Analysis • Transport behavior on long‑RTT paths

Reno vs Cubic vs BBR on high‑latency satellite links

Concrete observation: on a geostationary‑style satellite hop the round‑trip time (RTT) dominates transport responsiveness. That long control loop makes loss‑based reactions, probing patterns, and model assumptions diverge in predictable ways — this page maps those differences to a simple timeline of events so network engineers can reason about throughput, fairness, burstiness, and recovery without operational tuning steps.

Mechanism primer

loss‑based: Control by packet loss events (or ECN): window growth until loss, then multiplicative decrease. Examples in this page use the family labels Reno and Cubic.
model‑based: Attempts to maintain operating point by estimating bottleneck bandwidth and minimum RTT, then probing with limited, paced phases — represented here by BBR.
delay/RTT effects: High RTT amplifies the time between a control action and its feedback; bandwidth‑delay product (BDP) grows, making buffer interactions and burstiness more consequential.

Timeline legend: small right‑margin nodes mark events: slow‑start, loss, steady, path‑change, recovery.

Reno‑style loss‑based TCP

Claim: reactive growth and multiplicative decrease make Reno simple and predictable, but long RTTs cause slow recovery and pronounced throughput swings after loss.

Annotation: growth is exponential until loss detection; multiplicative decrease applied at high RTTs means many round‑trips are needed to regain prior window.

Counterexample/limitation: if the satellite link has large buffers, loss may be deferred until a large queue forms; Reno then reacts late and leaves queues high for long periods while recovery is slow.

Takeaway: predictable but slow to recover on long‑RTT links.

Cubic (modern loss‑based variant)

Claim: Cubic's cubic growth shape accelerates probing after a loss epoch and is optimized for high BDP; on satellite links it can probe more assertively than Reno but may generate bursts.

Annotation: cubic growth aims to close the gap to prior peak faster; on large‑BDP satellite hops that can translate into more aggressive bursts when probing.

Counterexample/limitation: in the presence of small policing queues or strict AQM, Cubic's bursts can be clipped and result in repeated losses instead of stable ramping.

Takeaway: faster peak recovery than Reno but higher burstiness risk on buffered links.

BBR‑like model‑based control

Claim: BBR tries to lock to estimated bottleneck bandwidth and minimum RTT, pacing packets; this removes loss as a primary control signal but depends on stable measurement windows — a challenge on high‑RTT variable satellite links.

Annotation: pacing and bandwidth estimation aim for low queuing; periodic probe phases increase sending rate briefly to test capacity rather than relying on loss signals.

Counterexample/limitation: when measurement windows are dominated by long RTTs or when path capacity fluctuates within a single RTT, bandwidth estimates lag and probes can either underutilize or briefly overshoot the link.

Takeaway: can reduce average queuing and latency but depends on accurate bandwidth/RTT estimates that are harder to obtain on long‑RTT variable paths.

Comparative synthesis

Qualitative comparison keyed to the timeline events above. Cells represent directional tendencies, not measured values.

Property	Reno	Cubic	BBR
Steady‑state throughput stability	Moderate; oscillatory after loss	Higher peak recovery; more aggressive closing to prior peak	Stable when estimates hold; can be conservative during uncertainty
Latency sensitivity / queuing	Tends to inflate queues after growth	Similar or higher queuing during probes	Lower queuing when pacing works as intended
Fairness vs loss‑based flows	Wins or shares with similar loss‑based peers	May take more share vs Reno in BDP scenarios	Can starve or be starved depending on estimator interactions
Reaction to transient loss	Full multiplicative decrease; slow recovery	Less time to previous peak but still loss‑triggered	May not reduce rate; relies on model to avoid loss
Buffering interaction	High risk of long queues on large BDP links	Bursts can cause clipping and repeated losses	Lower queue footprint if pacing and measurement are reliable

Evidence orientation: the entries summarize expected behavioral tendencies derived from control‑loop design differences and reported observational patterns; they do not assert universal numeric outcomes.

Design implications (selection considerations)

When fairness matters: Consider whether the environment will host mixed loss‑based and model‑based flows; fairness goals should guide policy at the link rather than presuming one transport will behave fairly across classes.
When latency matters: Model‑based pacing (BBR) tends to keep queues smaller when its measurements are stable, which favors latency‑sensitive traffic — but confirm estimator robustness in long‑RTT, variable capacity conditions before relying on it.
When buffer policies dominate: Link‑layer buffering and AQM shape observed behavior more than the choice of algorithm alone: buffer size and queue management can convert gentle probes into large losses on long‑RTT links.
When recovery speed is critical: Cubic's aggressive peak probing can regain throughput faster than Reno; however, faster recovery can come with increased burstiness and loss risk.

Operational exclusions & caveats

This analysis is comparative and non‑operational. It intentionally does not provide tuning commands, implementation steps, or configuration recipes for TCP stacks, routers, or satellite equipment.

Do not interpret qualitative tendencies as guaranteed numerical performance.
Observational pitfalls: single‑trial traces, unreported middleboxes, or local scheduling artefacts can mislead inference.
Policy decisions should be validated with controlled measurements and local safety reviews — not by following prescriptive tuning from this page.

Closing checklist

Have you identified whether fairness with legacy loss‑based flows or low latency is the primary objective?
Does your link have large BDP and buffering that will amplify burstiness? If so, prioritize buffer policy review.
Collect paced and unpaced traces across representative RTTs before committing to a transport‑level choice.

For further reading, consult standards and RFCs describing congestion control designs and measurement methodology; this page intentionally avoids operational rules and exact configuration examples.