跳转至

DISCUSSION

Mobile systems, especially 5G, exist at the intersection of many potentially impactful variables that operate within the control of cellular carriers. Furthermore, 5G is a maturing technology and may experience major changes in architecture, structure, and capabilities over the next few years. In this section, we discuss the limitations of our work, and the impact of our study in the context of future 5G. Limitations of measurement scope. Our work represents a rigorous examination of 5G mobility with respect to HOs. However, there are some factors we did not explore due to scope limitations or limited visibility into the carrier’s network. Regarding data-plane energy consumption, existing 5G studies investigate the observable differences by smartphone model [54, 65]. Our HO energy results compliment the existing work and our insights will hold true in general, regardless of model type. We conducted our study without any cooperation from cellular carriers. Hence, we did not explore disparity across base station vendors or manufacturers. Xie et al. show that the time of the day impacts user density [64], and thus the fair-share of bandwidth for each user. By experimenting at several locations (spatial diversity), and across multiple weeks (temporal diversity) and time-of-day (including night time: 12am-4am), we reduce the impact of crowds and congestion that may confound our QoE measurements.

Likewise, the impact of mobility speed and tower density on TCP performance, application QoE, and power consumption is well-explored by previous LTE studies [32, 37, 39, 63]. 5G mobility management is far more complex than LTE; HOs are more frequent and lead to higher QoE fluctuations (§4). Therefore, impact of factors such as speed and tower density intensify in 5G.

Applicability of measurement findings to future 5G and beyond. The current 5G infrastructure is still maturing, with much of the existing deployments being NSA 5G using LTE’s control plane. NSA 5G deployments are here to stay at least for a few years, but will eventually be replaced by SA 5G or future NSA 5G deployments. However, as these transitions happen, future NSA 5G will also evolve such that the control plane will be 5G, with LTE acting as data plane only. For example, 5G deployment option 4 enables carriers to continue using legacy 4G equipment while connecting to the 5G core [13]. Our findings will be relevant for these new NSA 5G deployments too. Moreover, our HO prediction system (Prognos) supports all 4G and 5G HO types, and therefore, can predict HOs for SA 5G deployments as well. Additionally, multiple 5G deployment options (e.g., NSA, SA, etc.) have been defined by the 3GPP to allow flexible (and easy) transition from 4G to 5G. In hindsight, studies like our work will help provide valuable insights in understanding the implications of adopting such transition strategies in future (e.g., 5G to 6G).

Delayed HO predictions during startup. Our system learns new HO patterns in real-time. In order to make reliable predictions, it first needs to collect a few initial HO patterns. The prediction score during the startup phase is typically low. From our analysis, the time until reliable prediction depends on multiple factors including but not limited to the density of cell towers, radio capability of the mobile device, and mobility speed. For our dataset D1 and D2, we observe that the prediction F1-Score goes above 0.9 after 14 and 11 mins, respectively. However, there are ways to improve predictions during the startup phase. For instance, bootstrapping the system with the most frequent pattern for each HO type can make predictions reliable. The most frequent patterns can be found empirically from our collected dataset. Fig. 15 uses a sample trace from dataset D1 to depict the benefit of bootstrapping Prognos with the most frequent pattern. It shows that F1-Score is typically low at the start if Prognos is not bootstrapped with frequent patterns. On the other hand, bootstrapping boosts the F1-Score to 0.8 within 1.5 mins. Another solution is to simply avoid making predictions during the startup phase and only learn the HO patterns for a while. Regardless, the question of how to orchestrate reliable HO predictions during the startup phase still remains open, and is left for future investigation.

The need for cross-layer communication for future 5G. Our work with Prognos rely upon information spanning several layers of the mobile network stack that is not accessible in its entirety without using special tools. Previous studies also used external tools to decode the lower layer information. Few examples are USRP-based control channel decoders [64], professional tools such as Accuver XCAL [15], and in-device solutions like MobileInsight [44], Mobilyzer [56], and LiveLab [62]. In future, the 5G Multi-access Edge Computing (MEC) will be able to gather and distribute control plane information through Radio Network Information (RNI) APIs [6]. We argue that exposing lower layer information through Android API calls can bring immense benefit to the mobile applications. This information can be leveraged for applications such as throughput and latency prediction, loss recovery, energy modeling, handover prediction, and more.

研究局限性

  1. 测量范围限制：

    • 本研究对5G移动性和切换进行了严格的检查，但由于 范围限制或对运营商网络的可见性有限 ，未能探索所有因素
    • 未涉及基站供应商或制造商之间的差异

  2. 数据收集多样性：

    • 通过在多个地点、多个星期和不同时间（包括夜间）进行实验， 减少了人群和拥堵对QoE测量的影响

未来5G的适用性

  1. 5G基础设施的演进：

    • 目前的5G基础设施仍在成熟中，现有部署主要是NSA 5G，未来将逐渐被SA 5G或新型NSA 5G取代
    • 本研究的发现仍将适用于未来的NSA 5G部署

  2. 切换预测系统的通用性：

    • Prognos 系统支持所有4G和5G切换类型，因此也适用于SA 5G部署

切换预测的启动阶段挑战

  1. 启动阶段预测延迟：
    • 系统需要收集初始切换模式才能进行可靠预测，启动阶段的预测分数较低
    • 可通过预先加载最常见模式来提高启动阶段的预测准确性

未来5G的跨层通信需求

  1. 跨层信息获取的重要性：

    • 本研究依赖于跨多个网络层的信息，这些信息通常需要特殊工具才能获取
    • 未来的5G多接入边缘计算（MEC）将能够通过RNI API提供控制平面信息

  2. API对应用的益处：

    • 通过Android API提供底层信息可以显著改善应用，如吞吐量预测、延迟预测、丢包恢复、能量建模 ...