Conclusion and Future Work

The increasing accessibility of space opens the orbital edge to new geospatial analysis applications. A limited downlink creates a need for on-orbit processing to extract value from sensor data. However, constraints on orbital edge computing limit the value of satellitebased applications. Kodan mitigates the downlink bottleneck and the computational bottleneck for space edge systems by leveraging geospatial contexts and specializing satellite computation to balance application processing time with precision. This approach contrasts with expensive, constellation-oriented techniques that extract value from data but require many satellites to do so. We implement and evaluate Kodan, which increases the density of valuable data downlinked from LEO between 89 and 97 percent without changing ground infrastructure or radio attributes despite bottlenecked bandwidth and computing.

空间可及性的不断提高为轨道边缘（orbital edge）带来了新的地理空间分析应用机遇。有限的下行链路催生了在轨处理的需求，以便从传感器数据中提取价值。然而，轨道边缘计算的限制反过来也制约了星上应用的价值。 Kodan 通过利用地理空间上下文并特化卫星计算，以平衡应用处理时间与精度，从而缓解了空间边缘系统的下行链路瓶颈和计算瓶颈。 这种方法与那些昂贵的、面向星座的技术形成对比，后者虽然也能从数据中提取价值，但需要部署大量卫星。我们实现并评估了 Kodan 系统，结果表明，在带宽和计算资源受限的情况下，Kodan 无需改变地面基础设施或无线电属性，便能将从低地球轨道（LEO）下传的高价值数据密度提升89%至97%

Orbital edge computing has an exciting future with many open research questions. Building on this work, computational space system designers should improve sensor coverage and computing capability of constellations through co-design of system-level optimizations and computer architecture while avoiding unfavorable cost-scaling of high device counts. Future constellations will feature heterogeneous sensors, computational capabilities, and actuators. Some satellites may share data via crosslinks and distribute processing so that each satellite need not contain telescope optics, precision pointing, laser communication, and a high-end GPU. Instead, a heterogeneous constellation supports hardware specialization (as opposed to processing specialization based on sample context), allowing individual satellites to contain fewer subsystems are therefore be more simple. Energy constraints, orbital dynamics, and client mission goals place time-dependent constraints on satellite operations. Communication and computation will remain perennial challenges for space-based computer systems. Kodan demonstrates that these challenges are surmountable with new techniques tailored to the unique constraints of the orbital edge.

轨道边缘计算的未来前景广阔，存在许多开放的研究问题。基于本项工作，未来的计算型空间系统设计者应通过系统级优化与计算机架构的协同设计（co-design），来提升星座的传感器覆盖范围和计算能力，同时避免因设备数量过多而导致的不利成本扩展。未来的星座将具有异构的传感器、计算能力和执行器。一些卫星可以通过星间链路（crosslinks）共享数据并进行分布式处理，这样一来，每颗卫星就不再需要同时配备望远镜光学系统、精密的指向系统、激光通信和高端 GPU。取而代之的是，一个异构星座可以支持硬件特化（相对于本研究中基于样本上下文的处理特化），从而允许单颗卫星包含更少的子系统，结构也因此变得更简单。能源限制、轨道动力学和客户任务目标对卫星操作施加了随时间变化的约束。通信和计算将继续是天基计算机系统长期存在的挑战。Kodan 的研究表明，通过专为轨道边缘独特约束条件定制的新技术，这些挑战是可以克服的。