Starlink Latency Prediction Achieves 140ms Accuracy with 500Hz Measurements

Low Earth orbit (LEO) satellite networks are rapidly becoming vital for global communication, forming a cornerstone of future 6G infrastructure. Andreas Casparsen, Jonas Ellegaard Jakobsen, and Jimmy Jessen Nielsen, from the Department of Electronic Systems at Aalborg University, alongside Petar Popovski and Israel Leyva Mayorga et al., have undertaken a detailed investigation into the unpredictable latency experienced within these networks. Their research presents a statistical characterisation of end-to-end latency, utilising high-resolution experimental data to reveal a distinct 15-second periodic behaviour in Starlink systems. By identifying and isolating latency spikes caused by handovers, the team demonstrates that accurate short-term prediction is possible, significantly improving upon estimates derived from long-term averages. This granular understanding of LEO latency paves the way for adaptive transmission strategies and ensures a more reliable connection for latency-critical applications.

Understanding and predicting these patterns is crucial for improving the Quality of Service (QoS) for latency-sensitive applications. Researchers collected end-to-end latency data from Starlink users and identified a consistent 15-second periodic structure, with predictably high latency at the beginning and end of each cycle attributed to handovers. They explored both parametric and non-parametric models to predict and classify latency within the stable core of each period, excluding the handover regions, using Mean Squared Error (MSE) for regression accuracy and Area Under the PR Curve (AUPRC) for classification accuracy.

The study concentrated on the stable portion of the cycle, demonstrating that accurate predictions are possible with relatively simple models and limited data. This work provides a statistical foundation for building intelligent middleware that can optimize performance in LEO satellite networks and demonstrates that predictable patterns exist that can be exploited to improve QoS for various applications. The research highlights the potential for application-aware adaptation and efficient system management in these networks, with future work planned to investigate interactions between transport/application protocols and the characterized latency dynamics, validate the framework’s generalizability across different LEO constellations, and explore integration with multi-connectivity capabilities.

Starlink Latency Analysis via High-Resolution Measurement

Researchers undertook a detailed investigation of Low Earth Orbit (LEO) satellite network latency, crucial for the development of future 6G communication systems. The study pioneered a high-resolution statistical analysis of end-to-end latency, employing experimental bidirectional one-way measurements sampled at 500Hz. This enabled the team to dissect the deterministic 15-second periodic behaviour exhibited by the Starlink network, revealing nuances in its performance. Experiments involving 500Hz bidirectional one-way measurements revealed a distinct 15-second periodic behaviour in network latency. The team meticulously characterised handover-induced boundary regions, identifying latency spikes of approximately 140ms at the beginning and 75ms at the end of each 15-second cycle, followed by a stable intra-period regime.

This precise temporal segmentation allows for accurate short-term latency prediction, surpassing the limitations of long-term statistical estimates. Results demonstrate that isolating these boundary regions and applying both parametric and non-parametric models to intra-period latency distributions achieved 99th-percentile latency prediction errors below 50ms. The study collected 7500 samples within each 15-second interval, partitioning each period into discrete time bins. A refined start phase was identified through edge detection and histogram analysis, remaining stable across months of data collection. Mean-centering techniques highlighted intra-period dynamics, revealing a sharp initial latency peak averaging 74ms above the mean, followed by a decline and secondary rise near the period’s end.

Starlink Latency Reveals Predictable Periodic Behaviour

This research details a statistical analysis of end-to-end latency in Low Earth Orbit satellite networks, specifically focusing on the Starlink system. Through extensive measurement and analysis of bidirectional communication, the authors demonstrate a consistent 15-second periodic behaviour in network latency, identifying distinct boundary regions at the start and end of each cycle where latency spikes occur. Characterisation of these boundary regions, lasting approximately 140ms and 75ms respectively, allows for accurate prediction of short-term latency within stable intra-period regimes.

The study’s key contribution lies in demonstrating that latency prediction based on the periodic structure significantly outperforms long-term statistical estimates, achieving 99th-percentile prediction errors below 50ms. By isolating these boundary regions and applying parametric and non-parametric models to intra-period latency, the authors reveal a predictable pattern previously obscured by broader statistical analysis. This period-level prediction capability enables adaptive transmission strategies, allowing systems to identify and potentially circumvent periods where latency requirements cannot be met. While the observed statistical structure remained consistent across multiple datasets, the research is limited to the specific characteristics of the Starlink network investigated, with future work planned to explore generalizability to other LEO constellations and investigate the underlying causes of the observed periodic behaviour.