LEO RECONFIGURATION IS CLOSELY RELATED TO NETWORK VARIATIONS

In this section, we introduce our analysis on the correlation between the reconfiguration of Starlink satellite links and end-perceived(终端感知) network variations. We find that the special LEO reconfiguration could be a useful indicator for discriminating non-congestion variations.

在这一部分，我们介绍了对Starlink卫星链路重新配置与终端感知网络变化之间相关性的分析。我们发现，特殊的低地球轨道（LEO）重新配置可能是区分非拥塞性变化的有效指标。

Reconfiguration of LEO satellite links is an important feature in today’s LSNs. Due to the high mobility of LEO satellites, the ground-to-satellite connections have to be reconfigured frequently to guarantee seamless and available satellite connectivity for terrestrial terminals. Taking Starlink as an example, in practice there are two mechanisms that jointly determine the reconfiguration in Starlink. First, according to the public document published by SpaceX [8], Starlink leverages a global re-configurator to plan connections on 15second intervals. This global re-configurator continuously re-generates and distributes a schedule of connection handovers between terminals and satellites, and between adjacent satellites. Several recent measurements (e.g., [26, 36, 46]) also confirm the existence of the global re-configurator from different vantage points around the world. Second, in addition to the global re-configurator, we find that there is a local re-configurator which may update the allocated satellite between two global reconfigurations. We observe that a local reconfiguration usually occurs when the terminal suffers from poor signal strength or is obstructed, as it should be too late to wait for a global reconfiguration.

低地球轨道（LEO）卫星链路的重新配置是当前低轨卫星网络（LSN）中的一个重要特征。由于LEO卫星的高机动性，地面与卫星之间的连接需要频繁地重新配置，以确保地面终端能够获得无缝且可用的卫星连接。以Starlink为例，实际上，Starlink的重新配置由两种机制共同决定。首先，根据SpaceX发布的公开文档[8]，Starlink利用一个全球重新配置器每15秒规划一次连接。该全球重新配置器持续生成并分发终端与卫星、以及相邻卫星之间的连接切换计划。若干近期的测量（例如，[26，36，46]）也从全球不同视角确认了全球重新配置器的存在。其次，除了 全球重新配置器 外，我们还发现存在一个 本地重新配置器，可能在两次全球重新配置之间更新分配的卫星。我们观察到， 当终端信号较弱或受阻时，通常会发生本地重新配置，因为此时等待全球重新配置已经为时过晚。

Note

存在两种重配置机制：

1. 全球范围重配置器
2. 本地重配置器

当终端信号较弱或受阻时，通常会发生本地重新配置（因为此时等待全球重新配置已经为时过晚）

Obtaining LEO reconfiguration on end host. In the live Starlink, the exact time when a reconfiguration occurs can be obtained on the end host. First, in our vantage points, the global reconfigurations occur at exactly the 12th, 27th, 42nd, and 57th second every minute. Second, local reconfigurations can be detected by analyzing the information provided by the satellite terminal. Concretely, Starlink’s terminal provides an internal gRPC service with APIs that expose some basic information to end host. Using the starlink-grpc-tools [9], we can extract the obstruction maps from the dish, which are 2-dimensional images plotting the trajectory of satellites allocated to the dish recently. The obstruction map updates every second and we can identify whether a local reconfiguration occurs by comparing the trajectory of connected satellite in two sequential obstruction maps.

在终端主机上获取LEO重新配置信息。在实际的Starlink网络中，可以在终端主机上获取重新配置发生的精确时间。首先，在我们的观察点上，全球重新配置每分钟的第12、27、42和57秒准确发生。其次，本地重新配置可以通过分析卫星终端提供的信息来检测。具体而言，Starlink的终端提供了一个内部gRPC服务，通过其API向终端主机暴露一些基本信息。利用starlink-grpc-tools [9]，我们可以从卫星接收盘中提取阻塞图，这些图像是二维的，绘制了最近分配给天线盘的卫星轨迹。阻塞图每秒更新一次，我们可以通过比较两个连续阻塞图中连接卫星的轨迹来识别是否发生了本地重新配置。

gRPC

What is gRPC?

在 gRPC 里客户端应用可以像调用本地对象一样直接调用另一台不同的机器上服务端应用的方法，使得您能够更容易地创建分布式应用和服务。与许多 RPC 系统类似，gRPC 也是基于以下理念：定义一个服务，指定其能够被远程调用的方法（包含参数和返回类型）。在服务端实现这个接口，并运行一个 gRPC 服务器来处理客户端调用。在客户端拥有一个存根能够像服务端一样的方法。

ref wiki

Note

获取LEO重配置信息（本地重配置）的方式有两种：

1. 在 Satellite Terminal 获取
2. 在 End Host 获取 (本质是Satellite Terminal提供gRPC服务，将信息传给用户终端)

The impact of reconfiguration on network variations is multifaceted. On the one hand, reconfiguration of satellite links may update the radio frame allocation for terminals which further changes the physical layer transmission rate of satellite links [19]. On the other hand, reconfiguration is also highly correlated to path fluctuations, because it may change the satellite allocated to the terminal and even update the path from the allocated satellite to a new ground station which is invisible to terminals [26]. Such path fluctuations can further exacerbate RTT variations and bursty loss.

Figure 6 plots a concrete example illustrating the correlation between Starlink reconfigurations and network variations. We obtain several interesting observations. First, the time-varying capacity/RTT can be approximately fitted as a step function divided by a sequence of reconfigurations. When switching to a new step, because the reconfiguration may change the link capacity and space segment path, both capacity and RTT may suddenly “jump” to a new state independent with the previous one. Second, inside a step, the capacity and RTT changes are relatively mild and smooth. This is likely because the network path does not fluctuate between two reconfigurations and network conditions are mainly affected by noises (e.g., wireless channel variations).

重新配置对网络变化的影响是多方面的。一方面，卫星链路的重新配置可能会更新终端的无线帧分配，从而进一步改变卫星链路的物理层传输速率[19]。另一方面，重新配置与路径波动高度相关，因为它可能会改变分配给终端的卫星，甚至更新从该卫星到新的地面站的路径，而这些路径对终端是不可见的[26]。这种路径波动可能进一步加剧RTT变化和突发丢包。

图6展示了一个具体示例，说明Starlink重新配置与网络变化之间的关联。我们得出了一些有趣的观察结果。首先， 时变的容量/RTT可以大致拟合为一个由一系列重新配置划分的阶梯函数 。当切换到新的阶梯时，由于重新配置可能会改变链路容量和空间段路径，容量和RTT可能会突然“跳跃”到一个与之前状态无关的新状态。其次，在 一个阶梯内，容量和RTT的变化相对温和和平滑 。这可能是因为网络路径在两次重新配置之间没有波动， 网络状况主要受到噪声（例如无线信道变化）的影响。

Note

重配置对卫星链路的影响：

1. 重配置信息 -> 更新终端 -> 卫星链路的物理层传输速率
2. 重配置信息 -> 路径波动 -> RTT突变 / Burst Pkt Loss

Statistical correlation analysis. Further, we perform a long-term statistical analysis to quantitatively analyze the correlation between LEO reconfiguration and the observed network variations. We run iperf3 to independently measure the maximum link capacity and non-congestion RTT on our vantage points for one month, based on which we obtain a large number of one-second samples of average capacity and min RTT measured in every second. We calculate the network variation Δ 𝑇 as the performance at 𝑇 second minus that in the previous 𝑇 − 1 second. Figure 7 plots the CDF of network variations associated with a global/local reconfiguration or not. We observe that an LEO reconfiguration is often associated with drastic network variations (e.g., significant capacity change or obvious propagation RTT variation). More than 80% of global reconfigurations involve high capacity variations with more than 10Mbps sudden increase or decrease. About 20% of the reconfigurations can even result in ≥40Mbps capacity variation between two adjacent seconds. About 90% of capacity variations during the non-reconfiguration period are less than 5Mbps. The similar observations can also be found in the RTT variations. The above correlation between reconfiguration and variations suggests that LEO reconfiguration is a critical factor that triggers non-congestion capacity and RTT variations in an LSN like Starlink.

统计相关性分析。进一步地，我们进行了一项长期统计分析，定量分析LEO重新配置与观察到的网络变化之间的相关性。我们使用 iperf3 在我们的观察点上独立测量最大链路容量和非拥塞RTT，测量时间为一个月，基于这些数据我们获得了大量的一秒钟样本，其中包括每秒测得的平均容量和最小RTT。我们将网络变化 Δ𝑇 定义为第 𝑇 秒的性能减去第 𝑇−1 秒的性能。图7展示了与全球/本地重新配置相关的网络变化的累积分布函数（CDF）。我们观察到，LEO重新配置通常伴随着剧烈的网络变化（例如，显著的容量变化或明显的传播RTT变化）。超过80%的全球重新配置涉及大于10Mbps的突发容量变化。约20%的重新配置甚至可能导致相邻两秒之间的容量变化达到 ≥ 40Mbps。在非重新配置期间，约90%的容量变化小于5Mbps。RTT变化也可以观察到类似的趋势。上述重新配置与网络变化之间的相关性表明，LEO重新配置是触发类似Starlink低轨卫星网络中非拥塞性容量和RTT变化的关键因素。

---

如何从fig7分析出上述结论的？如何分析CDF图？

图7分为两部分：

• (a) Link capacity variations (带宽变化)：展示了带宽差异（单位：Mbps）的CDF。
• (b) RTT variations (RTT变化)：展示了RTT差异（单位：ms）的CDF。

每条曲线对应不同的场景：

• Global Reconf（灰色实线）：全球重新配置时的网络变化。
• Local Reconf（橙色虚线）：本地重新配置时的网络变化。
• Non-Reconf（蓝色虚线）：没有重新配置时的网络变化。

CDF曲线的纵轴表示累计概率，横轴表示网络变化的幅度。例如，某点对应的纵轴值为0.8，意味着80%的样本数据在该点或更小范围内。

分析带宽变化 (图7a)

观察1：全球重新配置引发剧烈带宽变化

• 灰色实线（Global Reconf）在横轴为10 Mbps处，CDF值约为0.2。这意味着有80%的全球重新配置导致了超过10 Mbps的带宽变化。
• 在横轴为40 Mbps处，灰色实线的CDF值约为0.8。这说明约20%的全球重新配置导致了≥40 Mbps的带宽变化。

观察2：非重新配置期间带宽变化较小

• 蓝色虚线（Non-Reconf）在横轴为5 Mbps处，CDF值接近0.9。这表明非重新配置期间，约90%的样本带宽变化小于5 Mbps。

分析RTT变化 (图7b)

观察1：全球重新配置引发显著RTT变化

• 灰色实线（Global Reconf）在横轴为10 ms处，CDF值约为0.2。这说明80%的全球重新配置导致了超过10 ms的RTT变化。
• 在横轴为40 ms处，灰色实线的CDF值约为0.8。这表明约20%的全球重新配置导致了≥40 ms的RTT变化。

观察2：非重新配置期间RTT变化较小

• 蓝色虚线（Non-Reconf）在横轴为5 ms处，CDF值接近0.9。这意味着非重新配置期间，约90%的样本RTT变化小于5 ms。