Conclusion & Future Work

In this work, we develop orbital edge computing: edge computing on orbit using processing resources colocated with sensors inside small, low-cost satellites. The low cost of a nanosatellite compared to monolithic satellite systems makes large satellite constellations feasible for the first time. Applications of this emerging technology are impeded by existing, bent-pipe architectures. Orbital edge computing provides responsiveness, reliability, and scalability benefits. Future work should study energy collection and storage for orbital edge computing and radiation-hardened, machine-learning (ML) accelerators.

For example, we observe that incomplete ground track coverage stems from the energy-constrained, intermittent nature of these nanosatellites. Once a constellation, in aggregate, collects sufficient energy for full coverage, high processing latency can still limit coverage. Increasing energy availability with deployable solar panels is unsatisfactory, because this solution raises mission cost and mission risk. As an alternative solution, future work could investigate energyefficient, domain-specific accelerators (DSAs) for orbital edge computing workloads. Future work could evaluate the architectural vulnerability factors (AVFs) of recently-proposed ML accelerators and propose new ML accelerators that operate intermittently in the space environment.

在本作中，我们发展了轨道边缘计算：一种在小型、低成本卫星内部，利用与传感器并置的处理资源在轨进行边缘计算的技术。 与单体式卫星系统相比，纳米卫星的低成本首次使得大型卫星星座变得可行。 然而，这一新兴技术的应用受到了现有“弯管”架构的阻碍。轨道边缘计算在响应性、可靠性和可扩展性方面提供了优势。未来的工作应着重研究轨道边缘计算的能量采集与存储技术，以及抗辐射的机器学习（ML）加速器。

例如，我们观察到，不完整的星下点轨迹覆盖源于这些纳米卫星能量受限的间歇式特性。即使一个星座在总体上收集了足以实现完全覆盖的能量，高处理延迟仍然会限制覆盖率。使用可展开式太阳能帆板来增加能量可用性并非一个令人满意的方案，因为这种方案会提高任务成本和任务风险。作为一种替代方案，未来的工作可以研究适用于轨道边缘计算工作负载的、高能效的领域专用加速器（DSA）。未来的工作还可以评估近期提出的ML加速器的架构脆弱性因子（AVF），并提出能在空间环境中进行间歇式运行的新型ML加速器。

While this work has focused on nanosatellite constellations that share a single orbit and a single workload, future work may investigate heterogeneous systems and heterogeneous workloads. For example, a constellation operator may wish to serve many different clients over time. Clients will be interested in different features at different scales. As a result, different orbit altitudes and different hardware will be better suited to different clients. Supporting a dynamic set of workloads from a dynamic set of clients poses an interesting challenge, especially with regards to constellation reconfiguration. The high overhead of uplinking new ML models could be offset with federated learning techniques.

Looking forward, we expect deployments of satellites that are even smaller than nanosatellites. Chip-scale or gramscale satellites (“chipsats”) can be deployed more numerously and at even lower cost. Such devices are even more attritable than nanosatellites, but the smaller size places even tighter constraints on capability. Successful operation of these emerging devices will require the application of orbital edge computing techniques.

虽然本作专注于共享单一轨道和单一工作负载的纳米卫星星座，但 未来的工作可以研究异构系统和异构工作负载 。例如，一个星座运营商可能希望随时间推移为许多不同的客户服务。这些客户会对不同尺度下的不同特征感兴趣。因此，不同的轨道高度和不同的硬件将更适合不同的客户。支持一组动态变化的客户所提出的一组动态变化的工作负载，构成了一个有趣的挑战，尤其是 在星座重构方面。上传新ML模型的高昂开销可以通过联邦学习技术来抵消 。

展望未来，我们预计将会出现比纳米卫星更小的卫星部署。芯片级或克级卫星（“chipsats”）可以以更低的成本进行更大规模的部署。这类设备比纳米卫星更具可消耗性，但更小的尺寸也对其能力施加了更严格的约束。这些新兴设备的成功运行将需要轨道边缘计算技术的应用。