Source-linked AI summary
A Multifaceted Look at Starlink Performance
Nitinder Mohan, Andrew Ferguson, Hendrik Cech, Prakita Rayyan Renatin, Rohan Bose, Mahesh Marina, Jörg Ott
TL;DR
Starlink’s performance remains incompletely characterized at global scale and across real-time applications and internal network behavior. This paper combines crowdsourced, distributed, application-level, and controlled measurements to study those dimensions. It finds competitive performance with cellular networks under suitable conditions, but geographic variation and synchronized periodic reconfigurations produce important performance fluctuations.
Problem
Prior Starlink evaluations were limited by a lack of global vantage points and often used a few locations, simulations, or emulations.
Method
The paper combines 19.2M M-Lab measurements from 34 countries, RIPE Atlas probes, application tests, and controlled terminals in two European countries.
Results
Starlink is competitive with terrestrial cellular networks for real-time applications under optimal conditions, while performance varies geographically and synchronized 15 s reconfigurations cause latency and throughput degradations.
Takeaways & Limitations
Starlink can support real-time applications and may improve service for remote users through inter-satellite connections, but its observed performance depends on infrastructure and internal operations.
Takeaways & Limitations
The study’s application comparison used an unobstructed terminal under favorable weather, and mobility or other conditions may change Starlink’s relative performance.
Abstract
from arXiv · showhide
In recent years, Low-Earth Orbit (LEO) mega-constellations have emerged as a promising network technology and have ushered in a new era for democratizing Internet access. The Starlink network from SpaceX stands out as the only consumer-facing LEO network with over 2M+ customers and more than 4000 operational satellites. In this paper, we conduct the first-of-its-kind extensive multi-faceted analysis of Starlink network performance leveraging several measurement sources. First, based on 19.2M crowdsourced M-Lab speed test measurements from 34 countries since 2021, we analyze Starlink global performance relative to terrestrial cellular networks. Second, we examine Starlink's ability to support real-time web-based latency and bandwidth-critical applications by analyzing the performance of (i) Zoom video conferencing, and (ii) Luna cloud gaming, comparing it to 5G and terrestrial fiber. Third, we orchestrate targeted measurements from Starlink-enabled RIPE Atlas probes to shed light on the last-mile Starlink access and other factors affecting its performance globally. Finally, we conduct controlled experiments from Starlink dishes in two countries and analyze the impact of globally synchronized "15-second reconfiguration intervals" of the links that cause substantial latency and throughput variations. Our unique analysis provides revealing insights on global Starlink functionality and paints the most comprehensive picture of the LEO network's operation to date.
1 INTRODUCTION
Starlink has attracted global connectivity interest, but its performance has been studied from too few locations. This paper addresses that gap through comprehensive measurements spanning global performance, applications, access paths, and internal reconfiguration.
- Motivation: LEO megaconstellations promise ubiquitous low-latency coverage and could bridge connectivity gaps in remote regions.These networks use thousands of satellites orbiting at 300–2000 km altitudes.
- Research gap: Limited prior Starlink studies relied on a handful of locations, simulations, or emulations rather than global vantage points.Orbital coverage and ground infrastructure make geographically broad measurement important.
- Approach: The paper analyzes Starlink globally using ≈19.2 M M-Lab samples from 34 countries and active measurements from 98 RIPE Atlas probes in 21 countries.It also conducts controlled terminal experiments in two European countries.
- Findings: Starlink performs competitively with terrestrial cellular networks, while performance varies with ground-station and PoP infrastructure and latency rises under load.The study reports signs of bufferbloating alongside geographic variation.
- Findings: Under optimal conditions, Starlink supports Zoom and Luna cloud gaming at performance matching cellular networks, although periodic reconfigurations create artifacts.The paper also examines Starlink’s last-mile behavior and internal operations.
- Resources: The study publishes more than 300 GB of collected data and associated scripts to support reproducibility and future research.The release accompanies the paper’s multifaceted measurement analysis.
2 BACKGROUND
Starlink provides Internet access through a satellite-to-ground architecture centered on user terminals, satellites, ground stations, and terrestrial PoPs. Its constellation uses multiple orbital shells, with the 53° shell carrying most operational satellites and higher-inclination shells serving polar regions.
- Constellation: Starlink aims to provide global Internet coverage using satellites flying at approximately 500 km altitude.Most operational satellites lie within the 53° shell.
- Constellation: The 70° and 97.6° orbital shells serve regions near the poles but contain fewer satellites than the 53° shell.The 53° shell covers only parts of the globe.
- Access architecture: A user terminal communicates with visible satellites above 25° elevation through phased-array antennas using Ku-band user links.Satellites use multiple antennas and beams to connect with multiple terminals simultaneously.
- Access architecture: In direct bent-pipe connectivity, traffic travels from the terminal through a satellite to a ground station and then to a terrestrial PoP.PoPs route traffic onward through the terrestrial Internet.
- Inter-satellite links: When the terminal and ground station are not within one satellite’s coverage cone, laser inter-satellite links can extend the bent-pipe path to distant ground stations.Not all satellites support ISLs, and their use is difficult to estimate from IP-layer traceroutes.
3 MEASUREMENT METHODOLOGY
The study combines crowdsourced speed tests, distributed active probes, application experiments, and controlled terminals to measure Starlink from global scale to sub-second network behavior. The methodology pairs broad geographic coverage with targeted tests of applications and orbital access.
- Global measurements: The M-Lab analysis uses ndt7 TCP BBR measurements, retaining countries with at least 1000 samples since June 2021.The resulting dataset contains 19.2 M samples from 34 countries.
- Global measurements: M-Lab records throughput, goodput, RTT, losses, IP information, ASNs, and endpoint geolocation during end-to-end speed tests.NDT uses a single 10 s WebSocket TCP connection.
- Targeted measurements: The study uses 98 Starlink RIPE Atlas probes across 21 countries to measure paths to 145 data centers operated by seven major cloud providers.The targets are selected near Starlink PoPs.
- Application measurements: Zoom experiments connect an unobstructed Starlink terminal and high-speed terrestrial fiber endpoint to an AWS machine near the assigned Starlink PoP.Virtual cameras and microphones transmit a prerecorded bidirectional video.
- Application measurements: Cloud-gaming experiments evaluate Amazon Luna using a customized client that records media-stream information while a bot performs predefined in-game actions.The evaluation focuses on demanding throughput and delay requirements.
- Controlled experiments: Controlled experiments use terminals in Germany and Scotland, with metal shielding at the Scottish site restricting communication to the 70° and 97.6° orbital shells.The setup addresses RIPE Atlas’s lack of sub-second visibility.
4 GLOBAL STARLINK PERFORMANCE
Starlink latency and throughput are globally competitive but uneven, with performance shaped by ground infrastructure and degraded latency under load. Goodput is relatively homogeneous, while latency varies across regions and has stabilized over time.
- Global View: For most countries, terrestrial ISPs achieve better latency than Starlink, whose median latency is approximately 40–50 ms but varies geographically.Colombia is an exception, while Manila performs notably worse than established terrestrial networks.
- Global View: Proximity to Starlink ground stations and Points-of-Presence correlates with latency, producing lower delays in well-provisioned regions and longer delays elsewhere.Dublin, London, and Berlin have latencies comparable to the US, whereas Rome and Paris have 75th-percentile latency approximately 20 ms longer.
- Global View: 90%: Philippines latency to in-country servers fell after traffic was routed through a local Point-of-Presence deployed in May 2023.Before deployment, traffic reached Japanese PoPs and incurred an additional 50–70 ms RTT to in-country endpoints.
- Latency Under Load: ≈2–4×: Starlink RTT increases during active downloads, reaching almost 400–500 ms, while upload latency rises to ≤100 ms at the 60th percentile.The asymmetric inflation may reflect different queue behavior at the Dishy, ground station, or satellites.
- Goodput: Most Starlink clients achieve approximately 50–100 Mbps download and 4–12 Mbps upload at the 75th percentile, with relatively homogeneous goodput distributions.Goodput shows no correlation with baseline latency, but higher loss rates coincide with lower goodput.
5 REAL-TIME APPLICATION PERFORMANCE
Starlink can support demanding real-time applications under favorable conditions, performing competitively with 5G despite higher variability and recurring 15-second fluctuations. Zoom and Luna measurements reveal application-level effects that global speed tests can miss.
- Zoom Video Conferencing: Starlink supports Zoom video conferencing with approximately 27 FPS, while its uplink one-way delay averages 52±14 ms versus 27±7 ms terrestrially.Starlink sends more FEC traffic than terrestrial networks, helping address its slightly higher loss rate.
- Reconfiguration Effects: Every 15 s: Starlink one-way delay shifts at recurring interval points associated with satellite-path reallocation.The paper links these reconfiguration intervals to periodic performance fluctuations and reports that application-specific FEC can mitigate artifacts.
- Amazon Luna Cloud Gaming: 150 minutes: Across terrestrial, 5G, and Starlink cloud gaming, playback remained close to 60 FPS at approximately 20 Mbps bitrate.Starlink fell between terrestrial and cellular networks for bitrate fluctuations, frame drops, and freezes, but had the highest game delay.
- Amazon Luna Cloud Gaming: Starlink cloud gaming occasionally dropped below 20 FPS, with drops coinciding with Starlink reconfiguration intervals.These fluctuations appear only at sub-second granularity and therefore are absent from the global performance analysis.
- Overall Findings: Starlink remains competitive with 5G for demanding real-time applications, particularly in regions with less mature cellular infrastructure.This comparison was measured with an unobstructed terminal under favorable weather conditions; mobility and other conditions may change relative performance.
6 DISSECTING THE BENT-PIPE
The paper dissects Starlink’s bent-pipe performance using RIPE Atlas measurements and controlled experiments, linking latency differences to serving orbit, inter-satellite links, and synchronized reconfiguration intervals. These factors produce geographically varying performance and short-lived latency or throughput changes.
- 6.1 Global Bent-Pipe Performance: ≈60 ms improvement over fiber in Reunion Island indicates inter-satellite links can benefit remote regions with poor terrestrial connectivity.The comparison involves a Reunion-to-Frankfurt path of approximately 9000 km versus Germany-to-Frankfurt fiber of approximately 500 km.
- 6.1 Global Bent-Pipe Performance: ≈2× higher bent-pipe latency in Alaska, with intermittent connectivity, reflects sparse service from the 70° and 97.6° orbital shells.Sweden near the 53° boundary shows latency comparable to Canada, the UK, and Germany.
- 6.1 Global Bent-Pipe Performance: ≈40 ms bent-pipe latency remains consistent worldwide within Starlink’s dense 53° shell.
- 6.2 Controlled Experiments: Brief sub-second connection disruptions can occur during reconfigurations even when only a single candidate satellite is visible.The changes may reflect reassigned frequency or routing resources on the same satellite.
- 6.2 Controlled Experiments: 15-second reconfiguration intervals globally coordinate Starlink scheduling and produce latency and throughput variations at interval boundaries.The observations are synchronized across geographically separated terminals and are not attributed to satellite handovers.
7 RELATED WORK
Prior Starlink studies offered limited or narrow measurement coverage, leaving its geographically varying performance and black-box operation incompletely understood. This work addresses that gap with the most extensive examination to date, combining measurements from 34 countries and multiple vantage points.
- Existing Starlink performance studies were limited, often using only a few vantage points.
- Traceroute analysis revealed Starlink’s internal network topology, while other work identified a potential global network controller.
- The absence of global measurement sites has hindered comprehensive understanding of Starlink’s performance, which varies geographically with internal configurations.
- 34 countries and 19.2 million crowdsourced M-Lab measurements support this work’s extensive global analysis.The study also uses 1.8 million active RIPE Atlas measurements and two controlled terminals connected to different Starlink orbits.
- The combined approach provides deeper insight into Starlink’s bent-pipe architecture and overall performance.
8 CONCLUSIONS
The paper evaluates Starlink as a potential global Internet provider through a multifaceted investigation spanning global performance and internal network operations. It finds that Starlink is comparable to cellular networks for the examined real-time applications, although performance varies with proximity.
- Starlink is comparable to cellular networks for supporting the examined real-time applications, Zoom and Luna cloud gaming.
- The comparison varies based on proximity to Starlink and other relevant conditions.
- The investigation spans perspectives from global network performance to internal Starlink operations.
A STARLINK ORBITAL INFORMATION
Starlink is described as a megaconstellation because it combines multiple orbital shells, unlike single-shell systems such as Iridium. The paper presents its orbital-shell design and operational satellite counts in Table 2.
- Table 2 details Starlink’s orbital-shell design and number of operational satellites as of October 2023.
- Starlink and other emerging LEO constellations are termed megaconstellations because they combine multiple orbital shells.
- Single-shell systems such as Iridium provide the contrast to the megaconstellation design.
- The discussion simplifies Starlink’s orbital structure into circular orbits.
B DATA CENTER ENDPOINTS
The study uses a geographically diverse set of cloud data-center endpoints to reduce measurement bias from cloud-network routing differences.
- The endpoint selection spans cloud providers and continents to reduce bias from private WANs, peering agreements, and traffic steering.Each endpoint is a virtual machine in a compute-capable cloud data-center location.
C GLOBAL STARLINK PERFORMANCE
Starlink performance changes over time and across regions. Goodput decreases as user numbers increase, while latency is generally higher outside major operational areas but improves in some locations.
- Goodput decreases over one year across selected cities, largely attributed to increasing Starlink user numbers.
- RTT remains relatively stable but is higher in countries outside major Starlink operational areas.
- North American cities achieve lower latencies than the rest of the globe, with only slight variation between cities.
- Auckland and Perth show similarly low minRTT, while Sydney latency improved in June 2023.
D GLOBAL VIEW OF BENT-PIPE OPERATION
The paper characterizes Starlink’s global bent-pipe and ground-segment behavior using worldwide latency measurements, endpoint deployments, and controlled measurement procedures.
- Global GS ↔PoP latencies are generally ≤6 ms, with North America at ≥6 ms, likely reflecting dense GS and PoP deployment.
- The measurements use cloud data-center endpoints distributed across providers and continents to reduce routing-related bias.
- Figure 17 estimates Starlink last-mile latency by subtracting M-Lab server ↔PoP latency from overall measured minimum RTT.
- Figure 18 estimates the satellite-link latency fraction by dividing Figure 17 latency by overall latency.
- The controlled experiments replace interruption-sensitive TCP measurements with periodic ICMP pings and restart iperf across connectivity windows.
- Clock drift caused by interrupted connectivity required timestamp adjustment, while gRPC data came from a separate machine without clock drift.