1
Unveiling TCP BBR Dominance in Starlink Internet: Experimental Insights and Analysis Rakshitha De Silva
, Shiva Raj Pokhrel
Senior Member, IEEE and Jonathan Kua
arXiv:2607.07133v1 [cs.NI] 8 Jul 2026
Abstract—This experimental study delivers a global assessment of Google’s Bottleneck Bandwidth and Round-trip propagation time-version 3 (BBR-v3) Congestion Control Algorithm (CCA) over SpaceX’s Starlink network.Leveraging a strategically deployed six-city testbed across five continents, we systematically benchmark BBR-v3 against eight CCAs: Cubic, Hybla, Vegas, LeoCC, Copa, PCC, BBR-v1, and BBR-v2 under both dedicated and concurrent conditions. Our results demonstrate that BBRv3’s advantage is not aggressive bandwidth capture, but a more balanced fairness, loss, and delay trade-off over the Starlink Internet. We develop pragmatic mathematical models that capture Starlink’s complex network dynamics and characterize BBRv3 behavior to better explain the experimental observations. Our extensive evaluation of queue buildup and fairness further demonstrates BBR-v3’s capability to maximize throughput in high-latency, variable satellite environments, while maintaining a balance between aggressiveness and fairness. The findings establish BBR-v3 as a compelling CCA for Low Earth Orbit (LEO) satellite networks and provide a principled analytical foundation for next generation satellite Internet transport design. Index Terms—Bottleneck Bandwidth and Round-trip propagation time (BBR), Network measurements, Starlink, Transmission Control Protocol (TCP)
I. I NTRODUCTION
Low Earth Orbit (LEO) satellites have emerged as a cornerstone of next-generation global communications, providing wide-area broadband connectivity with significantly reduced latency compared to traditional Geostationary Equatorial Orbit (GEO) systems. Orbiting at altitudes below 2000 km, LEO mega-constellations bridge the digital divide by extending Internet access to underserved regions while offering competitive performance to terrestrial infrastructure. Among these, SpaceX’s Starlink represents the most mature deployment. Operating over 10,000 satellites across multiple orbital shells at ≈550 km altitude as of May 2026 and serving 2.7 million subscribers with 5 Tbps weekly capacity expansion [1], [2]. Its architecture integrates phased-array Ku-band terminals, Kaband feeder links, and laser Inter-Satellite Links (ISLs), enabling global connectivity even in gateway-sparse regions and transforming LEO constellations from supplemental backhaul into global Internet Service Providers (ISPs). This work is supported by SmartSat CRC, whose activities are funded by the Australian Government’s CRC Program. Authors are with the IoT & Software Engineering Research Lab, School of Information Technology, Deakin University, Geelong, VIC 3125, Australia (e-mail: [email protected]; [email protected]; [email protected]).
Member, IEEE
Fig. 1: A simplified illustration of our global experimental testbed over Starlink Internet. Google’s Bottleneck Bandwidth and Round-trip propagation time (BBR), introduced in 2017, represents a paradigm shift in Transmission Control Protocol (TCP) Congestion Control Algorithms (CCAs) by explicitly modeling bottleneck bandwidth and propagation delay rather than relying on lossbased congestion detection. While BBR version 1 (BBR-v1) demonstrated significant throughput improvements, it revealed fairness and efficiency challenges, particularly persistent queue buildup. BBR version 2 (BBR-v2) addressed these issues but continued to struggle in heterogeneous environments [3], [4]. The latest iteration, BBR version 3 (BBR-v3), refines probing and pacing strategies to optimize throughput, latency, and fairness across diverse network conditions, with ongoing research characterizing its behavior in various deployment scenarios [5]. The convergence of Starlink’s global LEO network with BBR’s model-based congestion control presents significant yet underexplored potential. Existing research on BBR over LEO networks primarily relies on simulations [6] or singleconnection analyses predating BBR-v3 [7], [8]. Studies using actual Starlink data have focused on video streaming application-level performance [9], [10], rather than BBR’s network-level performance. Furthermore, most work on TCP over LEO research does not cover physical layer evaluation, critical to transport layer performance. Addressing these gaps, this work makes the following contributions: • A decomposition of the Starlink network into user link, Inter-Satellite Links (ISLs), and feeder link components with mathematical models yielding end-to-end transmission path failure probability.
2
Design and implementation of a global Starlink Internet performance testbed across six cities: Ohio, São Paulo, London, Mumbai, Tokyo, and Sydney. • Extensive analyses of BBR-v3 performance comparison against eight CCAs: Cubic [11], Vegas [12], Hybla [13], LEO Network Congestion Control (LeoCC) [14], Copa [15], Performance-oriented Congestion Control (PCC) [16], BBR-v1, and BBR-v2. The evaluation covers four primary network conditions: dedicated uplink and downlink, inter-CCA concurrent uplink and downlink. • Present a fluid model of BBR-v3 congestion window modeling yielding end-to-end transmission path and transmission failure probability that capture Starlink LEO dynamics and explain experimental observations. • Queuing analysis and fairness evaluation of BBR-v3 encapsulating the captured data over the globally distributed testbed.
•
The remainder of this paper is organized as follows: Section II provides an overview of the Starlink network characteristics and derives a mathematical model of the end-toend link failure probability. Section III details the globally distributed testbed implementations and provides an evaluation of the collected results. Section IV presents the fluid model, queuing analysis, and provides a fairness evaluation of BBRv3 encapsulating the collected data, and Section V concludes the article. II. S TARLINK I NTERNET: S YSTEM AND S ETTINGS A. Overview of the Starlink Network As of May 2026, there are 10,408 active Starlink satellites in orbit, making SpaceX’s mega-LEO constellation by far the largest of its kind [2]. These satellites are categorized into four variants: V1, V1.5, V2-KU (utilizing Ku-band), and V2DTC, a specialized version of Direct to Cell (DTC) designed for direct smartphone connectivity [1]. According to Federal Communications Commission (FCC) filings, the constellation is structured into five distinct orbital shells with different orbital lanes and inclination angles [17]. This architecture ensures high satellite availability, with over 20 satellites simultaneously visible in densely populated mid-latitude regions upon full deployment [18]. Notably, Starlink deviates from the traditional Walker constellation designed to facilitate seamless coverage through its multi-shell architecture while maintaining consistent ground tracks [19]. Starlink’s network topology leverages optical ISLs (colloquially “space lasers”) that enable service provision without requiring gateway presence in the satellite’s coverage area. Each satellite incorporates three 200 Gbps optical ISLs, collectively forming a global mesh network [1]. This architecture significantly enhances coverage over oceanic and remote regions while reducing ground station infrastructure requirements. The globally distributed gateway network operates in Ka-band, with each gateway capable of simultaneously connecting to four satellites. Gateway downlinks utilize nine 250 MHz channels
Fig. 2: Traffic flow through “bent-pipes” in the Starlink network.
(17.8-19.3 GHz), while uplinks employ eight 500 MHz channels (27.5-30.0 GHz) [18]. User terminals (“Dishy”) employ ◦ Ku-band transmission, connecting to LEOs visible above 25 elevation angle. Through multi-beam antenna technology, a single satellite can simultaneously service multiple users. As illustrated in Fig. 2, user data traverses a ”bent-pipe” connection across ISLs [9], [20]. The Australian Communications and Media Authority (ACMA) has confirmed Starlink’s International Telecommunication Union (ITU) frequency allocations in Australia spanning 10.7–12.7 GHz, 14–14.5 GHz, 17.8–18.55 GHz, 18.8–19.3 GHz, 27.5–29.1 GHz, and 29.5–30 GHz [21]. The ground terminals leverage phased-array antennas, comprising a grid of smaller antennas that enable electronic beam steering through differential phase manipulation. Each terminal contains five Ku-band phased array antennas and three dualband antennas (Ku and E bands) for user connectivity. To counteract Signal to Noise Ratio (SNR) variations caused by orbital dynamics, handovers, and environmental factors, Starlink dynamically adjusts modulation and coding schemes [22]. User downlinks operate across eight 250 MHz channels (each with 10 MHz guard bands), while uplinks utilize four 125 MHz channels [18], [22]. The transmission structure employs time division multiplexing with a 1/750 s frame period subdivided into 4.4 µs frames and guard intervals, with frame headers containing satellite, channel, and modulation information [22]. The Round Trip Time (RTT) traces captured using irtt for two internet paths presented in Fig. 3 exhibit a unique characteristic of the Starlink constellation. The repeated steplike RTT shifts are experienced close to 15 s boundaries, where each interval tends to remain within a relatively stable regime before transitioning to a different level in the subsequent window. This behavior is a result of planned “connection handoffs” between the Dishy and the LEO satellite to maintain a seamless connectivity amid the constellation dynamics, which is further detailed in Starlink’s FCC filing [23]. B. End-to-End Link Failure and Packet Drop Probability As illustrated in Fig. 2, an end-to-end Starlink transmission path may consist of three main wireless components subject to uncertainty: the gateway-to-LEO link, ISLs, and the LEO-toground-terminal link. The achievable end-to-end capacity can
3
(a) Sydney to Melbourne
Fig. 4: Globally distributed server locations (b) São Paulo to Melbourne
A. Testbed Setup
Fig. 3: RTT shift at 15 s boundaries in Starlink network
therefore be affected by propagation latency variations due to orbital dynamics, Adaptive Coding and Modulation (ACM)driven changes in response to time-varying SNR, satellite handovers, and bandwidth contention across shared ISLs [24]– [26]. Accounting for these factors and the distinct transmission legs of the Starlink link, we present an end-to-end link failure and packet drop probability model. The model encapsulates four key impairment terms, the packet drop probability due cap to LEO satellite to ground terminal capacity limitations (p ) [27], [28], the failure probability due to atmospheric attenuation at ho (p ) [29], [30], the handover drop probability (p ) [23], [31], isl and the ISL failure probability (pi ) [32]. The detailed analysis is provided under Appendix A. The probability of link failure causing packet loss in the tot transmission path (p ) can be given as: N
p
tot
= 1−(1−p
cap
)⋅(1−p
gw
isl
)⋅(1−p )⋅(1−p )⋅ ∏(1−pi ) at
ho
isl
i=0
(1) isl where N is the number of ISLs that a given packet is routed gw at through in the constellation, p = p (gw), and the other parameters represent the respective failure probabilities [28]. For simplicity, we assume that gateway links are negligibly affected by handover losses. If the number of ISLs in a isl given extended bent pipe over the Starlink network (N ) is deterministic, Eq. (1) becomes: p
tot
= 1−(1−p
cap
)⋅(1−p
gw
)⋅(1−p )⋅(1−p )⋅(1−p ) at
ho
isl N
isl
(2)
III. T ESTBED S ETUP AND CCA P ERFORMANCE OVER S TARLINK This section presents our globally distributed testbed implementation across six test sites and compares the performance of BBR-v3 against eight CCAs: Cubic [11], Vegas [12], Hybla [13], LeoCC [14], Copa [15], PCC [16], BBR-v1, and BBRv2 using it over the Starlink network.
As illustrated in Fig. 4, we set up Linux servers in six main cities distributed globally, namely Ohio, São Paulo, London, Mumbai, Tokyo, and Sydney. We leveraged Amazon Web Services (AWS) Elastic Compute Cloud (EC2) for this distributed server setup, and the local portion was hosted on the university premises in Burwood, Melbourne, Australia. AWS EC2 instances were set up with Ubuntu 24.04.4 LTS Linux distribution, and the default free-tier settings were used, apart from the necessary port security group configurations. Furthermore, we assume the AWS cloud instances to provide a consistent network connection throughout the geographically distributed cloud instances. The Starlink user terminal consists of the latest-generation standard kit with a UTA-232 model dish. The local Linux server (Ubuntu 24.04.4 LTS) was connected to the Starlink terminal via a Category 6 Ethernet connection, creating a continuous test environment. To test TCP CCA performance over the testbed, we leverage iperf3, an open-source network testing tool used to measure the maximum achievable bandwidth and performance over network connections. It provides detailed metrics such as throughput, packet loss, jitter, and retransmissions, making it widely used for network diagnostics and benchmarking. To identify the Starlink Point of Presence (PoP) associated with our connection, we performed repeated logging and DNS resolving during active flows to multiple AWS regions, and the reverse DNS consistently resolved to customer.mlbeaus1.isp.starlink.com, indicating a stable Starlink PoP association in the Melbourne region during the tests rather than destination-dependent PoP switching. We implemented an automated measurement framework using iperf3, in which each remote EC2 server and the local host were configured with the required CCA. We define tests with a single active TCP flow as the dedicated flow scenario, in which forward and reverse iperf3 tests were conducted to collect dedicated downlink and uplink measurements for each CCA. To evaluate inter-CCA coexistence under sharedpath contention, we define the concurrent flow scenario. In this test, multiple TCP flows run simultaneously over the same Starlink path, with each flow configured to use a different CCA. Under concurrent flows, nine simultaneous TCP flows were generated, corresponding to Cubic, Hybla, Vegas, LeoCC,
4
Copa, PCC, BBR-v1, BBR-v2, and BBR-v3. Each flow was isolated in a separate Linux network namespace, allowing its corresponding CCA to be configured independently while sharing the same end-to-end Starlink path. Special care was taken to synchronize server startup, ensuring data flushing before retrieval, and preserve namespace-level isolation across CCAs. This enabled fair and repeatable benchmarking under Starlink network conditions. For each CCA, destination, direction, and scenario combination, we collected 10 independent runs, each with a 300-second capture window. The full measurement campaign was carried out during April 2026. B. Evaluation of Downlink Streams 1) Dedicated Downlink Streams: The dedicated-flow downlink results in Fig. 5 show that BBR-v3 provides a strong but not universally dominant operating point over Starlink. As shown in Fig. 5(a), BBR-v3 achieves consistently high median throughput across all six locations, with values of 176.11 Mbps in Tokyo, 134.11 Mbps in São Paulo, 180.37 Mbps in Ohio, 157.14 Mbps in London, 197.21 Mbps in Mumbai, and 247.56 Mbps in Sydney. This is substantially higher than Cubic, Hybla, Vegas, and Copa in most locations, whose rates remain limited or highly variable. However, LeoCC, PCC, BBR-v1, and BBR-v2 exceed BBR-v3 in some paths, indicating that more aggressive and LEO-aware schemes can capture additional capacity under favorable Starlink conditions. This higher capacity capture is accompanied by higher retransmission counts and larger congestion windows, as presented in Fig. 5(b) and Fig. 5(c). Interestingly, BBR-v3 maintains a more controlled congestion-window profile while avoiding severe retransmission penalties. The RTT distributions in Fig. 5(d) reflect the geographic path differences across the six endpoints, and highlight BBR-v3’s RTT stability over Starlink Internet. The receiver-advertised window and RTT-variance associated with these captures are presented in Fig. 17, and the summary of the data captured from 10 dedicated download tests is presented in Fig. 21 under Appendix B. Those results further indicate that BBR-v3 offers a favorable throughput, loss, and delay trade-off behavior over the Starlink Internet in comparison to the evaluated CCAs. 2) Concurrent Downlink Streams: The concurrent downlink results in Fig. 6 show a more contested operating regime than the dedicated downlink case, as all nine CCAs simultaneously share the same Starlink path. In this setting, BBR-v3 remains competitive but does not dominate the bandwidth allocation. As shown in Fig. 6(a), BBR-v3 records median throughput values of 10.49 Mbps in Tokyo, 34.58 Mbps in São Paulo, 47.17 Mbps in Ohio, 46.11 Mbps in London, 11.54 Mbps in Mumbai, and 62.87 Mbps in Sydney. These values are generally higher than the conservative Cubic, Hybla, Vegas, and Copa flows, but lower than LeoCC and BBR-v1, indicating that more aggressive or LEO-aware schemes capture a larger share of the available downlink capacity under concurrent contention. This behavior is reflected in Fig. 6(c), where LeoCC, BBR-v1, and PCC often maintain larger congestion
windows, while BBR-v3 operates with a more restrained inflight volume. The retransmission results in Fig. 6(b) show the cost of this aggressive capacity capture: LeoCC and BBRv1 incur substantially higher retransmission counts in several paths, especially Tokyo, whereas BBR-v3 maintains moderate retransmissions and avoids a severe loss behavior. The RTT distributions in Fig. 6(d) continue to reflect the geographic path differences, while also highlighting that aggressive CCAs can result in increased delay. The receiver-advertised window and RTT-variance results for the same concurrent downlink captures are provided in Fig. 18 in Appendix B, together with the 10-run summary in Fig. 21. The results indicate that BBRv3 sacrifices some bandwidth share compared with the most aggressive contenders, but provides a more controlled loss and delay profile under heterogeneous CCA competition. C. Evaluation of Uplink Streams 1) Dedicated Uplink Streams: The dedicated-flow uplink results in Fig. 7 show that the Starlink uplink presents a more capacity-constrained operating regime than the downlink case. This is primarily due to the limited bandwidth availability on the Starlink uplink. As shown in Fig. 7(a), BBR-v3 achieves strong median uplink throughput in Tokyo, Ohio, London, and Mumbai, with values close to 44 Mbps, while recording a lower median, but a higher lower quartile in Sydney. This indicates that BBR-v3 can efficiently utilize the available uplink capacity when the path is stable, but its performance remains sensitive to path-specific Starlink Internet conditions. LeoCC, PCC, and BBR-v1 achieve comparable or higher throughput in several locations, but this is accompanied by larger congestion windows and higher retransmission counts, as shown in Fig. 7(c) and Fig. 7(b), respectively. In contrast, Cubic, Hybla, Vegas, and Copa generally remain more conservative, producing lower or less stable uplink throughput. The RTT results in Fig. 7(d) again reflect the geographic path differences, with Sydney showing the lowest delay and São Paulo the highest. These dedicated-flow uplink results show that BBR-v3 provides a competitive uplink operating point, balancing throughput with more controlled retransmission and congestion-window behavior compared with the most aggressive CCAs. The receiveradvertised window and RTT-variance results associated with these captures are provided in Fig. 19, and the 10-run summary is presented in Fig. 22 under Appendix B. 2) Concurrent Streams: The concurrent uplink results in Fig. 8 show that simultaneous CCA competition over the Starlink uplink creates a strongly capacity-constrained regime, where most algorithms experience intermittent or low median throughput. As shown in Fig. 8(a), BBR-v3 achieves non-zero median throughput in Tokyo, Ohio, Mumbai, and Sydney, but drops to near-zero median throughput in São Paulo and London. Underlining that its uplink performance under contention is sensitive to path-specific capacity variation and competingflow pressure. BBR-v2 and LeoCC achieve higher medians in selected locations, while Cubic, Vegas, Copa, and PCC remain limited or highly bursty in most cases. The retransmission and
5
Fig. 5: Summarized downlink observations over Starlink with dedicated CCAs for globally distributed locations
Fig. 6: Summarized downlink observations over Starlink with concurrent CCAs for globally distributed locations
congestion-window results in Fig. 8(b) and Fig. 8(c) show that the algorithms that capture more uplink capacity often do so with higher retransmission counts or larger in-flight volumes. Whereas BBR-v3 generally maintains a more restrained
congestion-window profile and avoids the most severe loss behavior. The RTT distributions in Fig. 8(d) mainly follow the geographic path differences, while also reflecting additional delay variation introduced to aggressive CCAs by simultaneous
6
Fig. 7: Summarized uplink observations over Starlink with dedicated CCAs for globally distributed locations
Fig. 8: Summarized downlink observations over Starlink with concurrent CCAs for globally distributed locations
uplink contention. The receiver-advertised window and RTTvariance associated with these concurrent uplink captures are presented in Fig. 20, and the 10-run summary is provided in Fig. 22 under Appendix B. To this end, the concurrent uplink results further highlight the fact that BBR-v3 does not
consistently dominate over Starlink Internet, but provides a controlled operating point with moderate throughput, restrained congestion-window growth, and lower loss exposure than the most aggressive competing alternatives.
7
Fig. 9: Throughput vs RTT inflation of dedicated transmissions.
D. TCP CCA Operating Regimes Fig. 9 summarizes the dedicated-flow behavior by jointly presenting mean throughput, RTT inflation, and retransmissions per GB evaluation, encapsulating the data collected over the total runs. The results show that model-based or capacityseeking schemes, including BBR-v3, LeoCC, PCC, BBR-v1, and BBR-v2, achieve substantially higher throughput than conservative loss- or delay-sensitive schemes such as Cubic, Hybla, Vegas, and Copa. However, this higher capacity capture often comes with increased retransmission cost, particularly for PCC, LeoCC, and BBR-v1. However, in the downlink, BBR-v3 lies in the high-throughput region while maintaining lower retransmissions than the most aggressive alternatives, and its predecessors. In the uplink, where the Starlink path is more capacity constrained, BBR-v3 remains competitive while avoiding the severe retransmission overhead observed for PCC, LeoCC, and BBR-v1. To get a better perspective of the Starlink constraints, we have presented a terrestrial network throughput comparison of the downlink and uplink paths in Appendix B, Fig 23. The results further strengthen that BBR-v3 provides a balanced dedicated-flow operating point over Starlink, sustaining high utilization without excessive loss penalties. On the other hand, Fig. 10 summarizes the concurrent-flow behavior by comparing throughput share, RTT inflation, and retransmissions per GB against the ideal fair-share line. The results show that Starlink contention separates the CCAs into three operating regimes. Conservative schemes such as Cubic, Hybla, and Vegas remain below the fair-share line in both
Fig. 10: Throughput share vs RTT inflation of concurrent transmissions. downlink and uplink because their loss or delay-sensitive responses limit their ability to compete under Starlink’s dynamic network conditions. In contrast, LeoCC, PCC, and BBR-v1 capture a larger share of the bottleneck capacity, especially in downlink, but this comes with substantially higher RTT inflation and retransmission intensity, indicating aggressive bandwidth acquisition with increased loss exposure. BBRv2 and BBR-v3 operate closer to the fair-share region and with moderate RTT inflation in both downlink and uplink. Importantly, BBR-v3 maintains much lower retransmissions per GB than LeoCC, PCC, and BBR-v1, while avoiding the severe throughput starvation observed for conservative CCAs. Thus, under concurrent-flow operation, BBR-v3’s advantage is not maximum bandwidth capture, but a more balanced fairness, loss, and delay trade-off over the Starlink Internet. IV. M ODELING BBR- V 3 S TARLINK I NTERNET A. BBR-v3 Fluid Model min
BBR-v3 probing occurs every 62 ⋅ RT Ti for low RTT flows, while for high RTT flows, it happens at randomly selected intervals between 2 and 3 seconds. To eliminate stochasticity, let us define the time interval between sequential ProbeBW events for any flow i ∈ {1, . . . , N } in a parallel set of N BBR flows as [3]: i pbw rtp t̄ = min (62 ⋅ RTT , 2 + ) (3) N where RTT RTT.
rtp
represents the estimated minimum propagation
8
We introduce the indicator variable Idwn as follows, which equals 1 when the BBR flow attempts to reduce its inflight data: dwn
I
={
¯ ∨ ptot (t) > 0.02, 1, v(t) > 54 BDP 0, otherwise,
(4)
where v(t) represents the inflight volume, which evolves according to sender injection and network delivery rates, and ¯ is the estimated Bandwidth-Delay Product (BDP). ptot BDP is the end-to-end probability of packet drop due to link failure crs defined in Eq. (2). Similarly, I indicates whether the flow is in cruising state: I
crs
={
1, 0,
¯ , v(t) ≤ BDP otherwise,
(5)
The pacing rate is modeled as: pbw
x
pcg
= x (1 + btl
σ(ti
− RTT )(1 − I 4 rtp
dwn
)−I
dwn
)
(6)
Fig. 11: Visualization of BBR fluid model inflight limits
x(t)RTT(t), encapsulating a measurement-based estimate of inflight, where x(t) is the measured delivery rate. To obtain a tractable model, we approximate the dynamics using a firstorder relaxation: dvmodel (t) = λ (vtarget (t) − vmodel (t)) , dt
(12)
where vtarget (t) is determined by the ProbeBW state and the inflight bounds derived above. Figure 11 provides a validation of the proposed BBRv3 fluid model using a Sydney to local terminal downlink data capture. The long-term inflight bound hi BDP (t) is derived from loss events inferred from retransmissions, enabling direct comparison between the measured inflight vactual (t) and the model-predicted inflight vmodel (t) under network-driven constraints. In the probing phase (t ≈ 0–20 s), the measured in¯ Preflight vactual (t) exhibits large excursions above 1 × BDP. dicted inflight vmodel (t) follows the upper envelope defined by dwn crs dwn pbw rtp hi ∆ = (1 − I )(1 − I ) σ(t − RTT ) BDP (t), capturing the dominant probing behavior of BBRv3. hi ¯ Around t ≈ 20 s, both vactual (t) and BDP (t) show a sharp de5BDP tot ) + σ(p − 0.02), 1) × min (σ (v − crease due to a change in network conditions. The model repro4 hi duces this transition through loss-driven updates of BDP (t), dwn ˆ − v) − I σ(BDP (7) demonstrating its ability to capture time-varying path effects. dwn crs Deactivation of I directly activates I , which is disabled in For t ≳ 25 s, both measured and modeled inflighthi converge toward the headroom-controlled region (1 − β)BDP (t); (β = ProbeBW states: 0.15 = queue draining margin/safety headroom) [5] and repbw crs crs dwn pbw (8) main close to 1 × BDP, indicating consistent steady-state ∆ = −∆ − σ(t − t̄i )Ii behavior. We conjecture that the short-term spikes in vactual (t) The ProbeBW congestion window is modeled as [3], [5]: arise from transient effects such as RTT variability and ACK ⎧ crs ¯ BDPlo ), ⎪ min (2BDP, I = 1 (inflight_lo), compression, which are not explicitly modeled. ⎪ pbw ⎪ w =⎨ ⎪ ¯ BDPhi ), Icrs = 0 (inflight_hi), B. Measurement-Based Interpretation of BBR-v3 Dynamics ⎪ min (2.5BDP, ⎪ ⎩ (9) BBR continuously estimates maximum delivery rate (Bθ ) while in ProbeRTT state, it is: and minimum RTT (RTTmin ), setting its sending rate as ¯ BDP Rsending = G ⋅ Bθ , where G is a phase-dependent gain factor. prt w = (10) We present a validation of the analysis using Sydney captures, 2 encapsulating these properties. A BBR reaction window is The inflight dynamics can be defined as: identified if an RTT spike or bandwidth drop is followed dv(t) pcg = x (t) − x(t), (11) by a congestion window drop within the reaction window. dt Measurement-based validation of the proposed Starlink–BBRpcg where x (t) (Eq. 6) and x(t) is the network delivery v3 coupled model using Sydney captures is illustrated in Fig. 12 rate. Under quasi-steady conditions, this yields: vactual (t) ≈ for both downlink and uplink. btl
where x represents the estimated bottleneck bandwidth. dwn btl When I = 0, the pacing rate increases to 54 ⋅ x ; otherwise, btl it reduces to 34 ⋅ x . dwn Phase transitions are triggered by probing observations. I tot 5 ¯ activates when v(t) > 4 ⋅ BDP ∨ p (t) > 0.02, and deactivates once inflight data drops to the conservative target ˆ = min (BDP, ¯ 0.85 BDPhi ), where BDPhi represents the BDP inflight_hi long-term upper bound. This ensures queue clearance without pipe under utilization. The transition dynamics are modeled as:
9
indicate temporary changes in propagation distance or queuing delay. Since the sender generates a continuous load, the measured throughput reflects the instantaneous effective capacity: Ceff (t) = min[C(t), CISL (t)](1 − p
tot
(t))
(14)
tot
where CISL is the ISL capacity and p (Eq.(2)) is the path uncertainty probability (See Appendix A for further information on Starlink capacity modeling). Therefore, the retransmission tot bursts shown in Fig. 12 are consistent with increases in p (t) or queue overflow. Intervals, where RTT excursions coincide with throughput degradation, are classified as bandwidth perturbation events, which are highlighted in Fig. 12 respective subplots. Under the proposed fluid model, the behavior observed in pbw dwn Fig. 12 can be interpreted through the state variables t̄ , I , crs and I . The RTT excursions and retransmission bursts indicate intervals in which the Starlink path departs from its nominal operating point, corresponding to conditions where either the ¯ or the end-to-end packet-drop inflight process exceeds 54 BDP probability becomes significant. These conditions correspond dwn exactly to the scenarios in which I is activated within the dwn fluid model. Once I = 1, the pacing rule (Eq.(6)) reduces pcg x from its probing level toward the conservative level, which appears as throughput degradation during disturbed intervals. The congestion-window variations in Fig. 12 provide the pbw measurement-side manifestation of the inflight limits w and prt crs w . When the path remains stable and I = 1, the sender operates near the cruising/probing envelope, when RTT inflation and retransmission bursts occur, the transition dynamics dwn crs ∆ and ∆ drive the flow away from cruising and toward a reduced inflight state. Thus, producing the repeated congestion window contractions marked as reaction intervals. Accordingly, the event strips in Fig. 12 summarize the phase evolution predicted by the fluid model, dynamic-path disturbance → bandwidth perturbation → down transition → BBR-v3 reaction. Moreover, the disturbance spacing is on the order of tens of pbw seconds, which is comparable to t̄ , confirming that BBRv3 over Starlink must be analyzed as a coupled non-stationary system rather than as a conventional fixed-bottleneck path.
(a) Downlink BBR-v3 transmission stream
(b) Uplink BBR-v3 transmission stream
Fig. 12: Measurement validation through dedicated downlink and uplink BBR-v3 transmission streams.
The RTT variation due to satellite motion, inter-satellite routing, and queue accumulation, and can be expressed as: 2ds (t) di (t) q(t) c + ∑ c + C(t) I
RT T (t) =
(13)
i
where ds (t) is the satellite distance and di is the signal th traveling distance of i ISL, q(t) is the instantaneous queue occupancy, C(t) is the instantaneous bottleneck capacity, and c is the speed of light. Therefore, the observed RTT excursions I
C. Queuing Analysis The end-to-end latency comprises baseline propagation delay and variable queuing delay as expressed in Eq. (13). To characterize queuing behavior over Starlink downlinks, we apply an M /G/1 queue approximation to our empirical data. With MSS m = 1500 bytes and sampling window (w) of 15 s, we derive b service rate µ = 8⋅m [packets/s] from observed throughput b. Since packet arrivals are not directly observable, we approximate the arrival rate λ using the observed service rate and the queue occupancy fraction. Defining Xj = I(qj > 0) as a queue occupancy indicator, we estimate the arrival rate as λi = µi ⋅ X̄i , where: i 1 (15) X̄i = w ∑ Xj j=i−w+1
10
Fig. 14: Normalized TCP fairness indexes of the evaluated CCAs over the Starlink Internet.
Fig. 13: Queue buildup in dedicated and concurrent downlink BBR streams for different server locations over the Starlink network
This assumes arrivals occur at the service rate during busy periods and proportionally less during idle periods, consistent with the utilization relation ρ = λ/µ. In an M /G/1 queue, the mean waiting time follows the Pollaczek–Khinchine formula: λ E[S ] 2 (1 − ρ) 2
Wq =
(16)
where ρ = λ E[S] represents utilization and S is the service 2 time distribution. With E[S] = 1/µ and E[S ] = (1 + 2 2 2 cs )(E[S]) , where cs is the squared coefficient of variation 2 of service time (assumed cs = 1 for exponential service times), the average queue size follows Little’s Law as Q = λ Wq . Fig. 13 depicts the temporal evolution of the estimated queue size under BBR-v3 for downlink flows across the six geographically distributed Starlink paths. A prominent feature across all locations is the recurrent saturation of the queue at approximately Q ≈ 100 packets. This characteristic emerges implicitly from the M /G/1 formulation, specifically through the bounded utilization constraint ρi = λi E[Si ] ≤ ρmax . As ρi → 1, the Pollaczek–Khinchine relation causes the expected waiting time to grow rapidly, and consequently, the queue size Qi = λi Wq,i approaches a finite but large value determined by the imposed upper bound on ρi (ρmax = 0.995), yielding the observed saturation near 100 packets.
Both dedicated and concurrent BBR-v3 flows show frequent transitions between ρ → 1 and average queue buildup of around 7 packets, as depicted in Fig. 13. Thus confirming that BBR-v3 frequently drives the Starlink downlink close to full utilization. However, the concurrent traces tend to show more irregular timing and shorter low-queue intervals in several cities. This is expected because, under concurrent operation, BBR-v3 shares the bottleneck with other active congestion-control flows, so the aggregate offered load keeps the bottleneck closer to saturation even when the BBR-v3 flow temporarily reduces its sending rate. In contrast, the dedicated BBR-v3 traces show clearer drain periods in some cities, because the queue evolution is governed mainly by a single BBR-v3 control loop. Hence, the result indicates that BBR-v3’s downlink behavior over Starlink is dominated by repeated transitions between queue-draining and near-saturation phases. At the same time, concurrent traffic increases the persistence and irregularity of the saturated queue state. D. BBR-v3 Fairness Analysis Considering the uplink and downlink data collected under concurrent-flow scenario from the distributed Starlink testbed, we evaluate BBR-v3 fairness relative to the other CCAs. Normalized Jain fairness index is defined as: J=
(∑i Ti )
2
N ∑i Ti2
,
(17)
where Ti denote the mean throughput of CCA i, N the number of competing CCAs. The normalized max–min closeness is: MM
Fi
=
Ti , maxj Tj
(18)
and normalized Proportional-Fairness (PF): PF
Fi
= min (
⋆
Ci S ), , S ⋆ Ci
Ci =
log(1 + Ti ) , ∑j log(1 + Tj )
(19)
where Si = Ti / ∑j=1 Tj its achieved throughput share, S = α=0 1/N . The α = 0 throughput-maximization share is: Fi = N
⋆
11
Fig. 15: Throughput realized by BBR-v3 connections (a) Simultaneous flows (b) Staggered flows.
Si , were calculated for each CCA under this evaluation. These metrics allow BBR-v3 to be assessed not only by its throughput share, but also by how closely it approaches equal sharing, how far it is from the dominant flow, and how its proportionalfairness contribution compares with the ideal allocation. BBR-v3 fairness evaluation results under above metrics are illustrated in Fig. 14. In the downlink case, BBR-v3 remains competitive and achieves one of the highest PF-normalized scores, indicating that its proportional-fairness contribution remains close to the ideal allocation. However, its fair-share score, max–min score, and α = 0 throughput share are lower than the more aggressive high-throughput CCAs, particularly LeoCC, PCC, BBR-v1, and BBR-v2. This shows that BBR-v3 does not consistently dominate the downlink bottleneck, instead, it operates in a less aggressive regime while still maintaining strong proportional-fairness behavior. In the uplink case, BBRv3 exhibits a stronger overall fairness profile, achieving a high fair share and PF-normalized scores while maintaining a throughput share close to the ideal allocation. Although its max–min score remains below that of LeoCC, PCC, Copa, and BBR-v1. These combined metrics suggest that BBR-v3 coexists effectively with concurrent CCAs flows without persistently over-capturing the shared Starlink path. Thus, highlighting its favorable fairness–efficiency tradeoff and robust coexistence characteristics across both Starlink directions. To further evaluate the fairness characteristics of BBR-v3 over Starlink, we conducted dual-flow experiments between an Amazon EC2 instance in Sydney and the local Starlink terminal over the downlink path. Two BBR-v3 TCP flows were generated under simultaneous and staggered arrival conditions, with the latter introducing a 30 s offset between flows. Their throughput dynamics were analyzed using the phase-plane representation which is presented in Fig. 15. In the illustration, each point denotes the instantaneous throughput pair of the two flows. Furthermore, fairness was quantified using the above detailed four complementary allocation metrics: Jain’s fairness index, max–min fairness ratio, PF score, and the α = 0 throughput-maximization utility for both BBR-v3 flows. In the simultaneous case, BBR-v3 achieves balanced sharing, with a mean instantaneous Jain’s index of 0.949, PF score of 0.946, and a mean flow throughput ratio of 0.992, indicating that the
two flows obtain nearly symmetric rates. The max–min score is 0.761, showing that short-term throughput imbalance still occurs even though the long-term mean rates are close. In the staggered case, the mean instantaneous Jain’s index remains relatively high at 0.909 and the PF score is 0.907, but the max–min score decreases to 0.637 and the throughput ratio increases to 1.258. This indicates that the earlier flow retains a measurable advantage after the second flow joins. However, the α = 0 throughput-maximization utility increases from 197.30 Mbps in the simultaneous case to 242.52 Mbps in the staggered case, suggesting that the staggered scenario improves aggregate throughput at the cost of reduced fairness. Overall, these results show that BBR-v3 maintains stable two-flow coexistence over the Starlink downlink path, but its fairnessefficiency tradeoff is sensitive to flow arrival order. E. New Insights on BBR Performance Our experimental results demonstrate that BBR-v3 achieves high throughput in dedicated flows while maintaining a more balanced fairness, retransmission, and delay trade-off than aggressive alternatives such as PCC, LeoCC, and BBR-v1 under concurrent-flow operation. This stems from its modeldriven approach of explicitly estimating bottleneck bandwidth and round-trip propagation delay, rather than reacting to loss events. While Starlink’s dynamic conditions limit loss-based algorithms such as Cubic and Hybla and remain below the fair-share line in concurrent flows, BBR-v3 sustains nearcapacity utilization without incurring the excessive retransmission overhead observed for more aggressive CCAs. This balance between high utilization, moderate RTT inflation, and low loss exposure distinguishes BBR-v3 as a well-suited CCA for both dedicated and concurrent Starlink traffic. BBR-v3’s exponential startup mechanism achieves rapid bottleneck utilization, contrasting sharply with the prolonged convergence characteristics of additive increase algorithms. Its periodic bandwidth probing through ProbeBW_UP with pacing gain exceeding unity ensures swift discovery of available capacity, while competing CCAs remain constrained by conservative additive increase mechanisms. By intelligently bounding inflight data to the estimated BDP via inflight_hi and inflight_lo parameters, BBR effectively mitigates persistent queue buildup and bufferbloat. Thus, maintaining comparable RTT variance to benchmark CCAs despite higher throughput. Furthermore, BBR-v3’s tolerance for approximately 2% loss per RTT during probing prevents the catastrophic rate collapse characteristic of traditional CCAs. The higher retransmission counts observed with BBR-v3 stem from its model-based probing strategy, which periodically paces above the estimated bottleneck bandwidth to refine its bandwidth model. Unlike loss-based algorithms that treat packet loss as a catastrophic signal requiring multiplicative window reduction, BBR-v3 uses it as an informational cue, allowing transmission to continue near capacity. Additionally, BBR’s ACK aggregation modeling can generate transient bursts that exceed buffer capacity, causing clustered retransmissions.
12
However, these behaviors also indicate that the default BBRv3 configuration is not fully optimized for Starlink’s nonstationary capacity, RTT shifts, and handover dynamics. As shown in Appendix C, fixed pacing and congestion-window gains introduce an inherent utilization–retransmission trade-off, suggesting that a Starlink-oriented BBR-v3 configuration could better balance throughput, fairness, and loss exposure. V. C ONCLUSIONS In this paper, we present a globally distributed experimental evaluation of Google’s BBR-v3 over SpaceX’s Starlink LEO satellite network. Using our six-city testbed, we evaluate BBRv3 performance against eight CCAs: Cubic, Hybla, Vegas, LeoCC, Copa, PCC, BBR-v1, and BBR-v2. The results demonstrate that BBR-v3’s advantage is not aggressive bandwidth capture, but a more balanced fairness, retransmission, and delay trade-off over the Starlink Internet. Our mathematical modeling of Starlink’s network, incorporating atmospheric effects and LEO dynamics, provides a robust framework for understanding transport layer behavior over the Starlink Internet. Furthermore, we introduce a fluid model, present M /G/1 queue investigation and fairness analysis of BBR-v3 over the Starlink network, expanding the understanding of how it operates over the world’s largest LEO constellation.As LEO satellite constellations continue rapid deployment, BBR-v3 represents a compelling solution for maximizing performance in highlatency, variable-bandwidth environments. Future work will focus on adapting BBR parameters specifically for satellite characteristics and developing new approaches that balance throughput maximization with retransmission efficiency. R EFERENCES [1] S. Starlink, “Starlink Updates,” May 2026. [Online]. Available: https://www.starlink.com [2] Satellite Map, “Find starlink,” https://satellitemap.space/constellation/ starlink, 2026, accessed: May 2026. [3] S. Scherrer, M. Legner, A. Perrig, and S. Schmid, “Model-based insights on the performance, fairness, and stability of bbr,” in Proceedings of the 22nd ACM Internet Measurement Conference, 2022, pp. 519–537. [4] A. Abrol, P. Murali Mohan, T. Truong-Huu, and M. Gurusamy, “Bbr congestion control algorithms: Evolution, challenges and future directions,” ACM Comput. Surv., vol. 58, no. 9, Feb. 2026. [Online]. Available: https://doi-org.ezproxy-f.deakin.edu.au/10.1145/3793537 [5] N. Cardwell, I. Swett, and J. Beshay, “Bbr congestion control,” IETF Congestion Control Working Group (CCWG), Internet-Draft draft-ietf-ccwg-bbr-03, Oct 2024, expires 19 April 2025. [Online]. Available: https://datatracker.ietf.org/doc/html/draft-ietf-ccwg-bbr-03 [6] G. Barbosa, S. Theeranantachai, B. Zhang, and L. Zhang, “A comparative evaluation of tcp congestion control schemes over low-earth-orbit (leo) satellite networks,” in Proceedings of the 18th Asian Internet Engineering Conference, 2023, pp. 105–112. [7] S. Claypool, J. Chung, and M. Claypool, “Comparison of tcp congestion control performance over a satellite network,” in International Conference on Passive and Active Network Measurement. Springer, 2021, pp. 499– 512. [8] W. Yang, L. Cai, S. Shu, and J. Pan, “Mobility-aware congestion control for multipath quic in integrated terrestrial satellite networks,” IEEE Transactions on Mobile Computing, 2024. [9] N. Mohan et al., “A multifaceted look at starlink performance,” in Proceedings of the ACM Web Conference 2024, 2024, pp. 2723–2734. [10] L. Izhikevich, R. Enghardt, T.-Y. Huang, and R. Teixeira, “A global perspective on the past, present, and future of video streaming over starlink,” Proceedings of the ACM on Measurement and Analysis of Computing Systems, vol. 8, no. 3, pp. 1–22, 2024.
[11] I. Rhee, L. Xu, S. Ha, A. Zimmermann, L. Eggert, and R. Scheffenegger, “CUBIC for Fast Long-Distance Networks,” RFC 8312, Feb. 2018. [Online]. Available: https://www.rfc-editor.org/info/rfc8312 [12] S. H. Low, L. L. Peterson, and L. Wang, “Understanding tcp vegas: a duality model,” Journal of the ACM (JACM), vol. 49, no. 2, pp. 207– 235, 2002. [13] C. Caini and R. Firrincieli, “Tcp hybla: a tcp enhancement for heterogeneous networks,” International journal of satellite communications and networking, vol. 22, no. 5, pp. 547–566, 2004. [14] Z. Lai, Z. Li, Q. Wu, H. Li, J. Li, X. Xie, Y. Li, J. Liu, and J. Wu, “Leocc: Making internet congestion control robust to leo satellite dynamics,” in Proceedings of the ACM SIGCOMM 2025 Conference, 2025, pp. 129– 146. [15] V. Arun and H. Balakrishnan, “Copa: Practical {Delay-Based} congestion control for the internet,” in 15th USENIX Symposium on Networked Systems Design and Implementation (NSDI 18), 2018, pp. 329–342. [16] M. Dong, Q. Li, D. Zarchy, P. B. Godfrey, and M. Schapira, “{PCC}: Rearchitecting congestion control for consistent high performance,” in 12th USENIX Symposium on Networked Systems Design and Implementation (NSDI 15), 2015, pp. 395–408. [17] N. Pachler, I. Del Portillo, E. F. Crawley, and B. G. Cameron, “An updated comparison of four low earth orbit satellite constellation systems to provide global broadband,” in 2021 IEEE ICC workshops. IEEE, 2021, pp. 1–7. [18] I. Del Portillo, B. G. Cameron, and E. F. Crawley, “A technical comparison of three low earth orbit satellite constellation systems to provide global broadband,” Acta astronautica, vol. 159, pp. 123–135, 2019. [19] T. Jing-shi, Q. Ying-ying, and W. Qi, “Analysis and design of starlinklike satellite constellation,” Chinese Astronomy and Astrophysics, vol. 48, no. 1, pp. 161–189, 2024. [20] L. Wang, Z. Wang, Z. Deng, J. Zhang, and Y. Gao, “Alcs: An adaptive latency compensation scheduler for multipath tcp in satellite-terrestrial integrated networks,” IEEE Transactions on Mobile Computing, 2025. [21] ACMA, “Update to Foreign Space Objects Determination,” Oct. 2019. [Online]. Available: https://www.acma.gov.au/sites/default/files/2019-10/ Consultation%20paper%20-%20Update%20to%20Foreign%20Space% 20Objects%20Determination%20docx.docx [22] T. E. Humphreys, P. A. Iannucci, Z. M. Komodromos, and A. M. Graff, “Signal structure of the starlink ku-band downlink,” IEEE Transactions on Aerospace and Electronic Systems, vol. 59, no. 5, pp. 6016–6030, 2023. [23] Starlink Services, LLC, “Petition of Starlink Services, LLC for Designation as an Eligible Telecommunications Carrier,” FCC, WC Docket No. 09-197, Feb. 2021. [24] B. Al Homssi et al., “Deep learning forecasting and statistical modeling for q/v-band leo satellite channels,” IEEE Transactions on Machine Learning in Communications and Networking, vol. 1, pp. 78–89, 2023. [25] W. U. Khan et al., “Rate splitting multiple access for cognitive radio GEO-LEO co-existing satellite networks,” in GLOBECOM 2022. IEEE, 2022, pp. 5165–5170. [26] K. Weththasinghe, Q. T. Ngo, Y. He, and B. Jayawickrama, “Optimising beam size in multibeam leo satellite networks: Addressing interbeam interference, doppler shift, and frequency reuse,” IEEE Transactions on Aerospace and Electronic Systems, 2024. [27] Z. Song, Y. Liu, L. Xi, B. Ma, and X. Ma, “Analysis of link failure performance in low earth orbit satellite networks,” in 2024 IEEE/CIC International Conference on Communications in China (ICCC). IEEE, 2024, pp. 705–710. [28] G. Pan, J. Ye, J. An, and M.-S. Alouini, “Latency versus reliability in leo mega-constellations: Terrestrial, aerial, or space relay?” IEEE Transactions on Mobile Computing, vol. 22, 2023. [29] International Telecommunication Union, “ITU-R Recommendations for Atmospheric Attenuation Models: P.676-12, P.840-8, and P.838-3,” ITUR, Geneva, Switzerland, Recommendations, 2005–2019. [30] ITU, “Attenuation by atmospheric gases and related effects – recommendation,” ITU-R, Recommendation P.676-13, Aug 2022. [31] W. Hreha et al., “Synchronization for satellite system,” U.S. Patent US10411362B2, Sep. 2019, assignee: Space Systems/Loral, LLC. [Online]. Available: https://patents.google.com/patent/US10411362B2 [32] Q. Zhu, H. Tao, Y. Cao, and X. Li, “Laser inter-satellite link visibility and topology optimization for mega constellation,” Electronics, vol. 11, no. 14, p. 2232, 2022.
13
[33] ITU, “Propagation data and prediction methods required for the design of earth-space telecommunication systems,” ITU-R, Recommendation P.618-14, Aug 2023.
(0, 1] if the SN R falls between k and k − 1 SNR thresholds. The corresponding spectral efficiency is : S
ηk (SN R ) = rk log2 Mk S
C(t) = B ⋅ ηk (SN R (t)) s
We model the Starlink communication channel as follows. The distance between a LEO satellite and ground terminal at time t is given by: √ R ds (t) = R2 + a2 − 2Ra sin (θ(t) + a sin [ a cos θ(t)]) (20) where θ(t) represents the elevation angle, R is Earth’s radius, and a = R + h with h denoting the LEO altitude [24]. The corresponding free-space propagation loss is: 2
(21)
where λs is the signal wavelength [25]. Signal attenuation encompasses multiple environmental factors. Cloud attenuation l s K M dc /10 follows Lc = 10 1 , where K1 is the attenuation l coefficient, M is liquid water density, and dc (km) is the signal traveling distance through clouds. Rain attenuation is modeled b s aR̄ dr /10 as Lr = 10 , with R̄ representing rain rate, dr (km) the path through rain, and parameters a, b determined by frequency and polarization [29]. Scintillation and multipath fading ′ ′ s a ⋅σ contribution is defined under ITU guidelines as Ls = 10 , ′ ′ where a and σ are defined in [33], while gaseous attenuation s Lg follows the ITU definition in [30]. The Doppler shift due to relative motion is fd = vc fc , where v is relative velocity, c is light speed, and fc is carrier frequency. The receiver antenna s ′ 2 ′ gain is Gr = 4πar /λd , where ar is receiver aperture area and s λd = c/(fc ± fd ) [26]. With transmitter gain Gt and respective s s antenna efficiencies ηt , ηr , the received power at time t is: s
s Pr (t) =
s
s
s
s
Pt ⋅ η t ⋅ η r ⋅ G t ⋅ G r Lsfs (t) ⋅ Lsc ⋅ Lsr ⋅ Lss ⋅ Lsg
(22)
yielding an SNR of SN R (t) = Pr (t)/Nth , where Nth is thermal noise. S
s
s
s
S
[bits/s],
(25)
In Starlink’s downlink operation, high SNR periods enable higher-order modulation, yielding increased spectral efficiency η and consequently higher capacity C(t). Conversely, SNR degradation due to increased slant range, atmospheric effects, or handovers forces ACM to select lower modulation orders, substantially reducing spectral efficiency. This creates a propagation chain where SNR variance directly influences spectral efficiency variance, and ultimately capacity variance. C. Packet Drop Probability Due to Capacity Limitation The time-varying Starlink capacity in Eq. (25) determines the instantaneous service rate available to the packet stream. For a link without buffering, the instantaneous capacity-induced packet drop probability can be approximated as the fraction of traffic that cannot be served: Pdrop (t) = [1 − cap
+
C(t) ] , Rin (t)
(26)
where [x] = max(x, 0) and Rin is the corresponding traffic demand. Thus, no packets are dropped when C(t) ≥ Rin (t), while drops occur when the instantaneous Starlink capacity falls below the offered traffic rate. For a finite-buffer system, let Q(t) denote the queue occupancy in packets and let Qmax be the buffer capacity. The queue evolves according to +
Q(t + ∆t) = min [Qmax , [Q(t) + A(t, ∆t) − S(t, ∆t)] ] , (27) where A(t, ∆t) is the number of arriving packets during interval ∆t, and +
S(t, ∆t) = ⌊
C(t)∆t ⌋ ℓp
(28)
is the number of packets that can be served by the link, where ℓp is the average packet size. The number of dropped packets in the interval is therefore +
Let us consider instantaneous downlink SNR as a random 2 process where µSN RS and σSN RS denote the mean and variance of the SNR, respectively. ACM implements a deterministic mapping from instantaneous SNR to a discrete Modulation and Coding Scheme (MCS). Let us define the ordered set of SNR thresholds as: S
S
D(t, ∆t) = [Q(t) + A(t, ∆t) − S(t, ∆t) − Qmax ] .
B. Steady State SNR Estimation
min
(24)
Thus, for system bandwidth B , the instantaneous capacity becomes:
A PPENDIX A L INK FAILURE AND PACKET D ROP P ROBABILITY IN S TARLINK T RANSMISSION A. Starlink Channel Model
4πds (t) s ) Lf s (t) = ( λs
[bits/s/Hz].
s
max
SN R1 < ⋯ < SN Rk < ⋯ <
S
SN RK
(23)
The system employs a MCS k with modulation order Mk (e.g., QPSK, 16-QAM, 64-QAM) and the associated coding rate rk ∈
(29)
Hence, the empirical packet drop probability due to capacity limitation is ∑ D(t, ∆t) cap Pdrop = t . (30) ∑t A(t, ∆t) Since C(t) is determined by the SNR-dependent MCS selection, the capacity-limited drop probability can also be expressed s over the discrete MCS states. Let Ck = B rk log2 Mk be the capacity under MCS state k, and let pk = Pr (SN Rk−1 ≤ SN R (t) < SN Rk ) S
S
S
(31)
14
be the probability that the channel operates under state k. For a constant offered traffic rate Rin , the average capacity-induced drop probability can be approximated as + K Ck cap ] . P̄drop = ∑ pk [1 − Rin k=1
(32)
Therefore, high-SNR states yield larger Ck and thus reduce the probability of capacity-induced packet loss, whereas low-SNR states caused by increased slant range, atmospheric attenuation, or handover-related degradation reduce Ck and increase the likelihood of queue buildup and packet drops. Fig. 16: ISLs between different orbits. D. Probability of Failure due to Atmospheric Attenuation The ITU proposes a comprehensive estimation model for total attenuation (AT ) incorporating rain, cloud, gaseous, and tropospheric scintillation effects in satellite-to-ground links. This combined attenuation is expressed as [33]: √ ⎧ AG (p) + (AR (p)+AC (p))2 + A2S (p), ⎪ ⎪ ⎪ ⎪ ⎪ ∶ 0.001% ≤ p ≤ 5%, √ AT (p) = ⎪ ⎨ ⎪ 2 2 ⎪ AG (p) + AC (p) + AS (p), ⎪ ⎪ ⎪ ⎪ ∶ 5% < p ≤ 50%, ⎩ (33) where p represents the probability that attenuation exceeds the specified threshold, and AR , AC , AG , and AS denote rain, cloud, gaseous, and scintillation attenuation components, respectively. To accurately model link availability across varying elevation angles, we define θmin as the minimum usable elevation angle for LEO-to-ground connections and partition the operational elevation range into bins Bi = [θi , θi+1 ) (typically 5° increments as recommended in [33]). Given a time horizon T > 0 and instantaneous satellite elevation θ(t), the temporal fraction that the link spends in bin Bi is: 1 T wi = ∫ I{θ(t) ≥ θmin } I{θi ≤ θ(t) < θi+1 } dt, T 0
at
i
This elevation-weighted approach accounts for the nonuniform distribution of satellite positions and the corresponding variation in atmospheric path lengths. E. Inter Satellite Links Analysis We now develop a model for optical inter-satellite links between satellites in different orbital planes. Consider two satellites Si and Sj as depicted in Fig. 16, with optical wavelength λ and telescope apertures Dt and Dr for transmitter and receiver,
(36)
I
where d is the signal traveling distance. √ I d (t) = ri2 + rj2 − 2ri rj cos β(t)
(37)
β = arccos [sin(ϕi ) sin(ϕj ) + cos(ϕi ) cos(ϕj ) cos(Ψi − Ψj )] (38) where (ϕi , Ψi ) and (ϕj , Ψj ) refers to latitude and longitude position information of Si and Sj satellites, respectively [32]. Further, ri = R + hi and rj = R + hj , where hi and hj are orbital altitudes of the two satellites of interest as detailed in Fig 16. For a continuous transmission through the ISL without interruptions, it should uphold rj cos(ϕ) > R and ri cos(β − ϕ) > R. In a given ISL, for transmit power Pt , the received optical power is: I
Pr (t) =
(34)
(35)
2
4πd (t) = ( ) λ I
I Lfs (t)
I
For a specified attenuation threshold M (dB), we define pi (M ) as the percentage of time within bin Bi during which total (θ ) atmospheric attenuation exceeds M , where AT i (pi ) = M . The overall probability of link failure due to excessive atmospheric attenuation is therefore: p (M ) ≈ ∑ wi pi (M ).
respectively. The free space path loss of the optical ISL link between Si and Sj can be defined as:
I
I
I
I
I
Pt ηt ηr Gt Gr LIfs (t)
I
I
(39)
I
where ηt , ηr are efficiencies, and Gt , Gr are gains of transmitter and receiver ISL antennas receptively. Thus, the ISL capacity is given as: CISL (t) = B log2 (1 + I
Pr (t) ). Nth
(40)
I
where B is the bandwidth of the optical ISL and Nth is the thermal noise. In the Starlink constellation, co-orbital and adjacent-plane neighbors are permanent (availability ≈ 100% over 24 h), while some non-adjacent planes are intermittent with multiple disconnections per day [32]. Since each satellite is equipped with only a limited number of laser terminals (typically 3) [1], the aggregate node capacity is constrained by the sum of the effective ISL capacities of its terminals. If we assume that all the ISLs in the constellation have equal probability of failure, and the link failure and the recovery are modeled as Poisson processes, the probability of an ISL failure
15
Fig. 17: Receiver advertised window and RTT variance of CCAs in dedicated downlink.
Fig. 18: Receiver advertised window and RTT variance of CCAs in concurrent downlink.
Fig. 19: Receiver advertised window and RTT variance of CCAs in dedicated uplink.
Fig. 20: Receiver advertised window and RTT variance of CCAs in concurrent uplink. isl
where λon is the ISL failure arrival rate or the rate at which a isl functional ISL transitions into the disrupted state, and λof f is
can be given as [27]: isl
isl
p =
λon isl λisl on + λof f
(41)
16
Fig. 21: Median downlink performance of CCAs across 10 experimental runs under individual and concurrent conditions. Each marker corresponds to a single experimental run across 10 repetitions. Circles indicate individual flows and stars indicate concurrent flows.
the recovery arrival rate or the rate at which a disrupted link transitions back into the functional state. F. Impact of Handover Handovers in the Starlink constellation can be characterized as a schedule-driven process, subject to a system-wide synchronization [23], [31]. The handover timing is determined in advance by the gateway processor, leveraging satellite location, beam pattern, beam-hopping plan, and user terminal position. Based on the information, the gateway broadcasts handover data containing the terminal identity, source beam, target beam, and handover time. The terminal stores this information and performs the transition at the scheduled instant by returning to the new beam and updating its beam-hopping state. Near the handover instant, the system may also operate with wider ACM/Time Division Multiplexing (TDM) margins to improve robustness. Thus, a handover failure can arise primarily due to: (i) failure to receive or decode the broadcast handover command correctly, (ii) timing misalignment among the satellite, gateway, and user terminal, and (iii) terminal retuning or switching failure during execution, assuming there is no failure to provision or activate the target beam and associated beamhopping state at the scheduled epoch. Therefore, the per-attempt handover failure probability of user terminal j at slot n can be expressed as: pj [n] = 1 − (1 − pctrl,j [n])(1 − psync,j [n])(1 − psw,j [n]), (42) ho
where pctrl,j [n] is the probability of handover-control decoding failure, psync,j [n] is the probability of synchronization failure, and psw,j [n] is the probability of switching failure at the terminal. Therefore, the probability of link failure causing packet loss tot in the transmission path (p ) can be modeled as: N
p
tot
= 1−(1−p
cap
)⋅(1−p
gw
isl
)⋅(1−p )⋅(1−p )⋅ ∏(1−pi ) at
ho
isl
i=1
(43) where N is the number of ISLs that a given packet is gw at ′ routed through in the constellation, p = p (M ) (Eq. (35)), and the other parameters represent the respective failure probabilities [28]. isl
A PPENDIX B E XTENSIVE E VALUATION Fig. 17 and Fig. 18 present the CCA receiver-advertised window and RTT variance of dedicated and concurrent downlinks, respectively. On the other hand, Fig. 19 and Fig. 20 illustrate the CCA receiver-advertised window and RTT variance of dedicated and concurrent in the given order. To explore the consistency of the testbed results, we repeat all four test cases 10 times over the Starlink Internet. Fig. 21 depicts the medians of the six key measured performance indicators, both dedicated and concurrent downlinks, for the CCAs of interest. In addition, Fig. 22 presents the medians of the same key parameters of dedicated and concurrent uplink measurements. Furthermore,
17
Fig. 22: Median uplink performance of CCAs across 10 experimental runs under individual and concurrent conditions. Each marker corresponds to a single experimental run across 10 repetitions. Circles indicate individual flows and stars indicate concurrent flows.
in order to compare the Starlink Internet performance with terrestrial networks, we repeat dedicated uplink and downlink tests over the University network. The same testbed was connected to the University network, and the measured throughput and congestion window are presented in Fig 23(a) and (b) for downlink and Fig 23(c) and (d) for uplink. The University network throughput is limited to 1 Gbps at the switch port, thus underscoring the maximum experienced throughput in both uplink and downlink. These throughput results further confirm that the Starlink transmission path is the primary bottleneck in our distributed testbed, not AWS network constraints.
A PPENDIX C S ENSITIVITY A NALYSIS OF BBR- V 3 PARAMETERS In order to examine whether Starlink-specific parameter tuning could improve BBR-v3, we perform a trace-driven sensitivity analysis using the proposed fluid model. Let gp and gc denote pacing-gain and congestion-window-gain perturbation factors, respectively. These are not intended to reproduce all internal Linux BBR-v3 state-dependent constants; rather, they represent controlled variations around the default operating point. The pacing-gain range gp ∈ {0.9, 1.0, 1.25} follows the drain, cruise, and probe intuition of BBR’s ProbeBW behavior. The congestion window gain range gc ∈ {1.5, 2.0, 2.5} varies
around the BDP scaling values. Using the measured delivery ̂ rate (B(t)) and measured RTT (RTT(t)) from the Sydney dedicated downlink traces, we estimate the time-varying path state as: ̂ ̂ B DP(t) = B(t)RTT (44) min (t), where RTTmin (t) is the rolling minimum RTT. The gaindependent pacing rate and target inflight volume can be modeled as: ̂ rpace (t; gp ) = gp B(t), (45) ̂ vtarget (t; gp , gc ) = min (gc B DP(t), rpace (t; gp )RTT(t)) . (46) The modeled inflight volume follows the same first-order relaxation principle used in the proposed fluid model, thus connecting the parameter sweep to the measured Starlink throughput and RTT dynamics. dvmodel (t; gp , gc ) 1 (v = (t; gp , gc ) − vmodel (t; gp , gc )) Ts target dt (47) For each gain pair, we evaluate utilization, queue pressure, and retransmission-sensitive down-transition risk. The estimated queue pressure is computed as the excess modeled inflight volume above the measured BDP. ̂ Q(t; gp , gc ) = [vmodel (t; gp , gc ) − B DP(t)]
+
(48)
18
Fig. 23: Dedicated downlink and uplink performance over terrestrial network for the globally distributed server locations with different CCAs
queue pressure and down-transition risk, but it underutilizes the measured downlink for a significant fraction of the trace. Increasing gp and gc moves the model toward near-full utilization. When utilization saturates, additional aggressiveness mainly increases Q(t) and Rdwn rather than producing meaningful throughput gain. This indicates that fixed aggressive gains are not well matched to Starlink’s non-stationary capacity, RTTshift, and handover-driven dynamics. Therefore, an improved QoS could be delivered through Starlink-oriented BBR-v3 configuration, reducing aggressiveness during high RTT variance, retransmission bursts, or queue-buildup intervals, and restoring higher gains only during stable high-capacity periods. Fig. 24: Trace-Driven BBR-v3 gain sensitivity analysis for Starlink Internet
while the down-transition risk is approximated by the fraction of samples where the modeled inflight exceeds the probing threshold or where the estimated queue becomes large: Rdwn (gp , gc ) =
1 ̂ ∑ I(vmodel (t; gp , gc ) > 1.25B DP(t) T t ∨ Q(t; gp , gc ) > Qthr ).
(49)
The results illustrated in Fig. 24 show a clear Starlink-specific tuning trade-off. Conservative pacing, e.g., gp = 0.9, reduces