跳转至

FjordLink: Comparison of Starlink and 5G Networks for Teleoperated Vessel Control

The rapid growth of Low Earth Orbit satellite networks, such as Starlink, is increasing global connectivity by enabling low-latency broadband access in regions where wired and cellular networks fall short. Prior research focuses on the performance of Starlink in terrestrial settings. Yet, there is limited research on the performance of Starlink in coastal and maritime environments, raising the question of how Starlink performs in the presence of waves and tides.

In this paper, we introduce the FjordLink, a combined Starlink and 5G dataset for coastal maritime connectivity. We collect over 500,000 measurements using a Flat High Performance dish and 5G modems on a research vessel for four months. Starlink and 5G networks achieve median RTTs of less than 50 ms and mean upload throughputs exceeding 35 Mbps. Our results show that Starlink operates similarly (e.g., with a 10 ms median latency difference) in both maritime and terrestrial environments, and improves the 99th percentile latency compared to 5G networks. As a case study, we utilize traces from FjordLink in emulation to evaluate BBR, CUBIC, and Reno congestion control algorithms, where BBR achieves 18% higher upload throughput than CUBIC and Reno.

以星链（Starlink）为代表的低地球轨道（LEO）卫星网络的快速发展，正通过在有线和蜂窝网络覆盖不足的地区提供低延迟宽带接入，来增强全球连接性。以往的研究主要关注星链在陆地环境中的性能。然而，关于星链在沿海和海域环境中性能的研究有限，这引出了一个问题： 星链在存在波浪和潮汐的情况下性能如何。

在本文中，我们介绍了 FjordLink，这是一个结合了星链和5G的沿海海域连接性数据集。我们在四个月的时间里，使用一艘研究船上的平板式高性能天线和5G调制解调器，收集了超过50万次测量数据。星链和5G网络实现了低于50毫秒的中位RTT和超过35 Mbps的平均上传吞吐量。

我们的研究结果表明，星链在海域和陆地环境中的运行表现相似（例如，中位延迟差异为10毫秒），并且与5G网络相比，其99百分位延迟更优。作为一个案例研究，我们利用FjordLink的轨迹数据在仿真环境中评估了BBR、CUBIC和Reno拥塞控制算法，其中BBR的上传吞吐量比CUBIC和Reno高出18%。

Introduction

The number of Low Earth Orbit (LEO) satellites, such as Starlink, is increasing in popularity and reaching more than 1 million active users worldwide [20]. LEO satellites provide significantly lower latencies than Geosynchronous Equatorial Orbit (GEO) satellites, making them a viable alternative for connectivity in remote regions where wired and cellular broadband networks fall short [7, 10]. While the coverage of cellular networks requires a wide deployment of base stations, the operation of LEO satellites requires a stable platform to mount the dish and an unobstructed view of the sky. LEO satellites not only provide extra coverage but can also integrate with existing and future cellular systems [1]. Thus, LEO satellites are a prime alternative to cellular networks for teleoperated driving, such as remotely controlled and autonomous vehicles [19] due to their low latency, sufficiently high bandwidth, and near-global coverage. Existing research mainly focuses on a terrestrial setting for the performance analysis of LEO satellites [3, 6, 12–14, 18] and tuning congestion control algorithms (CCAs) [8, 9]. Unlike land-based environments, the coastal maritime domain poses unique communication challenges, such as varying waves and tidal movements that affect line-of-sight due to unstable dish orientation. Therefore, we need new datasets to, for example, (1) compare the performance of LEO satellites and 5G networks in inland waterways and (2) test and develop CCAs.

In this paper, we introduce FjordLink 1 , a combined Starlink and 5G dataset, collected using the Flat High Performance (FHP) kit during an extensive measurement campaign over 4 months, with more than 500,000 data points collected using a research vessel driving up to 20 km/h. Using the FjordLink, we compare the performance of Starlink versus 5G networks in supporting remote control and monitoring operations in coastal maritime routes. Under mobility, results show median round-trip times of 46 ms, 38 ms, and 40 ms for Starlink, 5G NSA, and 5G SA, respectively. The median jitter is 3.78 ms, 6.42 ms, and 6.18 ms for Starlink, 5G SA, and 5G NSA, respectively. With a target bitrate of 50 Mbps, Starlink achieves a mean upload speed of 35 Mbps, compared to 49 Mbps for 5G NSA and 41 Mbps for 5G SA. Compared to being docked in a port, the performance of Starlink while moving stays relatively similar (e.g., 10% difference in median and mean latencies), but 5G networks suffer up to 10 times higher 99th percentile delays, thus making Starlink the better network when reaching the cellular edge. We also evaluate BBR, CUBIC, and Reno CCAs as a case study using the FjordLink. BBR achieves 18.6% and 19.4% higher upload throughput than CUBIC and Reno, respectively. However, this costs 5 and 2.3 times more than CUBIC and Reno, respectively, in terms of 99th-percentile retransmissions.

低地球轨道（LEO）卫星，如星链（Starlink），其数量正日益增多，全球活跃用户已超过100万[20]。LEO卫星提供的延迟远低于地球同步轨道（GEO）卫星，这使其成为有线和蜂窝宽带网络覆盖不足的偏远地区的一种可行替代方案[7, 10]。蜂窝网络的覆盖需要广泛部署基站，而LEO卫星的运行则需要一个稳定的平台来安装天线，并保证天空视野无遮挡。LEO卫星不仅提供了额外的覆盖范围，还可以与现有及未来的蜂窝系统集成[1]。因此，由于其低延迟、足够高的带宽和近乎全球的覆盖范围，LEO卫星成为远程操控驾驶（如远程控制和自动驾驶车辆）中蜂窝网络的主要替代方案[19]。现有的研究主要集中在陆地环境下对LEO卫星的性能分析[3, 6, 12–14, 18]以及拥塞控制算法（CCAs）的调优[8, 9]。 与陆地环境不同，沿海海域环境带来了独特的通信挑战，例如变化的波浪和潮汐运动会因天线朝向不稳定而影响视线。因此，我们需要新的数据集来（1）比较LEO卫星和5G网络在内陆水道的性能，以及（2）测试和开发拥塞控制算法。

在本文中，我们介绍了 FjordLink¹，一个结合了星链和5G的数据集。该数据集是通过为期4个多月的广泛测量活动，使用平板式高性能（FHP）套件在一艘以最高20公里/小时速度行驶的研究船上收集的，包含超过50万个数据点。利用FjordLink，我们比较了星链与5G网络在支持沿海海域航线的远程控制和监控操作方面的性能。在移动状态下，星链、5G NSA和5G SA的中位往返时间分别为46毫秒、38毫秒和40毫秒。星链、5G SA和5G NSA的中位抖动分别为3.78毫秒、6.42毫秒和6.18毫秒。在目标比特率为50 Mbps的情况下，星链的平均上传速度为35 Mbps，而5G NSA和5G SA分别为49 Mbps和41 Mbps。与停靠在港口时相比，星链在移动中的性能保持相对稳定（例如，中位和平均延迟差异为10%），但5G网络在到达蜂窝网络边缘时，其99百分位延迟会高出10倍之多，这使得星链成为更好的选择。我们还使用FjordLink作为案例研究，评估了BBR、CUBIC和Reno拥塞控制算法。BBR的上传吞吐量分别比CUBIC和Reno高18.6%和19.4%。然而，其代价是在99百分位重传次数上分别比CUBIC和Reno高出5倍和2.3倍。

Overall, this paper makes the following contributions:

• Introduces the FjordLink, a combined Starlink and 5G dataset with more than 500,000 data points from a research vessel in coastal maritime routes.

• Analyzes the FjordLink dataset regarding latency, jitter, and throughput, and compares it to 5G networks. Under mobility, median round-trip times are 46 ms, 38 ms, and 40 ms for Starlink, 5G NSA, and 5G SA, respectively. With a target bitrate of 50 Mbps, Starlink achieves a mean upload speed of 35 Mbps, compared to 49 Mbps for 5G NSA and 41 Mbps for 5G SA.

• Demonstrates the utility of the FjordLink by evaluating the performance of CCAs under Starlink traces in an emulation environment, where BBR achieves 18% higher upload throughput than CUBIC and Reno.

The remainder of this paper is organized as follows: Section 2 presents background and related work, Section 3 details the measurement campaign, Section 4 analyzes the FjordLink dataset, and Section 5 concludes the paper.

总的来说，本文的贡献如下：

  • 介绍了FjordLink，一个在一艘研究船上于沿海海域航线收集的、包含超过50万个数据点的星链与5G组合数据集。
  • 分析了FjordLink数据集在延迟、抖动和吞吐量方面的表现，并将其与5G网络进行比较。在移动状态下，星链、5G NSA和5G SA的中位往返时间分别为46毫秒、38毫秒和40毫秒。在目标比特率为50 Mbps的情况下，星链的平均上传速度为35 Mbps，而5G NSA和5G SA分别为49 Mbps和41 Mbps。
  • 展示了FjordLink的实用性，通过在仿真环境中使用星链的轨迹数据评估拥塞控制算法的性能，结果表明BBR的上传吞吐量比CUBIC和Reno高18%。

本文的其余部分安排如下：第2节介绍背景和相关工作，第3节详述测量活动，第4节分析FjordLink数据集，第5节进行总结。

In this section, we first provide a background on Starlink. Next, we examine the related work in Starlink measurements.

在本节中，我们首先提供关于星链的背景知识，然后探讨星链测量方面的相关工作。

SpaceX operates Starlink, a LEO satellite constellation, providing broadband Internet service for global network coverage. As of May 2025, Starlink includes over 6,750 satellites in orbit [20]. Starlink mainly connects to a new satellite every 15 seconds at fixed 12-27-42-57 seconds of a minute [9]. With latencies ranging from 25 to 60 ms, Starlink advertises download and upload speeds of up to 220 Mbps and 25 Mbps, respectively [21]. Starlink directs traffic to Points of Presence (PoP) locations where satellites connect to the Internet.

SpaceX公司运营的星链是一个LEO卫星星座，旨在提供覆盖全球的宽带互联网服务。截至2025年5月，星链在轨卫星已超过6,750颗[20]。星链主要在每分钟的固定时间点（第12、27、42、57秒）与新卫星建立连接，切换周期为15秒[9]。星链宣传的延迟范围为25至60毫秒，下载和上传速度分别高达220 Mbps和25 Mbps[21]。星链将流量导向接入点（PoP）位置，卫星在此处连接到互联网。

Prior works mainly analyze the performance of Starlink in static and mobile scenarios on terrestrial environments [3, 6, 11–18]. For example, Ma et al. [16] measure latency, UDP and TCP throughput, packet loss, and routing information on a driving van. López et al. [15] collect latency measurements on a car in a rural area. Hu et al. [6] present a measurement study of Starlink and cellular networks on a car, examining upload and download throughput, latency, and packet loss, and investigating the potential benefits of enabling multipath communications. Beckman et al. [3] analyze the performance of Starlink under mobility and compare it to a cellular network around the Arctic Circle. Mohan et al. [18] compare the performance of Starlink measurements across 34 countries against cellular and fiber infrastructure. Laniewski et al. [14] collect and analyze the performance of Starlink in a static setting across changing weather conditions. Laniewski et al. [12, 13] also analyze the performance of Starlink under mobility and compare it to prior static measurements.

Unlike prior works, the FjordLink introduces a Starlink dataset for performance analysis in coastal maritime routes and compares it to 5G NSA and 5G SA networks. We measure inbound, outbound, and round-trip latency at a higher resolution of 10 Hz using the Two-Way Active Measurement Protocol (TWAMP) [2] and bidirectional throughput using iPerf3 2 in UDP mode. We also show the usefulness of the FjordLink in emulation and compare CCAs under latency, jitter, and throughput constraints.

以往的工作主要分析星链在陆地环境下静态和移动场景中的性能[3, 6, 11–18]。例如，Ma等人[16]在一辆行驶的货车上测量了延迟、UDP和TCP吞吐量、丢包率以及路由信息。López等人[15]在一辆位于农村地区的汽车上收集了延迟测量数据。Hu等人[6]在一辆汽车上对星链和蜂窝网络进行了测量研究，检测了上传和下载吞吐量、延迟和丢包率，并探讨了启用多路径通信的潜在好处。Beckman等人[3]分析了星链在移动状态下的性能，并与北极圈周围的蜂窝网络进行了比较。Mohan等人[18]比较了34个国家中星链的测量性能，并与蜂窝和光纤基础设施进行了对比。Laniewski等人[14]在变化的的天气条件下，于静态环境中收集并分析了星链的性能。Laniewski等人[12, 13]还分析了星链在移动状态下的性能，并与之前的静态测量结果进行了比较。

与以往的研究不同，FjordLink引入了一个用于沿海海域航线性能分析的星链数据集，并将其与5G NSA和5G SA网络进行了比较。我们使用双向主动测量协议（TWAMP）[2]以10赫兹的更高分辨率测量了入向、出向和往返延迟，并使用iPerf3²在UDP模式下测量了双向吞吐量。我们还展示了FjordLink在仿真中的实用性，并在延迟、抖动和吞吐量约束下比较了不同的拥塞控制算法。

Measurement Campaign

In this section, we detail our measurement campaign. We show an overview of our measurement setup and location in Figures 1 and 2. We collect data from the research vessel MS Wavelab, which operates on the Kiel Fjord, Germany, along coastal maritime routes, see our GitHub repository for a detailed coverage map. The dataset spans over 4 months.

We base our cellular measurement campaign on a similar architecture outlined in the Fjord5G dataset [4]. We collect data from a Starlink Flat High Performance (FHP) dish and two 5G routers, one with Vodafone 5G SA service (operator #1) and the other with Deutsche Telekom 5G NSA service (operator #2). While Vodafone uses 5G bands n28, n3, and n78, Deutsche Telekom utilizes 5G bands n28, n1, and n78 in addition to LTE bands. While Starlink is mostly accessible with a clear line of sight, 5G performance degrades the further we are away from the coast and the base stations.

We show an overview of captured features in Table 1. We use TWAMP to measure one-way latency, round-trip time (RTT), and jitter. Additionally, we use iPerf3 to measure upload and download throughput as well as jitter using UDP. While TWAMP measures data at 10 Hz, both iPerf3 and Starlink dish provide data at 1 Hz. We measure latency, jitter, and throughput simultaneously for 30-second chunks, then wait for 15 seconds and repeat. Using 30-second chunks, we aim to get at least one satellite handover event, possibly two, in the logs for our analysis, see examples in Figures 3 and 6. Additionally, we provide weather data offered by the German Weather Service 3 from stations in Kiel. However, we leave the correlation analysis of weather data to future work.

在本节中，我们详细介绍我们的测量活动。我们在图1和图2中展示了测量设置和位置的概览。我们在研究船MS Wavelab号上收集数据，该船在德国基尔峡湾的沿海海域航线上运行，详细的覆盖地图请参阅我们的GitHub仓库。该数据集的采集时间跨度超过4个月。

我们的蜂窝网络测量活动基于Fjord5G数据集[4]中概述的类似架构。我们从一个星链平板式高性能（FHP）天线和两个5G路由器收集数据，其中一个使用沃达丰（Vodafone）的5G SA（独立组网）服务（运营商＃1），另一个使用德国电信（Deutsche Telekom）的5G NSA（非独立组网）服务（运营商＃2）。沃达丰使用5G频段n28、n3和n78，而德国电信除了LTE频段外，还使用5G频段n28、n1和n78。虽然星链在有清晰视线的情况下大多可用，但5G的性能会随着我们离海岸和基站越远而下降。

我们在表1中展示了所捕获特征的概览。我们使用TWAMP来测量单向延迟、往返时间（RTT）和抖动。此外，我们使用iPerf3来测量上传和下载吞吐量以及使用UDP时的抖动。TWAMP以10赫兹的频率测量数据，而iPerf3和星链天线均以1赫兹的频率提供数据。我们以30秒为时间块同时测量延迟、抖动和吞吐量，然后等待15秒再重复。通过使用30秒的时间块，我们的目标是在日志中至少捕获一次卫星切换事件，甚至可能两次，示例见图3和图6。此外，我们还提供了由德国气象局³从基尔气象站提供的天气数据。然而，我们将天气数据的相关性分析留作未来工作。