StarPerf: Characterizing Network Performance for Emerging Mega-Constellations

Danger

本文对于笔者而言, 重在学习其写作与表达方式

学习ICNP模板下, simu 类文章的写作方式

(1) 论文背景与动机

• 背景: 以 Starlink, OneWeb 为代表的"新太空"巨型星座 (Mega-Constellations) 正在兴起, 承诺提供全球高容量, 低延迟的通信
• 问题:
  • 目前对这些新兴系统的架构, 负载以及拓扑选择对网络性能的影响知之甚少
  • 由于卫星的高动态性 (High Mobility) 和部署的高成本, 直接测量或评估其性能非常困难
• 目标: 设计一个仿真平台, 帮助制造商和内容提供商评估不同星座设计下的网络性能

(2) STARPERF 仿真平台

论文提出了 STARPERF 平台, 主要包含以下关键技术与模块:

输入与建模

• 拓扑模型: 支持定义轨道参数 (倾角, 高度, 相位等), 地面站分布, 链路类型 (激光/微波, ISLs/SGLs)
• 路由策略: 支持分布式 (如 OSPF) 和集中式 (如 DSR) 路由
• 用户需求: 基于地理网格 (H3) 和人口分布生成流量矩阵

性能评估指标

• 覆盖率 (Coverage): 包括区域覆盖率和热点区域覆盖率
• 端到端延迟 (Area-to-Area Latency): 主要受限于光速和路径长度, 忽略处理延迟
• 吞吐量 (Area-to-Area Throughput): 基于最大流计算
• 弹性 (Resilience): 使用图论中的介数中心性 (Betweenness Centrality) 来评估抗毁性

(3) 典型星座评测

作者利用 STARPERF 对三个主流星座进行了对比分析: Starlink (SpaceX), OneWeb, TeleSat.

• 延迟: 巨型星座在长距离 (尤其是跨洲) 通信上比地面光纤具有更低的延迟潜力, 但由于卫星高速运动, 路径变化会导致较大的延迟抖动 (Jitter)
• 覆盖: 所有星座都能很好地覆盖主要人口密集区
• 弹性: Starlink 的卫星分布均匀, 节点介数中心性差异小, 网络弹性较好; 而 TeleSat 存在中心性极高的节点, 可能成为瓶颈

(4) 应用案例: 低延迟中继选择

• 场景: 实时通信 (RTC) 应用 (如视频会议)
• 方案: 提出了一种自适应中继选择算法, 在云平台和 LEO 卫星之间智能选择最佳中继节点
• 效果: 实验显示, 通过卫星-云集成基础设施, 端到端通信延迟最多可降低 62%

Introduction

Constructing "NewSpace" satellite networks (SN) is gaining popularity in recent years, as SNs offer the promising potential to provide low-latency, high-throughput global Internet connectivity. We have witnessed a gold run to build constellations consisting of a large number of low earth orbit (LEO) satellites, namely "mega-constellations", with players like OneWeb [7], Amazon [1] and SpaceX [12] entering the market. The latter one, which is the largest commercial satellite constellation operator in the world since January 2020, is actively constructing the Starlink constellation that consists of thousands of mass-produced small satellites in LEO, and will be initially available to customers in Canada and in the northern United States in 2020, with additional service expansion to other areas of the world throughout 2021 [11].

Fundamentally, mega-constellations facilitate the Internet by extending the connectivity of existing terrestrial networks (i.e., integrated satellite-terrestrial networks), and breaking the inherently physical constraints in today's Internet deployment. First, SNs expand the Internet coverage at broadband speeds to the remote area where access might be unreliable, expensive, or completely unavailable. Second, constellations with thousands of satellites working in LEO enable new opportunities for constructing a network in space to provide low-latency communication. Modern LEO satellites can equip optical inter-satellite links (ISLs) for inter-satellite communication. In the free-space, data packets can propagate in the speed of light in vacuum, which is much faster than that in the terrestrial fiber. Therefore the latency penalty in space might possibly be lowered by avoiding long-distance and meandering fiber routes. Finally, SNs are also expected to enhance the network throughput, since the rapid evolution of on-board technology and the increase in power generation have led to the evolution of high-bitrate satellite [39], which can provide tens and even hundreds of Gbps data rate.

While above exciting prospects depict a blooming picture of the future integrated satellite-terrestrial networks, the community still has very limited understanding of the topological characteristic and the attainable network performance of modern mega-constellations. Quantitatively profiling mega-constellations is meaningful for designing, using, and optimizing SNs, but it also faces a series of practical challenges: (i) currently emerging satellite constellations such as Starlink are still under heavy development, and the deployment of satellites is costly and time-consuming. It is thus difficult to directly measure the network performance from a completely deployed constellation system; (ii) SNs are fundamentally different from terrestrial networks. Emerging satellite networks inherently expose two particular features: only relatively nearby satellites can connect to each other due to the limited range of ISLs, and satellites are moving in high-speed with respect to ground stations and each other [19]. Such particular features make it difficult for existing network profiling methodologies to accurately characterize the mega-constellations. Several existing works tried to model and analyze the characteristic of novel constellations but they mostly focus on modeling the system capacity under different physical layer payload decision [21], [43], which ignores the impact of using different constellation options or network policies on the user-perceived network performance.

In this paper we present the design and implementation of STAR PERF, a performance simulation platform that helps constellation manufacturers and content providers to estimate and understand the achievable performance under a variety of constellation options. The two key techniques behind the proposed platform are: (1) performance simulation for mega-constellation, which captures the impact of the inherent high mobility in satellite networks and profiles the area-to-area attainable network performance; and (2) constellation scaling, which synthesizes various topological options by scaling the architectural capability (e.g., number of satellite, link availability and capacity), and enables the exploration on multiple operating conditions that can not be easily reproduced.

To demonstrate the effectiveness of STAR PERF on understanding and optimizing satellite networks, we then leverage STAR PERF to evaluate and compare the performance of three state-of-the-art LEO constellations: Starlink, OneWeb and TeleSat, and perform what-if analysis, such as: what is the achievable latency between hosts located in London and NewYork respectively, if high-bitrate ISLs are fully deployed? The benchmark reveals a number of insights on using and optimizing existing mega-constellations like: emerging mega-constellations indeed offer low-latency opportunities for long-distance communications if ISLs are deployed, especially for communications between different continents. The constellation topology should be well designed to avoid high latency variation, as satellites move in high-speed and the path length in space is changing over time. The orbital decision and the scale of satellites can also significantly affect the resilience of the constellation.

Finally, to further show how applications or content providers can benefit from the proposed simulation platform, we illustrate STAR PERF's ability on assisting a future satellite-cloud integrated infrastructure that offers low-latency relay for delay-sensitive real-time communications (RTC). STAR PERF helps to intelligently choose the optimal relay located on either cloud platform or satellite to improve the quality of RTC. Evaluation results show that by properly selecting a relay in the satellite-cloud integrated infrastructure, end-to-end communication latency can be reduced by up to 62% for typical interactive traffic.

Conclusionally, this paper makes three key contributions:

• Presenting STAR PERF, a simulation platform for profiling and understanding the network performance of megaconstellations under a diversity of architectural options and network policies. (§III)

• Leveraging STAR PERF to benchmark three state-of-the-art mega-constellations and their possible topological extension, and highlighting insights on optimizing constellation designs to improve network performance. (§IV)

• Quantifying the potential benefits of a satellite-cloud integrated infrastructure, and proposing a low-latency relay selection algorithm that can effectively reduce the latency of interactive video applications. (§V)

The implementation of STAR PERF is mainly in Python and the code is open source 1 . To the best of our knowledge, STAR PERF is the first open-source simulator for characterizing the network performance of emerging mega-constellations under various constellation options and network policies. Moreover, today's mega-constellations like Starlink are still evolving rapidly. As our future work, we will keep upgrading STAR PERF to follow latest updates on Starlink and other emerging mega-constellations.

近年来, 构建"新太空 (NewSpace)"卫星网络 (SN) 正日益受到关注, 因为卫星网络有望提供低延迟, 高吞吐量的全球互联网连接. 我们见证了构建由大量低地球轨道 (LEO) 卫星组成的"巨型星座 (Mega-constellations)"的热潮, 诸如 OneWeb [7], 亚马逊 [1] 和 SpaceX [12] 等参与者纷纷进入这一市场. 作为自 2020 年 1 月以来全球最大的商业卫星星座运营商, SpaceX 正积极构建 Starlink 星座. 该星座由数千颗量产的小型 LEO 卫星组成, 已于 2020 年率先向加拿大和美国北部的客户提供服务, 并计划在 2021 年将服务范围扩展至全球其他地区 [11].

从根本上讲, 巨型星座通过扩展现有地面网络的连接能力 (即星地融合网络), 打破了当今互联网部署中固有的物理限制, 从而促进了互联网的发展. 首先, 卫星网络将宽带速度的互联网覆盖扩展到了那些接入不可靠, 昂贵或完全无法接入的偏远地区. 其次, 由数千颗在 LEO 运行的卫星组成的星座, 为构建空间网络以提供低延迟通信带来了新机遇. 现代 LEO 卫星可以配备光学星间链路 (ISLs) 进行星间通信. 在自由空间中, 数据包能够以真空中的光速传播, 这远快于在地面光纤中的传播速度. 因此, 通过避开长距离且蜿蜒曲折的光纤路径, 空间通信的延迟代价可能得以降低. 最后, 卫星网络也被期望能提高网络吞吐量, 随着星载技术的快速发展和发电功率的增加, 高比特率卫星 [39] 不断演进, 可提供数十甚至数百 Gbps 的数据速率.

虽然上述前景描绘了未来星地融合网络的繁荣景象, 但学术界对现代巨型星座的拓扑特性和可达到的网络性能的理解仍然非常有限. 定量地剖析巨型星座对于设计, 使用和优化卫星网络具有重要意义, 但也面临着一系列实际挑战: (i) 目前新兴的卫星星座 (如 Starlink) 仍处于紧锣密鼓的开发阶段, 且卫星部署成本高昂, 耗时漫长. 因此, 很难直接从已完全部署的星座系统中测量网络性能; (ii) 卫星网络与地面网络有着本质的区别. 新兴卫星网络由于星间链路范围有限, 仅相对邻近的卫星能相互连接, 且卫星相对于地面站及彼此之间都在高速运动 [19], 因而固有地表现出两个显著特征. 这些特殊属性使得现有的网络剖析方法难以准确表征巨型星座. 现有一些工作试图对新型星座的特性进行建模和分析, 但大多集中于不同物理层载荷决策下的系统容量建模 [21], [43], 而忽略了采用不同星座配置选项或网络策略对用户感知网络性能的影响.

本文介绍了 STARPERF 的设计与实现, 这是一个性能仿真平台, 旨在帮助星座制造商和内容提供商在多种星座配置选项下评估和理解可达到的性能. 该平台背后的两项关键技术是:

(1) 巨型星座性能仿真, 它捕捉了卫星网络固有 High Mobility 的影响, 并剖析了区域对区域 (Area-to-Area) 的可达网络性能

(2) 星座缩放, 它通过缩放架构能力 (如卫星数量, 链路可用性和容量) 来合成各种拓扑选项, 并支持对难以复现的多种运行条件进行探索

为了证明 STARPERF 在理解和优化卫星网络方面的有效性, 我们利用 STARPERF 评估并比较了三个最先进的 LEO 星座: Starlink, OneWeb 和 TeleSat 的性能, 并进行了假设分析 (What-if Analysis), 例如: 如果完全部署高比特率 ISL, 位于伦敦和纽约的主机之间可达到的延迟是多少? 此次基准测试揭示了关于使用和优化现有巨型星座的若干见解, 例如: 若部署 ISL, 新兴巨型星座确实为长距离通信 (特别是跨洲通信) 提供了低延迟的机遇; 星座拓扑应经过精心设计以避免高延迟抖动, 因为卫星处于高速运动中且空间路径长度随时间变化; 轨道决策和卫星规模也会显著影响星座的弹性

最后, 为了进一步展示应用或内容提供商如何从该平台获益, 我们阐述了 STARPERF 在辅助未来星云融合基础设施方面的能力, 该基础设施为延迟敏感的实时通信 (RTC) 提供低延迟中继. STARPERF 有助于智能选择位于云平台或卫星上的最佳中继节点以提升 RTC 质量. 评估结果表明, 通过在星云融合基础设施中合理选择中继, 典型交互式流量的端到端通信延迟最多可降低 62%

综上所述, 本文做出了三个主要贡献:

• 提出了 STARPERF, 一个用于在多种架构选项和网络策略下剖析和理解巨型星座网络性能的仿真平台 (§III)
  • "提出平台"
• 利用 STARPERF 对三个最先进的巨型星座及其可能的拓扑扩展进行了基准测试, 并强调了优化星座设计以提升网络性能的见解 (§IV)
  • "这个平台如何测试与查看性能"
• 量化了星云融合基础设施的潜在效益, 并提出了一种低延迟中继选择算法, 可有效降低交互式视频应用的延迟 (§V)
  • "这个平台如何帮助有关应用/理论落地"

STARPERF 主要使用 Python 实现, 且代码已开源. 据我们所知, STARPERF 是首个用于表征新兴巨型星座网络性能的开源模拟器, 支持各种星座选项和网络策略. 此外, 鉴于当今像 Starlink 这样的巨型星座仍在快速演进, 作为未来的工作, 我们将持续升级 STARPERF 以跟进 Starlink 及其他新兴巨型星座的最新动态.

Why Profiling Satellite Networks is Important?

A. Mega-constellations bring new opportunities for global low-latency and high-throughput communication

Quick primer for satellite networks: Typically, a satellite network (SN) built upon mega-constellations contains two primary components: (1) the space section which includes a large group of low-flying satellites running on low earth orbits (LEO), interconnected by inter-satellite links (ISLs); and (2) the terrestrial section that typically consists of a number of ground stations (GSes), which establish bidirectional satellite-to-ground links (SGLs) to connect the constellation in space.

New opportunities enabled by emerging constellations: The integration of satellite networks and the terrestrial Internet as a whole offers great new opportunities for improving the user-perceived network performance, which is not limited to broader Internet access. The rapid evolution of on-board technology has led to the development of the high-throughput satellites (HTS) [39], which promises to provide tens and even hundreds of Gbps transmission rate.

In addition to wider coverage and higher network capacity, mega-constellations also enables a promising potential for low-latency Internet communications. ISLs between satellites can use free-space lasers as the physical layer payload to communicate at the speed of light in a vacuum. Therefore, long-distance communications may attain lower latency via routing over LEO constellations [31]. Moreover, free-space mega-constellation breaks the geographical routing constraints that prolongs terrestrial paths. Figure 1 plots an example, showing the opportunity of leveraging inter-satellite links to reduce intercontinental communication latency. The traceroute result shows that current network deployments and routing policies forward data from Beijing to Sydney via Los Angeles by default. Such intercontinental detour incurs more hops and possibly larger delay than the shorter path built upon mega-constellations in space.

卫星网络背景简介: 通常, 基于巨型星座构建的卫星网络 (SN) 包含两个主要组成部分: (1) 空间段, 包含大量运行在低地球轨道 (LEO) 上的低轨卫星, 并通过星间链路 (ISLs) 相互连接; (2) 地面段, 通常由若干地面站 (GSes) 组成, 这些地面站建立双向星地链路 (SGLs) 以连接空间中的星座.

新兴星座带来的新机遇: 卫星网络与地面互联网的整体融合为提升用户感知的网络性能提供了巨大的新机遇, 这种提升不仅仅局限于更广泛的互联网接入. 星载技术的快速演进推动了高通量卫星 (HTS) [39] 的发展, 有望提供数十甚至数百 Gbps 的传输速率.

除了更广泛的覆盖范围和更高的网络容量外, 巨型星座还为低延迟互联网通信提供了广阔的潜力. 卫星间的 ISL 可以利用自由空间激光作为物理层载荷, 以真空中的光速进行通信. 因此, 通过 LEO 星座进行路由, 长距离通信有望实现更低的延迟 [31]. 此外, 自由空间巨型星座打破了延长地面路径的地理路由限制. 图 1 展示了一个示例, 说明了利用星间链路降低跨洲通信延迟的机遇:

Traceroute 结果显示, 当前的网络部署和路由策略默认将数据从北京经由洛杉矶转发至悉尼. 这种跨洲迂回相比于基于空间巨型星座构建的较短路径, 会产生更多的跳数, 并可能导致更大的延迟.

B. Understanding satellite networks is challenging

While SNs offer promising opportunities on improving network performance, it is very challenging to understand SNs' architecture, network performance and the impact of various design options (e.g., topological or routing potions).

First, SNs have the inherent "high mobility" property that differs from the terrestrial network, resulting in dynamic network topology and intermittent connectivity. In wired networks, network nodes such as routers and switches are typically static. Even in terrestrial Wi-Fi or cellular networks, mobile nodes (e.g., mobile phones or vehicles) are not moving so fast as satellites. In SNs, potential network nodes are moving in high speed with respect to the Earth and other satellites, and ISLs are limited by range. Only those relatively nearby satellites can be connected, and thus routes over SNs should be updated timely to adapt the dynamic connectivity.

Second, the real deployment for mega-constellations are significantly cost-intensive and time-consuming, and thus it is very difficult to directly measure the performance of a fully deployed constellation system. Finally, the design of SN consists of a large number of architectural and routing options. Such diversity on design options makes it meaningful but difficult to estimate and understand the impact of various options on the corresponding network performance.

尽管卫星网络在提升网络性能方面提供了广阔的机遇, 但要理解卫星网络的架构, 网络性能以及各种设计选项 (例如拓扑或路由选项) 的影响却极具挑战性.

首先, 卫星网络具有区别于地面网络的固有"高动态性"特征, 导致了动态的网络拓扑和间歇性的连接. 在有线网络中, 路由器和交换机等网络节点通常是静态的. 即使在地面 Wi-Fi 或蜂窝网络中, 移动节点 (如手机或车辆) 的移动速度也不如卫星快. 在卫星网络中, 潜在的网络节点相对于地球和其他卫星都在高速移动, 且 ISL 受距离限制. 只有相对邻近的卫星才能相互连接, 因此卫星网络上的路由需要及时更新以适应动态连接.

其次, 巨型星座的实际部署成本极高且耗时漫长, 因此很难直接测量一个完全部署的星座系统的性能. 最后, 卫星网络的设计包含大量的架构和路由选项. 这种设计选项的多样性使得评估和理解各种选项对相应网络性能的影响既有意义又充满困难.

C. Profiling mega-constellations is of significant importance

Summarily, when designing, operating and using emerging mega-constellations, it is often important and useful to profile the network performance of a constellation, specified by both the architectural options and network policies (e.g., routing scheme). Therefore, the goal of this paper is to design and implement such a simulation platform to model, analyze and understand the network performance and design trade-space of emerging mega-constellations. The usage of our platform includes: (i) guiding constellation operators to attain a function of the network capacity of the constellation topology. Such a function can help to understand and optimize the design of constellations; (ii) guiding content providers who want to deploy their contents upon LEO satellites to provide low-latency services globally. Next we present the details of such a platform, STAR PERF.

总之, 在设计, 运营和使用新兴巨型星座时, 剖析由架构选项和网络策略 (例如路由方案) 共同指定的星座网络性能, 往往是非常重要且有益的. 因此, 本文的目标是设计并实现这样一个仿真平台, 以建模, 分析和理解新兴巨型星座的网络性能及设计权衡空间.

我们平台的用途包括:

1. 指导星座运营商获取星座拓扑网络容量的函数关系, 这种关系有助于理解和优化星座的设计
2. 指导希望将其内容部署在 LEO 卫星上以提供全球低延迟服务的内容提供商

接下来, 我们将详细介绍这一平台 -- STARPERF

The StarPerf Platform

A. STAR PERF overview.

System overview. Figure 2 plots the overview of our STAR PERF platform. The STAR PERF platform takes network topology, network policy and traffic pattern as the platform input. The input of STAR PERF describes the composition and scale of the constellation, how satellites are connected to each other and ground stations, and how user requests are scheduled and forwarded over satellites. The main components inside STAR PERF is a suit of models which quantitatively describe a LEO satellite constellation together with its performance estimation. At runtime, STAR PERF loads the input and calculates the performance output based on the built-in models.

Collectively, STAR PERF includes two key techniques: (1) performance simulation for mega-constellation, which captures the impact of inherent high mobility in satellite networks and profiles the area-to-area attainable network performance; (2) resource scaling, which synthesizes various constellation topologies and network policies by scaling the space resource (e.g., number of satellite, link availability and capacity), and enables exploring multiple operating conditions that can not be easily reproduced.

Runtime workflow. To evaluate a satellite constellation by STAR PERF, first the user specifies the constellation options and network policy (e.g., traffic scheduling or routing strategy), together with a configured traffic pattern. The input is then loaded by the STAR PERF platform and is used to generate a simulated satellite network according to the built-in model inside the platform. The traffic pattern is then loaded and applied in the network. Performance metrics such as network latencies are measured and finally used to compute and quantify the network performance.

平台概览 (Overview)

输入与输出:

• 平台接收星座拓扑, 网络策略 (如路由) 和流量模式作为输入
• 输出为量化的网络性能指标

核心技术:

1. 性能仿真: 捕捉卫星网络的高动态性, 剖析区域对区域 (Area-to-Area) 的性能
2. 资源缩放: 通过缩放空间资源 (如卫星数量, 链路容量) 合成不同的拓扑和策略, 支持探索难以复现的运行条件

工作流程: 用户配置星座选项, 策略和流量 -> 平台生成仿真网络 -> 加载流量 -> 计算并输出性能指标

B. Characterizing network topology.

Our STAR PERF platform profiles the network topology of a LEO satellite constellation by modeling three primary aspects: (1) orbit property; (2) ground station distribution; and (3) link type and connectivity among satellites.

Orbit and constellation elements. The design option for constellation orbit can significantly affect the coverage and route stability in SNs. STAR PERF leverages five primary continuous parameters to describe the orbit design: (1) Inclination (Inc), which is the angle between an orbit and the Equator as the satellite travels northward. The value of Inclination for polar orbits is 90 ◦ ; (2) Altitude(Alt), which is measured over sea level and it determines the orbital velocity. Recent constellations typically consist of low-flying satellites with an altitude of 2,000 km or less; (3) Orbit phase shifts(OPS), which capture the relative placement of satellites in a constellation. The orbit phase offset between orbital planes indicates when satellites in consecutive orbits cross the equator; (4) Number of orbits(NoO); and (5) Number of satellites(NoS) in each orbit.

Link options. The link options include both the band allocation and payload type of inter-satellite and satellite-toground links. The link band is very critical to the network performance, as the link data rate largely depends on the band selected. For instance, data rate higher than 512Mbps is only doable if high bands (like Ka- or higher) are used. STAR PERF describes the link type between arbitrary two nodes in the SN as one option selected form S-band, X-band, Ku-band, Ka-band or optical. In addition, the payload type refers to the type of architecture implemented, which includes bent-pipe, circuit-switched or packet-switched.

Satellite connectivity pattern. Satellites connect to GSes and other satellites. The connectivities are mainly limited by the visibility and power supplement, as well as the ability to quickly establish links between fast-moving satellites via radio or laser alignment. The approach of establishing connectivities in satellite constellation also significantly affects the network performance, as it determines the basic network topology. In particular, existing non-GEO constellations like Iridium use a grid-like approach for their ISLs. Recently works have proposed a Grid+ [19] connectivity pattern, in which each satellite has four bi-directional ISLs with its nearby neighbors, two in the same orbit, and other two with immediate neighbors in the 2 adjacent orbits. STAR PERF supports grid and Grid+ connectivity pattern by default, and also allows customized connectivity design as the platform input.

Table I summarizes the design options for LEO mega-constellations supported by the STAR PERF platform.

网络拓扑建模 (Characterizing Network Topology)

平台从三个维度对 LEO 星座拓扑进行建模:

1. 轨道特性: 包括倾角 (Inc), 高度 (Alt), 轨道相位偏移 (OPS), 轨道数量 (NoO) 和每轨道卫星数 (NoS)
2. 地面站分布: 涉及地面站的位置和配置
3. 链路与连接:
   • 链路选项: 支持多种频段 (S/X/Ku/Ka/Laser) 和载荷类型 (弯管/电路交换/分组交换)
   • 连接模式: 默认支持 Grid (网格状) 和 Grid+ (每颗卫星连接前后同轨卫星及相邻轨道的卫星) 模式, 也支持自定义

参数设计如图所示:

C. Options for routing strategy.

The network performance over SNs is affected not only by the design of constellation, but also by the network policies running on SNs. In order to accommodate all expected customers, a LEO satellite network has to determine how to optimally (if possible) route and forward demand traffic. Typically, we define such routing and forwarding strategies as the routing policy, which will significantly affect the path performance including latency, throughput, reachability and resilience in SNs.

By default, STAR PERF models two kinds of routing strategies: (1) distributed routing strategies, such as OSPF and GPSR [36] which leverage the local information obtained on each node (i.e., satellite) to calculate the routing table (e.g., using djisktra); and (2) centralized routing strategies, like Dynamic Source Routing (DSR) [31], [35].

路由策略选项 (Options for Routing Strategy)

平台模拟了两类路由策略以评估其对性能的影响:

1. 分布式路由: 利用本地信息计算路由 (如 OSPF, GPSR)
2. 集中式路由: 如动态源路由 (DSR)

D. Characterizing network performance

STARPERF characterizes three key performance metrics of a satellite constellation system:

1. Coverage rate, which indicates the available range covered by satellites;
2. Latency, which is defined as the delay of sending a small packet from a source to its destination via satellites; and
3. Throughput, showing the ability of delivering content via the LEO constellation.

Since the ultimate goal of a mega LEO constellation is to provide better Internet accessibility and communication quality, these metrics can quantify the main network aspects concerned by both constellation designers and content providers who aim to deploy on-satellite contents.

Further, to model the network performance geographically, STARPERF builds a grid system upon the Earth surface for modeling and analyzing geographic information to measure the coverage, latency and throughput. The grid system used in STARPERF buckets user requests and satellites into hexagonal areas based on H3 [5]. STARPERF discretizes the Earth surface into hexagonal areas for several reasons:

(1) Satellites in LEO are often in high-speed motion, and hexagons minimize the quantization error introduced when satellites move in high speed.

(2) Hexagons have good scalability, since the size of a hexagon can be dynamically adjusted by setting its resolution. A higher resolution indicates a smaller hexagon.

(3) It is easy to use hexagonal areas to approximate radiuses, since they well fit the circle coverage of satellites.

Using the hexagonal hierarchical grid system, STARPERF groups a set of nearby locations into a hexagonal area, and maps a certain location (specified by its latitude and longitude) to a 64-bit area index.

We denote \(A_i^R\) as the \(i\)-th area in STARPERF, under a certain resolution \(R\), and the total number of hexagonal area is denoted as \(N_{\text{numarea}}^R\). Next we formulate the network performance upon the grid system.

Coverage

Given a LEO constellation containing a number of satellites in high speed motion, the coverage is time-varying and depends on the constellation topology. Let \(c_{it}\) denote a binary parameter and is set to true if area \(A_i^R\) is covered by the satellite constellation in slot \(t\). We then define the coverage of the constellation in slot \(t\) as the fraction of covered area:

\[ \frac{\sum_{i=1}^{N_{\text{numarea}}^R} c_{it}}{N_{\text{numarea}}^R}. \]

Thus, the coverage rate of constellation \(C\) over a period \(T\) can be formulated as \(CR_T^C\):

\[ CR_T^C = \left( \sum_{t=1}^{T} \sum_{i=1}^{N_{\text{numarea}}^R} c_{it} \big/ N_{\text{numarea}}^R \right) \big/ T \tag{1} \]

The above equation quantifies the fraction of covered area. However, many areas on the Earth are built on ocean and mountains with rare communication requirements. To model the ability of providing services for "necessary" areas, we use binary parameter \(h_{it}\) to indicate whether there is at least one communication request in area \(i\) in time slot \(t\). Then the hotspot coverage rate can be formulated as:

\[ \left( \sum_{t=1}^{T} \frac{\sum_{i=1}^{N_{\text{numarea}}^R} c_{it} \cdot h_{it}} {\sum_{i=1}^{N_{\text{numarea}}^R} h_{it}} \right) \big/ T \tag{2} \]

Equation (2) captures the coverage rate of hotspot areas, and a higher value of \(HCR_T^C\) indicates better satellite accessibility of the constellation during the period \(T\).

Area-to-area latency

STARPERF focuses on the attainable latency via routing over SNs, which is mainly constrained by the network topology and the speed of light. Therefore, in the design of STARPERF we refrain other impacts on the latency (e.g., packet processing time, encoding/decoding time) and assume the latency here is dominated by the propagation latency. Moreover, note that the prior work [38] has shown that Internet latencies to any particular data center are similar from users in the similar location. Users in the same area have nearby locations and similar distance from the user terminals to the connected LEO satellite. Hence we assume that the end-to-end delays via the LEO constellation are similar for the same area pair, and we do not add an excessive constraint on latencies of all paths of the same area pair. STARPERF then formulates the area-to-area latency from \(i\)-th area to \(j\)-th area in slot \(t\) as:

\[ L_{ij} = f_L(\textit{Tpl}, RS, t) \tag{3} \]

where \(\textit{Tpl}\) is the network topology of the LEO constellation and \(RS\) is the routing strategy used to route packets from the source area to the destination area. Once the area-to-area path is determined by the given routing strategy, the latency is estimated as the length of the path divided by \(c\), where \(c\) is the light speed in space. The value of area-to-area latency suggests the ability of providing low latency communication of a LEO constellation, which is useful for delay-sensitive interactive applications, as we will show in later sections.

Area-to-area throughput

Another important performance metric is the area-to-area throughput. The rapid evolution of on-board technology and an increase in power generation lead to the development of high-throughput satellites, which are able to provide tens and even hundreds of Gbps bandwidth. Therefore, the area-to-area throughput which indicates the achievable rate of successful data delivery between two areas over the SN, is critical for content providers who leverage satellites to distribute important content in real time. Particularly, we define the area-to-area throughput \(B_{ij}\) between area \(i\) and area \(j\) as the total throughput of all paths that have the similar latency with the shortest path. Therefore, \(B_{ij}\) quantifies the ability of a certain constellation topology to deliver contents from \(i\) to \(j\) by routing traffic without congestion and with low latency. The function of \(B_{ij}\) is denoted as:

\[ B_{ij} = f_B(\textit{Tpl}, RS, t, \beta) \tag{4} \]

where \(\textit{Tpl}\) and \(RS\) are the network topology and routing strategy respectively. \(\beta\) is a parameter to represent the latency requirement [30]. Calculating the area-to-area throughput provided by a certain constellation design follows the next stages. First, building a network with users in the source and destination areas, and all visible satellites. Second, running a routing algorithm to obtain the shortest path which consists of a set of sequential ISLs and satellite-to-ground links. Third, calculating the latency of the shortest path, denoted as \(D_{\min}\). Then identify all similar paths with the same source and destination that have latency \(< \beta \cdot D_{\min}\). Finally, constructing a sub-network that includes all these similar paths and compute the max-flow from the source to the destination.

Resilience

Resilience indicates the ability of mega-constellations to provide and maintain an acceptable level of service in the face of faults and challenges to normal operation. When using SNs to extend terrestrial networks and support Internet service, constellation operators have to consider the vulnerabilities and resilience of the proposed constellations. The constellation topology should be reasonably designed to provide resilient and affordable capabilities to preserve stable connectivity in space. STARPERF uses the betweenness centrality [2] from graph theory to quantify the resilience of a constellation. Betweenness is a metric that describes the centrality of a graph based on shortest paths, and it is widely used in telecommunications networks, e.g., a node with higher betweenness centrality would have more traffic passing through that node. Moreover, a node with a high betweenness centrality may also be a potential bottleneck node, since the failure of this node will affect all flows relying on it. Specifically, the betweenness of a satellite \(\textit{sat}\) which works as a node in the SN is calculated as:

\[ \sum_{s \ne d \ne \textit{sat}} \frac{p_{sd}(\textit{sat})}{p_{sd}} \tag{5} \]

where \(p_{sd}\) is the total number of the shortest paths from source \(s\) to destination \(d\) in the SN, and \(p_{sd}(\textit{sat})\) is the number of those paths that pass through \(\textit{sat}\).

网络性能表征 (Characterizing Network Performance)

平台基于 H3 六边形网格系统 对地球表面进行划分, 定义了四个关键性能指标:

• 覆盖率 (Coverage):
  • 普通覆盖率: 星座覆盖区域占总区域的比例
  • 热点覆盖率 (Hotspot Coverage): 针对有人口或通信需求的区域计算的覆盖率, 更能反映实际服务能力
• 端到端延迟 (Area-to-Area Latency): 主要考虑光速传播延迟, 忽略处理延迟, 用于评估低延迟通信能力
• 吞吐量 (Area-to-Area Throughput): 基于最大流计算, 评估在特定延迟约束下, 两区域间可达到的最大传输速率
• 弹性 (Resilience):
  • 使用图论中的 介数中心性 (Betweenness Centrality) 来量化
  • 介数高的节点可能是潜在瓶颈, 其故障对网络影响大

E. Characterizing user requests

Typically, in satellite network, user handsets can connect to satellites directly (e.g., like Iridium) or connect by very-small aperture terminal (VSAT), which is a two-way satellite ground station with a small dish antenna. Emerging constellation systems like Starlink claim that it will be linked to flat user terminals in the size of a pizza box which have phased array antennas and track the satellites. The terminals can be mounted anywhere, as long as they can observe the sky. Therefore, STARPERF assumes each end user can connect to satellites directly or connect via a pre-purchased VSAT, if the user is in the sight of view of the satellites.

User requests are formulated as a traffic matrix in the grid system. Let \(R_{ijt}\) denote a traffic task that requires to send \(R_{ijt}\) bytes data from \(i\)-th area to \(j\)-th area in slot \(t\). Thus \(R_{ijt}\) describes the traffic distribution between different areas. In particular, the traffic distribution can be estimated by the population of different cities, or generated according to dedicated applications.

通常, 在卫星网络中, 用户手持终端可以直接连接卫星 (例如铱星系统), 或者通过甚小口径终端 (VSAT) 进行连接, VSAT 是一种配备小型碟形天线的双向卫星地面站. 像 Starlink 这样的新兴星座系统宣称, 它将连接到披萨盒大小的扁平用户终端, 这些终端配备相控阵天线并能追踪卫星. 只要能通视天空, 这些终端可以安装在任何地方. 因此, STARPERF 假设只要用户处于卫星的视距范围内, 每个终端用户就可以直接连接卫星, 或者通过预先购买的 VSAT 进行连接.

用户请求在网格系统中被建模为流量矩阵. 设 Rijt​ 表示一个流量任务, 即在时隙 t 需从第 i 个区域发送 Rijt​ 字节的数据到第 j 个区域. 因此, Rijt​ 描述了不同区域间的流量分布. 具体而言, 流量分布可以通过不同城市的人口数量来估算, 或者根据特定的应用程序生成.

F. Constellation scaling.

Constructing and deploying mega-constellations is costintensive and time-consuming. Therefore, in addition to study on a certain constellation pattern, it is meaningful but difficult to explore the impact of various architectural design decisions on the achievable network performance. The STAR PERF platform has the capability of automatically scaling and enumerating all the possible design options outlined in previous sections. For example, the user can increase the number of satellites in the constellation, or enable/disable ISLs, by tuning the configuration files of STAR PERF. All possible topologies and routing policies can be automatically enumerated and evaluated using this platform, by iterating the Cartesian product of all options in the trade-space listed in Table I.

构建和部署巨型星座不仅成本高昂, 而且耗时漫长. 因此, 除了研究特定的星座模式外, 探索各种架构设计决策对可达到的网络性能的影响虽具有重要意义, 但也极具难度. STAR PERF 平台具备自动缩放和枚举前文所述所有可能设计选项的能力. 例如, 用户可以通过调整 STAR PERF 的配置文件来增加星座中的卫星数量, 或启用/禁用星间链路 (ISLs). 通过对表 I 所列权衡空间 (Trade-space) 中所有选项进行笛卡尔积迭代, 该平台可以自动枚举并评估所有可能的拓扑结构和路由策略.

G. Implementation of STAR PERF platform.

Figure 3 plots the key components in the implementation of STAR PERF. The STAR PERF platform loads input manifest which describes network topology, flow scheduling policy, and traffic pattern to generate a satellite network graph. The constellation simulation is partially implemented based on third-party orbit analysis tools (e.g., STK [8]), which help to simulate the movement of satellites over time. For each constellation, STAR PERF calculates the constellation decisions and orbit parameters. Once all the nodes of the network have been loaded and scaled to the desired size, user traffics are generated and applied to the network and finally STAR PERF calculates the corresponding network performance. As emerging constellations like Starlink are still under heavy deployment and it is difficult to collect the real number of its users, we follow the approach used in [21] to estimate the geo-distributed user requests, based on the real population in different areas. Specifically, the user requests are generated according to the Gridded Population of the World v4 dataset [4]. We assume that a satellite network operator will capture about 5% of the total Internet traffic of each area, and each user in an area has a 500Kbps data rate requirement.

图 3 展示了 STAR PERF 实现中的关键组件. STAR PERF 平台加载描述网络拓扑, 流调度策略和流量模式的输入清单 (Manifest), 以生成卫星网络图.

星座仿真部分基于第三方轨道分析工具 (如 STK [8]) 实现, 这有助于模拟卫星随时间的运动轨迹:

对于每个星座, STAR PERF 计算其星座配置决策和轨道参数. 一旦网络中的所有节点被加载并缩放至预定规模, 系统将生成用户流量并将其应用于网络, 最后由 STAR PERF 计算相应的网络性能.

鉴于像 Starlink 这样的新兴星座仍处于密集部署阶段, 难以收集其实际用户数量, 我们采用文献 [21] 中的方法, 基于不同地区的实际人口数据来估算地理分布的用户请求.

具体而言, 用户请求是依据"世界网格化人口 v4 (Gridded Population of the World v4)"数据集 [4] 生成的. 我们假设卫星网络运营商将承载每个区域总互联网流量的约 5%, 且区域内的每个用户具有 500Kbps 的数据速率需求.

Limitations and Future Work

Characterizing network performance of future hybrid SNs. Unlike previous SN simulators that simulate communications over GEO satellites (e.g., SNS3 [9]), the current implementation of our STAR PERF platform mainly focuses on characterizing the network performance of emerging LEO megaconstellations. However, constructing a hybrid constellation that integrates satellites working in various kinds of orbit (e.g., LEO, GEO and MEO) to collaboratively provide global network access, is another blooming picture in the evolution of SNs. We will extend STAR PERF to model and profile such kind of hybrid SNs in the future.

Improving the fidelity of STAR PERF. Like other recent works that study on the network performance of emerging constellations [19], [27], [31], [32], our performance results are obtained from the model-based estimation, in which satellite and orbital configurations are based on public data released by satellite operators or the astronomy community [10]. At the time of this submission (August, 2020), Starlink is still under heavy deployment and we have no public access to run Internet traffic over real Starlink constellation. Hence it is very difficult to compare the network performance obtained by STAR PERF with the corresponding value measured from real Starlink. However, we will track the evolution of Starlink and other similar megaconstellations. We will keep upgrading STAR PERF to follow the latest updates in Starlink and other constellations, combine STAR PERF with more fine-grained physical layer models and improve the fidelity by calibrating the performance results, if Starlink offers available public access in the future.

表征未来混合卫星网络的性能

与以往模拟地球静止轨道 (GEO) 卫星通信的卫星网络模拟器 (如 SNS3 [9]) 不同, STAR PERF 平台的当前实现主要侧重于表征新兴低地球轨道 (LEO) 巨型星座的网络性能. 然而, 构建集成多种轨道 (如 LEO, GEO 和 MEO) 卫星以协同提供全球网络接入的混合星座, 是卫星网络演进中另一幅蓬勃发展的蓝图. 未来, 我们将扩展 STAR PERF 以对该类混合卫星网络进行建模和剖析.

提升 STAR PERF 的仿真保真度

与近期其他研究新兴星座网络性能的工作 [19], [27], [31], [32] 类似, 我们的性能结果源于基于模型的估算, 其中的卫星和轨道配置依据的是卫星运营商或天文学界 [10] 发布的公开数据.

在本文提交之时 (2020 年 8 月), Starlink 仍处于密集部署阶段, 我们无法获得公开接入权限以便在真实的 Starlink 星座上运行互联网流量. 因此, 很难将 STAR PERF 获得的网络性能结果与真实 Starlink 系统的实测值进行比对.

尽管如此, 我们将持续追踪 Starlink 及其他类似巨型星座的演进过程. 我们将不断升级 STAR PERF 以跟进 Starlink 和其他星座的最新动态, 将 STAR PERF 与更细粒度的物理层模型相结合; 如果未来 Starlink 提供公开接入服务, 我们将通过校准性能结果来进一步提升仿真的保真度.

Related Works

We briefly discuss related works in this section.

Modeling and analyzing satellite networks. A body of previous literatures have studied on the modeling and analysis on satellite networks [20]–[22], [27], [29], [39], [40], [46]. Del Portilo et al. have studied on the architectural design [20], [22] and conducted technical comparison [21] for large LEO satellite constellations. Sanchez et al. conducted a stakeholder analysis to identify the main stakeholders of NASAs Space Communication and Navigation (SCaN) program systems, and explored the architectural trade-space of the system [43]. These existing works mainly focus on modeling and analyzing the physical layer performance under different physical payloads, while STAR PERF characterizes the achievable network performance, such as area-to-area latency and throughput, under various constellation options and routing schemes. Brian et al. proposed to leverage SDN applications to optimally and autonomously handle aerospace network operations, including steerable beam control and network routing updates [17]. Moreover, authors in [27] studied cost-performance tradeoffs in the design space for Internet routing, and proposed a CDN-inspired routing mechanism. The cost analysis in [27] complements our study, and in addition to the design option for routing, STAR PERF also explores the impact of various constellation options on the final network performance. In addition, the community also has many simulators for SNs. The European Space Agency (ESA) provided a list of open source software resources for developing space downstream applications [3]. Most of these open source projects are designed for positioning and navigation, or earth observation, while STAR PERF focuses on characterizing the network performance of emerging constellations. SNS3 [9] is a high-fidelity ns3-based simulator for satellite communications. However SNS3 is built on a static system configuration, with only one geostationary satellite and does not support LEO constellations in its current version.

Routing protocols in satellite networks. Existing studies working on routing in satellite networks typically fall into two key categories: (1) inter-domain satellite routing [24], [33], [37], [48], and (2) intra-domain satellite routing [18], [19], [23], [25]–[28], [31], [32], [41], [42]. Authors in [33], [48] have studied and analyzed the inter-domain routing instability that is caused by the high-speed movement of satellite. In addition, routing inside a satellite constellation is also a well studied problem [23], [25], [26], [28], [41]. More recently, as the topic of using large commercial constellations of LEO satellites has re-gained popularity, several works have revisited the topology design and routing in emerging mega-constellations [18], [19], [27], [31], [32]. Giuliari et al. [27] studied the cost-performance tradeoffs in the design of routing over satellite networks. These recently works are mainly built on a certain constellation pattern (e.g., Starlink). STAR PERF complements above researches as it provides an open platform to explore the performance benefit of various routing designs and topological decisions.

Relay selection for delay-sensitive applications. Optimizing the server selection to attain low-latency communication is a much studied topics in the terrestrial Internet [34], [44], [47]. While existing works focus on using cloud infrastructures to construct low-latency path, our work in this paper further explores the benefits of integrating on-satellite node as the relay options to reduce end-to-end latency.

Satellite mobility management. Satellite networks represent a new category of wide-area network where thousands of satellites move in high speed but connect to each other. The mobility of satellites is also a well studied problem [16], [45]. Tsunoda et al. proposed a handover-independent mobility management scheme specifically designed for IP/LEO satellite networks. The basic idea of the proposed approach is to make IP addresses independent of logical locations and associated to only geographical location information.

研究领域 (Category)	主要研究内容/焦点 (Focus/Description)	代表性工作/文献 (Representative Works)	局限性或与 STARPERF 的区别 (Limitations / Difference)
卫星网络建模与分析 (Modeling and Analyzing Satellite Networks)	- 架构设计与对比: 研究大型 LEO 星座的架构设计及技术对比. - 利益相关者分析: 如 NASA SCaN 系统的架构权衡空间分析. - 物理层性能: 主要关注不同载荷下的物理层性能. - SDN 应用: 利用 SDN 处理航空网络操作 (波束控制, 路由更新).	- Del Portillo et al. [20]-[22]: LEO 星座架构与技术对比. - Sanchez et al. [43]: 系统利益相关者与架构权衡分析. - Brian et al. [17]: SDN 在航空网络操作中的应用. - Giuliari et al. [27]: 互联网路由的设计空间与成本性能权衡.	- 现有工作局限: 大多关注物理层载荷性能, 而非网络层性能 (如端到端延迟, 吞吐量). - 仿真器局限: ESA [3] 的开源软件多用于定位导航; SNS3 [9] 基于 ns3 但仅支持静态 GEO 卫星, 不支持 LEO 星座. - 本文区别: STARPERF 专注于表征多种星座选项下的网络性能.
卫星网络路由协议 (Routing Protocols in Satellite Networks)	- 域间路由 (Inter-domain): 研究卫星高速运动导致的路由不稳定性. - 域内路由 (Intra-domain): 研究星座内部的路由问题. - 巨型星座路由: 针对 Starlink 等新兴星座的拓扑设计与路由重访.	- 域间路由: [24], [33], [37], [48] (关注不稳定性). - 域内路由: [18], [19], [23], [25]-[28] 等. - Giuliari et al. [27]: CDN 启发的路由机制及成本权衡.	- 现有工作局限: 通常基于特定的星座模式 (如仅针对 Starlink) 进行研究. - 本文区别: STARPERF 提供了一个开放平台, 支持探索各种路由设计和拓扑决策的性能效益, 互补了现有研究.
延迟敏感应用的中继选择 (Relay Selection for Delay-sensitive Applications)	- 服务器选择优化: 在地面互联网中优化服务器选择以实现低延迟通信.	- 地面互联网研究: [34], [44], [47] (基于云基础设施).	- 现有工作局限: 主要关注利用地面云基础设施构建低延迟路径. - 本文区别: 进一步探索了将星载节点 (On-satellite node) 集成作为中继选项, 以降低端到端延迟.
卫星移动性管理 (Satellite Mobility Management)	- 移动性处理: 解决成千上万颗高速移动卫星的互联问题. - 切换管理: 如与切换无关的 IP 移动性管理方案 (IP 地址与地理位置绑定而非逻辑位置).	- Tsunoda et al. [45]: 针对 IP/LEO 网络的独立于切换的移动性管理. - 其他研究: [16].	- 主要贡献: 针对卫星网络特有的高动态性 (High Mobility) 提出的具体管理方案.