IMPACT OF MOBILITY ON APPLICATION PERFORMANCE

In this section, we use a combination of latency-sensitive and bandwidth-hungry applications to understand QoE fluctuation during mobility. We exclude SA 5G from our analysis as it is not fully mature to achieve high downlink throughput (similar to recent findings in [54]) required by applications under study.

Quantifying App QoE under Mobility

We consider the following three applications as case studies. Live Video Conferencing. We run Zoom while driving around a loop in a downtown area with NSA 5G coverage. Fig. 4 shows a representative trace collected during our study. We extract a 1-second time window around the UE’s HO timestamps (HOs annotated using green arrows). We find the average latency is 2.26× higher compared to no-handover periods (up to 14.5× higher in the worst case). Likewise, the average packet loss rate increases by 2.24×. Prior studies show that Zoom requires a minimum bandwidth of 0.6-0.95 Mbps for a one-on-one call as in our case [34, 47]. Low-band NSA 5G offers much higher bandwidth than what Zoom requires. Despite this, we show that video conferencing over today’s 5G remains challenging, especially during mobility as frequent HOs cause network fluctuations and increase latency impacting the QoE. Additionally, NSA 5G requires the UE be connected to both the eNB and the gNB. This causes HOs to occur on both radios. In today’s 5G, the frequency of HOs are far higher than LTE (§5.1), thus the impact is amplified.

Real-time Cloud Gaming. Using a cloud-gaming application, we show the impact of HO type on QoE. We select two key metrics: (i) network (or transmission) latency, and (ii) dropped frames. The other latency (encoding, decoding, rendering, etc.) stays at the same level and the network latency dominates the overall latency during experiments. In our setup, the game fetches the streams at 4K@60FPS, thus in addition to being latency-sensitive, our setup also had high bandwidth requirement. As shown in Fig. 5, the network latency increases by an average 2.26× (up to 14.5×) during HOs. Likewise, HOs increase the dropped frame rate by 2.6× for a game running at 60FPS.

Considering NSA handles 4G and 5G radios at the same time, both NSA-4C (defined in §1) and 5G-NR HOs can be triggered on the UE. 5G-NR HOs in NSA 5G e.g., SCG Modification (SCGM) have lower impact on the QoE than NSA-4C HOs, e.g., MeNB HO (MNBH): compared to SCGM, MNBH averages 16.8ms higher network latency and a 65% increase in the number of dropped frames (see Fig. 5). Since SCGM only involves a HO between gNB cells over 5G, whereas MNBH changes the LTE primary cell (see Table 2), the QoE degradation of SCGM is relatively less than MNBH. This is also observed in volumetric video streaming experiments. Hence, we conclude that the QoE fluctuation level depends on the HO type in NSA 5G.

Volumetric Video Streaming. 5G-NR supports a wide range of radio frequencies (up to 100 GHz). The diversity of bands has a cascading impact on application performance especially under user mobility. To quantitatively capture this impact, we consider a volumetric streaming application (ViVo [40]), which is a key building block of telepresence [24], and compare HOs across two 5G-NR bands (low-band and mmWave). We focus on two key performance metrics: video bitrate and network latency. From our experiments, we note that high frequency bands usually cause more QoE degradation than low frequency bands. Fig. 6 contrasts the perceived QoE metrics between the radio frequency bands. Although low-band HOs result in a lower video quality, the degradation is significantly higher for mmWave HOs. In the median case, the video bitrate reduces by 31% for low-band HOs whereas it degrades by 58% for mmWave HOs. Similarly, the network latency increases by 41% for low-band HOs while mmWave HOs see a stark 107% increase in latency. The mmWave 5G performance fluctuates wildly especially during HOs, sometimes incurring a ∼2Gbps drop in throughput (see §6.2). On the other hand, the throughput degradation during HOs is comparatively lower for low-band 5G [65]. All in all, the above results suggest that the level of QoE fluctuation under mobility is determined by a combination of HO type, radio access technology, and radio frequency band.

5G-only vs. dual traffic mode in NSA

In NSA, 5G-NR radio resources (such as radio data bearers) are added to the ongoing 4G/LTE connection to increase data plane bandwidth for users. The user data can be exchanged on the LTE radio interface, 5G interface, or both. The NSA deployment scheme of a carrier typically decides the proportion of data arriving on each interface. A dual mode (MCG Split bearer [4]) splits the traffic across both 4G and 5G radio interfaces. In contrast, the 5G-only mode employs the 5G interface for all data traffic (SCG bearer [4]). During mobility, the NSA traffic mode can differ from one area to another.

To understand how HOs and traffic mode affect network performance, we use a simple TCP application and measure its round-triptime (RTT). We conduct a driving experiment in areas with two different traffic modes. The traffic mode information is extracted from the PDCP layer messages [9]. There are three key takeaways from the results in Fig. 7. First, 5G-only mode results in a comparatively lower RTT than dual mode when there is no HO (w/o HO case). Second, the median RTT does not change significantly during HOs in dual mode as 4G radio is not impacted by 5G-NR HO interruptions. This allows 4G radios to continue transmission during HOs. In the median case, we only observe a 1-4% change in RTT for dual mode which can be due to HO latencies [63, 65]. Finally, in 5G-only mode, HOs have a relatively higher impact on RTT since there is no secondary interface. To be precise, the RTT can inflate by up to 37-58% in the median case. Although the results are only shown for TCP BBR, Cubic also behaves in a similar manner. Notably, the dual mode absorbs HO fluctuations while the 5G-only mode does not. However, the dual mode has comparatively lower performance (higher RTT) when there is no HO. In dual mode, the core network first sends 5G data to the eNB which is then forwarded to the gNB (before getting transmitted to the UE). Whereas, in 5G-only mode, the 5G data is directly sent to the gNB from the core network, resulting in lower RTT compared to dual mode. We believe that a combination of 5G-only and dual modes can get carriers the best of both worlds; they can employ dual mode where core network sends 5G data directly to the gNB. This can lead to a similar performance as 5G-only mode while also minimizing HO fluctuations.

TL; DR

5G移动性对应用程序服务质量（QoE）的影响

为了研究5G移动性对应用程序QoE的影响，我们选择了三个案例：实时视频会议、实时云游戏和体积视频流媒体。

1. 实时视频会议：我们使用Zoom在NSA 5G覆盖的市中心区域进行测试。结果显示，切换（HO）期间的平均延迟增加了2.26倍（最坏情况下高达14.5倍），平均包丢失率也增加了2.24倍。尽管低频NSA 5G提供了远高于Zoom所需的带宽，但频繁的切换仍导致网络波动，影响了视频会议的QoE。

2. 实时云游戏：我们评估了不同类型切换对云游戏QoE的影响。结果表明，切换期间网络延迟增加了2.26倍（最坏情况下高达14.5倍），丢帧率增加了2.6倍。NSA 5G中的5G-NR切换（例如SCG修改）对QoE的影响较小，而4G LTE切换（例如MeNB切换）则导致更高的网络延迟和丢帧率。

3. 体积视频流媒体：我们比较了低频段和毫米波5G-NR频段对体积视频流媒体的影响。结果显示，毫米波切换导致视频码率降低58%，网络延迟增加107%，远高于低频段切换（分别为31%和41%）。这表明毫米波5G在切换期间的性能波动更为显著。

NSA 5G中的双模式与单模式流量

在NSA 5G中，用户数据可以通过4G和5G接口传输（双模式），或仅通过5G接口传输（单模式） 。

我们的实验表明， 在没有切换的情况下，单模式的TCP往返时间（RTT）较低，但在切换期间RTT会显著增加（最高达37-58%）。相比之下，双模式可以吸收切换波动，RTT变化较小（1-4%），但在没有切换时RTT较高 。

我们认为，将两种模式结合起来可以实现最佳效果：在核心网络直接将5G数据发送到gNB，同时减少切换波动。