by ≈25% with equivalent queueing delays.

3.3 Delay-BBR

Delay-BBR (Zhang et al., 2019) targets multipath, real-time scenarios. When smoothed RTT exceeds a fraction (β=1.2) of the base RTT, the controller switches to a “drain” phase (pacing_gain=0.75), holding inflight to a BDP-based limit until queues clear. Modified ProbeBW gain cycles ({1.11, 0.9, 1, …}) induce smooth rate evolution. Utility-maximization packet scheduling is integrated on top, selecting per-packet subpaths so as to minimize empirical per-path “costs” (queue, propagation, inverse rate). Experimental benchmarking demonstrates the approach simultaneously reduces delay and loss.

4. Multi-Path, Scheduler-Integrated and Machine Learning-Optimized BBR

4.1 Multi-Path and Multipath Scheduling

MP-DCCP with CCID5 (Moreno et al., 2021) ports BBR into DCCP, exposing fine-grained per-path model variables to the multipath scheduler. This enables aggregation frameworks to react more quickly to per-path bandwidth changes and improves both latency and load balancing. Coupled BBR for MPTCP (Han et al., 2020) employs per-subflow BBR instances, with long-term average pacing rates split according to a weighted coupling law ensuring aggregate throughput does not exceed the best constituent subflow, with path allocation proportional to squared bandwidth share. Companion adaptive schedulers (e.g., AR-based) minimize out-of-ordering and latency, further exploiting BBR’s explicit rate control.

4.2 Machine Learning and Data-Driven Tuning

LLM-driven systematic search produces BBR variants with substantial (up to 27% in QUIC implementation) throughput improvement, primarily via adaptive gain scaling mechanisms and hybrid direct/model-based estimates (He et al., 22 Aug 2025). Patches produced by LLMs often introduce early congestion window reductions linked directly to loss and RTT spikes, providing more immediate responsiveness.

Graph neural networks and classical ML classifiers have additionally been explored to steer phase gain parameters dynamically (Mhaske et al., 2023). Here, periodic feature vectors (block size, throughput, latency) are classified into operational classes guiding pacing gain adjustment, potentially improving fairness in multi-flow contexts.

Reinforcement learning approaches dynamically retune BBR model filter parameters (such as RTprop window and BtlBw sample windows), yielding faster convergence to fair share and lower tail latency in multi-tenant, large-scale deployments (Ketabi et al., 2023). Agents observe network states and learn optimal filter updating strategies to balance throughput and latency trade-offs.

5. Fairness, Stability, and Theoretical Perspectives

5.1 Intra- and Inter-Protocol Fairness

Canonical BBRv1 is unfair to both short-RTT BBR flows and to loss-based flows (Reno, Cubic), especially in shallow buffers and with significant RTT heterogeneity. The fairness index (Jain’s J) can drop as low as ~0.55 when coexisting with Cubic (BBR typically dominates), whereas BBRv2 and BBRPlus, BBQ, FaiRTT, and similar algorithms approach J≈0.95–0.99 (Scherrer et al., 2022, Zhang, 2019, Abrol et al., 2024, Ma et al., 2017). BBRv3 further closes the gap, delivering inter-protocol fairness on par with best-in-class AIMD variants in well-provisioned topologies (Abreu et al., 25 Oct 2025).

5.2 Stability and Oscillation under Competition

Fluid and control-theoretic analyses formalize the (in)stability of BBR when coexisting with CUBIC or under aggressive gain cycles (Scherrer et al., 26 Oct 2025, Scherrer et al., 2022). The oscillation arises from probe cycles and underdamped feedback, causing multi-second rate swings and share fluctuations. BBRv2/v3 address the instability via flattened gain functions, restricted probe amplitude, and tighter window bounds, restoring local stability. Sufficient conditions for oscillation and stability are rigorously derived and validated.

6. Application-Specific Adaptations and Operational Considerations

6.1 Real-Time Video, Data Center, and Wireless/Satellite Scenarios

Delay-BBR and utility-driven scheduling demonstrate the ability to halve end-to-end delays and reduce losses by more than 5× compared to WebRTC-BBR and GCC in multipath real-time transmission (Zhang et al., 2019). In data centers, pure BBR variants are usually superseded by ECN-based schemes (e.g., DCTCP, HPCC) due to even lower achievable tail-latency, but BBRv2/v3 offer competitive Gbps-scale bulk-throughput with moderate queuing (Abreu et al., 25 Oct 2025). Wireless and satellite links expose BBR's limitations when faced with high random loss or frequent RTT spikes, where cross-layer or predictive controllers (QCC, OnlineAEPC) are preferable.

6.2 Performance Trade-offs and Best-Practice Guidelines

Key trade-offs between throughput, latency, fairness, robustness, and loss responsiveness are summarized in deployment studies (Abreu et al., 25 Oct 2025, Zhang, 2019, Ivanov, 2020). For bulk transfers and homogeneous high-BDP environments, BBRv3 is generally optimal; low-latency or legacy domains benefit from AIMD/ECN. Hybrid stacks that dynamically select or blend BBR and classic CCAs by detected workload properties are increasingly prominent.

Variant	Intra-Fairness	RTT-Fairness	Low Latency	Inter-Fairness	Loss Robustness	Responsiveness	Recommended for
BBR v1	✗	✗	◯	✗	✓	✓	Avoid on shared, shallow-buffered
BBRPlus, BBQ, FaiRTT	✓	✓	✓	✓	✓	✓	General internet, fairness needs
BBR v2/v3	✓	◯/✓	◯	✓	✗/✓	◯/✓	Mixed-CCA, Internet default

Legend: ✓=excellent, ◯=moderate, ✗=undesirable.

7. Future Directions and Open Challenges

Further ground remains for robustly resolving RTT-unfairness, especially in dynamic multi-bottleneck or deeply asymmetric environments. Fine-grained integration of learning techniques, further theoretical refinement (Lyapunov-stable controllers), and principled congestion signal fusion (hybrid delay/loss/ECN) remain active research areas. Machine-optimized variants (via LLM or RL-driven evolution) have demonstrated the capacity for rapid, adaptive protocol innovation but require deeper evaluation for fairness and convergence in heterogenous deployments (He et al., 22 Aug 2025, Ketabi et al., 2023).

Emerging use cases, including multipath aggregation, cloud interconnects, and next-generation wireless, are driving further BBR derivatives and architectural experimentation, often with merger of predictive, learning-aided, and explicit scheduling mechanisms to optimize both classic and novel utility functions.

Key references: