MOSAIC: Piecing Together 5G and LEOs for NTN Integration Experimentation

Despite the rapid growth of 5G technologies, geographical network coverage remains a significant challenge. In certain areas - notably rural - it is anticipated that removing these technologies will result in a complete lack of service. To address this, standards bodies, such as 3GPP, have begun advancing toward 5G-and-beyond architectures incorporating Non-Terrestrial Networks (NTNs), most notably using Low-Earth Orbit (LEO) satellite constellations to expand coverage and improve resilience of 5G terrestrial networks (TN). However, the integration of 5G and NTN introduces new challenges due to the nature of mobility, network characteristics, and deployment costs. To support the development of new 5G-NTN integration architectures, we propose MOSAIC (MObile-SAtellite Integration Cradle) , a realistic end-to-end 5G-NTN emulation platform that can recreate the unique features and software of emerging mobile infrastructures. MOSAIC offers a reproducible environment for recreating realistic 5G NTN experiments, utilizing unmodified, off-the-shelf software components. MOSAIC models NTN network characteristics using a Generalized Additive Model for Location, Scale, and Shape (GAMLSS), evaluating it against open-source satellite link measurement data from Starlink. Additionally, using our platform, we assess the performance of the Multipath TCP (MPTCP) protocol to support seamless handover scenarios between TN and NTN. We believe MOSAIC provides a holistic and open environment for experimentation with beyond 5G technologies.

尽管5G技术发展迅速，但地理网络覆盖仍然是一个重大挑战。在某些地区——特别是农村地区——预计这些技术的退役将导致服务完全中断。为解决此问题，3GPP等标准组织已开始推动包含非地面网络（NTN）的5G及未来架构，其中最引人注目的是利用低地球轨道（LEO）卫星星座来扩展5G地面网络（TN）的覆盖范围并提高其弹性。

然而，由于移动性、网络特性和部署成本的性质，5G与NTN的集成引入了新的挑战。

为了 支持新型5G-NTN集成架构的开发， 我们提出了MOSAIC（MObile-SAtellite Integration Cradle，移动-卫星集成平台）， 一个现实的端到端5G-NTN仿真平台， 能够重现新兴移动基础设施的独特功能和软件。

MOSAIC利用未经修改的现成软件组件，为复现真实的5G NTN实验提供了一个可复现的环境。MOSAIC使用位置、尺度和形状的广义可加模型（GAMLSS）对NTN网络特性进行建模，并利用来自星链的开源卫星链路测量数据对其进行评估。

此外，我们利用该平台评估了多路径TCP（MPTCP）协议在支持TN与NTN之间无缝切换场景中的性能。我们相信，MOSAIC为未来5G技术的实验提供了一个全面且开放的环境。

Introduction

5G technologies have transformed mobile infrastructures from a telephony network into a programmable platform that supports a wide range of network services. Although 5G capacity has drastically increased, coverage remains a significant challenge, particularly in rural areas, considering that mobile networks cover only 20% of habitable areas [4]. The forthcoming 2/3G decommission will further compound the problem.

To address these challenges, 3GPP explores the integration of NTN into the 5G ecosystem. NTNs can extend mobile network coverage to regions where terrestrial infrastructure is either unavailable, difficult to deploy, or economically impractical, such as rural, remote, or disaster-prone areas. NTNs can fill coverage gaps and provide truly ubiquitous connectivity using satellites, high-altitude platforms, and other spaceborne systems. Recent developments have moved beyond theoretical models, with real-world test deployments already underway as of 2025, showcasing practical feasibility and growing industry adoption [22].

5G TN/NTN integration requires architecture extension to improve resource utilization and guarantee user experience. For example, 3GPP proposes a series of RAN architectures (e.g. regenerative payload, O-RAN) [1] to improve spectral efficiency using on-board satellite compute resources for signal processing. Similarly, handovers between access technologies in TN/NTN 5G services pose novel challenges for service continuity on the 5G Core. Dual/ fallback connectivity proposes that each User Equipment (UE) simultaneously connects to both a TN and NTN using two parallel PDU sessions, thus enabling resiliency [2, 21]. During periods of reduced coverage from the TN, the UE uses the NTN link to maintain connectivity. Network applications can use MPTCP backup subflows to maintain connectivity during handovers and manage UE multi-homing [24].

Exploring design options towards an integrated 5G TN/NTN architecture requires a flexible, realistic, and extensible experimentation platform. Practitioners and network operators can develop reproducible experiments to assess the trade-offs between end-to-end performance, costs, and operational capabilities, thereby guiding the development of future mobile and NTN technologies.

Efforts to develop emulation platforms for 5G/NTN integration remain relatively limited. Existing solutions, such as StarryNet [11] and the framework presented by [17], primarily focus on modelling satellite constellation dynamics but require substantial effort to incorporate 5G network functions. Consequently, they focus less on specific link characteristics in their model, such as latency spikes during handover [16].

5G技术已将移动基础设施从一个电话网络转变为一个支持广泛网络服务的可编程平台。尽管5G容量已大幅增加，但覆盖范围仍然是一个重大挑战，特别是在农村地区，考虑到移动网络仅覆盖了20%的宜居区域[4]。即将到来的2G/3G网络退役将进一步加剧这一问题。

为应对这些挑战，3GPP探索将NTN集成到5G生态系统中。NTN可以将移动网络覆盖扩展到那些地面基础设施不可用、部署困难或经济上不切实际的地区，如农村、偏远或灾害多发地区。NTN可以利用卫星、高空平台和其他空基系统填补覆盖空白，提供真正无处不在的连接。最近的发展已超越理论模型，截至2025年，真实世界的测试部署已在进行中，展示了其实际可行性和日益增长的行业采纳度[22]。

5G TN/NTN集成需要架构扩展 以提高资源利用率并保障用户体验:

例如，3GPP提出了一系列RAN架构（如再生有效载荷、O-RAN）[1]，利用星上计算资源进行信号处理以提高频谱效率。

同样，TN/NTN 5G服务中不同接入技术间的切换也为5G核心网的服务连续性带来了新的挑战。

双连接/回退连接方案提出，每个用户设备（UE）通过两个并行的PDU会话同时连接到TN和NTN， 从而实现弹性[2, 21]。 在TN覆盖减弱期间，UE使用NTN链路维持连接。 网络应用可以使用MPTCP的备份子流在切换期间维持连接并管理UE的多宿主特性[24]。

探索集成5G TN/NTN架构的设计选项需要一个灵活、真实且可扩展的实验平台。从业者和网络运营商可以开发可复现的实验，以评估端到端性能、成本和运营能力之间的权衡，从而指导未来移动和NTN技术的发展。

开发用于5G/NTN集成的仿真平台的工作仍然相对有限。现有解决方案，如StarryNet [11]和[17]提出的框架，主要侧重于对卫星星座动态进行建模，但需要大量工作才能整合5G网络功能。因此，它们在其模型中较少关注特定的链路特性，例如切换期间的延迟尖峰[16]。

Moving towards the goal of developing open experimentation 5G TN/NTN platforms, we present MOSAIC , an emulation platform for testing and evaluation of 5G NTN integration in a controlled environment. MOSAIC automates the deployment of custom 5G network topologies, using off-the-shelf software components. In parallel, we develop a link emulation model using open measurement from LEO satellites, which accurately captures the characteristics of NTN links, including latency and packet loss. The contributions of this paper are the following:

• We present MOSAIC , a novel emulation platform that enables testing and evaluation of end-to-end performance in 5G NTN integration scenarios.

• We develop an open link emulation model that captures the characteristics of the NTN link, including latency and packet loss.

• We utilize MOSAIC to investigate the MPTCP protocol and its congestion control (CC) algorithms in realistic TN-NTN handover scenarios, assessing their impact on throughput and latency.

For the paper organization, we first discuss related research on open network emulation platforms (§ 2) and elaborate on the limitations of 5G experimentation. Furthermore, we present MOSAIC design (§ 3), and evaluate the performance of the platform (§ 4). Finally, we conclude the paper and discuss future work (§ 5).

为了实现开放式5G TN/NTN实验平台的目标， 我们提出了MOSAIC，一个用于在受控环境中测试和评估5G NTN集成的仿真平台。 MOSAIC使用现成的软件组件，自动化部署定制的5G网络拓扑。同时，我们利用来自LEO卫星的公开测量数据开发了一个链路仿真模型，该模型能准确捕捉包括延迟和丢包在内的NTN链路特性。本文的贡献如下：

• 我们提出了MOSAIC，一个新颖的仿真平台，能够在5G NTN集成场景中进行端到端性能的测试与评估。
• 我们开发了一个开放的链路仿真模型，该模型能捕捉包括延迟和丢包在内的NTN链路特性。
• 我们利用MOSAIC在真实的TN-NTN切换场景中研究了MPTCP协议及其拥塞控制（CC）算法，评估了它们对吞吐量和延迟的影响。

关于本文的组织结构，我们首先讨论开放网络仿真平台的相关研究（§2），并阐述5G实验的局限性。接着，我们介绍MOSAIC的设计（§3），并评估该平台的性能（§4）。最后，我们总结本文并讨论未来的工作（§5）。

Related Work

Initial NTN experimentation platforms, like ns-3-LEO [20], SCNE [10] and Hypatia [7], employed simulation-based experimentation. The use of simulation, however, creates nonrealtime execution times and requires extended application rewrites to support the simulation environment runtime.

StarryNet [11] is the first attempt to emulate networking over large satellite constellations. The platform utilizes analytical models to predict satellite mobility and translate them into link characteristics and topology changes. StarryNet offers an experimentation API, integration with hardware acceleration, and pre-built models for production LEO constellations. Nonetheless, the use of analytical satellite models increases experimental execution times, ranging from 1.5 minutes to 14 minutes to start the topology, and captures partially link characteristics. Xeoverse [6] improves the scalability by optimising satellite mobility computation and precomputing topology changes. Similarly, OpenSN [13] further improves scalability by optimizing the integration with the Docker service and improving state management. Celestial [18] is a LEO experimentation platform for edge cloud services via emulating resources onboard satellites.

It is worth highlighting that 5G service coverage is typically limited geographically to a single country, and thus satellite mobility patterns are simpler than a constellation offering global coverage. Furthermore, precise link emulation has a higher impact on experimental fidelity in such scenarios than constellation dynamics. In parallel, configuring 5G functions in an emulation environment is not trivial and exhibits increased complexity. Effective 5G/NTN emulation requires new tools that simplify 5G configuration.

Commercial satellite broadband services, like Starlink, have motivated several measurement studies of NTN link characteristics. The LENS dataset [25] provides a large and open dataset of long-term RTT and packet loss data from Starlink terminals in various locations worldwide. The study of LENS has revealed that Starlink RTTs are highly variable, whilst Inter-Satellite Links (ISL) significantly impact RTT. In parallel, Mohan et al. [16] conducted a similar study by measuring network applications’ performance over Starlink and highlighted that control plane synchronization processes create noticeable periodic performance degradation. Similarly, the WetLinks dataset [12] combined satellite network performance measurements with weather data to study the impact of environmental factors on LEO satellite connectivity. These datasets provide empirical foundations for validating and refining emulator models, thereby enhancing the realism of recent simulation and emulation platforms. By incorporating trace-based characteristics from LENS and WetLinks, systems like MOSAIC can better replicate real-world link behavior, including latency and intermittent packet loss, leading to more accurate performance evaluations.

OpenAir Interface (OAI) is an open-source experimental 5G gNodeB implementation, frequently used for NTN emulation. ESA has deployed OAI instances on LEO and GEO satellites to experiment with the impact of NTN links on 5G connectivity [23], while hardware emulators are used to inject NTN link impairments to accurately replicate NTN links in 5G testbeds [8]. OAI offers enhanced RAN realism, but NTN integration remains limited [15], and requires significant computer resources to scale for large deployments. In parallel, precise NTN emulation using the OAI stack depends on hardware acceleration.

"模拟":

早期的NTN实验平台，如ns-3-LEO [20]、SCNE [10]和Hypatia [7]，采用基于模拟的实验方法。然而，使用模拟会导致非实时的执行时间， 并需要对应用程序进行大量重写以支持模拟环境的运行时。

"仿真":

StarryNet [11] 是首个尝试在大型卫星星座上进行网络仿真的平台。该平台利用分析模型来预测卫星移动性，并将其转化为链路特性和拓扑变化。StarryNet提供了一个实验API，集成了硬件加速，并为生产级LEO星座提供了预建模型。然而，使用分析性卫星模型增加了实验执行时间，启动拓扑需要1.5分钟到14分钟不等，并且只能部分捕捉链路特性。

Xeoverse [6] 通过优化卫星移动性计算和预计算拓扑变化来提高可扩展性。

同样，OpenSN [13] 通过优化与Docker服务的集成和改进状态管理进一步提高了可扩展性。

Celestial [18] 是一个LEO实验平台，通过仿真星上资源来支持边缘云服务。

值得强调的是， 5G服务的覆盖范围通常在地理上仅限于单个国家 ，因此卫星移动模式比提供全球覆盖的星座要简单。此外， 在这种场景下，精确的链路仿真对实验保真度的影响比星座动态更大。 同时，在仿真环境中配置5G功能并非易事，且复杂性日益增加。有效的5G/NTN仿真需要新的工具来简化5G的配置。

"数据集":

像星链这样的商业卫星宽带服务，催生了多项关于NTN链路特性的测量研究。

LENS数据集[25] 提供了一个大型开放的数据集，包含了全球各地星链终端的长期RTT和丢包数据。对LENS的研究揭示，星链的RTT变化极大，而星间链路（ISL）对RTT有显著影响。

同时，Mohan等人[16]通过测量网络应用在星链上的性能进行了类似的研究，并指出控制平面的同步过程会造成明显的周期性性能下降。

类似地，WetLinks数据集[12] 将卫星网络性能测量与天气数据相结合，研究环境因素对LEO卫星连接的影响。

这些数据集为验证和完善仿真器模型提供了经验基础，从而增强了近期模拟和仿真平台的真实性。通过整合来自LENS和WetLinks的基于轨迹的特性，像MOSAIC这样的系统可以更好地复制真实世界的链路行为，包括延迟和间歇性丢包，从而实现更准确的性能评估。

"开销很大的NTN集成":

OpenAir Interface (OAI) 是一个开源的实验性5G gNodeB实现，常用于NTN仿真。欧洲航天局（ESA）已在LEO和GEO卫星上部署了OAI实例，以实验NTN链路对5G连接的影响[23]，同时使用硬件仿真器注入NTN链路损伤，以在5G测试平台中精确复制NTN链路[8]。OAI提供了增强的RAN真实性，但NTN集成仍然有限[15]，并且需要大量的计算资源才能进行大规模部署。同时，使用OAI协议栈进行精确的NTN仿真依赖于硬件加速。

System Design

5G/NTN integration unlocks many parameters and architectural options that influence the performance and reliability of mobile networks. MOSAIC provides an integrated and automated emulation platform that enables researchers and network experimenters to recreate realistic 5G topologies with NTN connectivity, allowing them to evaluate end-to-end service performance. This section presents the MOSAIC architecture (§ 3.1), the design of the satellite emulation mechanism (§ 3.2), and its implementation (§ 3.3).

5G/NTN集成解锁了许多影响移动网络性能和可靠性的参数和架构选项。MOSAIC提供了一个集成且自动化的仿真平台，使研究人员和网络实验者能够重现具有NTN连接的真实5G拓扑，从而评估端到端服务性能。本节介绍MOSAIC的架构（§3.1）、卫星仿真机制的设计（§3.2）及其实现（§3.3）。

3.1 Emulation Platform

MOSAIC leverages containerization to virtualize network functions and efficiently build complex 5G topologies on a single host machine. This architecture enables scalable, modular experimentation without the overhead of deploying on physical infrastructure. A schematic overview of the system architecture is shown in Figure 1. Each MOSAIC experiment is defined by three key configuration files: the topology file, the NTN model parameters file (discussed in Section 3.2), and the evaluation scenario file.

The topology file extends the standard Docker Compose format to include experiment-specific details. It defines the layout and components of the 5G network, specifying which network functions (NFs) are deployed, how they are distributed across TN and NTN segments, the number of UPFs, SMFs, and other 5G functions, and how they are interconnected across networks. This file acts as a blueprint for the experimental setup. The Topology Controller processes this information and creates a virtual topology using Docker containers.

The evaluation scenario file, on the other hand, defines the logic for testing and monitoring within the emulated environment. It includes parameters and scripted actions for executing network tools (e.g., iperf, ping, or custom probes), collecting performance data, and orchestrating evaluation workflows. One of the strengths of MOSAIC is its ability to automate experiments beyond initial topology deployment—scenario actions can be composed into pipelines where the output of one task serves as the input for the next. Additionally, parallel action execution is supported, allowing for realistic and complex testing scenarios that facilitate efficient and repeatable experimentation workflows. The Life-cycle Controller coordinates the execution of an experimental scenario by executing actions within the namespace of each container instance.

The MOSAIC Manager coordinates experimental execution. It adopts a modular design approach, offering built-in modules to interface with the Docker daemon and the Linux network stack for deploying the topology and configuring experiment hosts. The platform manages the full lifecycle of all major 5G components—including core network functions, gNodeBs, and User Equipment (UEs)—which are deployed as Docker containers. This containerized architecture allows for highly flexible and repeatable testbed configurations, making it well-suited for a wide range of experimental scenarios.

MOSAIC利用容器化技术虚拟化网络功能，并在单台主机上高效构建复杂的5G拓扑。这种架构实现了可扩展、模块化的实验，而无需在物理基础设施上部署的开销。系统架构的示意图如图1所示。每个MOSAIC实验由三个关键配置文件定义：拓扑文件、NTN模型参数文件（在3.2节讨论）和评估场景文件。

拓扑文件扩展了标准的Docker Compose格式，以包含实验特定的细节。它定义了5G网络的布局和组件，指明了部署哪些网络功能（NF）、它们如何在TN和NTN段之间分布、UPF、SMF和其他5G功能的数量，以及它们如何在网络间互连。该文件作为实验设置的蓝图。 拓扑控制器 处理这些信息，并使用Docker容器创建一个虚拟拓扑。

另一方面，评估场景文件定义了在仿真环境中进行测试和监控的逻辑。它包括用于执行网络工具（如iperf、ping或自定义探针）、收集性能数据和编排评估工作流的参数和脚本化操作。MOSAIC的优势之一是它能够将实验自动化扩展到初始拓扑部署之外——场景操作可以组合成流水线，其中一个任务的输出可作为下一个任务的输入。此外，还支持并行操作执行，从而允许进行真实而复杂的测试场景，促进高效和可重复的实验工作流。 生命周期控制器 通过在每个容器实例的命名空间内执行操作来协调实验场景的执行。

MOSAIC管理器 协调实验的执行。它采用模块化设计方法，提供内置模块与Docker守护进程和Linux网络协议栈接口，用于部署拓扑和配置实验主机。该平台管理所有主要5G组件的完整生命周期——包括核心网功能、gNodeB和用户设备（UE）——这些组件都作为Docker容器部署。这种容器化的架构允许高度灵活和可重复的测试平台配置，使其非常适合广泛的实验场景。

3.2 Satellite Link Modelling

One of the MOSAIC key strengths is its ability to accurately emulate real-world network conditions for 5G TN/NTN networks. In contrast to satellite emulation platforms, like [11], which simulate entire satellite constellations and orbital dynamics to derive network behavior, MOSAIC adopts a fundamentally different approach: it focuses directly on the end-to-end characteristics of the NTN link—such as latency, jitter, and packet loss—based on real-world measurements. This makes MOSAIC significantly more lightweight, scalable, and easier to integrate into experimental workflows. MOSAIC achieves by integrating satellite link characteristics sourced from the LENS dataset [25], a repository of Starlink RTT latency measurements from multiple geographically dispersed hosts. We utilize the Linux kernel netem queue in the traffic control subsystem to emulate latency, loss, and bandwidth characteristics, thereby modeling critical link properties and providing a realistic and controlled networking environment across both terrestrial and satellite paths. To apply these link characteristics, we embed a link adaptor function in the experimental topology to connect different network segments. A link controller daemon dynamically and in real-time manages the adaptor network configuration using a netlink socket, according to the link model.

The input values for the network emulation tools require upper and lower bounds for the latency, and values for the jitter and packet loss. This creates a modeling challenge distinct from the usual time series forecasting problem, as we need to estimate the distribution of time series values at a given time – to obtain the quantiles – rather than a single expected value. To further complicate this problem, there is a need to capture nonstationarity in the model parameters over time, as the data exhibits a time-varying trend. In practice, it is possible to show that the data are trend-stationary per one-minute segment using a KPSS test [9], however, and we show this is true for a period of LENS data in our evaluation. We exploit this property in our model, eliminating the need to fit the model for each recorded packet, as this would quickly become computationally impractical.

To address these issues, we propose a statistical modelling approach based on a Generalized Additive Model for Location, Scale, and Shape (GAMLSS) as this offers a convenient way to capture the nonstationarity and simultaneously estimate both point values and the distribution over time.

Let \(t\) index the trend stationary interval of time, for example a minute, and let \(t \le T\) where \(T\) is the time horizon. Suppose that \(T_t\) latency values are recorded during this interval, and let \(y_{t1}, \dots, y_{tT_t}\) denote these latencies. We assume that each is independent and identically distributed following the data-generating process \(Y_t \sim N(\mu(t), \sigma(t))\). Note that the mean and variance are themselves nonstationary over the different time intervals. Under this, we have the following model for the latencies:

\(y_t = \mu(t) + \epsilon_t\), (1)

where \(\epsilon_t \sim N(0, \sigma^2(t))\).

The nonstationary mean and variance are modelled using a B-spline expansion as follows:

\(\mu(t) = \sum_{k=1}^{K} \beta_k B_k(t) \quad \log(\sigma^2(t)) = \sum_{k=1}^{K} \gamma_k B_k(t)\).

Here \(K\) is the number of splines in the system, and \(\{B_k(t)\}\) are the K columns of the spline system design matrix. Note that the reason the log-variance is smoothed rather than the variance directly is that this ensures the estimated variance is positive. As described in Rigby and Stasinopoulos [19] this model can be calculated using a variant of penalised maximum likelihood estimation. Once the GAMLSS has been estimated for each trend stationary period, it is possible to identify the NTN model parameters. We use the 2.5% and 97.5% quantiles of the \(N(\mu(t), \sigma^2(t))\) distribution for the latencies bound, and the estimated variance for the jitter.

The last parameter in the model is the packet loss proportion, which we propose to incorporate by modelling the loss proportion on each stationary interval. We denote this proportion as \(p_t\), and assume they are independent and identically distributed as a Beta distribution with parameters \(a\) and \(b\); that is, \(P_t \sim Beta(a,b)\). This can be estimated from the LENS data using maximum likelihood estimation.

As a final comment, we remark that the GAMLSS model in equation (1) can be fitted to the LENS round-trip time (RTT) data, giving a model for the RTT. The advantage of doing this is that one obtains a data-driven model for the link latencies that captures both nonstationarity and packet loss. Furthermore, one can use the Mathis model to estimate the link's TCP throughput based on this RTT model:

\(T = \frac{MSS}{RTT} \times \frac{C}{\sqrt{p}}\). (2)

Here, \(MSS\) is the maximum segment size, \(C\) is a constant, and \(p\) denotes the packet loss.

MOSAIC的关键优势之一是其能够为5G TN/NTN网络精确仿真真实世界的网络条件。与像[11]这样的卫星仿真平台（它们模拟整个卫星星座和轨道动力学来推导网络行为）相比，MOSAIC采用了一种根本不同的方法：它基于真实世界的测量，直接关注NTN链路的 端到端特性 —— 如延迟、抖动和丢包。这使得MOSAIC更加轻量、可扩展，并且更容易集成到实验工作流中。MOSAIC通过整合来自 LENS数据集[25] 的卫星链路特性来实现这一点，该数据集是来自多个地理位置分散的主机的星链RTT延迟测量的大型存储库。我们利用Linux内核中的netem队列在流量控制子系统中仿真延迟、丢包和带宽特性，从而为跨地面和卫星路径的关键链路属性建模，并提供一个真实且受控的网络环境。为了应用这些链路特性，我们在实验拓扑中嵌入了一个 链路适配器 功能来连接不同的网络段。一个 链路控制器 守护进程根据链路模型，使用netlink套接字动态地实时管理适配器的网络配置。

网络仿真工具的输入值需要延迟的上限和下限，以及抖动和丢包的值。这产生了一个与通常的时间序列预测问题不同的建模挑战，因为我们需要估计给定时间的 时间序列值的分布 —— 以获得分位数 —— 而不是单个期望值。使这个问题进一步复杂化的是，需要在模型参数中捕捉 非平稳性 ，因为数据显示出随时间变化的趋势。在实践中，可以使用KPSS测试[9]证明数据在每分钟的段内是趋势平稳的，并且我们在评估中也证明了LENS数据在一段时间内确实如此。我们在模型中利用了这一特性，从而无需为每个记录的数据包拟合模型，因为那样做很快就会变得计算上不切实际。

为解决这些问题，我们提出了一种基于 位置、尺度和形状的广义可加模型（GAMLSS） 的统计建模方法，因为它提供了一种便捷的方式来捕捉非平稳性，并同时估计点值和时间上的分布。

让 \(t\) 表示趋势平稳的时间区间索引，例如一分钟，并设 \(t \le T\)，其中 \(T\) 是时间范围。假设在该区间内记录了 \(T_t\) 个延迟值，并让 \(y_{t1}, \dots, y_{tT_t}\) 表示这些延迟。我们假设每个值都是独立同分布的，遵循数据生成过程 \(Y_t \sim N(\mu(t), \sigma(t))\)。注意，均值和方差本身在不同的时间区间上是非平稳的。在此之下，我们有以下延迟模型：

\(y_t = \mu(t) + \epsilon_t\)， (1)

其中 \(\epsilon_t \sim N(0, \sigma^2(t))\)。

非平稳的均值和方差使用B样条展开进行建模，如下所示：

\(\mu(t) = \sum_{k=1}^{K} \beta_k B_k(t) \quad \log(\sigma^2(t)) = \sum_{k=1}^{K} \gamma_k B_k(t)\)。

这里 \(K\) 是系统中的样条数量，\(\{B_k(t)\}\) 是样条系统设计矩阵的K列。注意，平滑处理的是对数方差而不是方差本身，这是为了确保估计的方差为正。如Rigby和Stasinopoulos [19]所述，该模型可以使用一种惩罚最大似然估计的变体进行计算。一旦为每个趋势平稳周期估计了GAMLSS，就可以确定NTN模型参数。我们使用 \(N(\mu(t), \sigma^2(t))\) 分布的2.5%和97.5%分位数作为延迟的界限，并使用估计的方差作为抖动。

模型中的最后一个参数是丢包比例，我们建议通过对每个平稳区间上的丢包比例进行建模来整合它。我们将此比例表示为 \(p_t\)，并假设它们是独立同分布的，服从参数为 \(a\) 和 \(b\) 的贝塔分布；即 \(P_t \sim Beta(a,b)\)。这可以从LENS数据中使用最大似然估计来估算。

最后，值得一提的是，方程(1)中的GAMLSS模型可以拟合到LENS的往返时间（RTT）数据，从而得到一个RTT模型。这样做的好处是，可以获得一个数据驱动的链路延迟模型，该模型同时捕捉了非平稳性和丢包。此外，可以使用马西斯（Mathis）模型基于此RTT模型来估计链路的TCP吞吐量：

\(T = \frac{MSS}{RTT} \times \frac{C}{\sqrt{p}}\)。 (2)

这里，\(MSS\) 是最大报文段长度，\(C\) 是一个常数，\(p\) 表示丢包率。

3.3 Implementation

The MOSAIC manager, written in Go, contains modules to interact with the Docker and NetLink APIs and dynamically control the network topology. MOSAIC supports out-of-the-box the Fraunhofer FOKUS Open5GCore [3] platform, used in our evaluation, the UERANSIM UE and RAN emulator, and the open Free5GCore 5G Core. The codebase offers a range of 5G-NTN topologies, and the baseline topology features a dual-path setup in which the UE connects to the internet via two independent Data Networks (DN) served by a dedicated UPF: a ground DN (TN path) and a satellite DN (NTN path). Experimenters can utilize this dual connectivity scenario to evaluate approaches that enhance connectivity resilience in 5G/NTN services. For example, the satellite DN acts as a fallback when the ground DN is unavailable – whether due to a network failure or during transient events such as a handover procedure, where the UE may temporarily lose connectivity.

All experiments were conducted on a Dell PowerEdge server equipped with an Intel Xeon E5-2630 v3 CPU (2.40 GHz, 16 cores) and 32GB of RAM, running Ubuntu 24.04 LTS (Noble). Throughout the entire lifecycle—from topology setup to UE-to-server traffic tests using iPerf—the platform maintained a lightweight resource footprint. CPU usage remained below 20 % per core, and total memory consumption did not exceed 3GB, demonstrating the system’s efficiency and suitability for scalable experimentation.

用Go语言编写的MOSAIC管理器包含与Docker和NetLink API交互并动态控制网络拓扑的模块。MOSAIC开箱即用地支持Fraunhofer FOKUS Open5GCore [3]平台（在我们的评估中使用）、UERANSIM UE和RAN仿真器，以及开放的Free5GCore 5G核心网。代码库提供了一系列5G-NTN拓扑，基线拓扑具有双路径设置，其中UE通过两个由专用UPF服务的独立数据网络（DN）连接到互联网：一个地面DN（TN路径）和一个卫星DN（NTN路径）。实验者可以利用这种双连接场景来评估在5G/NTN服务中增强连接弹性的方法。例如，当由于网络故障或在切换等瞬态事件期间（UE可能暂时失去连接）地面DN不可用时，卫星DN可作为后备。

所有实验均在一台戴尔PowerEdge服务器上进行，该服务器配备了英特尔至强E5-2630 v3 CPU（2.40 GHz，16核）和32GB内存，运行Ubuntu 24.04 LTS（Noble）。在整个生命周期中——从拓扑设置到使用iPerf进行UE到服务器的流量测试——该平台都保持了轻量级的资源占用。CPU使用率保持在每核20%以下，总内存消耗不超过3GB，证明了该系统的效率及其对可扩展实验的适用性。

MOSAIC Evaluation

In this section, we present the evaluation of MOSAIC , focusing on the accuracy of our link emulation model and evaluating the performance during handovers using different congestion control algorithms in the MultiPath TCP (MPTCP) protocol, a recommended approach to improve resilience.

The evaluation uses the topology in Figure 2. Our scenario considers two gNodeBs (a terrestrial and a non-terrestrial operating on board the satellite), each served by a dedicated User Plane Function (UPF) and offering coverage for our test UE. The UE establishes two parallel PDU sessions via both gNodeBs, using different Subscription Permanent Identifiers (SUPI). We enable MultiPath TCP (MPTCP) support on the UE, using the TN link for the primary subflow and the NTN link for a backup link. We use the Open5Gcore software stack [3] to emulate the 5G core, RAN, and UE components. The 5G Core is configured with two Data Networks, each serving the TN and NTN links respectively. The link emulator applies a fixed latency between the TN RAN and the Core of 10 msec, while the link between the NTN RAN and the Core is emulated using the model described in Section 3.2. We also operate an iperf3 server (IPERF) and an echo service (ECHO) in the core network to support transport layer measurements of bandwidth and latency.

在本节中，我们对 MOSAIC 进行评估，重点关注我们链路仿真模型的准确性，并评估在使用多路径TCP（MPTCP）协议（一种推荐用于提高弹性的方法）中不同拥塞控制算法时的切换性能。

评估采用了图2所示的拓扑结构:

我们的场景考虑了两个gNodeB（一个地面gNodeB和一个在星上运行的非地面gNodeB）

每个gNodeB都由一个专用的用户面功能（UPF）提供服务，并为我们的测试用户设备（UE）提供覆盖

UE通过这两个gNodeB建立两个并行的PDU会话，并使用不同的订阅永久标识符（SUPI）。我们在UE上启用了多路径TCP（MPTCP）支持，使用TN链路作为主子流，NTN链路作为备份链路

我们使用Open5Gcore软件栈[3]来仿真5G核心网、RAN和UE组件。5G核心网配置了两个数据网络，分别服务于TN和NTN链路。链路仿真器在TN RAN与核心网之间应用了10毫秒的固定延迟，而NTN RAN与核心网之间的链路则使用第3.2节中描述的模型进行仿真。我们还在核心网中运行一个iperf3服务器（IPERF）和一个echo服务（ECHO），以支持对带宽和延迟的传输层测量。