Source-linked AI summary
Beyond Idealized Orbits: Realistic Network Topology and Spatio-Temporal Traffic Emulation for LEO Satellite Constellations
Suryansh Aryan, Josh Smith, Alexa Houck, Rahul V. Chintalapati, Alexander Kedrowitsch, Jeffrey H. Reed, Nishith D. Tripathi, Samantha Parry Kenyon
TL;DR
Dynamic orbital changes and environmental effects complicate LEO network routing, creating a need for realistic testbeds. This paper extends SpaceNet to multi-shell and terrestrial topology with congestion modeling, then compares Starlink emulation with real measurements. Across evaluated scenarios, its predictions closely match measured performance, with regional throughput error substantially below the low-fidelity model.
Problem
Orbital motion, handovers, Doppler shifts, and atmospheric impairments complicate network management, motivating testbeds that emulate realistic dynamics, environmental conditions, and congestion.
Source: §1 · p. 1Method
SpaceNet combines multi-shell topology and congestion-informed routing with stochastic traffic generation and evaluates emulated RTT and throughput against Starlink measurements.
Source: §1 · p. 2, §5 · p. 11, §6.1 · p. 13Results
Across scenarios, SpaceNet closely matched measured performance and outperformed low-fidelity; regional throughput prediction error fell by approximately 92–95%, with MAE decreasing from over 73 Mbps to less than 6 Mbps.
Source: §7 · p. 20Takeaways & Limitations
Predictions for inter-continental lower-traffic and regional higher-traffic cases closely followed ground truth within a 95% confidence level over the simulation duration.
Source: §7 · p. 20The model assumes uniform user behavior and interoperable subnet protocols, and uses a simple ISL and GSL handover mechanism.
Source: §5 · p. 12
Abstract
from arXiv · showhide
Large low-earth orbit (LEO) satellite constellations are revolutionizing network access globally by delivering potential high-capacity, low-latency communication, fueling innovation and accelerating research in advanced space-based non-terrestrial network architectures. Consequently, a research testbed that allows flexibility in designing realistic network routing schemes, traffic load balancing and inter-shell communication methods to emulate real-world multi-shell network becomes an important tool for industry developers and academic researchers. This study broadens the scope of the SpaceNet testbed, now including emulation of multi-shell constellations and computation of routes with the existing global web of terrestrial terminals. Furthermore, the testbed produces accurate trends in return trip time (RTT) and throughput metrics due to its support of high-level congestion modeling and the necessary local and global congestion routing. To evaluate the efficacy of the testbed, the network performance of emulated results for the Starlink constellation is evaluated and compared with data from real Ping and iPerf tests between Starlink terminals. Two important communication scenarios have been discussed that highlight the testbed's potential for realistic multi-shell topology generation and handling various types of traffic distribution with well-informed congestion routing. Predictions of network performance using SpaceNet accurately addresses temporal behaviors for different traffic levels with strong temporal predictability in its throughput and RTT time-series, as well as enhanced ground truth coverage within a 95% confidence interval.
1. Introduction
LEO network dynamics and environmental effects complicate routing and link management, motivating high-fidelity testbeds that model orbital dynamics, environmental conditions, and congestion. This work extends SpaceNet with multi-shell topology, terrestrial connectivity, and congestion routing.
1. Introduction
- Dynamic orbital motion, handovers, Doppler shifts, rain fade, and ionospheric scintillation complicate network management and motivate realistic NTN testbeds.The testbeds are intended to emulate orbital dynamics, environmental conditions, and congestion scenarios.Source: §1 · p. 1
- Existing testbeds trade off dynamic traffic adaptability, multi-shell packet-flow emulation, constellation scope, protocol support, or terrestrial-node behavior.The cited comparison spans Hypatia, NPVT, StarPerf, Celestial, and LSNS.Source: §1 · pp. 1–2
- SpaceNet adds high-fidelity GSL and ISL modeling, scalable multi-shell topology generation, and modular congestion distribution and routing.The work extends prior SpaceNet experiments on single-shell Walker-type constellations.Source: §1 · p. 2
- SpaceNet evaluates RTT and client throughput over simulated Walker-type and Starlink constellations, comparing predictions with real Starlink network data.The evaluation includes routing dynamics and global congestion scenarios.Source: §1 · p. 2
2. SpaceNet Testbed Architecture
SpaceNet combines satellite and terrestrial data in a two-phase simulation and emulation testbed. Its architecture reconstructs multi-shell constellations, builds terrestrial and non-terrestrial topologies, and updates links and traffic over simulation time.
2. SpaceNet Testbed Architecture
- SpaceNet uses open-access satellite, terrestrial-network, and weather data in a two-phase architecture separating simulation from emulation.The simulation phase reconstructs multi-shell constellations from TLEs and uses weather data to calculate GSL signal-to-noise ratios.Source: §2 · pp. 2–3
2.1. Multi-Shell Constellation Reconstruction
- Multi-shell reconstruction parses shell-specific satellites from TLEs and uses Fisher-Jenks clustering to identify orbital planes for non-uniform constellations.For the Starlink TLE example, SpaceNet reconstructed shells despite shell satellite counts differing from FCC disclosures.Source: §2.1 · p. 3
2.2. Terrestrial Sub-Topology Generator
- The terrestrial topology models customer terminals, gateways, and internet endpoints, connecting gateways to endpoints and deriving terrestrial link metrics from communication models and databases.The default terrestrial links are fully connected between gateways and internet endpoints.Source: §2.2 · pp. 3–4
2.3. Network Topology Processes
- Topology processes track visible satellites for GSL-capable ground stations and establish links using shortest distance or maximum visibility association criteria.The default shortest-distance criterion selects nearby visible satellites; maximum visibility favors longer connection windows.Source: §2.3 · pp. 4–5
2.4. Mobility Processes
- Mobility processes update the non-terrestrial topology as simulation time advances, handling ISL and GSL handovers through separate processes.ISL updates are managed by mobility processes, while network topology processes compute GSL handovers.Source: §2.4 · p. 5
2.4.1. Non-Terrestrial Sub-Topology Generator
- The non-terrestrial generator propagates satellite states, validates ISLs using communication models, constructs the network mesh, and passes it to traffic mobility processes.The described propagation options include SGP4 for LEO and a prototype HPOP for MEO, GEO, and hypothetical xGEO regimes.Source: §2.4.1 · p. 5
2.4.2. Tra
- Traffic mobility processes model background-user loads and their effects on shared network resources across the topology.The modeled traffic can reflect differences between densely populated low-inclination shells and more sparsely populated high-inclination shells.Source: §2.4.2 · pp. 5–6
3. Non-Terrestrial Handover Grid Schemes
SpaceNet constructs connected multi-shell topologies by generating intra- and inter-shell ISLs, then associating ground stations through GSLs. Its default link criteria favor nearby visible satellites, while the paper identifies a resilience-related limitation in ISL selection.
3. Non-Terrestrial Handover Grid Schemes
- TIM +Grid models Starlink-like ISLs by matching satellite transceiver orientations, with each satellite able to establish up to four links simultaneously.The scheme scans integer-labeled orbits to establish links.Source: §3 · p. 6
- Along-track links connect each satellite to its two closest same-orbit neighbors; cross-track links scan adjacent orbits for a candidate with a free interface, yielding two to four ISLs per satellite.The cross-track rule departs from the conventional +Grid scheme by requiring the candidate satellite’s corresponding interface to be free.Source: §3 · p. 7
- The default ISL criterion selects the closest visible satellite within an interface’s field of view, while maximum-visibility selection is harder for non-uniform constellations.The testbed’s predicted future satellite states and interface orientation model allow integration of the more resilient criterion.Source: §3 · p. 7
- GSLs use the shortest-distance visible satellites, with up to four connections per ground station; elliptical orbits (e > 0.05) trigger additional geometric validation.The candidate threshold assumes near-circular orbits and uses operational altitude and minimum elevation angle; validation considers the minimum elevation angle, satellite half-angle, and current altitude.Source: §3 · p. 7
4. Communication modeling
The communication models replace idealized link assumptions with RF link budgets for ISLs and environment-dependent loss modeling for GSLs. Both derive signal quality for network performance metrics.
4.1. Inter-Satellite-Link modeling
- The ISL model replaces connectivity-and-bandwidth assumptions with physics-based RF link budgets that account for signal attenuation.The earlier approach treated each link as achieving its full input bandwidth.Source: §4.1 · p. 7
- ISL losses combine free-space path loss with polarization, system, and misalignment losses before calculating received signal strength.The model assumes constant transmitter EIRP and receiver gain, and allocates one fifth of total throughput to each of five simultaneous users.Source: §4.1 · p. 8
4.2. Ground station-Satellite-Link modeling
- Building-penetration loss depends on building type and materials, signal frequency, elevation angle, and noise, with distributions representing typical and worst-case loss.The model also applies building-specific coefficients in calculating overall penetration loss.Source: §4.2 · p. 9
- Atmospheric absorption is selected or interpolated by frequency at zenith, corrected for elevation angle, and capped at 10 dB; scintillation varies by frequency regime.Ionospheric scintillation is stronger below 6 GHz, while tropospheric scintillation is stronger above 6 GHz.Source: §4.2 · p. 9
- GSL modeling sums free-space, shadow-fading, clutter, building-penetration, atmospheric-absorption, and scintillation losses to calculate SNR for throughput and latency metrics.The model derives link performance from the radio-channel parameters between a ground element and a satellite.Source: §4.2 · pp. 8, 10
5. Spatio-Temporal Traffic Spread Model
SpaceNet models geographic traffic hotspots over time, spreads stochastic user demand across the satellite network, and uses the resulting load and network metrics for congestion-aware routing. The model assumes equal user service demands and has other realism boundaries in its routing assumptions.
5. Spatio-Temporal Traffic Spread Model
- SpaceNet identifies traffic hotspots from geographic rush-hour trends, spreads their traffic across relevant satellites, and supports centralized congestion routing each simulation epoch.The model can represent bottlenecks in terrestrial and non-terrestrial sub-topologies.Source: §5 · p. 10
- The default normalized traffic curve divides local time into low-, medium-, and high-demand zones and uses a fifth-order polynomial to evolve the rush-hour window.Traffic load is stochastic; the mean varies with the diurnal curve, while the regional user load is evenly distributed among active ground stations.Source: §5 · pp. 10–11, Figure 9 · p. 10
- Traffic demand drives latency and throughput estimates using log-normal interarrival and service times for large-timescale simulations; a Poisson process is also available for short, light-traffic conditions.The log-normal model allows traffic aggregation over time and preserves heavy-tailed behavior.Source: §5 · p. 11
- A recursive traffic-spreading process updates link usage across ISL and GSL pathways, removes loops and returns to source ground stations, and can carry traffic to secondary ground stations.Intermediate-node loads are included when calculating latency and throughput; the default spreading balances outflow evenly across connections.Source: §5 · p. 11, Figure 10 · p. 12
- The model assumes unbiased user service demands and uniform user behavior, and its ideal inter-subnet protocol and simple ISL/GSL handovers limit routing realism.HTTP requests and file transfers are treated equally in the current version.Source: §5 · p. 12
- After updating network latency and throughput, SpaceNet uses Floyd–Warshall routing to select paths by latency, throughput, or minimum hop count.The resulting routing tables and connectivity matrices support Phase 2 ICMP emulation.Source: §5 · p. 12
6. Results & Discussion
The evaluation compares SpaceNet’s Starlink emulations with real Ping and iPerf measurements across an intercontinental multi-shell route and a regional route. Prediction accuracy and temporal behavior vary with traffic zone and routing conditions, with larger errors under regional traffic and high uncertainty.
6. Results & Discussion
- SpaceNet predictions were compared with repeated Starlink Ping and iPerf measurements, using shell2 and shell4 TLEs and routes minimizing cumulative propagation and system latency.The real measurements were collected at one-second intervals; the emulations used Starlink TLEs extracted on 19 September 2025.Source: §6 · pp. 12–13
6.1. Test case 1 Analysis: Multi-shell topology with back-
- About 90% of Zone-1 and Zone-2 intervals fell within the ground-truth 2-sigma level, while throughput MAE and RMSE were about 10–14% of ground truth.Zone-3 performed worst, with extreme RTT oscillations and sudden throughput changes that the model did not predict reliably.Source: §6.1 · pp. 14–15, Figure 13 · p. 15, Figure 14 · p. 15
- In the intercontinental case, traffic load and orbital orientation shaped bent-pipe routing; Zone-3 often used lower-load shell2 satellites, while Zone-1 had more stable paths.Zone-1’s RTT autocorrelation was higher, whereas Zone-3 throughput had the strongest short-term persistence; RTT autocorrelation showed high-frequency variation around zero lag.Source: §6.1 · pp. 15–16, Figure 17 · p. 16
6.2. Test case 2 Analysis: Regional tra
- For the regional route, high-fidelity routing penalized intermediate ground-satellite links as traffic-zone penalties increased, while throughput was limited to about 10 Mbps by loaded North American inter-satellite links.Low-fidelity Zone-3 throughput was substantially overestimated relative to the high-fidelity prediction.Source: §6.2 · pp. 17–18, Figure 20 · p. 18
- Regional high-fidelity errors were about 30–50% of ground truth, versus about 900–1200% for low fidelity; RTT fell within the 95% confidence region more often than throughput.Throughput remained overestimated, and the authors identify higher-fidelity inter-satellite-link and handover models as a possible way to reduce that bias.Source: §6.2 · p. 18, Table 8 · p. 18, Table 9 · p. 18
- Regional predictions showed throughput–RTT cross-correlation peaks near zero lag of about 0.85, 0.78, and 0.81 for Zones 1, 2, and 3, respectively, while throughput-ground-truth alignment peaked at about 0.6 in Zone-2.The authors report a distinct zero-lag peak for all zones but note that outliers limit prediction of higher temporal properties.Source: Figure 22 · p. 19, §6.2 · p. 19
7. Conclusion
High-fidelity SpaceNet closely matched measured network performance and substantially outperformed its low-fidelity implementation across the evaluated scenarios. Its temporal predictions tracked ground truth, but stable-state accuracy does not extend to sudden throughput surges.
7. Conclusion
- Under high traffic, throughput correlations indicated stable route states; regional routing showed throughput-stability opportunities, while strategic ground-station links could compensate for poor higher-shell geometry on inter-continental paths.The inter-continental path observation assumes a congestion-free network topology.Source: §7 · p. 19
- Throughput predictions were more temporally stable than RTT predictions in lower-traffic inter-continental and higher-traffic regional cases, tracking ground truth within the 95% confidence level throughout the simulation.Source: §7 · p. 20
- Regional throughput MAE fell 92–95% relative to low-fidelity, from over 73 Mbps to less than 6 Mbps; RTT prediction error fell by up to 58%.Across scenarios, SpaceNet closely matched measured performance and substantially outperformed the low-fidelity implementation.Source: §7 · p. 20
- Inter-continental throughput prediction errors were approximately 10–12 Mbps, with nearly 90% of predictions within the measured 95% confidence interval for representative traffic zones.Source: §7 · p. 20
- More robust handovers and higher-fidelity RF and optical ISL models are needed to address persistent biases and sudden dips in simulated throughput.The authors also identify limited data for topology and traffic hyperparameters, centralized routing that may not generalize to all providers, and missing queue buildup and congestion spikes that leave sudden throughput surges unpredicted.Source: §7 · p. 20