Better Fill Up Your Pipe - Revisiting Starlink’s Burst Characterization

We present a comprehensive analysis of Starlink’s physical layer performance based on a year-long UDP measurement campaign. Building upon a known methodology, we replicate and improve the approach to capture frame-level data rates with high granularity, identifying discrete modulation steps, and proposing an improved modulation model. Our approach reveals distinct symbol allocation patterns and validates modulation schemes ranging from robust low-order to higher-order Quadrature Amplitude Modulation (QAM). The findings provide a foundation for further research into the influence of environmental and network conditions on link performance. Additionally, we offer an open dataset to support ongoing studies of Low Earth Orbit (LEO) satellite networks.

本研究基于为期一年的UDP测量活动，对Starlink的物理层性能进行了全面分析。我们借鉴一种已知的方法论，通过复现与改进，实现了对帧级数据速率的高粒度捕获，从而识别出离散的调制阶梯，并提出了一个优化后的调制模型。

我们的方法揭示了独特的符号分配模式，并验证了从稳健的低阶调制到高阶正交幅度调制（QAM）的多种方案。这些研究成果为进一步探究环境与网络条件对链路性能的影响奠定了基础。此外，我们还提供了一个开放数据集，以支持低地球轨道（LEO）卫星网络领域的持续研究。

Introduction

Low Earth Orbit (LEO) satellite networks are on the rise. For example, with Starlink’s almost global coverage, users in remote areas, who had no or poor connectivity, can connect to the internet for prices comparable to terrestrial connections. Besides the reasonable price, these networks offer low latencies and high throughput rates [1].

Starlink is the first – and as of May 2025, also the largest commercially usable LEO satellite network. Assessing the systems performance capabilities has been of increasing interest to the research community. Because of the closed nature of the system, such performance assessments are primarily conducted via real-world measurements. Such measurement studies have been conducted from around the globe under various different conditions. These studies can be roughly categorized into stationary, e.g., [2, 3, 4], and mobile, e.g., [1, 5] measurements. The vast majority of those studies measure on the application layer. As a consequence, the achievable application layer performance and the influence of different impact factors such as weather [4, 6, 7], space weather [8, 9], recurring system reconfigurations [3], and obstruction [1, 10] have been relatively well understood.

In contrast, limited work has analyzed Starlink’s physical layer. Humphreys et al. [11] laid significant ground to understand Starlink’s physical frame structure and the properties of the used KU band. Garcia et al. [12, 13, 14] proposed a method that allows to determine Starlink’s used modulation based on analyzing packet timings and physical layer burst characteristics.

In this paper, we improve the approach of Garcia et al., leading to (partly) significantly different results and new insights into Starlink’s physical layer transmission rates.

Our contributions are as follows:

• We improve the approach of Garcia et al. [12, 13, 14] to infer Starlink’s used modulation via measurements of the physical layer packet timings and burst characteristics.

• We deploy the extended approach in a one-year-long measurement campaign conducted in Lindenberg, Germany.

• We extensively analyze our dataset and provide a better understanding of Starlink’s signal structure.

• We release our large dataset containing packet timings, inferred physical layer rates, transport layer throughput, and latency open-access to the research community.

The remainder of this paper is structured as follows. First, we describe the baseline approach proposed by Garcia et al. and our improvements in Sec. 2. Afterward, we describe our measurement setup in Sec. 3. We analyze our dataset in Sec. 4 and conclude our paper in Sec. 5. 低地球轨道（LEO）卫星网络正在迅速发展。例如，借助Starlink近乎全球的覆盖范围，过去网络连接缺失或不佳的偏远地区用户，如今能以媲美地面网络的价格接入互联网。除价格合理外，这类网络还提供低延迟和高吞吐速率的特性[1]。

Starlink是首个，同时截至2025年5月，也是规模最大的可商用LEO卫星网络。评估该系统的性能已成为科研界日益关注的焦点。由于该系统具有封闭性，此类性能评估主要通过实际测量来开展。全球范围内已在多种不同条件下开展了此类测量研究。这些研究可大致分为静态测量（例如[2, 3, 4]）和移动测量（例如[1, 5]）。这些研究绝大多数是在应用层进行测量。因此，关于可实现的应用层性能，以及天气[4, 6, 7]、空间天气[8, 9]、周期性系统重构[3]和信号遮挡[1, 10]等不同因素对其影响，学界已有了较为透彻的理解。

相比之下, 针对Starlink物理层的分析工作则较为有限

Humphreys等人[11]为理解Starlink的物理帧结构及所用KU波段的特性奠定了重要基础。Garcia等人[12, 13, 14]提出了一种方法，通过分析数据包时序和物理层突发特征来确定Starlink所使用的调制方式。

本文改进了Garcia等人的方法，得出了（部分）截然不同的结果，并对Starlink的物理层传输速率提出了新的见解。

我们的贡献如下：

• 我们改进了Garcia等人[12, 13, 14]的方法，通过测量物理层数据包时序和突发特征来推断Starlink所使用的调制方式
• 我们在德国林登贝格（Lindenberg）部署了这一扩展方法，并开展了为期一年的长期测量活动
• 我们对数据集进行了深入分析，并对Starlink的信号结构提供了更深刻的理解
• 我们 向科研界公开发布了我们的大型数据集，其中包含数据包时序、推断的物理层速率、传输层吞吐量和延迟等数据

本文的其余部分结构如下。首先，我们在第二节中描述Garcia等人提出的基准方法及我们的改进。之后，我们在第三节中介绍我们的测量设置。我们在第四节中分析我们的数据集，并在第五节中对全文进行总结。

Methodology

We will first introduce the base approach, followed by our improvements. Figure 1 provides a schematic overview of both approaches [12], with the methodology from the original approach in purple and our changes in green.

我们首先介绍基准方法，随后阐述我们的改进。图1提供了这两种方法的示意图[12]，其中原始方法的流程以紫色表示，我们的改进之处以绿色表示。

2.1 Base approach by Garcia et al

Using Inter Packet Delay (IPD) for inferring the physical layer rates was first published in [14], and then further analyzed and extended in [12]. Their data collection contains, 2195 runs of durations from 9ms to 15ms. Each run consists of multiple TCP connections to a destination server in Stockholm using the Ookla speed test tool [15]. The measurements were conducted using an Ethernet connected Starlink Standard Kit Gen2 [16] in Karlstad, Sweden. Traffic was collected using tcpdump and the IPDs were obtained using the tc-ebpf hook, by collecting hardware timestamps from the NIC.

The collected timings are then used to identify bursts. Within each burst, Garcia et al. identified clear periodic patterns in so-called sub-bursts (cf. Fig. 1). The authors analyzed these sub-bursts and found a relation to the physical layer OFDMA frame structure. The pacing of consecutive sub-bursts is exactly 1.33ms, which is the same value as the physical frame time reported in [11], concluding that sub-bursts directly correspond to individual frames. Next, they derived a set of commonly used burst rates and analyzed their usage frequency. Moreover, they built a modulation model and matched modulation types to their observed rates. The model yields 287 symbols on 1000 subcarriers with 4QAM as base modulation, increasing by 18 symbol steps to 16QAM. Thus, the resulting rate increments always occur in steps of 27 Mbit/s.

[12] provides an extensive analysis of burst rate transitions and comes up with a state graph to show which changes from current rates to higher or lower rates can be observed. Additionally, it performs a pattern mining on burst rate, burst duration, and interburst silent time. The resulting patterns were then used to train an Ngram model which can predict the upcoming burst characteristics.

利用 包间延迟（Inter Packet Delay, IPD） 来推断物理层速率的方法最早发表于[14]，随后在[12]中得到了进一步的分析和扩展。他们的数据集包含2195次测量，每次持续时间为9毫秒至15毫秒。每次测量都包含使用Ookla测速工具[15]与位于斯德哥尔摩的目标服务器建立的多条TCP连接。这些测量在瑞典卡尔斯塔德进行，设备为通过以太网连接的Starlink第二代标准套件[16]。流量通过tcpdump进行采集，而IPD则利用tc-ebpf钩子从网卡（NIC）收集硬件时间戳来获取。

采集到的时序数据随后被用于识别数据突发（bursts）。在每个突发内部，Garcia等人识别出了所谓的“子突发”（sub-bursts）中清晰的周期性模式（参见图1）。作者们分析了这些子突发，并发现了其与物理层OFDMA帧结构之间的关联。连续子突发之间的间隔恰好为1.33毫秒，这与[11]中报告的物理帧时长完全相同，由此得出结论：子突发与单个物理帧直接对应。接下来，他们推导出了一组常用的突发速率，并分析了其使用频率。此外，他们还建立了一个调制模型，并将不同的调制类型与观测到的速率进行匹配。该模型设定在1000个子载波上分配287个符号，以4QAM为基础调制，并通过18个符号的步长增加到16QAM。因此，最终的速率增量总是以27 Mbit/s的步长出现。

文献[12]对突发速率的转换进行了详尽分析，并构建了一个状态图来展示当前速率可以向更高或更低速率转换的各种情况。此外，该研究还对突发速率、突发持续时间和突发间静默时间进行了模式挖掘。挖掘出的模式被用于训练一个Ngram模型，该模型可以预测即将到来的突发特征。

2.2 Improved methodology

We use UDP instead of TCP for the throughput measurements. Using UDP ensures that the link is fully saturated, which is not guaranteed with TCP due to its congestion control and fairness mechanisms. Higher throughput can be achieved using UDP compared to TCP, indicating that TCP is often unable to fully utilize the available link capacity [1, 17]. Saturating the link is essential for IPD, as insufficient load can lead to bursts that are only partially filled. This, in turn, results in incomplete or irregular sub-burst patterns, undermining the accuracy of burst-based inference. We illustrate the issue of underutilized frames and the resulting sub-bursts, in Fig. 1 and indicate the improvements we applied in our study.

In total, we measured UDP throughput for one year, while collecting IPD data using the eBPF application provided by Garcia et al. This high number of captured packets allows us to analyze the data on a frame-based level, instead of a burst level, as the number of outliers in individual frames is comparatively low compared to the number of relevant frames. We ground our following analysis on this extensive dataset, providing a more in-depth characterization of the Starlink physical layer properties.

在吞吐量测量中，我们使用UDP代替TCP。

• 使用UDP可以确保链路被完全饱和，而TCP由于其拥塞控制和公平性机制，无法保证这一点
• 与TCP相比，使用UDP可以实现更高的吞吐量，这表明TCP常常无法充分利用可用链路容量[1, 17]

对于IPD分析而言，饱和链路至关重要，因为负载不足会导致数据突发仅被部分填充。这进而导致子突发模式不完整或不规则，从而损害了基于突发的推断方法的准确性。我们在图1中展示了帧利用率不足的问题及其导致的子突发形态，并标明了我们在研究中应用的改进之处。

我们总共进行了一年的UDP吞吐量测量，同时使用Garcia等人提供的eBPF应用程序收集IPD数据。如此大量的捕获数据包使我们能够在帧级别而非突发级别上分析数据，因为与有效帧的总数相比，单个帧中的异常值数量相对较少。我们将后续分析建立在这个庞大的数据集之上，从而对Starlink物理层的特性进行更深入的刻画。

Measurement Setup

To accurately assess the impact of environmental conditions on Starlink performance, we established a long-term measurement setup monitoring network performance. The setup comprises a Raspberry Pi 5 at the measurement site and a virtual machine running an iPerf server at a local datacenter, which guarantees a minimum bandwidth of 5 Gbit/s to the public internet, ensuring the physical Starlink setup remains the bottleneck. To avoid obstructions, the whole setup was installed outdoors on a small hill free from trees or comparable obstacles. Owing to the scientific nature of the site, interference from consumer devices or aircraft can be excluded. The system was operational from May 2024 to May 2025, with brief interruptions for technical maintenance. A Starlink Standard Kit Gen2 [16] with a Starlink Router configured in bypass mode is used, disabling local NAT and routing functions – though the precise network-layer details of this bypass mode remain undocumented. The router was connected via an additional Starlink Ethernet adapter to the Raspberry Pi’s Gigabit Ethernet port. Measurements were conducted within an isolated network namespace containing this interface exclusively, preventing side traffic from OS processes. An auxiliary USB network interface facilitated remote management and data transfer to a central database without disrupting the ongoing measurements. We observed terminal operating software versions via the gRPC-API from a49fe058-7b35-4c55-aa40-1fbeeb59fe65 over 2024.10.13.mr44429 (when the numbering scheme changed) to 2025.05.11.mr55451.1. We continuously measured raw network throughput using the standard iPerf3 tool in UDP mode. By transmitting packets at a rate (set via the -b option) exceeding the observed limits – 100 Mbit/s upload and 500 Mbit/s download – we ensured full network utilization. Round-trip times and dish parameters were concurrently collected to track infrastructural changes and support analysis. The campaign transferred over 360 billion packets and roughly 500 TB of data, enabling detection of both transient and persistent performance variations, and yielded a total of 1.7 TB of measurement results.

为精确评估环境条件对Starlink性能的影响，我们建立了一个长期测量平台以监测网络性能。该平台由位于测量点的一台树莓派5（Raspberry Pi 5）和一台在本地数据中心运行iPerf服务器的虚拟机组成。该数据中心保证了至少5 Gbit/s的公网带宽，从而确保物理上的Starlink设备始终是整个链路的瓶颈。为避免信号遮挡，整套设备安装于室外一处没有树木或类似障碍物的小山丘上。由于该测量点具有科研性质，可以排除来自消费级设备或飞行器的干扰。

该系统从2024年5月至2025年5月持续运行，期间仅因技术维护有过短暂中断。实验使用了一套Starlink第二代标准套件[16]，其路由器被配置为旁路模式（bypass mode），从而禁用了本地NAT和路由功能——尽管该旁路模式精确的网络层细节尚无公开文档说明。该路由器通过一个额外的Starlink以太网适配器连接到树莓派的千兆以太网端口。为防止操作系统进程产生无关流量干扰，所有测量均在一个仅包含该以太网接口的隔离网络命名空间（network namespace）内进行。一个辅助的USB网络接口用于远程管理和向中央数据库传输数据，且不会干扰正在进行的测量。

我们通过gRPC-API观测到终端操作软件版本从 a49fe058-7b35-4c55-aa40-1fbeeb59fe65 演进至 2024.10.13.mr44429（此时版本号命名方案发生变更），最终更新至 2025.05.11.mr55451.1。我们使用标准的iPerf3工具，在UDP模式下持续测量原始网络吞吐量。通过以超过观测极限（上行100 Mbit/s，下行500 Mbit/s）的速率（通过-b选项设置）发送数据包，我们确保了网络得到充分利用。往返时延（Round-trip times）和碟形天线参数也被同步收集，以追踪基础设施的变化并为分析提供支持。整个测量活动共传输了超过3600亿个数据包和约500 TB的数据，从而能够检测到瞬时性和持续性的性能变化，并最终产生了总计1.7 TB的测量结果。