TL;DR
This paper introduces a novel edge-side measurement technique to distinguish Starlink's service tiers and throttling from incidental performance drops. By correlating terminal telemetry with host-side probes, they identify unique "fingerprints" for different service states, enabling a lightweight detector to audit network policies without operator visibility.
What Problem Does This Paper Address?
Starlink, a prominent Low-Earth Orbit (LEO) satellite internet provider, offers various service tiers and data quotas. When users experience a sudden drop in internet speed, it's challenging to determine if it's due to policy enforcement (like throttling after exceeding a data cap) or other factors such as temporary network congestion, satellite obstruction, or path variations. This ambiguity is exacerbated by the lack of direct access to Starlink's internal policy metadata. This paper aims to bridge this gap by developing a method for edge observers (users or researchers) to infer these internal service states using only readily available terminal telemetry and active network probes.
Methods
The core of the methodology lies in a "policy-aware, cross-layer measurement" approach. They deployed a residential Starlink terminal in the UK and ran a 232.8-hour campaign where the service plan was intentionally hopped between different states (priority, post-quota throttling, stay-active, residential).
Key techniques include:
- Terminal Telemetry Collection: Using
starlink-gRPC-tools, they collected 1 Hz time-series data including internal downlink throughput, PoP RTT, and obstruction flags. - Host-Side Active Probing: A Raspberry Pi simultaneously ran ICMP pings (for RTT) and
speedtest-cli(for end-to-end download goodput, denoted asTuser). - Internal-to-User Ratio (R): A crucial metric,
R = Cint / Tuser, was calculated by comparing the terminal's internal downlink throughput (Cint) with the host-side goodput (Tuser). This ratio acts as a proxy for how effectively the internal throughput translates to user-perceived performance, potentially revealing policy-related inefficiencies. - Policy-Aware Labeling: The campaign was designed to trigger specific Starlink policies. Portal-reported service states were used as ground truth to label the collected data, allowing for the identification of distinct "fingerprints" for each state.
- Lightweight Detector: Based on the identified fingerprints, a simple rule was developed using windowed medians of download goodput and the
Rratio to classify service states.
Datasets
The paper utilizes a newly generated dataset from a single residential Starlink terminal over 232.8 hours. This dataset is characterized by intentional "plan-hopping" to exercise different service tiers and quota policies. While not a public benchmark, the authors suggest the methodology is reusable for creating similar datasets for other terminals, locations, or LEO systems.
Key Insights and Ideas
- Distinct Edge-Side Fingerprints: Different Starlink service states (high-speed, low-rate stay-active, and post-quota throttling) manifest as unique, low-dimensional signatures in the combined space of download goodput and the internal-to-user ratio (
R). - The Power of the
RRatio: TheRratio proves to be a powerful indicator. High-speed states show a stableRaround 10.7, while low-rate states exhibit significantly higherRvalues (around 21.6 for stay-active and 18.1 for post-quota). This upward shift inRin low-rate regimes is a key differentiator. - Quota Depletion Dynamics: Quota depletion isn't instantaneous. There's a noticeable "enforcement-delay window" (around several minutes) between the portal reporting zero quota and the actual throttling to the lower speed tier. This delay is clearly visible in the joint evolution of throughput and
R. - Simple Detection is Possible: A straightforward rule based on windowed median download goodput (e.g., > 50 Mbps) and the
Rratio (e.g., < 14.5) can accurately distinguish high-speed from low-rate states on this dataset.
Key Results
- The study identified three dominant operating signatures: two low-rate regimes (stay-active around 0.5 Mbps and post-quota throttling around 1 Mbps) and a high-speed regime (tens to hundreds of Mbps).
- The internal-to-user ratio
Rshowed stable, state-dependent characteristics: medianRof ~10.7 for high-speed (S2/S4), ~21.6 for stay-active (S1), and ~18.1 for post-quota (S3). - A simple detector rule achieved accurate classification of service states on the labeled windows, separating high-speed from low-rate regimes without errors on this trace.
- The enforcement-delay window
Gwas observed to be a minutes-scale gap between portal quota status and throttling onset, captured by lightweight measurements.
Limitations and Future Work
- Single-Site, Single-Terminal: The study is based on a single residential Starlink terminal in the UK. The numerical thresholds and specific values identified are deployment-specific and may vary across different locations, terminal models, or network configurations.
- No Adverse Weather/Obstruction: The dataset did not include adverse weather conditions or sustained obstructions, which could influence performance and potentially complicate the fingerprinting.
- Transferability: While the methodology is presented as reusable, applying the same procedure to additional terminals, locations, firmware versions, and other LEO systems is left for future work. This would involve validating and potentially recalibrating the thresholds.
- Future Work: The authors suggest extending this work to other LEO systems and exploring more complex policy scenarios. The core contribution is a foundational methodology for edge-side auditing of LEO network policies.
Summary updated