PaperlayerView exact revision on arXiv

Source-linked AI summary

Democratizing LEO Satellite Network Measurement

Liz Izhikevich, Manda Tran, Katherine Izhikevich, Gautam Akiwate, Zakir Durumeric

arXiv:2306.07469v2cs.NI

TL;DR

Real-world LEO network research faces a high data-collection barrier, and HitchHiking measures satellite links by probing publicly exposed services that use LEO Internet. Across Starlink customers, the measurements reveal latency patterns tied to POP distance, ground-station routing, and inter-satellite links.

  • Problem

    Real-world LEO measurement requires expensive specialized satellite hardware or recruiting satellite Internet customers, limiting studies to few vantage points and unvalidated theoretical models.

    Source: §1 · p. 1

  • Method

    HitchHiking probes Internet-exposed services that use LEO connectivity to identify satellite links and measure their characteristics without specialized hardware or recruitment.

    Source: §1 · p. 1

  • Results

    Customer distance to a POP is associated with more than threefold growth in minimum latency, ground-station changes cause sustained near-order-of-magnitude RTT spikes, and ISL reliance correlates with increased RTT.

    Source: §7.4 · p. 17

  • Takeaways & Limitations

    ISLs increase connectivity, including to ships, but can lengthen routing paths and bring some Starlink RTTs closer to GEO latency.

    Source: §9 · p. 18

    HitchHiking does not provide full coverage of the LEO network.

    Source: §8 · p. 18

Abstract

from arXiv · show

Low Earth Orbit (LEO) satellite networks are quickly gaining traction with promises of impressively low latency, high bandwidth, and global reach. However, the research community knows relatively little about their operation and performance in practice. The obscurity is largely due to the high barrier of entry for measuring LEO networks, which requires deploying specialized hardware or recruiting large numbers of satellite Internet customers. In this paper, we introduce HitchHiking, a methodology that democratizes global visibility into LEO satellite networks. HitchHiking builds on the observation that Internet-exposed services that use LEO Internet can reveal satellite network architecture and performance, bypassing the need for specialized hardware. We evaluate HitchHiking against ground truth measurements and prior methods, showing that it provides more coverage and accuracy. With HitchHiking, we complete the largest study to date of Starlink network latency, measuring over 2,400 users across 13 countries. We uncover unexpected patterns in latency that surface how LEO routing is more complex than previously understood. Finally, we conclude with recommendations for future research on LEO networks.

1 INTRODUCTION

HitchHiking lowers the barrier to collecting real-world LEO measurements by probing publicly exposed satellite-routed devices. Applied to Starlink, it measures latency at scale and reveals routing behavior that differs from prior assumptions.

1 INTRODUCTION

  • Researchers face a high barrier to real-world LEO data collection because it requires expensive satellite hardware or recruiting satellite Internet customers.The authors argue that lowering this barrier can help researchers accurately understand worldwide LEO networks in practice.Source: §1 · p. 1
  • HitchHiking identifies LEO-connected endpoints, locates satellite hops in their paths, then measures satellite-link characteristics without specialized hardware or participant recruitment.The approach addresses a barrier that has limited prior studies to a small number of vantage points and unvalidated theoretical models.Source: §1 · p. 1
  • Starlink latency peaks are not caused by distant satellites; nearby POP availability bounds customer latency, while ISL use increases path length and latency.HitchHiking measures 2.4K customers across 13 countries and reports latency nearly identical to physical-dish ground truth.Source: §1 · pp. 1–2
  • HitchHiking’s authors call for diverse perspectives on LEO routing and latency and release the measurement pipeline and collected data under Apache 2.0.The release is intended to lower the barrier for future LEO research.Source: §1 · p. 2

2 BACKGROUND

LEO satellites’ proximity enables lower theoretical latency than GEO satellites, but reduces coverage and requires satellites to move relative to Earth. Their routing can use a direct ground-station path or inter-satellite links.

2 BACKGROUND

  • GEO satellites orbit over 22,000 miles above Earth and provide wide coverage, but their minimum packet round-trip time is physically bounded at 480 ms.LEO networks emerged in response to lower latency requirements and cheaper satellite technology.Source: §2 · p. 2
  • LEO satellites orbit 180 to 1,300 miles above Earth, with approximately 10 ms theoretical RTT, but offer reduced coverage and are not geostationary.Their high-speed motion takes them around Earth every 90 minutes.Source: §2 · p. 2, Figure 1 · p. 2
  • In the basic bent-pipe architecture, a dish sends packets via satellite to a ground station, which forwards them to a provider POP and then onto the Internet.Some newer architectures use inter-satellite links to relay packets until a ground station is in view, enabling service where a client can see a satellite but not a ground station.Source: §2 · pp. 2–3
  • As of 2023, Starlink was the only consumer-targeted operational LEO network; OneWeb served businesses, while Kuiper and Telesat were expected to deploy consumer services.Starlink operates satellites both with and without inter-satellite-link capability.Source: §2 · p. 3

3 RELATED WORK

Prior LEO-network studies use physical hardware, recruited customers, or theoretical models, each with trade-offs in real data, coverage, labor, and cost. Existing remote measurement techniques also face challenges from satellite mobility and unverified simulations.

3 RELATED WORK

  • LEO measurement methods trade off real data, vantage-point coverage, labor, and monetary cost across physical hardware, recruited users, and theoretical models.Physical dishes offer real data but little coverage; recruitment expands coverage with more labor; models provide broad coverage without real data.Source: §3 · p. 3
  • Physical measurement can be costly and geographically limited: dishes cost $500–$23K, subscriptions exceed $100 monthly, and worldwide coverage may require travel.Examples include studies using one dish in Belgium and four dishes around Canada; RIPE Atlas had probes at 58 Starlink users across 13 countries.Source: §3 · p. 3
  • Recruitment and data-collection infrastructure consume labor and provide limited geographic coverage; RIPE Atlas measurements have minimum intervals of 60 seconds, too coarse for significant fluctuations.Recruiting users has supported broader studies, including browser-performance measurements with 18 Starlink users worldwide.Source: §3 · p. 3
  • LEO simulators support flexible studies of congestion, attacks, routing variability, and inter-satellite links, but popular simulations have not been verified and can misrepresent deployed architecture.StarryNet was reported as 20 times less accurate for 90th-percentile than 70th-percentile latency predictions.Source: §3 · p. 4
  • Prior network-measurement tools estimate paths, latency, bandwidth, and loss, but satellite mobility adds a non-deterministic delay source often attributed by existing tools to router queuing.Prior work assumed inter-satellite links would yield low-latency, direct routing; the paper argues they can increase path length and latency.Source: §3 · p. 4

4 LEO HITCHHIKING OVERVIEW

HitchHiking measures LEO links from publicly exposed services rather than requiring access to a satellite dish. Its workflow identifies satellite-routed endpoints, isolates satellite hops, and then applies suitable link-measurement experiments.

4 LEO HITCHHIKING OVERVIEW

  • HitchHiking is an outside-in method that probes publicly exposed satellite-routed devices to infer network architecture and performance wherever satellite clients are located.Unlike inside-out approaches, it requires neither specialized hardware nor painstaking recruitment.Source: §4 · p. 4

4.1 Finding Satellite-Routed Endpoints

  • To find satellite-routed endpoints, HitchHiking identifies relevant networks and scans their exposed services, then filters out services that may not traverse a satellite link or may use performance-enhancement proxies.A service inside a LEO network’s address space does not necessarily use a satellite link.Source: §4.1 · p. 5

4.2 Isolating Satellite Links

  • To isolate satellite links, HitchHiking needs enough of the route visible to identify the hop before and the hop after the satellite path.The method removes enough terrestrial routing artifact to make meaningful inferences about satellite links.Source: §4.2 · p. 5

4.3 Conducting An Experiment

  • After identifying satellite links, researchers can apply existing link-measurement strategies, such as comparing pings across satellite and terrestrial hops to identify outages.For bandwidth, increasing packet sizes sent to hops before and after the link can estimate capacity without flooding it; this paper focuses on Starlink latency.Source: §4.3 · p. 5

4.4 Ethical Considerations

  • HitchHiking experiments should send the minimum packets needed and avoid bandwidth-overloading tools such as iperf to protect users’ service quality.The authors note that LEO links often have lower capacity than terrestrial links and call for guardrails in experiments on exposed satellite services.Source: §4.4 · pp. 5–6

5 HITCHHIKING STARLINK TO MEASURE LATENCY

HitchHiking measures Starlink latency by sending TTL-limited pings over links to exposed services. The daily automated pipeline finds measurable Starlink services using public Internet scans.

5 HITCHHIKING STARLINK TO MEASURE LATENCY

  • HitchHiking measures Starlink latency by continually sending TTL-limited pings over hitchhiked LEO links.The implementation focuses on Starlink, the only commercially available consumer-targeted LEO network described here.Source: §5 · p. 6
  • The authors run a daily automated pipeline to measure latency.Source: §5 · p. 6
  • Censys scans IPv4 and IPv6 exposed services in Starlink's AS 14593 to find measurable Starlink-hosted services.Researchers could also scan with tools such as ZMap, Masscan, LZR, and GPS.Source: §1 · p. 6

1. Collect Exposed Services

The measurement pipeline identifies Starlink customer endpoints, locates satellite-related hops from an outside-in traceroute, and estimates latency while filtering terrestrial and short-lived artifacts.

1. Collect Exposed Services

  • 4,521 exposed services across 2,051 unique IPs were identified, with 1,790 unique IPs remaining after customer-endpoint filtering.The experiments used data collected on May 10, 2023.Source: §1 · p. 6
  • Filtering removes services with Peplink TLS certificates, then geolocates services and assigned POPs using Starlink feeds and PTR-record location names.Peplink filtering removes 9% of all services, leaving 1,629 IPs.Source: §1 · p. 6
  • The outside-in traceroute locates the satellite link between the last and second-to-last responsive hops, using a latency jump; HitchHiking does not see all satellite routing.Hop 16 is identified as the last visible pre-satellite hop, while the service-adjacent Hop 17 is not visible.Source: §1 · p. 7, Figure 3 · p. 7, §10 · p. 7
  • For five minutes, the method sends two ICMP pings per second with TTLs targeting terrestrial-router and exposed-service hops; only one probe crosses the satellite link.TTL-specific pings increase coverage by an order of magnitude compared with TCP, UDP, and non-TTL-specific pings.Source: §1 · p. 7
  • Subtracting terrestrial-router RTT from exposed-service RTT isolates satellite-link latency, and a 15-second smoothing window removes short-lived artifacts.The window matches the time-step at which Starlink dishes stay connected to a satellite before considering a switch.Source: §1 · p. 7

10. Validating Incomplete Visibility

Validation indicates HitchHiking's incomplete routing visibility is comparable to what physical Starlink customers can see, while worldwide vantage points detect sustained latency spikes. The pipeline is designed to adapt to new users and locations.

10. Validating Incomplete Visibility

  • The validation dish is configured to respond to pings so an external server can probe its public Starlink IP address.The setup uses bypass mode and a separate router connected to the dish.Source: §10 · p. 7
  • HitchHiking from different worldwide servers detects long-lasting latency spikes.The validation comparison uses ground truth measurements.Source: Figure 4 · p. 8
  • Physical Starlink customers have near-identical visibility into routing, despite not knowing the satellite count, satellite identities, or ground stations used.The authors use physical equipment to validate whether HitchHiking's incomplete visibility is sufficient for meaningful inferences.Source: §10 · p. 8
  • The pipeline is designed to adapt quickly to new users and geographic locations, and is released with collected data under Apache 2.0.The paper links adaptability to frequent changes in LEO network architecture.Source: §10 · p. 8

6 EVALUATION

The evaluation tests HitchHiking against dish ground truth, Hypatia simulations, and RIPE Atlas coverage, finding strong latency agreement and broader measurable reach.

6 EVALUATION

  • HitchHiking captures 100% of sustained spikes, is up to 80% more accurate than the simulator, and finds 28 times more measurable links than other methods.Source: §6 · p. 8

6.1 Comparison with Ground Truth

  • 96% of RTTs fall within 10 ms of ground truth, 50% within 3 ms, and HitchHiking captures 100% of sustained spikes with one false positive over 10,000 seconds.Source: §6.1.1 · p. 9

6.1.1 HitchHiking.

  • Two of five sustained spikes occur without a satellite switch; Starlink-validated measurements attribute the spikes to routing-path changes exacerbated by ISLs, not satellite location or congestion.Source: §6.1.1 · pp. 9–10, Figure 5 · p. 10

6.1.2 LEO Simulations.

  • Hypatia misses sustained RTT spikes, while HitchHiking RTTs are on average 1.8 times more accurate; Hypatia's second-nearest-ground-station prediction averages 7.6 ms error.StarryNet could not be evaluated because simulating Starlink requires over 2 TB RAM and the simulator cannot run in cloud environments.Source: Figure 6 · p. 10, §6.1.2 · pp. 10–11

6.2 HitchHiking Coverage

  • Table 1 lists exposed-service and RIPE Atlas POP locations: HitchHiking spans 20 cities in 13 countries, versus RIPE Atlas services across five countries.Source: Table 1 · p. 11, §6.2 · p. 12

7 WORLDWIDE LATENCIES IN THE WILD

Starlink latency varies with distance to assigned POP, ground-station routing, and ISL use; measurements across POPs reveal substantial geographic and temporal differences.

7 WORLDWIDE LATENCIES IN THE WILD

  • From May 18–June 23, 2023, HitchHiking measurements filtered 3.5K exposed Starlink IPs to 2.4K after excluding jittery pre-satellite hops.The resulting service set followed a geographic distribution similar to Table 1.Source: §7 · p. 12

7.1 POP Distance

  • Minimum RTT rises with POP distance; remote customers can exceed same-POP peers by over threefold, while mandatory routing through the assigned POP adds unavoidable latency.A US Virgin Islands customer assigned to Atlanta experienced a minimum RTT nearly twice that of customers closer to their POP.Source: §7.1 · p. 13

7.3 Geographic and Temporal Patterns

  • Across POPs, average RTT ranged from 28 ms in Mexico to 149 ms in Nigeria; Nigerian median latency fell 25% from May to June, unlike Brazil, while Georgia remained stable.Brazil-POP customers averaged 104 ms RTT; Georgia-POP customers averaged 58 ms, and their latency patterns remained relatively stable over time.Source: Figure 11 · p. 16, §7.3 · pp. 16–17, Figure 12 · p. 16

7.2 Routing Changes

  • At least 70% of customers experienced a sustained latency spike daily during month-long sampling; one-third of the Nigerian customer’s RTTs rose 2–5 fold the median, likely from distant-ground-station ISL routing.Starlink engineers confirmed that sustained spikes were due to customers using ISLs to reach ground stations that may be far away.Source: §7.2 · pp. 13, 15, 16

7.4 Summary

  • POP distance, ground-station changes, and ISL reliance align with higher latency, while diverse customers and routing infrastructure produce large worldwide variation in RTT.The paper’s summary notes that reliance on ISLs correlates with increased RTT and sustained latency increases.Source: §7.4 · p. 17

8 LIMITATIONS AND FUTURE WORK

The authors position HitchHiking as a complement to existing measurement methods and outline future applications to other LEO networks, longitudinal studies, simulations, and broader coverage.

8 LIMITATIONS AND FUTURE WORK

  • HitchHiking complements rather than replaces existing methods for collecting data about LEO links worldwide.The authors suggest integrating it into future LEO research.Source: §8 · p. 17

Measuring Other LEO Networks

  • HitchHiking can measure other LEO networks once their customers expose services, enabling comparisons across architectures in practice.The authors expect these networks to exhibit both similar and different behaviors from Starlink.Source: §8 · pp. 17–18
  • The authors identify roughly 20 measurable OneWeb endpoints, but cannot validate their latency measurements against a dish they control.OneWeb dishes are business-only and can cost upwards of $23,000.Source: §8 · p. 18

Measuring LEO Networks Over Time

  • Daily HitchHiking data can support studies of how changes in Starlink’s architecture and routing policies relate to network measurements over time.The authors pose questions about congestion, bandwidth, latency, and user experience as users and routing policies change.Source: §8 · p. 18

Data-Driven Simulations

  • Future work could train predictive models or build replay models from HitchHiking data, since simulation accuracy is constrained by unknowns about real LEO networks.Replay models could test new algorithms against past real-world data.Source: §8 · p. 18

LEO Network Coverage

  • HitchHiking measures over 10 times more of a LEO network than the leading alternative, but only an estimated 0.1% of Starlink customers.Its exposed-service sampling may introduce confounding factors; the authors propose seeking measurements of customers who do not expose services.Source: §8 · p. 18

9 CONCLUSION

The conclusion presents HitchHiking as a scalable, accurate way to study Starlink latency and reports routing-related latency patterns that differ from prior assumptions. The authors also describe their ethics review and harm-minimization measures.

9 CONCLUSION

  • HitchHiking measures over 2,400 Starlink customers across farms, boats, and remote regions, with accuracy and coverage exceeding alternative solutions.The method uses exposed services on LEO-based Internet connections without physical hardware.Source: §9 · p. 18
  • Inter-satellite-link routing creates the widest latency variance and lengthens paths between ground stations and POPs, despite being advertised as a low-latency solution.The authors report that sustained latency peaks are not caused by distant satellite location.Source: §9 · p. 18
  • By increasing connectivity, inter-satellite links can leave Starlink customers with occasional RTTs approaching GEO latency.The authors compare this connectivity-latency trade-off with international roaming in mobile networks.Source: §9 · p. 18

A ETHICS

  • The work involved no human subjects under the institution’s IRB policies, and the authors discuss harm-minimization measures in Section 4.4.Source: A ETHICS · p. 22

B MAPPING THE STARLINK NETWORK WITH TRACEROUTE

The authors test traceroute tools and protocols to map Starlink, then use physical-dish measurements to interpret the visible route and its latency.

B MAPPING THE STARLINK NETWORK WITH TRACEROUTE

  • From Stanford, 60.3% of endpoints were reachable using ICMP, 17.2% using UDP, and 24.6% using TCP; hop counts were unchanged 96% of the time.Source: B MAPPING THE STARLINK NETWORK WITH… · p. 22
  • Starlink’s consistent ICMP route crosses the customer’s POP, while TCP and UDP use changing IP addresses; MPLS hides some routing, and TNT finds no new paths.Across traceroute tools and protocols, the number of Starlink routing hops nearly always remains unchanged.Source: B MAPPING THE STARLINK NETWORK WITH… · p. 22

B.1 Mapping Starlink’s Routing

  • Dish-based egress traceroutes show an RTT spike after the LAN, indicating the first visible hop includes the satellite link and likely groundstation-to-POP terrestrial latency.Starlink hops after the POP add negligible latency in the reported experiment.Source: §B.1 · p. 22

C HITCHHIKING IMPLEMENTATION DETAILS CONTINUED

The implementation sections describe TTL-based pings, satellite-location inference from a dish obstruction map, worst-case Hypatia simulations, and a OneWeb measurement adaptation with no ground-truth validation.

C HITCHHIKING IMPLEMENTATION DETAILS CONTINUED

  • TTL pings to terrestrial routers had 0% complete packet loss, unlike regular pings, which never responded 99.2% of the time.The method relies on stable hop numbers, which remained consistent in 96.1% of exposed-service hops and 99.9% of terrestrial-router hops.Source: C HITCHHIKING IMPLEMENTATION DETAILS… · p. 23
  • Table 2 locates the dish-to-server LEO link between the first and second hops, where the traceroute shows a single RTT spike.Source: Table 2 · p. 23

D USING A STARLINK DISH OBSTRUCTION MAP TO INFER SATELLITE LOCATION

  • The dish’s obstruction map records its connected satellite location; the authors infer changes by subtracting consecutive maps after rebooting the dish.They poll the map every second and visually validate the inferred locations.Source: D USING A STARLINK DISH OBSTRUCTION… · p. 23

F USING HITCHHIKING TO MEASURE ONEWEB

  • For OneWeb, the authors collect exposed services with Censys, filter customer endpoints using business-name records, and geolocate services by registrant location.They report that most identified exposed services are in northern regions, including Alaska.Source: F USING HITCHHIKING TO MEASURE ONEWEB · p. 24

E HYPATIA WORST CASE SATELLITE SELECTION PREDICTIONS

F USING HITCHHIKING TO MEASURE ONEWEB

  • OneWeb satellite latency is estimated from the last-minus-second-to-last ingress-traceroute hop, but no controlled-dish ground truth validates the estimate.The corresponding table identifies the LEO link between those hops; OneWeb dishes are business-only and prohibitively expensive.Source: Table 3 · p. 24, F USING HITCHHIKING TO MEASURE ONEWEB · p. 24

Related papers

Loading 2306.07469v2…