跳转至

Cellular Networks vs. Starlink in Non-Contiguous US Regions

In the previous two sections, we showed that both cellular networks and Starlink face unique challenges in the non-contiguous US regions and their performance is significantly lower than in the mainland US. In this section, we compare the performance of the two technologies in the non-contiguous US regions, and discuss their spatiotemporal diversity.

DL Performance. Fig. 15a compares the TCP DL throughput of Starlink (SL) and the 3 cellular operators in AK and HI. We make two observations. First, Starlink consistently outperforms AT&T and Verizon in both regions and outperforms T-Mobile in HI 71% of the time. However, T-Mobile’s extensive 5G midband coverage in HI allows it to achieve a peak throughput of 937 Mbps, whereas Starlink’s throughput is capped at 237 Mbps. Second, in both regions, Starlink has a lower fraction of zero-throughput samples than the 3 cellular operators – 9% vs. 24-32% in AK, 22% vs. 40-47% in HI. These numbers suggest that Starlink delivers more consistent connectivity than cellular networks in the non-contiguous regions, but does not completely eliminate outage periods.

In Fig. 16, we break down the performance based on area type. In AK (Fig. 16a), Starlink significantly outperforms both cellular networks in rural areas, with a median DL throughput of 64 Mbps vs. only 11/6 Mbps for AT&T/Verizon. In urban areas, Starlink performs similarly to AT&T (27 Mbps in the median case) despite its 5G service (recall from Fig. 8a that AT&T’s 5G-low service performs poorly in AK urban areas), but performs worse than Verizon’ LTE service (27 Mbps vs. 41 Mbps in the median case).

In HI (Fig. 16d), Starlink significantly outperforms AT&T and Verizon in both urban and rural areas, yielding a median throughput of 94/65 Mbps in urban/rural areas vs. 37/0 Mbps for Verizon and 25/0 Mbps for AT&T. It also reduces the fraction of zero-throughput samples significantly in rural areas (26% vs. 53/52% for Verizon/AT&T), but it has a higher fraction of zero-throughput in urban areas, (9.7% vs. 5.5/6.0% for Verizon/AT&T). On the other hand, T-Mobile ’s extensive 5G-mid coverage outperforms Starlink in urban areas (130 Mbps vs. 94 Mbps in the median case) and offers higher peak rates (above the 80-th percentile) than Starlink in rural areas. However, Starlink offers much more consistent coverage in rural areas, with a median throughput of 65 Mbps vs. 0 Mbps for T-Mobile and an outage rate of 26% vs. 61%.

下行链路对比:

(1) 整体表现:

  • Starlink 在两个地区通常优于 AT&T 和 Verizon
  • 在夏威夷 (HI), Starlink 击败 T-Mobile 的概率为 71%, 但 T-Mobile 凭借广泛的中频段 5G (5G-mid) 覆盖, 峰值速率远超 Starlink (937 Mbps vs. 237 Mbps)

(2) 可靠性:

  • Starlink 的零吞吐量样本比例显著低于三大蜂窝运营商(AK: 9% vs. 24-32%; HI: 22% vs. 40-47%), 表明其连接性更稳定, 但仍无法完全消除中断

(3) 城乡差异:

  • 农村地区: Starlink 在 AK 和 HI 的农村地区均显著优于蜂窝网络, 提供更高的中位数吞吐量(例如 AK 农村: 64 Mbps vs. 11/6 Mbps)和更低的断连率
  • 城市地区: Starlink 优势减弱.
    • 在 AK 城市, 其表现与 AT&T 相似但不及 Verizon 的 LTE
    • 在 HI 城市, T-Mobile 的中位数吞吐量 (130 Mbps) 高于 Starlink (94 Mbps).

UL Performance. Fig. 15b compares the UL throughput of Starlink and the 3 cellular networks. Unlike in the DL case (Fig. 15a), where Starlink largely outperformed the cellular networks, we observe a different picture in the UL case. Starlink’s UL throughput is very low (only 5/3 Mbps in the median case in AK/HI) and lower than the cellular throughput about 36-38% of the time. Additionally, Starlink’s TCP UL throughput peaks at 41/35 Mbps in AK/HI, while cellular TCP UL throughput can reach up to 47-111 Mbps in AK and 96-174 Mbps in HI. Recall that our UL TCP tests were done with Cubic, which is less resilient to packet losses compared to BBR (used for DL tests), and is probably contributing to Starlink’s low UL performance.

Nonetheless, Starlink still eliminates a significant fraction of outage (or near-outage) periods compared to cellular networks. Starlink’s fraction of zero-throughput samples is 6%/27% in AK/HI vs. 39-46%/52-53% for cellular networks. The performance breakdown based on area type (Figs. 16b, 16e) shows that cellular networks significantly outperform Starlink in urban areas in both noncontiguous US regions, while the trends for the rural areas are similar to those observed for the overall performance (in Fig. 15b).

上行链路对比:

(1) Starlink 劣势:

  • 与下行不同, Starlink 的上行吞吐量非常低(中位数仅 3-5 Mbps), 且在约 36-38% 的时间里低于蜂窝网络
  • 峰值速率也远低于蜂窝网络
    • Starlink 约 40 Mbps vs. 蜂窝网络 100+ Mbps
  • 原因: 可能与测试使用的 TCP Cubic 协议对丢包更敏感有关(相比下行测试用的 BBR)

(2) 连接稳定性:

  • 尽管速度慢, Starlink 在减少完全中断(零吞吐量)方面仍优于蜂窝网络
  • AK: 6% vs. 39-46%; HI: 27% vs. 52-53%

(3) 区域对比:

在城市地区, 蜂窝网络的上行性能显著优于 Starlink

Network Latency. Fig. 15c compares the latency of Starlink and cellular networks in AK and HI. We observe that Starlink’s latency is significantly higher than the cellular latency in AK (100 ms vs. 75-77 ms in the median case) and falls between the best and worst latency observed among the cellular operators in HI (109 ms vs. 93/110/121 ms for Verizon/T-Mobile/AT&T in the median case).

As we mentioned previously, Starlink’s latency is not affected by the area type (Fig. 13), while the cellular network RTT is affected significantly in HI, due to different distribution of cellular technologies in rural vs. urban areas (Fig. 10). Fig 16c shows that Starlink’s RTT is always worse than the cellular network RTT in AK, regardless of area type. In HI urban areas (Fig 16f), AT&T and T-Mobile have more consistent RTTs than Starlink, outperforming Starlink about 40% of the time, but lagging behind about 60% of the time. In rural areas, AT&T has the worst RTT among all four operators, while T-Mobile and Starlink have similar RTT, although T-Mobile has higher worst case RTT (586 ms vs. 160 ms at the 95-th percentile). On the other hand, Verizon has the lowest RTT among the four operators in both area types.

Network Diversity. Up till now, our analysis has focused on the individual performance of Starlink and cellular networks in the two non-contiguous US regions. However, we have not yet explored how cellular and Starlink performance vary at a given time and location. To gain deeper insights into this spatiotemporal variation, we analyze the throughput difference of concurrent throughput samples for a given operator pair in Fig. 17.

Fig. 17 shows a significant diversity in performance for any given network pair in both traffic directions and both regions. The throughput difference between any network pair is often higher than 50 (10) Mbps and can exceed 150 (20) Mbps in the DL (UL) direction. Additionally, the throughput difference between Starlink and any cellular operator for concurrent throughput tests is negative 15-35% of the time in the DL direction and 45-50% of the time in the UL direction, suggesting that Starlink is unable to consistently provide better performance. This result indicates a great potential for multipath transport to improve overall performance.

Additionally, Table 1 presents the fraction of time during which both networks in a given network pair simultaneously experience an outage (0 Mbps throughput), as well as the fraction of outage times for each individual network. We clearly observe that the fraction of concurrent outage for any network pair is much lower than when we consider any network in isolation, suggesting again that leveraging multipath transport can greatly enhance connectivity.

Table 1 further shows that combining Starlink with a cellular network nearly eliminates outage events in AK, but not in HI, where the faction of concurrent outage events still ranges from 14% to 19%, indicating areas where no network is available. Nonetheless, in both regions, combining Starlink with a cellular network results in lower numbers of concurrent outage events compared to combining any two cellular networks, suggesting that the use of two different network technologies has a greater potential to improve connectivity than relying solely on cellular infrastructure.

Remarks. In summary, our results demonstrate Starlink’s potential to close the digital divide in non-contiguous US regions, offering improved connectivity and substantially higher DL throughput compared to cellular networks that rely on LTE or 5G-low service. Nonetheless, there is still a lot of room for improvement. Zero-throughput periods are still present (although less frequent compared to cellular networks), Starlink’s DL throughput is lower than 5G-mid’s throughput and its UL throughput is often lower than that of cellular networks, and higher RTTs compared to cellular networks are often observed, particularly in AK. These limitations are the result of both special terrain and geographical characteristics of the two non-contiguous US regions (e.g., dense rain forests in HI and lower satellite availability in AK, which falls outside the 43 𝑜 /53 𝑜 inclination orbit) and inherent challenges in LEO satellite networks, such as packet loss and congestion control inefficiencies, which are less pronounced in terrestrial networks. Our spatiotemporal analysis revealed substantial diversity in performance for any given pair of networks in both non-contiguous US regions, suggesting that multipath transport has a great potential to improve the overall user performance. We explore this potential in the next section.

  • Starlink 延迟较高:

    • 在 AK, Starlink 的延迟(中位数 100 ms)显著高于蜂窝网络(75-77 ms)
    • 在 HI, Starlink 的延迟(中位数 109 ms)处于蜂窝运营商的最佳和最差表现之间

  • 稳定性:

    • Starlink 的延迟不受区域类型(城市/农村)影响, 而蜂窝网络(特别是 HI)受区域技术分布影响很大
    • Verizon 在两个地区的所有区域类型中均保持最低延迟

  • 时空多样性:

    • 同一时间和地点下, Starlink 与蜂窝网络的性能差异巨大
    • Starlink 并非总是更优(下行约 15-35% 的时间, 上行约 45-50% 的时间表现不如蜂窝网络)

  • Multi-path潜力:

    • 并发中断率低: 任意两个网络同时发生中断的概率远低于单个网络
    • 最佳组合: 将 Starlink 与蜂窝网络结合(如 Starlink + Verizon), 在 AK 几乎消除了中断事件, 在 HI 也显著降低了中断率, 效果优于结合两个蜂窝网络