Low Earth Orbit (LEO) satellite networks, such as Starlink, are transforming global Internet access by delivering high-speed connectivity to underserved and remote regions. Despite extensive research into Starlink’s performance, latency characteristics remain under-explored. This study presents a comprehensive analysis of one-way delay components in the Starlink network using high-frequency, high-precision measurement probes. Over a 10-day period, more than 500 million probe packets were collected and analyzed. The results reveal minor diurnal latency variation and provide means to separate out the delay components contributing to the observed one-way delay, and we sketch a delay model and provide empirical distributions. By measuring both uplink and downlink paths, the study uncovers significant differences in scheduling behavior, with uplink delays more affected by Starlink’s periodic 15-second reconfiguration cycles. The results also highlight the limitations of using too coarse measurement intervals, which can introduce aliasing effects. Our OWD data set and traffic generation tool are made available to support further research in the area. 

Leveraging a dataset of almost half a billion packets with high-precision packet times and sizes, we extract characteristics of the bursts emitted over Starlink’s Ethernet interface. The structure of these bursts directly reflects the physical layer reception of OFDMA frames on the satellite link. We study these bursts by analyzing their rates, and thus indirectly also the transition between different physical layer rates. The results highlight that there is definitive structure in the transition behavior, and we note specific behaviors such as particular transition steps associated with rate switching, and that rate switching occurs mainly to neighboring rates. We also study the joint burst rate and burst duration transitions, noting that transitions occur mainly within the same rate, and that changes in burst duration are often performed with an intermediate short burst in-between. Furthermore, we examine the configurations of the three factors burst rate, burst duration, and inter-burst silent time, which together determine the effective throughput of a Starlink connection. We perform pattern mining on these three factors, and we use the patterns to construct a dynamic N-gram model predicting the characteristics of the next upcoming burst, and by extension, the short-term future throughput. We further train a Deep Learning time-series model which shows improved prediction performance. 

We examine the performance of 14 different Linux TCP congestion control (CC) variants over Starlink connectivity. We then focus on the two most commonly used CCs, BBR and Cubic, and the best performing other CC, Illinois. Multiple measurement campaigns were conducted to evaluate throughput performance for single and multiple flows, examine how the performance of the CCs relate to the length of the TCP flow, and investigate how the CCs share resources over Starlink. The results show that BBR (both v1 and v3) outperform Cubic in the current Starlink environment. Cubic, which is the Linux default CC, performs considerably worse than BBR, particularly for single flows. Cubic with HyStart also has additional slowstart performance issues, which can be mitigated by disabling HyStart. The Illinois CC is a promising alternative to BBR and although being less performant for single flows, it has better fairness properties. Leveraging earlier similar measurements for a longitudinal study, we further show that changes made to the Starlink infrastructure during 2023/2024 are likely to have resulted in significantly reduced Cubic throughput. 

Starlink satellite internet has emerged as a viable solution for extending internet connectivity, particularly in remote or underserved areas, complementing terrestrial networks. With the increasing use of Starlink for latency-critical realtime multimedia communication such as live streaming, video conferencing, virtual reality, and online gaming, understanding and predicting Starlink performance becomes increasingly relevant. However, the dynamic nature of satellite constellations, atmospheric conditions, bandwidth availability, and network congestion present hurdles to maintaining steady network performance, which is crucial for real-time multimedia applications. This study investigates the predictability of Starlink downlink throughput, which can aid in proactive scheduling decisions and intelligent traffic management across multiple paths, thereby optimizing Starlink for real-time multimedia applications. The research develops a comprehensive approach for throughput prediction by leveraging historical time series data and a diverse array of machine learning and deep learning models over time slot sizes of 1 s, 100 ms, 50 ms, 20 ms, 10 ms, and 5 ms. Furthermore, the work includes three baseline models to assess the performance enhancement that the machine learning and deep learning models provide. The study proposes a prediction scheme that allows a set of predictors to perform single-step predictions over these prediction intervals. Model evaluation results show that predicting downlink throughput at short time slot sizes of 5 ms and 10 ms for different history sizes offers substantial benefits relative to the baseline models.

Leveraging a data set of almost half a billion packets with high-precision packet times and sizes, we process it to extract characteristics of the bursts emitted over Starlink's Ethernet interface. The structure of these bursts directly reflect the physical layer receipt of OFDMA frames on the satellite link. We study these bursts by analyzing their rates, and by proxy the transition between different physical layer rates. The results highlight that  there is definitive structure in the transition behavior, and we note specific behaviors such as  particular transitionsteps associated with rate switching, and that rate switching occurs mainly to neighboring rates. We also study the joint burst rate and burst duration transitions, noting that transitions occur mainly within the same rate, and that changes in burst duration are often performed with an intermediate short burst in-between.Finally, we examine the configurations of the three factors burst rate, burst duration, and inter-burst silent time, which together determine the effective throughput of a Starlink connection.

Although Starlink has been deployed for several years, a detailed understanding of system internals is still lacking. In this work we employ precise per-packet timestamps obtained from a hardware-timestamp capable NIC connected to a Starlink terminal.We find that Starlink frame timing details are readily observable at the network layer by analyzing the packet timing patterns.Based on a one-week measurement campaign we collect around half a billion of packet size and timing observations. Processing these observations yields 2.3 million transmission bursts. To learn details on the radio resource management we develop a methodology to infer the effective physical layer sending rate. Our findings show that although Starlink throughput can vary widely over multiple time-scales, there are a small number of fundamental physical layer transmission rates. We employ Gaussian Mixture Modeling to determine 14 such fundamental transmission rates, and relate the obtained rates to previous knowledge of the Starlink OFDMA frame structure. Our empirical observations provide an excellent match for a radio resource configuration where a Starlink frame employs 1000 subcarriers and 287 symbols per frame for user traffic transmission, which for uniform 4-QAM modulation yields a base rate of 430.5 Mbps. This physical layer base rate appears to mostly be varied by multiples of 27 Mbps, in several instances likely by modifying the modulation of a subset of the symbols in multiples of 18 symbols. 

Low Earth Orbit (LEO) satellite networks such as Starlink can provide broadband connectivity in sparsely populated regions, including those located in arctic and subarctic regions. However, only satellites in a subset of inclination planes can serve these northern regions. In order to evaluate Starlink connectivity in these geographically extreme regions, we perform a drive test across the arctic circle measuring Starlink and cellular connectivity. We analyze the results with regards to Starlink throughput characteristics including the effects of the 15 second reconfiguration interval, and compare the throughput and connectivity. Starlink generally provide better throughput, with around 10x better median throughput than our comparison cellular network. However, the low tail of the distribution is less impressive with more frequent connectivity losses. Our results also show that the combination of Starlink and cellular has the potential to provide considerable gains in aggregated connection reliability and tail throughput. We additionally discuss policy implications relating to state subsidy policies designed to ensure that the whole population has access to broadband connectivity. 

With a growing user base and the deployment of new systems, such as 5G and Starlink, a deep understanding of the varying link throughput for wireless systems is highly important. While detailed analysis of link throughput can be done on packet traces, collecting extensive packet traces often faces storage and privacy challenges. Instead, we propose using traces of link-wide inter-packet delay (IPD) to enable highly granular link throughput characterization on a wider scale. To this end, we present an eBPF-based tool designed to capture IPDs, and evaluate the accuracy of captured IPDs with the IPD tool and tcpdump, both with and without access to hardware timestamps. While hardware provided timestamps provide accurate IPDs, we find that software based timestamps lead to IPD values which are very inaccurate, but still useful in aggregate form to characterize throughput at millisecond timescales. Furthermore, we show that concurrent packet processing incurs a significant amount of packet reordering, which necessitates the consideration of several previous packets when computing the link-wide IPD. Finally we present an example use case of IPD collection, characterizing frequent silent periods during a speedtest over a 5G link.

Although Starlink has been rolled-out for several years, there is still a lack of knowledge regarding system details and performance characteristics. To address this, we perform a network layer measurement campaign utilizing precise times-Tamping to analyze throughput variations at multiple time scales, and infer system timing details. On larger timescales we quantify the diurnal variations where the throughput varies with the time of day. On the medium timescales we establish the likely frequency allocation and beam switching period to be 15 seconds. The associated connectivity disturbances contribute to severe link underutilization for single long-lived TCP flows, which typically reach only 46% of the estimated link capacity. On the sub-millisecond timescale our network layer measurements corroborate recent physical layer investigations of the Starlink frame timing, which is confirmed to be 1.33 ms. © 2023 Owner/Author(s).

The detailed performance characteristics of networking equipment is to a large extent a function of the software that controls the underlying hardware components. Most networking equipment is regularly updated with new software versions. By studying performance changes related to such changes in software, it is possible to identify particular software versions that affect the performance of the system. Consequently, having automated methods for detecting changes in network equipment performance is crucial. In this work we study the change point detection problem arising when the placement in time of software updates is known a priori, but the presence of any performance implications on any of the thousands of performance indicators that can be collected is unknown. The ability to improve the automated detection of such change points by clustering the monitored systems according to the set of collected indicators has not been fully evaluated. We here report our experience with employing clustering, together with a bootstrap-based change point detection, across a range of performance indicators. We evaluate four variations of clustering approaches, and demonstrate the resulting improvement in change point detection sensitivity.