In-orbit Computing: An Outlandish thought Experiment?

outlandish 异乎寻常的; 异想天开的

Space industry upstarts are deploying thousands of satellites to offer global Internet service. These plans promise large improvements in coverage and latency, and could fundamentally transform the Internet. But what if this transformation extends beyond network transit into a new type of computing service? What if each satellite, in addition to serving as a network router, also offers cloud-like compute, making the new constellations not just global Internet service providers, but at the same time, a new breed of cloud providers offering “compute where you need it”?

We examine, qualitatively and quantitatively, the opportunities and challenges of such in-orbit computing. Several applications could benefit from it, including content distribution and edge computing; multi-user gaming, co-immersion, and collaborative music; and processing space-native data. Adding computing hardware to a satellite does not seem prohibitive in terms of weight, volume, and space hardening, but the required power draw could be substantial. Another challenge stems from the dynamics of low Earth orbit: a specific satellite is only visible to a ground station for minutes at a time, thus requiring care in managing stateful applications.

Our exploration of these trade-offs suggests that this “outlandish” proposition should not be casually dismissed, and may merit deeper engagement from the research community.

航天产业的新兴企业正在部署数千颗卫星以提供全球互联网服务。这些计划有望在覆盖范围和延迟方面带来巨大改进，并可能从根本上改变互联网。但如果这种变革超越了网络传输，扩展为一种新型的计算服务呢？如果每颗卫星在作为网络路由器的同时，也提供类似云的计算能力，使得这些新星座不仅是全球互联网服务提供商，同时也是一种提供“随需随算”服务的新型云提供商，情况又会如何？

我们从定性和定量的角度，探讨了这种在轨计算的机遇与挑战。一些应用可能会从中受益，包括内容分发和边缘计算；多用户游戏、共同沉浸式体验和协同音乐创作；以及处理空间原生数据。在卫星上增加计算硬件，在重量、体积和空间加固方面似乎并非不可行，但所需的功耗可能相当可观。另一个挑战源于低地球轨道的动态特性：地面站仅能在数分钟内看到一颗特定的卫星，因此在管理有状态应用时需要格外小心。

我们对这些权衡的探索表明，这个“异想天开”的提议不应被轻易忽视，或许值得研究界进行更深入的探讨。

Introduction

“New Space” companies are gearing up to offer global broadband Internet using satellites. At least 3 of these proposals — Starlink [17], Kuiper [34], and Telesat [50] — envision constellations with more than 1,000 satellites each. Starlink’s plans are the most ambitious, with tens of thousands of satellites, and by far the most mature, with them having already deployed more than 400 satellites [15].

Unlike today’s satellite Internet services [30, 53], which operate in geostationary orbit (GEO) at 35,786 km above the Earth’s surface, the proposed constellations will operate in low Earth orbit (LEO), below 2,000 km. This will allow the new networks to offer low latency, with round-trip times between satellites and ground stations potentially in single-digit milliseconds. Further, with their scale, these networks can provide truly global Internet coverage.

This promise of a new breed of global, low-latency, high-bandwidth Internet service providers has generated tremendous interest in both the popular press [14, 25, 43, 52], and among networking researchers [8, 9, 23, 26, 27, 32].

“新航天”公司正积极准备使用卫星提供全球宽带互联网。在这些计划中，至少有三个——Starlink [17]、Kuiper [34] 和 Telesat [50]——都构想了拥有超过1,000颗卫星的星座。其中，Starlink的计划最为宏大，计划部署数万颗卫星，并且目前最为成熟，已经部署了超过400颗卫星 [15]。

与当今在距离地表 35,786 km的地球静止轨道（GEO）运行的卫星互联网服务[30, 53]不同，这些规划中的星座将在低于 2,000 km的低地球轨道（LEO）运行。这将使新网络能够提供低延迟服务，卫星与地面站之间的往返时间有望达到个位数毫秒级别。此外，凭借其庞大规模，这些网络可以提供真正全球性的互联网覆盖。

这种新一代全球性、低延迟、高带宽互联网服务提供商的前景，已在主流媒体[14, 25, 43, 52]和网络研究人员[8, 9, 23, 26, 27, 32]中引起了巨大兴趣。

However, we posit that there is a potentially overlooked opportunity here beyond just network service: What if mega-constellations also offer in-orbit compute as a service, much like cloud computing from today’s terrestrial data centers? What if we have thousands of networked satellite-servers, with some of them accessible from anywhere on Earth within milliseconds at almost all times? We explore the ramifications of this thought experiment.

Given that such in-orbit compute would be much more limited and pricier than today’s cloud compute, it should be obvious that this is a terrible idea for most cloud applications. However, we find that several new and interesting use cases might be well-served by such an unusual infrastructure.

然而，我们认为这里存在一个可能被忽视的潜在机遇，它超越了单纯的网络服务：如果巨型星座也提供在轨计算即服务（in-orbit compute as a service），就像今天地面数据中心提供的云计算一样，会怎样？如果我们拥有数千个联网的卫星服务器，并且在几乎任何时候，地球上任何地方都能在几毫秒内访问其中一些服务器，又会带来怎样的影响？我们在此探讨这一思想实验的深远意义。

鉴于在轨计算会比当今的云计算更为有限且昂贵，显而易见，对于大多数云应用而言，这将是一个糟糕的主意。然而，我们发现一些新颖有趣的应用场景可能非常适合这种非同寻常的基础设施。

In-orbit compute would extend the cloud’s promise of computing when you want, to computing wherever you want. Current cloud data center maps are relatively sparse, with hardly any sites in many geographies, such as South America, Africa, and large parts of Asia. Even CDN edge locations, that offer more limited services, incur 100+ ms latencies in many places [18, 19, 24, 45]. In contrast, a large LEO constellation can be within a few milliseconds from everywhere on Earth, including locations unsuitable for terrestrial facilities, e.g., due to poor power and support infrastructure, or prohibitive political and legal concerns. In-orbit compute can thus offer ubiquitous “edge computing” capabilities, without the many hurdles in deploying terrestrial infrastructure in many locations. There is clear demand for such anywhere-compute: Amazon’s recently announced Snowcone [4] is targeted explicitly at edge computing in environments without accessible compute. Snowcone is a ruggedized small form-factor server, that provides cloud synchronization by shipping it back and forth. In-orbit compute would alleviate the long delays for such data movement, especially from regions with poor transport connectivity.

The flip-side of LEO constellation ‘omnipresence’ is that a large amount of LEO network infrastructure is largely idle at any given time: with most of Earth’s surface being very sparsely populated, many satellites are ‘useless’ at any time, and will not even be transiting data most of the time. Making this infrastructure more continually useful could thus be attractive. One way of doing this is to use it for computing, particularly in cases where the data computed upon is space-native, like for remote sensing and aerial imagery. For such applications, the amount of actually interesting or actionable data is often a minute fraction of the data gathered, but the volume of data generated can overwhelm the down-links from satellites [22, 39]. In-orbit processing can help save the limited bandwidth between satellites and ground stations. The ample satellite-satellite bandwidth can also allow collective data processing across satellites.

Perhaps the most interesting use cases arise for multi-party interactive applications, like games, music collaboration, virtual reality immersion in the same environment with friends, etc. Such applications require a “meetup server”, which provides low latency to all involved clients; in some applications, such as gaming, it may also be necessary to have this latency be as uniform across clients as possible. If in-orbit compute is available, a meetup server can be picked specifically for the set of users involved, guaranteeing the (nearly) lowest-possible latency for them, with a more consistent latency experience across users, compared to terrestrial servers.

在轨计算将云服务“按需计算”的承诺，扩展为“随处计算”。当前云数据中心的分布相对稀疏，在南美、非洲以及亚洲大部分地区等许多区域几乎没有站点。即使是提供更有限服务的CDN边缘节点，在许多地方也会产生超过 100 ms 的延迟[18, 19, 24, 45]。 相比之下，一个大型LEO星座可以与地球上任何地方都只相距几毫秒，包括那些因电力和支持基础设施薄弱，或因政治和法律问题而无法部署地面设施的地点。因此，在轨计算可以提供无处不在的“边缘计算”能力，而无需克服在许多地区部署地面基础设施的重重障碍。 市场对此类“随处计算”有明确需求：亚马逊最近发布的Snowcone[4]就明确针对那些缺乏计算设施环境下的边缘计算。Snowcone是一款坚固耐用的小型服务器，通过物理寄送往返来实现云同步。在轨计算将能缓解此类数据传输的长时间延迟，特别是对于那些运输连接不便的地区。

LEO星座“无处不在”的另一面是， 大量LEO网络基础设施在任何特定时间都处于基本闲置状态：由于地球大部分表面人口稀少，许多卫星在任何时刻都是“无用的”，大部分时间甚至不传输数据。因此，让这些基础设施能够持续发挥作用可能具有吸引力。 一种方式是将其用于计算，特别是当被计算的数据是空间原生数据时，例如遥感和航空影像。对于这类应用，真正有价值或可操作的数据量往往只占采集数据总量的极小部分，但生成的数据量可能会压垮卫星的下行链路[22, 39]。在轨处理有助于节省卫星与地面站之间有限的带宽。充足的星间带宽也支持跨卫星进行协同数据处理。

或许最有趣的应用场景出现在多方交互式应用中，如游戏、音乐协作、与朋友在同一环境中进行虚拟现实沉浸等。这类应用需要一个能为所有参与客户端提供低延迟的“汇合点”服务器；在某些应用（如游戏）中，可能还需要所有客户端的延迟尽可能一致。 如果存在在轨计算，便可以为特定的用户群体专门选择一个“汇合点”服务器，保证他们获得（近乎）最低的延迟，并且与地面服务器相比，用户之间的延迟体验更加一致。

We flesh out these possibilities in greater detail, quantifying some of the relevant aspects, where possible. We also ruminate on the challenges and downsides of in-orbit compute, including limits on the types of applications that may benefit, and the weight, power budget, life-cycle, cost, and space-readiness of the computing hardware used. While these are substantive concerns, thus far, surprisingly, none of these have seemed entirely prohibitive.

我们将尽可能地量化相关方面，更详细地阐述这些可能性。我们也会深入探讨在轨计算的挑战与弊端，包括可能受益的应用类型限制，以及计算硬件的重量、功率预算、生命周期、成本和空间适用性。尽管这些都是实质性的顾虑，但到目前为止，它们似乎没有一个是完全无法克服的。

A key technical challenge arises from the high orbital velocity of satellites: unlike GEO satellites, which appear stationary from Earth, any ground station sees a particular LEO satellite for only a few minutes, followed by a hand-off to another satellite. For stateful applications, e.g., a multiplayer game, having such ephemeral satellite-servers is challenging. Presenting applications the needed persistence requires planning ahead to pick a suitable satelliteserver, and timely state migration to a suitable successor. If this can be done, large LEO constellations can offer a powerful abstraction: GEO-like stationarity, simultaneously above all terrestrial locations, with ∼65× lower latency than GEO orbits. We thus explore the constraints on such state migration, and how suitable servers and successors may be picked to relax these constraints as much as possible.

We fully acknowledge how unusual this proposal is: although there have been computers in space for many decades now, the idea of in-orbit compute as a service is, very literally, outlandish. Nevertheless, we found this to be an interesting thought experiment with non-trivial use cases and challenges, and hope that the community’s engagement with it will either raise killer objections, or a different and interesting set of applications that we did not foresee.

一个关键的技术挑战源于卫星的高轨道速度：与从地球上看似乎静止的GEO卫星不同，任何地面站观测一颗特定的LEO卫星仅有几分钟时间，随后便需要切换到另一颗卫星。对于有状态的应用（例如多玩家游戏），这种短暂存在的卫星服务器构成了挑战。要为应用提供所需的持久性，就需要提前规划，选择合适的卫星服务器，并及时将状态迁移到合适的后继者。如果能做到这一点，大型LEO星座可以提供一个强大的抽象：类似GEO的静止特性，即同时位于所有地面位置的上空，而其延迟比GEO轨道低约65倍。因此，我们探讨了这种状态迁移所面临的约束，以及如何选择合适的服务器和后继者来尽可能地放宽这些约束。

我们完全承认这个提议非同寻常：尽管几十年来太空中一直有计算机，但在轨计算即服务的想法，毫不夸张地说，是异想天开的。尽管如此，我们发现这是一个有趣的思想实验，其中包含了重要的应用案例和挑战。我们希望，学术界的参与能对此提出决定性的反对意见，或者激发出我们未曾预见到的、另一组新颖有趣的应用。

低地球轨道（LEO）卫星在距离地表低于 2,000 km的高度运行。根据轨道力学原理，卫星的速度和轨道周期由其高度决定。例如，在 550 km的高度上（SpaceX迄今部署的星链卫星所采用的高度），卫星以 27,306 km/h的速度飞行，每 95 分 39 秒完成一次轨道运行。

与地球静止（GEO）卫星相比，LEO卫星提供一种截然不同的服务类型。GEO卫星相对于地球是静止的，而实现这种静止性要求其在 35,786 km的高度运行。通过在比GEO卫星更低的高度飞行，LEO卫星能提供低得多的传播延迟——在 550 km的例子中，延迟降低了65倍——但代价是失去了静止性。另一个区别是，GEO卫星较高的轨道使其拥有更大的覆盖锥域，仅需少量几颗卫星便足以覆盖全球。相比之下，LEO卫星的覆盖区域必然小得多，实现全球覆盖需要使用数量远超于此的卫星。LEO卫星的覆盖和速度特性意味着一个地面站只能在几分钟内观测到一颗特定的LEO卫星。此后，如果需要持续连接，地面站必须执行一次连接切换，转接到另一颗进入可达范围的LEO卫星。

为了提供这种持续的低延迟连接，数家公司已提出建设由数百到数万颗卫星组成的大型LEO星座。特别是，SpaceX的星链（Starlink）、亚马逊的柯伊伯（Kuiper）和Telesat公司，都在规划超过一千颗卫星的星座。SpaceX的计划最为宏大，计划部署 42,000 颗卫星。SpaceX现已发射超过400颗卫星，使星链成为有史以来规模最大的卫星星座。

除了大量卫星连接到地面站之外，大多数规划中的网络还具备 星间链路（Inter-Satellite Links, ISLs）功能，使得两个遥远地面站之间的连接可以经由一条上行链路、一系列星间链路和一条下行链路进行传输。上行和下行链路计划采用无线电技术，其带宽较为有限，约为 10 Gbps量级，而星间链路则可能实现更高的带宽[2, 36]。

这种设计方法使LEO巨型星座能够提供低延迟的宽带互联网连接。通过合理设计的卫星轨道，LEO星座可以提供真正的全球覆盖：在地球上任何地点、任何时间，都有一颗或多颗卫星处于可达范围之内。