跳转至

Satellite and Cellular Network Synergy in the Non-Contiguous US

Given the inherent limitations of both cellular deployments and the Starlink network in the noncontiguous US regions, it is natural to consider multipath transport as a means for enhancing both performance and reliability in these regions. In this section, we evaluate the benefits of multipath transport using MPTCP [52], the most widely used multipath transport protocol [12, 13, 19, 32, 34, 39, 59, 60, 62, 64]. We compare the performance of MPTCP vs. single-path TCP via trace-driven emulation using iperf on MpShell [13], a variant of Mahimahi [57], following the methodology in [34].

Experimental methodology. We use the DL and UL throughput traces from our driving datasets in AK and HI and convert them to packet traces for replay on MpShell. Since the traces were collected simultaneously with different phones over different networks, they reflect the network conditions experienced by a user at the same location and time. We ran two types of emulation in the MpShell environment: (1) Single-path emulation, where MPTCP is disabled and two iPerf clients communicate with an iPerf server individually over different MpShell interfaces; (2) MPTCP emulation, where a single iPerf client communicates with an iPerf server over both MpShell interfaces using MPTCP.

For MPTCP, we experimented with different schedulers and congestion control algorithms. We found that the packet scheduler has minimal impact on performance, but BBR significantly outperforms all the other congestion control algorithms (Cubic, LIA [11], OLIA [39], BALIA [59]). Hence, we show in the following the results with the default scheduler, BLEST [19], and BBR. For a fair comparison, we also use BBR for the single-path TCP results. Similar to [34], we found that MPTCP brings minimal gains over single-path TCP with the default buffer size. Hence, we ran the emulation with a buffer of 450 MB for both MPTCP and single-path TCP, same as in our driving tests. Finally, for each operator, we set the RTT equal to the minimum RTT extracted from our AK and HI ICMP ping measurements. This represents the default RTT (propagation delay). Changes to the RTT, e.g., due to congestion or packet losses, are reflected in the packet timestamps used to drive the emulation.

  • 目标: 鉴于蜂窝网络和 Starlink 在非本土地区(AK 和 HI)各有局限, 现在我们希望利用 "多路径传输协议(MPTCP)" 来增强性能和可靠性
  • 方法:
    • 使用路测收集的下行(DL)和上行(UL)吞吐量轨迹, 在 MpShell 仿真环境中进行回放
    • 对比了单路径 TCP 与 MPTCP 的性能
  • 配置优化: 实验发现 BBR 拥塞控制算法表现最佳, 因此在 MPTCP(配合 BLEST 调度器)和单路径测试中均使用 BBR, 并设置了 450MB 的大缓冲区以适应长延迟环境

DL throughput. Fig. 18 shows boxplots of the single-path DL throughput of all networks and the MPTCP throughput for every pair of networks in AK and HI, respectively. We also evaluate the difference between the MPTCP throughput and the single-path throughput over the best path (𝑀𝑃𝑇𝐶𝑃 − 𝑀𝑎𝑥). Recall that one of the three main goals in the design of MPTCP (RFC 6536 [11]) is that the MPTCP throughput should be higher than the single-path throughput over the best path. Finally, to evaluate how close MPTCP’s performance is to the optimal performance, we plot the difference between the sum of the two single-path throughputs and MPTCP’s throughput (𝑆𝑢𝑚 − 𝑀𝑃𝑇𝐶𝑃).

In AK (Fig. 18a), MPTCP provides a substantial increase in performance when Starlink is combined with either of the cellular operators, compared to using a single network, while it also reduces the outage time to less than 1%. The median MPTCP throughput is 85 Mbps for SL+VZ and 92 Mbps for SL+AT vs. 64/15/23 Mbps for SL/VZ/AT alone. More importantly, the 25-th percentile of the MPTCP throughput is 64 Mbps for SL+VZ and 65 Mbps for SL+AT vs. 37/0.3/6 Mbps for SL/VZ/AT alone. These improvements are particularly valuable in rural areas, where cellular networks frequently fail to provide connectivity. On the other hand, combining two cellular operators is less effective; the median MPTCP throughput for VZ+AT is only 55 Mbps, much lower than when one of the two cellular networks is combined with Starlink. More importantly, the 25-th percentile remains very low (12 Mbps) due to the large fraction of concurrent zero-throughput samples (Table 1).

We make two additional observations from Fig. 18a for every pair of networks: (i) MPTCP consistently provides performance better than or almost as good (within 2-3 Mbps) as the singlepath performance of the best network. (ii) MPTCP’s performance is very close to the theoretical optimal; the 75-th percentile of 𝑆𝑢𝑚 − 𝑀𝑃𝑇𝐶𝑃 is less than 3 Mbps for all three pairs of networks.

In HI (Fig. 18b), we observe trends similar to AK, but with two key differences. First, both Starlink and cellular networks in HI experience a higher fraction of zero-throughput samples and a higher fraction of concurrent outage time compared to AK (Table 1). As a result, combining Starlink with Verizon or AT&T brings moderate benefits in the median case compared to using Starlink alone, increasing the median throughput from 71 Mbps (with SL) to 93/92 Mbps (using SL+VZ/SL+AT). Second, T-Mobile’s 5G mid-band coverage in HI leads to significantly higher gains for multipath transport. MPTCP with T-Mobile as one of the networks provides the highest bestcase performance, surpassing 200 Mbps at the 75-th percentile and offering peak throughput of 500-600+ Mbps. However, combining T-Mobile with another cellular network still yields lower median throughput and significantly lower worst-case throughput (0-2 Mbps at the 25-th percentile) compared to combining Starlink with Verizon or AT&T. This result suggests that combining Starlink with cellular networks has the best potential to simultaneously boost both performance and reliability.

Fig. 18b shows that, similar to AK, MPTCP in HI consistently improves the performance over the best path (the 25-th percentile of 𝑀𝑃𝑇𝐶𝑃 − 𝑀𝑎𝑥 is 0 for all pairs of networks), although there are some notable exceptions, especially in the case of AT+TM. Also, MPTPC generally performs again quite close (but not as close as in AK) to the optimal most of the time, although again there are some notable exceptions. These exceptions include T-Mobile as one of the two networks, which often offers much higher throughput than the other operators (Fig. 15a) thanks to its 5G-mid service, confirming the well-known results that MPTCP performs suboptimally in the case of highly heterogeneous paths.

  • Starlink + 蜂窝网络效果最佳:

    • 在阿拉斯加(AK), 将 Starlink 与任一蜂窝运营商(如 SL+Verizon)结合, 相比单网络能带来显著的性能提升, 并将中断时间减少到 1% 以下
      • MPTCP 的中位数吞吐量和尾部(第 25 百分位)吞吐量均大幅提高, 这对于经常断连的农村地区尤为重要
      • 在夏威夷(HI), 尽管两类网络的中断率都较高, 结合 Starlink 与蜂窝网络仍能提升中位数吞吐量(从 71 Mbps 提升至 92-93 Mbps)

  • 蜂窝网络 + 蜂窝网络效果较差:

    • 结合两个蜂窝网络(如 Verizon+AT&T)的效果不如结合 Starlink, 尤其是在同时发生零吞吐量(断连)的情况下, 性能提升有限

Multi-path TCP Takeaways:

  1. 接近理论最优: MPTCP 的性能通常优于或接近最佳单路径的性能, 且非常接近理论上的吞吐量之和(Sum-MPTCP 差距很小)
  2. 异构路径挑战: 在 HI, T-Mobile 的 5G-mid 速率极高, 导致路径间差异巨大, 虽然 MPTCP 仍有提升, 但不如在 AK 那么接近最优

UL throughput. Fig. 19 shows the results for UL throughput. Similar to DL, multipath transport also shows great potential in improving UL performance. In contrast to DL, here combining two cellular networks yields similar median performance and higher best-case performance (above the 75-th percentile) compared to combining Starlink with a cellular network, as cellular networks often achieve higher UL throughput than Starlink (Fig. 15b). On the other hand, combining Starlink with a cellular network still offers higher reliability, thanks to the higher Starlink availability (lower fraction of zero-throughput samples). In particular in HI, Fig. 19b shows that all 3 pairs of cellular networks yield 0 Mbps at the 25-th percentile with MPTCP, while the combination of Starlink with a cellular network yields non-zero throughput at the 25-th percentile. Interestingly, MPTCP always achieves near-optimal performance in the UL direction in both regions. The max value for 𝑆𝑢𝑚 − 𝑀𝑃𝑇𝑃𝐶 never exceeds 2.5 Mbps.

性能与可靠性的权衡:

  • 速度优势:
    • 结合"两个蜂窝网络"在上行方向通常能获得比"Starlink + 蜂窝"更高的中位数和峰值吞吐量
      • 因为蜂窝网络的上行速度通常优于 Starlink

  • 可靠性优势: 结合 "Starlink与蜂窝网络"提供了更高的可靠性

    • 例如在 HI, 仅靠蜂窝网络组合在第 25 百分位时吞吐量为 0, 而加入 Starlink 后则能保持非零吞吐量

  • 接近最优: 在上行方向, MPTCP 始终能实现接近最优的聚合性能

Remarks. In summary, multipath transport can provide significant improvements in terms of both performance and reliability, leveraging the complementary characteristics of Starlink and cellular networks. In the DL direction, combining Starlink with a cellular network shows the greatest potential, as it can simultaneously boost throughput and reliability, thanks to Starlink’s superior performance and higher availability compared to cellular networks. On the other hand, in the UL direction, Starlink’s lower TCP throughput compared to cellular networks introduces an interesting tradeoff between performance and reliability. Combining two cellular networks shows greater potential in improving the best-case performance, while combining Starlink with a cellular networks still offers higher reliability.

  • 互补性强: MPTCP 利用了 Starlink 和蜂窝网络的互补特性
    • Starlink 的覆盖/下行优势
    • 蜂窝网络的上行峰值速率优势
  • 最佳策略:
    • 下行方向: 结合 Starlink + 蜂窝网络 是最佳选择, 能同时大幅提升吞吐量和可靠性
    • 上行方向: Trade-off
      • 若追求极速可选择"两个蜂窝网络"
      • 若追求连接的稳定性(可靠性)则应选择"Starlink+蜂窝网络"