跳转至

Starlink in Non-Contiguous US Regions

LEO satellite networks, offering higher throughput and lower latency than traditional geostationary satellites, have recently gained traction as an alternative Internet access technology, particularly in remote and underserved regions with poor wired/wireless connectivity. In this section, we conduct the first study of the performance of Starlink, the most popular LEO satellite network, in non-contiguous US regions, and compare it against the performance in the mainland US.

4.1 Background

The routing topology of the Starlink network consists of a mix of space and terrestrial paths. The Starlink architecture has four primary components: (a) Starlink constellation with thousands of LEO satellites connected via inter-satellite links (ISLs) for communication between satellites, (b) Ground stations (GSs), which connect satellites to terrestrial infrastructure via fiber, (c) Starlink backbone network, which consists of a hierarchical structure of Points of Presence (PoPs) and Internet Exchange Points (IXPs), enabling internal and Internet-bound connectivity, and (d) user terminals, which directly communicate with satellites to gain Internet access.

Starlink relies on a "bent-pipe" design, where satellites reflect signals between terminals and ground stations. However, when there is no GS in the coverage area of a LEO satellite, ISLs may be utilized to relay the traffic to a neighboring satellite and eventually transmit the data to a nearby GS. The Starlink constellation operates on multiple orbital planes at varying inclinations and altitudes.

Different orbital planes have different number of satellites deployed. As of July 2025 [48], the 43º and 53º inclination orbits, which serve the mainland US and HI, have 7,121 satellites deployed. In contrast, AK is primarily served by the 70º and 97.6º inclination orbits, which have 617 and 293 satellites deployed, respectively.

  • 架构: Starlink 采用"弯管"(bent-pipe)设计, 依赖卫星, 地面站(GS), 骨干网(PoP/IXP)和用户终端.
  • 轨道差异:
    • 本土与夏威夷 (HI): 拥有 7,121 颗卫星, 覆盖密集.
    • 阿拉斯加 (AK): 仅有 617 和 293 颗卫星, 覆盖稀疏.

For network performance analysis, knowing the locations of PoPs is important, as they indicate where packets enter or exit an operator’s network. This, in turn, determines the routing distance and the minimum round-trip time to target servers. However, obtaining PoP locations in cellular networks is challenging due to the absence of publicly available information and the limited availability of DNS PTR records [70]. In contrast, Starlink offers greater transparency. It provides DNS PTR records for user-assigned IPs and includes location hints of the associated PoPs in the hostnames, as discussed in § 2. In this section, we examine how Starlink PoPs are allocated to UEs along our measurement route across different regions of the US.

We visualize links from UE’s locations to PoP locations in Fig. 11. In general, all observed Starlink PoPs are located within the mainland US. For Alaska, only a single PoP, located in Seattle, was used. In contrast, for the island of Maui, Hawaii, we observe two PoPs, located in Los Angeles and San Jose, respectively. Interestingly, the allocation of these two PoPs to Maui appears evenly split across the measured points on the island. We conjecture that this may be due to similar distances to both PoPs (approximately 3,800–4,000 km), or possibly due to Starlink employing load-balancing strategies. Along the route from Los Angeles, CA to Omaha, NE, we observe three PoP locations: Los Angeles, Salt Lake City, and Denver.

Starlink appears to assign each PoP to cover a distinct surrounding area, with each PoP acting as a central hub. Moreover, we observe no overlap in coverage between different PoPs.

alt text

  1. 位置依赖本土: 所有观测到的 Starlink 接入点(PoP)均位于美国本土.
  2. 具体的分配:
    • AK: 仅连接到西雅图(Seattle)的 PoP.
    • HI: 连接到 Los Angeles 和 San Jose 的 PoP, 且分配均匀.
    • 本土: 连接到附近的城市(如 LA, SLC, Denver), 覆盖区域互不重叠.

"Starlink appears to assign each PoP to cover a distinct surrounding area, with each PoP acting as a central hub. Moreover, we observe no overlap in coverage between different PoPs."

注意!

透明度: 相比蜂窝网络, Starlink 更透明, 通过 DNS PTR 记录可识别 PoP 位置.

DNS-PTR解析: 反向查找

简单来说, 它的作用与我们常见的 A 记录(将域名解析为 IP 地址)正好相反:

  • A 记录: 域名 → IP 地址(例如: google.com → 142.250.x.x)
  • PTR 记录: IP 地址 → 域名(例如: 142.250.x.x → google.com)

在网络日志分析中, PTR 记录可以将难以记忆的 IP 地址转换为可读的域名, 帮助管理员识别流量来源

本文利用 DNS PTR 记录来推断 Starlink 网络的基础设施位置

alt text

具体做法: 作者对 Starlink 分配给用户的公共 IP 地址执行反向 DNS 查找(即查询 PTR 记录)

发现: Starlink 的 PTR 记录中包含了接入点 (PoP) 的位置信息!

  • 例如: 一个阿拉斯加用户的 IP 地址反向解析后的主机名为 customer.sttlwax1.pop.starlinkisp.net
  • 其中的 sttlwax1 遵循通信行业标准, 表明该 IP 对应的接入点 (PoP) 位于 Seattle

结论: 通过这种方式, 研究人员能够确定 Starlink 用户的数据流量是从哪里进入互联网主干网的, 从而绘制出 Starlink 在不同区域的 PoP 分配图

4.3 Network Performance

Fig. 12 provides a comprehensive comparison of Starlink performance, including DL throughput, UL throughput, and latency, in AK, HI, and mainland US. Figs. 12a, 12b show that the median DL/UL throughput is comparable in both non-contiguous US regions (HI exhibits slightly higher DL throughput and slightly lower UL throughput) and Fig. 12c shows that Starlink’s RTT is higher in HI than in AK. This is expected, since the UE-server distance is longer in HI than in AK. These figures also show that Starlink on the mainland route demonstrates exceptionally strong performance in terms of all metrics compared to the non-contiguous regions. First, it maintains near-perfect connectivity. The fraction of zero-throughput samples on the mainland route is only 2.6%/2.4% in the DL/UL direction, in stark contrast to the 9.4%/6.2% of zero-throughput samples observed in AK and 22.4%/27.5% in HI. Second, it offers superior performance compared to AK and HI, particularly in the DL direction. The median improvement is 85/3.2 Mbps in the DL/UL direction compared to AK and 69/5.2 Mbps compared to HI. Third, it achieves the lowest latency among the three regions – lower than 50 ms about 30% of the time, with a 50-th/75-th percentile of 56/74 ms vs 100/125 ms in AK and 109/133 ms in HI.

The lower Starlink performance in AK compared to the mainland US is probably due to the different inclination orbits serving the two regions. As mentioned in §4.1, AK is served by the 70º and 97.6º inclination orbits, which have only a few 100s of satellites, in contrast to the 43º and 53º inclination orbits serving the mainland US, which have 7,000+ satellites. The much sparser satellite deployment in AK possibly results in coverage gaps and service interruptions, as well as high queuing delays as the user terminal queues packets until it locates a nearby satellite. On the other hand, HI is served by the same inclination orbits as the mainland US. However, the high number of zero-throughput samples in HI suggests areas of heavy occlusion, which disrupt performance. Finally, the RTT in the mainland is lower than in AK/HI due to the much shorter UE-server distances on the mainland route compared to the non-contiguous regions. Using rough calculations and the PoP locations in Fig. 11, we estimate the distance from Maui to the AWS server in Boardman, OR via the PoP in San Jose/LA to 4,752/5,326 km, corresponding to a one-way propagation latency of 24-27 ms, 2 and the UE-server distance from Anchorage/Fairbanks in AK via the PoP in Seattle to 2,589/2,723 km (13-14 ms). In contrast, the UE-server distance on the mainland route varies between 2,159 km (when the UE is in Omaha via the PoP in Denver) and 847 km (when the UE is in Salt Lake City using the PoP in the same city), corresponding to a propagation delay of only 4-11 ms. The much higher RTT in the non-contiguous US, particularly in HI, can also affect TCP performance.

The much higher fraction of zero-throughput samples in HI compared to AK is counter-intuitive, as HI is served by the 43/º53º inclination orbits consisting of a much denser constellation of satellites, compared to the 70º and 97.6º inclination orbits serving AK. To investigate this, we break down Starlink’s performance based on area type in Fig. 13. It shows that Starlink throughput exhibits different trends for the two regions. In AK (and the mainland US), Starlink DL/UL performance is higher in rural areas. This is expected and in line with the findings of previous studies [34], as tall buildings in urban areas cause occlusion to the satellite signal and degrade the performance. In contrast, in HI, the performance is higher in urban areas. We conjecture that this counterintuitive result stems from HI’s unique terrains. Unlike AK’s urban areas (Anchorage and Fairbanks), the urban areas of Maui, HI are characterized by the absence of vertical constructions. On the other hand, a large part of Maui’s rural areas are covered by dense rain forests, leading to higher levels of occlusion.

Fig. 14 supports our conjecture, showing the outage rate reported by the Starlink gRPC API – the fraction of outage duration relative to the total measurement duration in each region. The results show a significantly higher outage rate of 11.8% in rural HI, compared to just 0.3-2.5% across other combinations of region and area type. Note that an outage is reported only when the signal is 100% blocked. Given that rural areas in HI have a higher outage rate, we conjecture that the overall signal degradation was much higher due to occlusion, leading to worse performance in the rural areas of HI and a higher fraction of zero-throughput samples.

Fig. 13 also shows that the RTT is unaffected by the area type in all three regions. This is in contrast to cellular networks, where operator prioritization policies for different cellular technologies can result in drastic differences between the urban and rural RTTs (Fig. 10).

Remarks. In summary, similar to cellular networks, our analysis in this section reveals significant disparities in Starlink performance between the non-contiguous US regions and the mainland US, due to area-specific challenges, including unique terrains in Hawaii and sparse satellite deployment in Alaska. Additionally, the longer UE-server distances in the non-contiguous US compared to the mainland US significantly inflate the RTT in those regions.

一眼看结论:

  • 本土表现优异: Starlink 在美国本土的各项指标(吞吐量, 延迟, 可靠性)均显著优于非本土地区. 本土路线的零吞吐量样本仅为 2.6% (DL), 且延迟极低(中位数 56ms).
  • 非本土地区性能下降:
    • 可靠性差: AK 和 HI 的零吞吐量样本比例很高(AK: 9.4%, HI: 22.4%), 远高于本土.
    • 延迟高: AK (100ms) 和 HI (109ms) 的延迟约为本土的两倍, 主要原因是用户到本土 PoP/服务器的物理距离更远.

性能下降的具体原因分析:

  • 阿拉斯加 (AK) - 卫星稀疏: 性能下降主要是由于卫星部署稀疏(仅数百颗), 导致覆盖缺口, 服务中断和排队延迟增加.
  • 夏威夷 (HI) - 地形遮挡:
    • 尽管卫星覆盖密集, 但 HI 表现出极高的零吞吐量和中断率(农村地区中断率达 11.8%).
    • 这是由于 HI 农村地区茂密的热带雨林造成了严重的信号遮挡(Occlusion).

城乡差异的反直觉发现:

  • 通常情况 (AK & 本土): 农村地区性能优于城市地区, 因为城市高楼会阻挡卫星信号.
  • 夏威夷 (HI) 的反常: 城市地区性能优于农村地区. 原因是 HI 的城市缺乏高层建筑(视野开阔), 而农村地区被茂密植被覆盖(信号受阻).

延迟特征:

  • 与蜂窝网络不同, Starlink 的 RTT(往返时延)不受区域类型(城市 vs. 农村)的影响, 主要取决于物理距离.