FaiRTT Fairness in BBR Control

• FaiRTT is an algorithm family that defines per-flow RTT fairness thresholds to adjust the congestion window, ensuring balanced bandwidth distribution.
• It computes a dynamic multiplier based on real-time RTT measurements to penalize aggressive flows while maintaining network utilization.
• Simulation results indicate FaiRTT nearly eliminates elephant flow dominance, achieving Jain's fairness index up to 0.99 without compromising throughput.

FaiRTT refers to a family of algorithms and frameworks aimed at improving fairness in various computational contexts. Distinct FaiRTT systems have been introduced in networking (specifically TCP congestion control), fairness-aware classification in machine learning, and differentiable fairness regularization for neural models. This entry focuses on the canonical FaiRTT associated with enhancing RTT fairness and bottleneck throughput in BBR congestion control (Abrol et al., 2024), and provides context for the related fairness-centric variants that bear the same or very similar acronym.

1. Motivation: RTT Unfairness in BBR Congestion Control

In packet-switched networks, achieving Round-trip Time (RTT) fairness is required for equitable bandwidth allocation among TCP flows with heterogeneous path delays. Google's Bottleneck Bandwidth and Round-trip Time (BBR) congestion control algorithm estimates the bottleneck bandwidth (BtlBw) and propagation delay (RTprop) to dynamically pace transmissions. BBRv1, however, systematically advantages long-RTT flows ("elephants")—allowing them to monopolize buffer space and starving short-RTT flows ("mice") due to aggressive in-flight inflation (≈1.5×BDP). BBRv2 introduced explicit congestion notification (ECN) handling and refined probing logic but continues to exhibit intra-protocol RTT unfairness, with persistent throughput dominance by elephant flows, especially in high-buffer or large RTT-disparity regimes. Attempts to mitigate unfairness by uniformly lowering pacing rate degrade overall bottleneck utilization (Abrol et al., 2024).

2. FaiRTT Algorithmic Principles and Mathematical Formulation

FaiRTT's central innovation is to embed RTT-sensitivity into the Bandwidth–Delay Product (BDP) estimate for each flow. Rather than using a global BDP, FaiRTT computes a per-flow RTT fairness threshold, $\alpha_t$, and adapts the sending window accordingly.

Key terms and computations:

• $BtlBw_t$: estimated bottleneck bandwidth in window $t$
• $RTprop_t$: estimated propagation RTT in window $t$
• $lastRTT_t$, $minRTT_t$: most recent and per-window minimum RTT samples
• $Wfcount_t$: estimated number of concurrent BBR flows (via unique $minRTT$ samples)
• $\gamma=0.99$: stability discount
• $BtlBw_t$0: smoothing constant for running average of $BtlBw_t$1
• $BtlBw_t$2: phase-dependent multiplier ($BtlBw_t$3 in ProbeRTT, $BtlBw_t$4 otherwise)

FaiRTT's BDP Computation:

$BtlBw_t$5

$BtlBw_t$6

$BtlBw_t$7

$BtlBw_t$8

$BtlBw_t$9

$t$0

This dynamic scaling penalizes flows whose instantaneous RTT exceeds the fairness threshold, curbing over-aggression by elephants while leaving mice unaffected.

3. Algorithmic Flow and Update Logic

On every ACK arrival, FaiRTT executes the following steps (see pseudocode in (Abrol et al., 2024)):

1. Update $t$1 by counting unique $t$2 samples.
2. Compute $t$3 as the flow-specific RTT fairness threshold.
3. If in PROBE_BW and $t$4: set $t$5; adjust $t$6 via $t$7.
4. Set $t$8 using new $t$9 and $RTprop_t$0.
5. Derive the corresponding pacing rate for the next transmission cycle.

This aggressive per-ACK feedback loop, exploiting real-time RTT statistics, distinguishes FaiRTT from naïve rate/pacing cap mechanisms that induce excessive link underutilization.

4. Experimental Evaluation: Topology, Metrics, and Results

Simulation Configuration:

• NS-3 dumbbell: two senders, two receivers, shared 10 Mbps bottleneck, Drop-Tail, 1 kB packets.
• Default: queue size = 10×BDP, tests cover 0.5–100×BDP.
• Flows: "elephant" (default RTT = 15 ms) versus "mice" (RTT = 5 ms), RTT sweeps cover 5–30 ms.
• All results are averaged across 5 trials per scenario, reporting 95% confidence intervals.

Performance Metrics:

Metric	BBRv2	FaiRTT
Avg. throughput ratio (elephant/mice) [queue=10 BDP]	1.44	1.05
Throughput ratio (queue size sweep)	>1.4 (large queues)	≈1.04 (stable, 2–100)
Throughput ratio (RTT sweep, elephant 5–30 ms)	1.32–1.82	1.02–1.18 (avg 1.11)
Jain's fairness index (RTT sweep)	0.94	0.98
Jain's fairness index (overall avg)	0.95	0.98–0.99
Bottleneck link utilization	97.21%	98.78%

These results demonstrate nearly ideal fairness and eliminate consistent elephant dominance without degrading — indeed, slightly improving — link utilization.

5. Trade-offs, Limitations, and Practical Implications

Trade-offs:

• FaiRTT adds modest per-ACK computation (for $RTprop_t$1, $RTprop_t$2, running averages).
• Introduction of the dynamic $RTprop_t$3 only mildly slows elephant flows; capacity is reallocated to mice flows, so aggregate throughput is maintained or increased.
• No extra network state or modifications to routers required.

Limitations:

• Robustness requires accurate RTT measurement. Packet reordering or jitter can degrade $RTprop_t$4/$RTprop_t$5 fidelity, leading to suboptimal scaling.
• In severe congestion, classical ECN/loss signals will still dictate fundamental backoff (FaiRTT does not override BBR’s loss/ECN framework).
• The method's fairness gains depend on the presence of moderate to large buffers and significant RTT heterogeneity.

Best-case deployment scenarios:

• Heterogeneous flows with disparate RTTs and moderate/large bottleneck buffers.
• Next-generation (B5G) network fabric, e.g., AR/VR telepresence or telemetry with concurrent high-bandwidth persistent traffic and short RT signaling.

6. FaiRTT Nomenclature in Other Fairness Contexts

Variants bearing similar acronyms operate in distinct domains:

• FairTTTS: A decision-tree–specific post-processing method that increases the probability of flipping traversal direction at protected-attribute nodes for unprivileged groups. Unlike threshold-shifting baselines, FairTTTS applies a distance-based heuristic within tree internals, yielding reductions in Equalized Odds Difference (EOD) and Disparate Impact (DI), often improving accuracy (+0.55% over baselines) (Cohen-Inger et al., 14 Jan 2025).
• fairret: A differentiable fairness regularization toolkit for neural models, where fairness constraints—expressed as linear-fractional statistics—are incorporated into the loss via violation or projection-based strict regularizers. This framework achieves near-zero fairness violation for linear notions (Demographic Parity, Equal Opportunity) with minor predictive performance loss (Buyl et al., 2023).

7. Summary and Broader Context

FaiRTT as introduced for BBR congestion control (Abrol et al., 2024) provides an RTT-sensitive in-flight window adjustment mechanism, resolving intra-protocol RTT unfairness in scenarios with diverse flow RTTs and non-negligible buffering. The algorithm is strictly empirical, leveraging real-time RTT measurements to dynamically adjust aggressiveness, tightly bounding throughput disparity while preserving or boosting bottleneck utilization. Extensions of FaiRTT’s underlying philosophy to decision trees (FairTTTS) and neural nets (fairret/fairrets) indicate a broader trend: fairness-aware post-processing and plug-in regularization are becoming practical, measurable, and hybridizable with high-accuracy state-of-the-art models.