NovaPlan: An Efficient Plan of Renting Ground Stations for Emerging LEO Satellite Networks

Abstract

Low-Earth Orbit (LEO) Satellite Networks (LSNs) are promising to provide wide Internet access and high-quality network services. Such potentials attract more operators to construct their own LSNs, e.g.,Amazon Kuiper and Telesat. Typically, in an LSN the ground stations (GSs) enable the spaceto-ground transmissions. However, due to the long construction period, high cost, and geographical restrictions, it is difficult for those operators whose LSNs are under construction to build sufficient geo-distributed GSs in the short term. Therefore, earlystage LSNs may suffer from poor network reachability under the quality of service (QoS) constraints due to the lack of GSs. Based on the emergence of “Ground-Station-as-a-Service (GSaaS)”, we present NOVA PLAN, a novel GS planning mechanism that can guide operators of developing LSNs to rent GSaaS services and schedule their traffic to guarantee high network availability in a cost-effective manner. Specifically, to efficiently solve the renting GS problem (RGP) with performance constraints, NOVA PLAN aggregates adjacent timeslots without drastic topology changes to decrease the GS combinatorial space and determines the renting policy of individual GS to lower the total cost. Moreover, NOVA PLAN judiciously schedules traffic to the path within latency constraints. Extensive evaluations based on real-world constellation information show that NOVA PLAN outperforms existing solutions with nearly 100% service rate and the lowest cost.

低地球轨道（LEO）卫星网络（LSN）有望提供广泛的互联网接入和高质量的网络服务。这些潜力吸引了诸如亚马逊Kuiper和Telesat等更多运营商构建自己的LEO卫星网络。通常，在LEO卫星网络中，地面站（GS）负责实现星地传输。然而，由于建设周期长、成本高昂及地理位置限制，那些其LEO卫星网络尚在建设中的运营商，很难在短期内建成足够数量且地理分布广泛的地面站。因此，处于发展初期的LEO卫星网络可能会因缺少地面站，而在满足服务质量（QoS）约束的前提下面临网络可达性差的问题。基于“地面站即服务（GSaaS）”这一新兴模式，我们提出了NovaPlan，一种新颖的地面站规划机制。该机制能够指导发展中的LEO卫星网络运营商租赁GSaaS服务并调度其业务流量，从而以经济高效的方式保证高网络可用性。具体而言，为有效解决带性能约束的地面站租赁问题（RGP），NovaPlan通过聚合拓扑结构无剧烈变化的相邻时隙来减小地面站的组合搜索空间，并确定单个地面站的租赁策略以降低总成本。此外，NovaPlan能够在延迟约束内审慎地调度流量路径。基于真实星座信息的大量评估表明，NovaPlan的性能优于现有解决方案，能达到近100%的服务率且成本最低。

Index Terms — LEO Satellite Network, Ground-Station-as-aService (GSaaS), Optimization.

Introduction

Mega Low-Earth Orbit (LEO) Satellite Networks (LSNs) are promising to provide wide Internet access and low-latency services to the global users [1]. As the cost of launching a satellite is getting lower [2], more operators are dedicated to deploying their LSNs, e.g., Amazon’s Kuiper [3] and Telesat [4]. These LSNs usually consist of tens to thousands of LEO satellites flying around the world. Thanks to the technology of high-speed and high-capacity laser inter-satellite links (ISLs) [5], long-haul and even aerial [6], [7] traffic can be forwarded to the terrestrial networks enabled by the ground stations (GSs) that work as gateways connected to satellites.

However, building sufficient terrestrial GSs globally is a struggle in the short term due to the long construction period, high building cost [8], and geographical restrictions [9], leading to a collision between wide-covered space segment and ground segment with scarce GSs, e.g., Amazon plans to launch 3236 satellites covering the majority of the earth while only owning 12 GSs [10]. Such imbalance may cause explicit problems: low coverage and high end-to-end latency.

From our simulation utilizing parameters of two earlystage constellations: Kuiper [3] and Telesat [4], and their GS distribution [10], [11], it shows that these early-stage LSNs suffer from extremely poor network reachability and high user-perceived latency. None of the globally chosen city pairs can continuously communicate through the two LSNs in one day. Specifically, the one-way latency can reach 284.75ms and

268.68ms in Kuiper and Telesat in the presence of ISLs.

In our further analysis, these problems are caused by two reasons: dynamics of LEO satellites and lack of GSs. The highly mobile LEO satellites fly around the earth, incurring frequent network interruptions. During a period, serviceable GSs cover certain orbits, however, after a while, satellites fly away with no GSs within their reachable range. Besides, the private GSs owned by individual operators are located in places with high user density, e.g., North America and Europe. Hence, quantitative satellites over other continents access GSs with difficulty. Moreover, the perceived latency of user pairs whose packets are forwarded to GSs is prolonged, especially between two distant GSs that are connected by wandering, lower-speed fibers or submarine optical cables, as the latency and physical distance between two terrestrial nodes showcase a linear relationship [12].

Fortunately, just like terrestrial Internet service providers, the LSN operators are witnessed to provide services to each other [9], [13]. “Ground-Station-as-a-Service” (GSaaS) [10],

[15] that enables operators to rent services like computing and transmission is more cost-effective than building a new GS. Hence, we design N OVA P LAN, an effective GS planning mechanism, enabling operators deploying new LSNs to rent GSs considering the coverage and latency constraints, thereby remedying the aforementioned deficiencies.

At a high level, N OVA P LAN combines offline and online stages. Initially, it rents GSs and schedules traffic based on the historical user distribution and traffic patterns and then determines the renting policy of each GS given its usage time and cost model, i.e., rented in a reserved or on-demand mode [16],

[17]. Then, it dynamically schedules the incoming traffic to those GSs that generate the lowest total cost while satisfying the constraint of end-to-end latency. Besides, to cope with newly-coming users that no viable GSs can serve, N OVA P LAN enables quick responsiveness by renting temporary GSs from global leasable GSs, utilizing the predictable satellite trajectory and estimated satellite-GS connectivity.

In the offline stage, we model the renting GS problem (RGP) which is NP-hard, whose objective is to rent and use GSaaS with the minimum cost under the constraints of service rate and quality of services (QoS). The large search space and high computational complexity due to the scale of leasable GSs and users and the dynamics of LSNs motivate us to design a lightweight solution. Specifically, N OVA P LAN computes the serviceable GSs of historical users within QoS constraints with the estimated satellite-GS connectivity and then aggregates the continuous timeslots to prune the unrelated GSs. To avoid renting redundant GSs that only satisfy the QoS of a few users, N OVA P LAN selects GSs with larger latency contributions (LC). LC measures the GS potential to reduce ground latency when it connects with the GSs of other users.

After that, for each flow, N OVA P LAN dynamically schedules traffic to the private GSs and those rented in a reserved way in priority to reduce the cost. If a newcomer cannot find GSs that satisfy its QoS demand, N OVA P LAN computes the potential GSs from global leasable GSs and immediately rents them in a pay-as-you-go mode to serve it.

Since the early-stage LSNs have not been put into use today, we carry out the simulations based on real constellation parameters (Kuiper and Telesat), distribution of global GSs [10], [11], [15], [18]–[20], and LSN traffic patterns collected from Cloudflare [21]. The results show that N OVA P LAN can achieve the near-optimal service rate and latency with the minimum cost, and scales well in the online stage compared with other state-of-the-art algorithms [22]–[24].

To sum up, our contributions are as follows:

• We reveal that the LSNs deployed in the early stage face poor network reachability and QoS problems and point out the root causes: dynamics of LSNs and lack of GSs. (§II)

• We present N OVA P LAN , a prospective mechanism tailored for those operators that are planning to build new LSNs to rent GSs economically and schedule traffic judiciously, providing wide Internet access and low-latency service in dynamic LSNs. (§III,§IV)

• Based on the results of the simulation that uses realworld constellation parameters, distribution of global GSs, cost models, and traffic patterns of operational LSNs [21], N OVA P LAN achieves nearly 100% service rate and scales well, reducing much more cost than other algorithms. (§V)

超大规模低地球轨道（LEO）卫星网络（LSN）有望为全球用户提供广泛的互联网接入和低延迟服务[1]。随着卫星发射成本的不断降低[2]，越来越多的运营商致力于部署自己的LEO卫星网络，例如亚马逊的Kuiper项目[3]和Telesat[4]。这些LEO卫星网络通常由数十到数千颗环绕地球飞行的LEO卫星组成。得益于高速、高容量的激光星间链路（ISL）技术[5]，长途甚至空中[6], [7]的业务流量可以通过作为网关连接卫星的地面站（GS）转发至地面网络。

然而， 在全球范围内建设充足的地面站，在短期内是一项艰巨的任务。这源于其漫长的建设周期、高昂的建造成本[8]以及地理位置的限制[9]，导致了覆盖广泛的空间段与地面站稀疏的地面段之间的矛盾 。例如，亚马逊计划发射3236颗卫星以覆盖地球大部分区域，但其目前仅拥有12个地面站[10]。这种不平衡可能引发明显的问题：即 低覆盖率和高昂的端到端延迟。

我们利用两个早期星座项目（Kuiper [3] 和 Telesat [4]）的参数及其地面站分布[10], [11]进行了仿真，结果表明，这些处于发展初期的LEO卫星网络面临着极其糟糕的网络可达性和极高的用户感知延迟。在全球范围内选择的城市对中，没有一对能够通过这两个网络在一天内实现持续通信。具体而言，在有星间链路的情况下，Kuiper和Telesat网络的单向延迟可分别高达284.75毫秒和268.68毫秒。

在我们的进一步分析中，这些问题由两个原因造成：LEO卫星的动态性和地面站的缺乏。高速移动的LEO卫星环绕地球飞行，导致频繁的网络中断。在某个时间段内，可服务的地面站能覆盖特定轨道；然而片刻之后，卫星便飞离，其可达范围内再无地面站。此外，运营商自建的私有地面站通常位于用户密度高的地区，如北美和欧洲。因此，位于其他大洲上空的卫星难以接入地面站。更有甚者，对于那些数据包需要通过地面站转发的用户对，其感知延迟会被拉长，尤其是当两个地面站相距遥远，需要通过迂回且速度较慢的光纤或海底光缆连接时，因为地面两节点间的延迟与物理距离呈现出线性关系[12]。

幸运的是，正如地面互联网服务提供商之间会互相提供服务一样，我们观察到LEO卫星网络运营商之间也存在此类合作[9], [13]。“地面站即服务”（GSaaS）[10], [15]模式允许运营商租赁计算和传输等服务，这比新建一个地面站更具成本效益。因此，我们设计了NovaPlan，一个高效的地面站规划机制，旨在帮助部署新兴LEO卫星网络的运营商在考虑覆盖范围和延迟约束的条件下租赁地面站，从而弥补上述缺陷。

从宏观上看，NovaPlan结合了离线和在线两个阶段。在初始阶段，它根据历史用户分布和流量模式租赁地面站并调度流量，然后依据每个地面站的使用时间和成本模型（即以预留模式或按需模式租赁[16], [17]）来确定其租赁策略。随后，在动态调度阶段，它将传入的流量导向那些在满足端到端延迟约束的同时能产生最低总成本的地面站。此外，为了应对新出现的、没有可用地面站服务的用户，NovaPlan能够利用可预测的卫星轨迹和估算的星地连接性，从全球可租用的地面站中快速租赁临时站点，以实现快速响应。

在离线阶段，我们将地面站租赁问题（RGP）建模为一个NP难问题，其目标是在满足服务率和服务质量（QoS）约束的前提下，以最低成本租赁和使用GSaaS。可租赁地面站和用户规模的庞大，以及LEO卫星网络的动态性，共同导致了巨大的搜索空间和高计算复杂性，这促使我们设计一种轻量级解决方案。具体而言，NovaPlan首先利用估算的星地连接性计算出满足QoS约束的历史用户的可用地面站，然后聚合连续的时隙以剔除不相关的地面站。为避免租赁那些仅能满足少数用户QoS的冗余地面站，NovaPlan会选择具有较大“延迟贡献度”（Latency Contributions, LC）的地面站。LC度量了一个地面站在与其他用户所连接的地面站建立通信时，其降低地面段延迟的潜力。

此后，对于每一个数据流，NovaPlan会优先将其动态调度至私有地面站和以预留方式租赁的地面站，以降低成本。如果一个新用户找不到满足其QoS需求的地面站，NovaPlan会从全球可租赁的地面站中计算出潜在可用的站点，并立即以“按需付费”（pay-as-you-go）模式租用它们来提供服务。

由于这些处于发展初期的LEO卫星网络至今尚未投入使用，我们基于真实的星座参数（Kuiper和Telesat）、全球地面站的分布[10], [11], [15], [18]–[20]以及从Cloudflare收集的LEO卫星网络流量模式[21]进行了仿真。结果表明，与其它先进算法[22]–[24]相比，NovaPlan能够以最低的成本实现近乎最优的服务率和延迟，并在在线阶段展现出良好的可扩展性。

综上所述，我们的贡献如下：

• 我们揭示了 处于发展初期的LEO卫星网络面临着糟糕的网络可达性和服务质量问题，并指出了其根本原因：LEO卫星网络的动态性和地面站的缺乏
• 我们提出了NovaPlan，一个为计划建设新兴LEO卫星网络的运营商量身定制的前瞻性机制，旨在帮助他们经济地租赁地面站并审慎地调度流量，从而在动态的LEO卫星网络中提供广泛的互联网接入和低延迟服务
• 基于使用真实星座参数、全球地面站分布、成本模型以及在运行的LEO卫星网络流量模式[21]所进行的仿真结果，NovaPlan实现了近100%的服务率，并表现出良好的可扩展性，相比其他算法节省了更多的成本

Background and Motivation

A. Preliminaries for LSNs

Low-Earth Orbit (LEO) Satellite Networks (LSNs). An LSN usually consists of tens to thousands of LEO satellites flying around the world, and the orbital altitude is less than 2000km [25], which is promising to serve global users and further lower the end-to-end latency in long-haul communication compared to the terrestrial networks. Besides those operational LSNs like SpaceX [26] and OneWeb [20], many other LSN operators are planning to deploy their own constellations. For example, Kuiper [3] is said to consist of 3236 satellites in three shells with different orbital inclinations and altitudes, and Telesat [4] consists of 2340 satellites in two different shells. Architecture of LSNs. Fig. 1 plots the architecture of the current LSNs [27], [28]. An LSN is composed of Space segment and Ground segment. Dynamic satellites working as forwarders compose the space segment. In the ground segment, users access the LSN through the user-satellite links (USLs). Satellites transmit data to ground stations (GSs) that function as gateways and enable this space-to-ground transmission through ground-satellite links (GSLs). Many providers are planning or have deployed their own GSs globally [10], [11], [15], [18]–[20] as shown in Fig. 2. Nowadays, most satellites forward packets in a bent-pipe way [29], i.e., a satellite directly transmits data to a GS within its visible range. Additionally, some operators have already equipped [30] or plan to equip their satellites with high-speed laser inter-satellite links (ISLs) [31], [32], enabling long-haul traffic transmission.

低地球轨道（LEO）卫星网络（LSNs）

LEO卫星网络通常由数十到数千颗环绕地球飞行的卫星组成，其轨道高度低于2000公里[25]。与地面网络相比，它有望为全球用户提供服务，并进一步降低长途通信的端到端延迟。除了像SpaceX [26]和OneWeb [20]这样已在运行的LEO卫星网络外，许多其他运营商也正计划部署自己的星座。例如，据称Kuiper [3]星座将包含3236颗卫星，分布在三个具有不同轨道倾角和高度的轨道层上；而Telesat [4]星座则由2340颗卫星构成，分布在两个不同的轨道层。

LEO卫星网络的架构

图1 描绘了当前LEO卫星网络的架构[27], [28]。一个LEO卫星网络由空间段（Space segment）和地面段（Ground segment）构成。作为转发节点的动态卫星组成了空间段。在地面段，用户通过用户-卫星链路（USLs）接入LEO卫星网络。卫星则通过地面-卫星链路（GSLs）将数据传输到作为网关的地面站（GSs），从而实现星地传输。如图2所示，许多服务商正在计划或已经全球性地部署了他们自己的地面站[10], [11], [15], [18]–[20]。如今，大多数卫星以“弯管”（bent-pipe）方式转发数据包[29]，即卫星直接将数据传输至其可视范围内的地面站。此外，一些运营商已经[30]或计划[31], [32]为其卫星配备高速激光星间链路（ISLs），以支持长途流量传输。

B. Early-stage LSNs face challenges

Considering long construction periods, deployment cost [8], rugged terrains and political factors [33], etc., operators whose LSNs are under construction have difficulty in building plenty of GSs like Starlink in the short term. Hence, although LSNs are promising to provide high-quality Internet services, these LSNs deployed in an early stage cannot attain continuous network connectivity and service accessibility. Further, even though some users can access an LSN, they still suffer from high end-to-end latency.

Dataset and methodology. We randomly select 40 communication city pairs from around the world where the potential customers of current LSNs [21] are located. We divide a continuous period of 24 hours into discrete timeslots with an interval of 2 minutes. We simulate certain shells of two constellations under construction based on the real parameters: Telesat (40 orbits, 33 satellites per orbit) and Kuiper (34 orbits, 34 satellites per orbit). Their operators now own 14 [11] and 12 [10] GSs, respectively. Fig. 2 depicts the distribution of globally planned or deployed GSs. Since some operational satellites only emit at most 3 ISLs [30] or work in the absence of ISLs, i.e., bent-pipe (bp) [29], we set two LSN topologies: ring (two ISLs connect the neighboring satellites in the same orbit considering the intra-orbit connectivity is stable, and the remaining one is the backup) and bp. We follow the routing methodology in [34], i.e., to schedule each flow to the path with the lowest latency. We analyze the network reachability by calculating the proportion of timeslots when a satellite can reach a GS directly or through ISLs. Besides, we observe the end-to-end one-way latency of city pairs who are served.

考虑到建设周期长、部署成本高[8]、地形崎岖以及政治因素[33]等，那些其LEO卫星网络尚在建设中的运营商，很难在短期内建成像Starlink那样数量庞大的地面站。因此，尽管LEO卫星网络有望提供高质量的互联网服务，这些处于发展初期的网络却无法实现持续的网络连通性与服务可达性。更有甚者，即使用户能够接入网络，他们也仍需忍受高昂的端到端延迟。

数据集与研究方法。 我们从全球范围内随机选取了40个通信城市对，这些城市是当前LEO卫星网络潜在客户的所在地[21]。我们将连续的24小时划分为时间间隔为2分钟的离散时隙。我们基于真实的星座参数，对两个建设中的星座的部分轨道层进行了仿真：Telesat（40个轨道，每轨道33颗卫星）和Kuiper（34个轨道，每轨道34颗卫星）。它们的运营商目前分别拥有14个[11]和12个[10]地面站。图2描绘了全球已规划或已部署的地面站分布。由于部分在轨卫星最多只发射3条星间链路[30]或在没有星间链路的情况下工作（即“弯管”模式，bent-pipe, bp）[29]，我们设定了两种LEO卫星网络拓扑：环状（ring）拓扑（考虑到轨道内连接的稳定性，用两条ISL连接同轨道内的相邻卫星，剩下的一条作为备份）和弯管（bp）拓扑。我们遵循[34]中的路由方法，即将每个数据流调度至延迟最低的路径。我们通过计算卫星能够直接或通过ISL到达地面站的时隙比例来分析网络可达性。此外，我们观察了被服务城市对的端到端单向延迟。

Based on the experiments, our observations are as follows.

i) Poor network reachability. In our simulation, we define the service time of a satellite as the length of time when a satellite can find serviceable GSs covering its orbit during a period, i.e., it can forward the packets to these GSs. The service time demonstrates the ability of an LSN to access the Internet. Fig. 3 plots the CDF of the proportion of service time of two LSNs under different topologies during one day. In Kuiper and Telesat under ring topology, a satellite can find serviceable GSs only 50.82% and 45.43% of a day on average, respectively, and the performance is even worse in bp with only less than 1% in both constellations.

ii) High end-to-end latency. Fig. 4 plots the average one-way latency (right side) of 10 city pairs of the selected 40 pairs in the presence of ISLs during one day and their service rates, i.e., the proportion of service time (left side). In two constellations, none of these communication pairs can access an LSN continuously. The pair of Santiago-Atlanta nearly cannot access any of the two LSNs because Santiago is only surrounded by a few GSs. Besides, most of these ten pairs perceive a long end-to-end latency, e.g., the maximum one-way latency from Singapore to Perth reaches 227.93ms during a day due to the long distance between two Kuiper GSs. Takeaways. Operators that aim at deploying new LSNs while lacking sufficient GSs have difficulty in achieving the goal of providing worldwide, continuous, and low-end-to-end latency network services. The phenomena of these early-stage LSNs are widely found around the world.

i) 网络可达性差。 在我们的仿真中，我们将卫星的“服务时间”定义为：在一段时间内，一颗卫星能找到覆盖其轨道的可用地面站的时长，即它能够将数据包转发至这些地面站。服务时间展示了LEO卫星网络接入互联网的能力。图3描绘了两种LEO卫星网络在不同拓扑下，一天内服务时间占比的累积分布函数（CDF）。在环状拓扑下，Kuiper和Telesat的卫星平均每天只有50.82%和45.43%的时间能找到可服务的地面站；而在弯管拓扑下，性能更差，两个星座的服务时间占比均不足1%。

ii) 端到端延迟高。 图4展示了所选40个城市对中的10对，在有星间链路的情况下，一天内的平均单向延迟（右侧）及其服务率（即服务时间占比，左侧）。在这两个星座中，没有任何一个通信对能够持续接入网络。圣地亚哥-亚特兰大这对组合几乎无法接入任何一个网络，因为圣地亚哥周围只有极少数的地面站。此外，这十对中的大多数都承受着很长的端到端延迟，例如，由于两个Kuiper地面站之间距离遥远，从新加坡到珀斯的最高单向延迟在一天内达到了227.93毫秒。

结论要点。 那些旨在部署新兴LEO卫星网络但缺乏足够地面站的运营商，难以实现其提供全球性、持续且低端到端延迟网络服务的目标。这种发展初期LEO卫星网络的现象在世界范围内普遍存在

C. Root cause analysis

i) Dynamics of LEO satellites. LEO satellites move dynamically around the world. For only a few minutes, a satellite connects to a certain GS within its visible range or reaches the GS through ISLs. After a while, it flies away with no serviceable GSs covering its orbit. This frequent topology variation causes a low proportion of satellite service time in Fig. 3 that leads to a low service rate in Fig. 4. In our further analysis, the average maximum continuous service time of a satellite is only about one hour and 2.5 hours in Kuiper and Telesat, respectively. Such frequent interruptions inevitably lead to the unavailability of LSNs.

ii) Lack of ground stations. As illustrated in Fig. 2, due to the restricted quantity of GS owned by an individual LSN operator, only a few GSs cover the orbit of a satellite. This situation becomes even more pronounced under bentpipe topology. Besides, due to the uneven GS distribution, e.g., Asia and South Africa lack GSs while Europe has dense GSs, packets are forwarded to distant GSs through ISLs. Such detours prolong the latency, especially in the ground segment as the latency between two terrestrial nodes has a linear relationship with their physical distance [12]. For example, in Fig. 5a, the terrestrial latency occupies 68.90% of the one-way latency from Singapore to Perth. Because under the ring topology, the closest GS that covers the orbit of Perth’s access satellite is in North America. The terrestrial distance between two served GSs is such a long way that it crosses continents. Unfortunately, challenges, such as the long construction period and terrain restrictions [8], [33], make it difficult to deploy widely distributed ground infrastructure.

i) LEO卫星的动态性。 LEO卫星在全世界范围内高速移动。一颗卫星可能仅在几分钟内连接到其可视范围内的某个地面站，或通过星间链路到达该地面站。片刻之后，它便飞离，其轨道再无可用地面站覆盖。这种频繁的拓扑变化导致了图3中卫星服务时间的占比很低，进而造成了图4中的低服务率。在我们的进一步分析中，Kuiper和Telesat中一颗卫星的平均最长连续服务时间分别仅为约1小时和2.5小时。如此频繁的中断不可避免地导致了LEO卫星网络的不可用。

ii) 地面站的缺乏。 如图2所示，由于单个LEO卫星网络运营商拥有的地面站数量有限，仅有少数地面站能覆盖一颗卫星的轨道。这种情况在弯管拓扑下尤为明显。此外，由于地面站分布不均（例如，亚洲和南非缺少地面站，而欧洲的地面站很密集），数据包需要通过星间链路被转发到遥远的地面站。这种绕行延长了延迟，尤其是在地面段，因为两个地面节点间的延迟与其物理距离呈线性关系[12]。例如，在图5a中，从新加坡到珀斯的单向延迟中，地面延迟占了68.90%。这是因为在环状拓扑下，覆盖珀斯接入卫星轨道的最近地面站位于北美。两个被服务地面站之间的地面距离非常遥远，以至于跨越了多个大洲。不幸的是，建设周期长和地形限制等挑战[8], [33]，使得部署广泛分布的地面基础设施变得困难。

Opportunities. Although an individual LSN operator has difficulty in providing continuous and low-latency services, renting GSs is an efficient and economical way to improve the performance of its LSN, owing to: i) lower renting cost. Thanks to the emergence of “Ground Station as a Service (GSaaS)” [10], [14], [15], billing by antenna usage time brings an opportunity to rent GS services which is much more economical compared with constructing global GSs; and ii) more choices of serviceable GSs. Integrated GSs of different providers demonstrate more abundant choices of serviceable GSs, compensating for the scarce GSs of a single operator. For example, when renting a Starlink GS near Perth, the one-way end-to-end latency is largely decreased by 79.51% in Fig. 5b, as the two access satellites can directly connect to their nearby GSs.

尽管单个LEO卫星网络运营商难以提供持续和低延迟的服务，但租赁地面站是提升其网络性能的一种高效且经济的方式，这得益于:

i) 更低的租赁成本。 “地面站即服务”（GSaaS）[10], [14], [15]的出现，带来了按天线使用时间计费的模式，这为租赁地面站服务提供了机会，远比建设全球性的地面站经济得多

ii) 更多的可用地面站选择。 集成不同提供商的地面站可以提供更丰富的可用地面站选择，弥补了单个运营商地面站稀缺的短板。例如，在图5b中，当在珀斯附近租用一个Starlink地面站后，由于两个接入卫星能直接连接到邻近的地面站，单向端到-端延迟大幅降低了79.51%

TL; DR

(1) 离线规划阶段

它不看具体的某个包裹，而是看历史数据信息

1. 聚合时隙: 没必要为全天的每一分钟做计划
   1. 它会把一天分为几个大时段，比如“早高峰”、“午间平峰”、“晚高峰”
   2. 设计思考: 在同一个时段内，卫星的飞行路线 和 用户的包裹数量 大体上是相似的
2. 剔除无关仓库: 换言之, 剪枝
   1. 对于一个在亚洲的客户，很显然我们要直接忽略掉 远在北美 的仓库
3. 筛选“黄金节点”仓库: 面对很多很多的仓库, 我们基于 延迟贡献度 (LC) 进行筛选
   1. LC: 比如对于 Beijing - NewYork 这条线路，现在我们只能选择一个"中继仓库"。很明显, 选择一个北极的节点要优于一个欧洲的节点。因此, LC(北极) > LC(欧洲)
4. 制定“租赁合同”:
   1. 预留模式（包年/包月）: 对于那些“延迟贡献度”极高、天天都会用到的“黄金节点”仓库，直接签长期合同。虽然每天都付钱，但单价便宜
   2. 按需模式（临时清单）: 对于其他一些可能偶尔会用到的仓库，不签长租，而是把它们放进一个 “备用清单” 里

(2) 在线调度阶段

1. 成本优先原则
   1. 能不能用上我们自己的私有地面站
   2. 或者那些已经包了年的"黄金仓库"（预留模式租赁）
2. 应急响应机制
   1. 立刻扫描“备用临时仓库清单”
   2. 挑一个原来的备用仓库, 选择“激活”
   3. 保障灵活性: 虽然这次租赁的单价贵一点，但成功地完成了任务，保证了服务质量（QoS）