Related Work

This work spans computer systems, space systems, space networks, and systems ML. Section 2 provides an overview of the emerging computational space systems domain, and Section 2.1 characterizes major challenges of this research area. Recent works in orbital edge computing provide context for this work. The space networking challenge of the bent pipe bottleneck is quantified in [7], which also introduces the concept of a computational satellite deploying machine inference to address this challenge. Orbital edge computing [8] aims to address the downlink bottleneck and introduces the computational bottleneck of the inelastic space edge. Challenges and opportunities in this domain are examined in [30].

The computational bottleneck poses a major space systems challenge. Although monolithic satellites cost hundreds of millions of US dollars each, this high cost does not provide high-performance onboard processors. These systems must operate for decades to justify the high cost, which means that computer hardware must be low-risk and highly-reliable. Often, these systems use decades-old, “flight heritage” CPUs. After two or more decades of operation, onboard processors could be nearly half a century behind the state-of-the-art.

For example, LEON processors — which receive significant support from the European Space Agency (ESA) — implement the 32-bit SPARC V8 instruction set architecture (ISA). A recent implementation achieved a clock frequency of 250 MHz [1]. The EO-1 makes use of a 12 MHz Mongoose-V central processing unit (CPU), which implements the 32-bit MIPS ISA, to demonstrate autonomous science [6]. The RAD5500 implements a 64-bit PowerPC ISA operating at 466 MHz [2]. The limited performance of these CPUs stems from larger technology nodes and functional unit duplication, which help to provide reliability in the space environment. Extremely high costs demand extremely low risk and long-duration missions at the expense of performance. Recent trends in space systems have started to consider COTS embedded computer systems [28, 29].

Several works consider models optimized for accuracy or speed in terrestrial applications [4, 21, 22, 34, 37, 41]. Works on embedded, terrestrial, wireless sensor systems study the tradeoff between computation and communication for energy-harvesting devices [10, 11, 16, 17, 35]. These works identify the relatively high energy cost of communication and quantify benefits to spending energy on computation instead. While these terrestrial systems can transmit data at any time within energy constraints, satellites can transmit data only while near a ground station. In this work, we focus on model specialization for geo-spatial contexts at the orbital edge to improve the data value density of a saturated downlink.

Space networking is a growing field of research [3, 19, 24]. Much work focuses on inter-satellite communication, which is a challenging engineering question encompassing control theory, orbital dynamics, robotics, and energy-performance tradeoffs. Less attention has been paid to the challenges of the downlink bottleneck and the saturated downlink, which we examine in this work. Alternate approaches to addressing the downlink bottleneck [39] are complementary to this work by enabling higher performance with even smaller satellite constellation populations.

这项工作横跨计算机系统、空间系统、空间网络和系统化机器学习（systems ML）等多个领域。第 2 节概述了新兴的计算型空间系统领域，第 2.1 节描述了该研究领域的主要挑战。近期在轨道边缘计算（orbital edge computing）方面的工作为本研究提供了背景。文献[7]量化了空间网络中“透明转发管道”（bent pipe）瓶颈的挑战，并引入了部署机器学习推理的计算型卫星概念以应对此挑战。轨道边缘计算[8]旨在解决下行链路瓶颈问题，并引出了缺乏弹性的空间边缘（inelastic space edge）所带来的计算瓶颈。文献[30]则探讨了该领域的挑战与机遇。

计算瓶颈是空间系统领域的一个重大挑战。尽管单体式卫星（monolithic satellites）每颗成本高达数亿美元，但如此高的成本并未带来高性能的星上处理器。这些系统必须运行数十年才能证明其高昂成本的合理性，这意味着其计算机硬件必须是低风险且高可靠的。通常，这些系统使用的是有数十年历史、“经过飞行验证”（flight heritage）的CPU。在运行二十多年后，这些星上处理器可能比当前最先进的技术落后近半个世纪。

例如，受到欧洲航天局（ESA）大力支持的 LEON 处理器，实现了 32 位的 SPARC V8 指令集架构（ISA）。其最近的一个实现版本时钟频率达到了 250 MHz[1]。EO-1 卫星则使用了一颗 12 MHz 的 Mongoose-V 中央处理器（CPU），该处理器实现了 32 位的 MIPS 指令集架构，用于展示自主科学决策能力[6]。RAD5500 处理器则实现了 64 位的 PowerPC 指令集架构，运行频率为 466 MHz[2]。这些 CPU 的性能之所以有限，是因为它们采用了较大的工艺节点和功能单元冗余设计，这有助于在空间环境中提供高可靠性。极高的成本要求极低的任务风险和超长的任务周期，而这正是以牺牲性能为代价的。近年来，空间系统的发展趋势已开始考虑使用商用现货（COTS）嵌入式计算系统[28, 29]。

一些研究工作考虑了在地面应用中为准确度或速度优化的模型[4, 21, 22, 34, 37, 41]。在嵌入式、地面无线传感器系统方面的研究，探讨了能量收集设备在计算与通信之间的权衡[10, 11, 16, 17, 35]。这些工作指出了通信相对较高的能量成本，并量化了将能量用于计算所带来的益处。然而，这些地面系统可以在能量允许的范围内随时传输数据，而卫星只有在靠近地面站时才能传输数据。在本文中，我们专注于在轨道边缘针对地理空间上下文进行模型特化，以提高饱和下行链路的数据价值密度。

空间网络是一个不断发展的研究领域[3, 19, 24]。许多工作聚焦于星间通信，这是一个涵盖了控制理论、轨道动力学、机器人学以及能量与性能权衡的挑战性工程问题。而对下行链路瓶颈和饱和下行链路的挑战则关注较少，这正是本研究探讨的重点。其他旨在解决下行链路瓶颈的方法[39]与本研究是互补的，它们通过更小的卫星星座规模来实现更高的性能。