INTRODUCTION

With 5G’s support for diverse radio bands, mobility management becomes far more complex. Moreover, with generally smaller and denser cells compared to its predecessors, 5G handovers (HOs) between cells are more frequent. Given that 4Gand 5G are expected to co-exist, 3GPP has introduced a number of 5G non-standalone (NSA) deployment architectures and the 5G standalone (SA) mode [2]. All these further complicate the 5G HO procedure: besides horizontal HOs between cells within the same technology (e.g., 5G-to-5G lowband, mid-band and high-band), there are also vertical HOs across the technologies (e.g., 4G-to-5G and 5G-to-4G).

Previous studies in 4G/LTE [35, 43, 63, 66] and recently in 5G [50, 51, 53, 54, 65] have shown that frequent HOs can lead to wild fluctuations in 5G throughput, and in the worst case, complete service “outages”. These impairments will translate to poor application performance in particular for low-latency applications that 5G is supposed to support, such as AR/VR, edge offloading, and vehicleto-everything (V2X) communication. The impact will be further aggravated by improper HO configurations that are observed in 3G/4G [35, 36, 59].

随着5G支持多样化的无线频段，并且与其前身相比，具有更小、更密集的蜂窝，5G的切换（HO）变得更加频繁。鉴于4G和5G将共存，3GPP引入了多种5G非独立（NSA）部署架构和独立（SA）模式，这进一步复杂化了5G的切换过程：除了同一技术内的水平切换（例如5G低频、中频、高频之间的切换），还存在跨技术的垂直切换（例如4G到5G和5G到4G）。

之前的4G/LTE研究和最近的5G研究表明， 频繁的切换会导致5G吞吐量出现剧烈波动，甚至导致服务中断 。这些问题会对应用程序性能产生负面影响，尤其是对于5G支持的低延迟应用，如AR/VR、边缘卸载和车辆与一切（V2X）通信。这些影响将因不当的切换配置而进一步加剧，这些配置在3G/4G中已被观察到。

5G网络的水平与垂直切换

水平切换: 5G低频、中频、高频 之间切换

垂直切换: 4G->5G 和 5G->4G

Study Goal, Challenges, and Data Collection. Given the importance and complexity of 5G HOs, it is imperative to gain a thorough understanding of the current 5G HO mechanisms and practices adopted by commercial carriers. With this goal, we conduct – to our knowledge – a first comprehensive, in-depth study of 5G mobility management. Unlike in-lab experiments, measuring 5G HOs in the wild faces numerous challenges: How to obtain key control-plane signaling events from unrooted smartphones? How to thoroughly survey various 5G architectures (SA vs. NSA), radio bands, and carriers under limited human resources and budgets? How to orchestrate data collection tasks at different layers? How to accurately profile the HO effect on UE (user equipment) energy consumption? To overcome these challenges, we set up a measurement platform comprising of: (1) multiple 5G smartphones with access to three major 5G carriers in the U.S., (2) a custom-built software that captures mobility-related information on unrooted smartphones, (3) a professional measurement tool that collects cellular controlplane events, and (4) a physical power monitor with an external power bank for accurately profiling UE’s battery drain.

Using this platform, we carry out a cross-country data collection field trip, conducting measurements along highways (5560 km+) and within several major cities (712 km+). With over 600GB+ of logs collected, we observe 47,000+ handovers in our datasets that span multiple dimensions: (1) carriers (denoted as OpX, OpY, and OpZ), (2) radio technologies (5G vs. 4G), (3) 5G architectures (NSA vs. SA), and (4) 5G bands – low-band, mid-band, mmWave (highband). This constitutes – to our knowledge – the largest (in terms of the mileage) cross-layer driving test of commercial 5G networks.

Leveraging our unique driving dataset summarized in Table 1, we conduct a detailed analysis to obtain key insights regarding 5G HOs and uncover their impacts. Our findings reveal that there indeed exist significant disparities among the HO mechanisms adopted by the major 5G carriers with considerable performance implications as detailed below.

研究目标、挑战和数据收集。鉴于5G切换的重要性和复杂性，深入理解当前5G切换机制和商用运营商的实践至关重要。为此，我们开展了一项全面、深入的5G移动性管理研究。与实验室实验不同，在野外测量5G切换面临多个挑战：如何从未root的智能手机中获取关键控制平面信令事件？如何在有限的人力和预算下彻底调查各种5G架构（SA与NSA）、无线频段和运营商？如何协调不同层次的数据收集任务？如何准确地描述切换对用户设备（UE）能耗的影响？为了克服这些挑战，我们建立了一个测量平台，包括：（1）多部具有访问权的5G智能手机，连接到美国三家主要5G运营商；（2）自定义软件，用于在未root的智能手机上捕获与移动性相关的信息；（3）专业测量工具，用于收集蜂窝控制平面事件；（4）物理电源监控器，带有外部电源，用于准确地描述UE的电池耗尽情况。

使用这个平台，我们进行了一次跨国数据收集的实地考察，沿着高速公路（超过5560公里）和多个主要城市（超过712公里）进行测量。我们收集了超过600GB的日志，观察到数据集中47,000多次切换，这些切换涵盖了多个维度：（1）运营商（分别标记为OpX、OpY和OpZ）；（2）无线技术（5G与4G）；（3）5G架构（NSA与SA）；（4）5G频段——低频、中频、毫米波（高频）。这构成了我们所知的最大规模（以行驶里程计算）的跨层次商用5G网络驾驶测试。

利用我们独特的驾驶数据集（如表1所示），我们进行了详细的分析，以获取有关5G切换的关键见解，并揭示其影响。我们的发现表明，主要5G运营商采用的切换机制之间确实存在显著差异，这些差异对性能有着重要的影响。

How do 5G HOs Impact Applications? (§4) To study the impact of 5G HOs on application QoE (quality-of-experience), we consider three case studies: i) live video conferencing, ii) real-time 3D volumetric video streaming, and iii) cloud gaming. Our experiments suggest that 5G HOs adversely affect application QoE. For example, a HO event during a live video conferencing application causes the average frame loss-rate to increase by 2.24×, and the end-to-end latency increases by 2.26× (up to 14.5×). For 4K cloud gaming at 60 FPS, we observe an average 3.64× increase in dropped frames due to HOs.

Based on both our experimental results and prior studies of 3G/4G mobility [63, 66], we note that 5G HOs exert a far severe impact on application QoE than their 4G counterparts — the severity hinges on HO types, radio bands, and radio access technologies. For instance, most of today’s 5G deployment is NSA that uses 4G as the control plane and 5G New Radio (5G-NR) as the high-speed data plane – referred to as NSA-4C thereafter. NSA-4C and 5G-NR incur separate HOs over 4G eNodeBs (eNB) and 5G gNodeBs (gNB) respectively, leading to more frequent HOs. In particular, due to the directionality and shorter range of mmWave radio, applications over mmWave 5G suffer far higher performance fluctuations compared to mid-band and low-band 5G due to mmWave HOs (between beams). On the positive side, applications employing the dual mode in NSA 5G, where user data can be delivered over both 4G and 5G, mitigate the negative impact of HOs, thanks to its flexible multiradio paradigm.

5G切换如何影响应用程序？（§4）为了研究5G切换对应用程序服务质量（QoE）的影响，我们考虑了三个案例研究：i）实时视频会议；ii）实时3D体积视频流媒体；iii）云游戏。我们的实验表明，5G切换对应用程序QoE产生负面影响。例如，在实时视频会议应用程序中，切换事件导致平均帧丢失率增加2.24倍，端到端延迟增加2.26倍（最高达14.5倍）。对于4K云游戏（60帧每秒），我们观察到由于切换导致的平均丢帧增加了3.64倍。

基于我们的实验结果和之前的3G/4G移动性研究，我们注意到 5G切换对应用程序QoE的影响比其4G对应的更为严重 —— 这种严重程度取决于切换类型、无线频段和无线接入技术 。例如，大多数今天的5G部署是NSA，它使用4G作为控制平面，使用5G新无线（5G-NR）作为高速数据平面——我们将其称为NSA-4C。

NSA-4C和5G-NR分别在4G基站（eNB）和5G基站（gNB）上发生独立的切换，导致更频繁的切换。特别是，由于毫米波无线的方向性和较短的范围，毫米波5G上的应用程序比中频和低频5G更容易受到切换的性能波动影响，尤其是在毫米波束之间的切换。另一方面，NSA 5G中使用双模式的应用程序可以通过其灵活的多无线范式来减轻切换的负面影响，因为用户数据可以通过4G和5G同时传递。

Danger

1. 5G切换对应用程序QoE的影响比其4G对应的更为严重

   • 这种严重程度取决于切换类型、无线频段和无线接入技术

2. 目前大多5G是 NSA (NSA-4C) 的:

   • 以 4G 作为 ctrl. plane
   • 以 5G-NR 作为 data. plane
   • 这两者是独立切换的, 因此导致更频繁的切换

What are the Key Characteristics of 5G HOs? (§5) Motivated by the above findings, we conduct an in-depth, measurement-driven investigation of 5G HOs to uncover their key characteristics. We focus on three aspects: HO frequency, duration, and UE energy consumption. We find that 5G HOs are indeed triggered frequently. While driving over freeways, we experience a 5G HO occurs every 0.4 km on average, compared to every 0.6 km for 4G. The HO frequency depends on the 5G architecture and band: HOs occur more frequently in NSA (every 0.4 km) compared to 5G low-band SA (every 0.9 km) due to NSA’s separate HO procedures for NSA-4C and 5G-NR; 5G NSA HOs are particularly excessive in mmWave 5G (every 0.13 km) compared to mid/low-band 5G (every 0.35/0.4 km) given mmWave gNBs’ much smaller coverage. In terms of HO duration, an average HO in NSA 5G takes 167 ms to complete, about 1.19× longer than a HO in 4G.

To understand why 5G HOs take a longer time, we break down a 5G HO into multiple stages. We find that the HO preparation stage – during which base stations make HO decisions (before executing them) – accounts for 41% of the overall HO duration in NSA 5G. Compared to 4G, NSA 5G causes on average a 48% increase in HO preparation stage. This increase contributes to a longer dataplane interruption time (1.4× longer than 4G). This points to the complexities of NSA 5G HOs that involve both gNBs and eNBs as the plausible culprit. Somewhat surprisingly, we also observe high preparation time in many SA 5G HOs, likely attributed to the technical immaturity of today’s SA 5G that is still in its early stages of commercial deployment.

We also examine the UE energy overhead incurred by 5G HOs. This turns out to be non-trivial: a smartphone traveling at 130 km/h for 1 hour (without user data transmission or reception) can witness on average 553 5G HOs that drain 34.7 mAh energy. 4G HOs, on the other hand, only consume 3.4 mAh energy. This hints at the importance of reducing the number of HO-related signaling messages, which is found to be positively correlated with the increased energy consumption during 5G HOs.

5G切换的关键特征是什么？（§5）受上述发现的启发，我们进行了一项深入的、基于测量的5G切换调查，以揭示其关键特征。我们关注三个方面：切换频率、持续时间和UE能耗。我们发现5G切换确实被频繁触发。在高速公路上行驶时，我们平均每0.4公里经历一次5G切换，而4G则是每0.6公里。切换频率取决于5G架构和频段：NSA切换更频繁（每0.4公里），而5G低频SA切换则较少（每0.9公里），因为NSA需要为NSA-4C和5G-NR分别执行切换；在毫米波5G中，切换尤其频繁（每0.13公里），比中频和低频5G（每0.35/0.4公里）更快，因为毫米波基站的覆盖范围更小。

在切换持续时间方面，NSA 5G中的平均切换需要167毫秒才能完成，比4G长约1.19倍。为了了解为什么5G切换需要更长时间，我们将5G切换分解为多个阶段。我们发现切换准备阶段——在此阶段基站做出切换决定（在执行之前）——占NSA 5G切换总持续时间的41%。与4G相比，NSA 5G的切换准备阶段平均增加了48%，这导致数据平面中断时间延长了1.4倍。这表明NSA 5G切换的复杂性——涉及gNB和eNB——可能是导致切换时间延长的原因。令人惊讶的是，我们还观察到许多SA 5G切换的准备时间较高，这可能是由于今天的SA 5G仍处于商业部署的早期阶段所致。

我们还检查了5G切换对UE能耗的影响，这被证明是非微不足道的：一部智能手机以每小时130公里的速度行驶1小时（不进行用户数据传输或接收），平均会经历553次5G切换，耗电34.7毫安时。相比之下，4G切换仅消耗3.4毫安时。这表明减少与切换相关的信令消息的数量至关重要，因为这些消息与能耗增加之间存在正相关关系。

Note

1. 对于 SA 和 NSA: NSA 的切换更频繁, 因为 NSA-4C 和 5G-NR 是独立切换的, 这在上文提了
2. 对于 5G 和 4G: 5G 的切换时间比 4G 长
   • NSA 5G 时间长: 基站的准备时间长（eg. 如何做决策）
   • SA 5G 时间长: 今天的 SA 5G 仍处于商业部署的早期阶段
3. 切换次数多 (单纯是切换, 不存在任何数据传送), 导致能耗增加

What are 5G HOs’ Implications on Carriers? (§6) Our analysis also sheds light on potential improvements on the carrier side. We highlight three key findings. First, our extensive drive test helps depict a landscape of 5G cell coverage that is closely relevant to HOs. We find that for NSA 5G, the average coverage (diameter) of a single 5G cell is 1.4 km, 0.73 km, and 0.15 km for low-band, mid-band, and mmWave, respectively. In particular, for low-band NSA 5G, although the data plane (5G-NR) operates on the low-band, its coupled control plane (NSA-4C) still uses the mid-band, which reduces the effective coverage of low-band 5G-NR since an NSA-4C HO always triggers a 5G-NR HO. Second, HOs are performed with the goal to improve the received signal strength of UE and hence its throughput. However, we find that a 5G→5G HO between two gNBs often worsens the performance, with a median bandwidth reduction of 14% after HOs. This is because NSA 5G does not support direct HOs between gNBs; the UE instead experiences a 5G→4G and then a 4G→5G HO where each HO is performed independently without accounting for the overall (5G→5G) signal strength improvement. Third, we find that for NSA HOs where the (origin or destination) gNB and eNB are co-located at the same physical tower, their duration is significantly shorter than HOs whose gNB and eNB are not co-located where the cross-tower communications incur delays. These findings not only identify new inefficiencies of NSA 5G, but also provide valuable hints on how NSA carriers can mitigate the impact of 5G HOs, such as accounting for 4G/5G antenna locations and considering the overall HO sequence when making HO decisions.

5G切换对运营商的影响是什么？（§6）我们的分析还为运营商提供了潜在的改进方案。我们强调了三个关键发现。首先，我们的广泛驾驶测试有助于描绘与切换密切相关的5G蜂窝覆盖图景。我们发现，对于NSA 5G，单个5G蜂窝的平均覆盖范围（直径）分别为低频1.4公里、中频0.73公里和毫米波0.15公里。特别是，对于低频NSA 5G，尽管数据平面（5G-NR）在低频运行，但其耦合的控制平面（NSA-4C）仍使用中频，这降低了低频5G-NR的有效覆盖范围，因为NSA-4C切换总是触发5G-NR切换。其次，切换的目的是为了提高UE的接收信号强度和吞吐量。然而，我们发现5G到5G之间的切换（在两个gNB之间）往往会降低性能，中位数带宽减少了14%。这是因为NSA 5G不支持gNB之间的直接切换；UE会经历5G到4G然后4G到5G的切换，每次切换都是独立进行的，没有考虑到整体（5G到5G）信号强度的改善。第三，我们发现，当（源或目的）gNB和eNB位于同一物理塔时，切换的持续时间显著短于gNB和eNB不位于同一塔的切换，因为跨塔通信会引入延迟。这些发现不仅揭示了NSA 5G的新效率低下，也为NSA运营商提供了有价值的提示，例如考虑4G/5G天线位置，并在做出切换决策时考虑整体切换序列。

Can We Predict 5G HOs to Improve Application QoE? (§7) Last but not the least, we explore 5G HO prediction to help applications to accommodate and mitigate the negative impact of frequent 5G HOs. For this, we develop a robust and effective 5G HO prediction framework (dubbed Prognos). It uses observed signal strength readings, UE-side measurement reports (MRs), and past HOs to predict future HOs and their types. Prognos can work with any 3GPP-compliant 5G deployment without requiring proprietary information from the carrier. Prognos consists of a novel two-stage prediction pipeline. It first predicts the future signal strength that determines UE’s MRs sent to the base station, and then learns the base station’s HO logic that produces the HO decision based on the MRs. Compared to a monolithic model, decoupling the UE MR inference and network side decision logic learning reduces the model complexity and improves accuracy by eliminating indirect or unnecessary features.

我们能否预测5G切换以改善应用程序QoE？（§7）最后，我们探索了5G切换预测，以帮助应用程序适应和减轻频繁5G切换的负面影响。

为此， 我们开发了一个强大的5G切换预测框架（称为Prognos）。它使用观察到的信号强度读数、UE侧测量报告（MR）和过去的切换来预测未来切换及其类型 。Prognos 可以与任何 3GPP 兼容的 5G 部署一起工作，无需运营商的专有信息。Prognos由一个新颖的两阶段预测流水线组成。它首先预测决定UE向基站发送的MR的未来信号强度，然后学习基站的切换逻辑，该逻辑基于MR生成切换决策。与单一模型相比，将UE MR推理和网络侧决策逻辑学习解耦可以降低模型复杂度并提高准确度，因为它消除了间接或不必要的特征。

We conduct extensive evaluation of Prognos using our dataset. Prognos achieves an F1-Score between 0.92–0.94 for predicting 4G/5G HOs, significantly outperforming existing HO prediction approaches developed for 4G/5G [49, 57] by 1.9×–3.8×. We incorporate Prognos into two applications, 16K panoramic video streaming and real-time volumetric video streaming, by modifying the throughput prediction algorithm used in the adaptive bitrate (ABR) adaptation modules. Prognos significantly boosts both applications’ QoE compared to using the default throughput prediction algorithm: a 34.6%–58.6% reduction in stall time without degrading video quality for 16K streaming, and an 15.1%–36.2% increase in the content quality without prolonging stalls for volumetric video streaming.

我们使用我们的数据集对Prognos进行了广泛的评估。Prognos在预测4G/5G切换时实现了0.92至0.94之间的F1分数，显著优于为4G/5G开发的现有切换预测方法，性能提高了1.9倍至3.8倍。我们将Prognos集成到两个应用程序中：16K全景视频流媒体和实时体积视频流媒体，通过修改自适应码率（ABR）适应模块中使用的吞吐量预测算法。Prognos显著提高了这两个应用程序的QoE，与使用默认吞吐量预测算法相比：16K流媒体的停顿时间减少了34.6%至58.6%，而不降低视频质量；体积视频流媒体的内容质量提高了15.1%至36.2%，而不延长停顿时间。

Contributions. We summarize our contributions as follows: (1) creation of a large cross-layer, multi-band, multi-carrier dataset of 5G mobility management, (2) a first comprehensive characterization of mobility management in commercial 5G networks, and (3) a new methodology of accurately predicting 5G HOs and demonstrations of its efficacy on real-world applications over 5G.

Artifacts. To support future research, we make our dataset, source code of analysis/proposed techniques, and results publicly accessible through our project website: https://github.com/SIGCOMM225GMobility/artifact.

Ethics: This work does not raise any ethical issues.

贡献。我们总结了我们的贡献如下：（1）创建了一个大型跨层次、多频段、多运营商的5G移动性管理数据集；（2）对商用5G网络的移动性管理进行了首次全面性描述；（3）提出了一种新的准确预测5G切换的方法，并展示了其在实际应用中的有效性。

文档。为了支持未来的研究，我们通过项目网站公开提供我们的数据集、分析/提出的技术的源代码以及结果：https://github.com/SIGCOMM22-5GMobility/artifact。

伦理问题：本研究不涉及任何伦理问题。