Kodan: Addressing the Computational Bottleneck in Space¶
Decreasing costs of deploying space vehicles to low-Earth orbit have fostered an emergence of large constellations of satellites. However, high satellite velocities, large image data quantities, and brief ground station contacts create a data downlink challenge. Orbital edge computing (OEC), which filters data at the space edge, addresses this downlink bottleneck but shifts the challenge to the inelastic computational capabilities onboard satellites. In this work, we present Kodan: an OEC system that maximizes the utility of saturated satellite downlinks while mitigating the computational bottleneck. Kodan consists of two phases. A one-time transformation step uses a reference implementation of a satellite data analysis application, along with a representative dataset, to produce specialized ML models targeted for deployment to the space edge. After deployment to a target satellite, a runtime system dynamically selects the best specialized models for each data sample to maximize valuable data downlinked within the constraints of the computational bottleneck. By intelligently filtering low-value data and prioritizing high-value data for transmit via the saturated downlink, Kodan increases the data value density between 89 and 97 percent.
将航天器部署至低地球轨道的成本不断降低,催生了大型卫星星座的出现。然而,卫星的高速度、海量的图像数据以及与地面站的短暂接触时间,共同构成了数据下行链路的严峻挑战。轨道边缘计算(Orbital Edge Computing, OEC)通过在空间边缘过滤数据来应对此下行链路瓶颈,但这也将挑战转移至卫星自身固有的、缺乏弹性的计算能力上。
在本文中,我们提出了Kodan:一个旨在最大化饱和卫星下行链路效用,同时缓解计算瓶颈的轨道边缘计算系统。Kodan包含两个阶段:
- 首先,一个一次性的转换步骤会利用卫星数据分析应用的参考实现和一个代表性数据集,来生成一系列专门用于空间边缘部署的机器学习模型
- 在部署到目标卫星后,一个运行时系统会为每个数据样本动态选择最优的专用模型,以便在计算能力受限的条件下,最大化下行传输的高价值数据
通过智能过滤低价值数据,并优先通过饱和的下行链路传输高价值数据,Kodan将数据价值密度提升了89%至97%
Note
一眼看上去,系统的构型特别特别常见:
- 先给个相对合理的“初始态”
- 后续每一步,在前面的基础上,动态变化 --> Optimize
Introduction¶
The proliferation of commercial space launch services [15] and nanosatellites [32, 36] over the past two decades makes low-Earth orbit (LEO) accessible to deployments of state-of-the-art, sensorequipped computer systems inside satellites. These satellites enable new and valuable geospatial sensing and computing applications, including disaster relief [18], agriculture [40], and infrastructure monitoring. The Earth-observation market remains dominated by large-scale, monolithic satellites costing hundreds of millions of US dollars each. However, the ascendance of inexpensive nanosatellites has led to large, commercial constellations of nanosatellites in LEO [5, 26]. Lower device cost and higher launch cadence decreases risk and enables new satellite applications [30].
在过去二十年中,商业航天发射服务 [15] 和纳米卫星 [32, 36] 的普及,使得在卫星内部署搭载了先进传感器的计算系统到近地轨道 (LEO) 成为可能。这些卫星催生了全新的、极具价值的地理空间感知与计算应用,涵盖了灾害救援 [18]、农业 [40] 及基础设施监测等领域。目前,对地观测市场仍由每颗耗资数亿美元的大型、单体卫星主导。然而,廉价纳米卫星的兴起已促使大型商业纳米卫星星座在近地轨道上涌现 [5, 26]。更低的设备成本和更高的发射频率降低了风险,并为新的卫星应用创造了条件 [30]。
Communication and computation ability constrain satellite utility. Today, most satellites operate as “bent pipes” [25] and are tasked manually to collect and downlink raw sensor data. These satellites face a downlink bottleneck stemming from a lack of communication opportunities. Recent work on “orbital edge computing” (OEC) processes data on satellites before downlinking [7, 8]. OEC mitigates the downlink bottleneck by identifying signals of interest and downlinking those signals only. While OEC addresses the downlink bottleneck, we observe that edge processing creates a computational bottleneck limiting the value of each satellite.
通信与计算能力限制了卫星的功用。 如今,大多数卫星作为“弯管”(bent pipes) [25] 运行,通过人工指令来收集并下传原始传感器数据。这些卫星因通信机会的缺乏而面临下行链路瓶颈。
近期的“在轨边缘计算”(Orbital Edge Computing, OEC) [7, 8] 研究主张在数据下传前于卫星上进行处理。OEC 通过识别感兴趣的信号且仅下传这些信号,从而缓解下行链路瓶颈。
然而,我们观察到, 尽管 OEC 解决了下行链路瓶颈,边缘处理本身却创造了一个计算瓶颈,从而限制了单颗卫星的价值。
Processing satellite sensor data at the edge is challenging. Satellite sensor samples (e.g., images) are large and arrive at a high rate compared to the rate at which embedded satellite hardware can process them. Geospatial image frames can contain hundreds of square kilometers and hundreds of millions of pixels. Depending on orbit altitude and camera characteristics, a LEO, Earth-observation satellite observes an entirely new frame every 1−30 seconds (the “frame deadline”). Applications typically tile these large frames and process each tile on the ground using, e.g., machine learning algorithms. However, not all samples are equally valuable; some observations are of high-value to an application and others are of low-value. A system faced with a saturated link should prioritize transmission of high-value data, but today’s bent pipes send data indiscriminately. Processing samples at the space edge distinguishes these categories. The computational bottleneck stems from an inability to process all tiles within the frame deadline. Prior OEC work manages the computational bottleneck by statically distributing tile processing across a constellation, using satellite-parallelism to meet the frame deadline [8]. Such a scheme distributes work across hundreds of satellites while addressing the computational bottleneck for just a single application — a relatively high-cost solution resulting in a vertically-integrated constellation aimed at a particular purpose.
在边缘处理卫星传感器数据具有挑战性。 与星上嵌入式硬件的处理速率相比,卫星传感器样本(例如,图像)尺寸大且采集速率高。地理空间图像帧可包含数百平方公里的范围和数亿像素。根据轨道高度和相机特性,一颗位于近地轨道的对地观测卫星每 1−30 秒便会观测一帧全新的图像(即“帧时限”)。应用通常在地面上将这些大尺寸图像帧切分为图块,并使用机器学习算法等方法逐一处理。
然而,并非所有样本都具有同等价值;某些观测对应用而言是高价值的,而另一些则是低价值的。一个面临链路饱和的系统应优先传输高价值数据,但如今的“弯管”模式却不加选择地发送所有数据。在太空边缘处理样本可以区分这些价值类别。计算瓶颈源于无法在帧时限内处理所有图块。
以往的 OEC 工作通过将图块处理任务静态地分布于整个星座来管理计算瓶颈,利用卫星并行处理来满足帧时限 [8]。这种方案将工作分散到数百颗卫星上,仅为解决单个应用的计算瓶颈——这是一种成本相对高昂的解决方案,其结果是形成一个针对特定目的的垂直整合星座。
Kodan is an orbital edge computing system that maximizes data value from satellites limited by communication and computation. Under a saturated satellite downlink, Kodan mitigates the computational bottleneck without the high cost of hundreds of satellite-parallel processors. Kodan uses a combination of techniques that modify applications based on unique, orbital data characteristics by trading between geospatial analysis precision and processing speed. After deployment to a satellite, the Kodan runtime system dynamically selects appropriate optimizations for each observation. Kodan decides how to process a sample based on its geospatial context. A geospatial context is a property of a data sample indicating its likelihood to contain certain features, e.g., the presence of ocean, forest, tundra, clouds, or high-value data. Highprecision value labels are computationally easier in some contexts and harder in others.
Kodan 是一个在轨边缘计算系统,旨在从受通信和计算双重限制的卫星中实现数据价值最大化。 在卫星下行链路饱和的情况下,Kodan 缓解了计算瓶颈,且无需承担数百个卫星并行处理器带来的高昂成本。
Kodan 采用一系列技术组合,通过在地理空间分析精度与处理速度之间进行权衡,基于独特的在轨数据特性来修改应用程序。
部署到卫星后,Kodan 运行时系统 (runtime system) 会为每次观测动态选择合适的优化策略。Kodan 根据数据样本的地理空间上下文来决定如何处理它。 地理空间上下文是数据样本的一种属性,用以表征其包含特定特征的可能性,例如,是否存在海洋、森林、苔原、云层或高价值数据。在某些上下文中,获取高精度的价值标签在计算上较为容易,而在其他上下文中则更为困难。
Kodan balances precision and execution time to maximize data value density. Software running on each satellite prioritizes decreased compute time when computationally bottlenecked and prioritizes precision otherwise. When computationally limited, Kodan uses tile context to select an action: the satellite downlinks data in high-value contexts, discards data in low-value ones, and executes an application to more thoroughly filter the rest. Whether or not computationally limited, Kodan uses context-specific models to increase precision and downlinked data value density, i.e., the fraction of a saturated downlink composed of high-value bits. Kodan recognizes that not all sensor data need equal care in processing. To trade precision for execution time, Kodan adjusts frame tile count to reduce data quantity at a cost in quality.
Kodan 通过平衡精度与执行时间来最大化数据价值密度。 运行在每颗卫星上的软件,在计算受限时优先缩短计算时间,反之则优先保证精度。
当计算能力有限时,Kodan 利用图块上下文来选择一个动作: 在高价值上下文中直接下传数据,在低价值上下文中丢弃数据,并对其余数据执行应用程序以进行更彻底的筛选。
无论计算能力是否受限,Kodan 都使用上下文特定模型来提高精度和下传数据价值密度,即饱和下行链路中高价值比特所占的比例。Kodan 认识到,并非所有传感器数据都需要同等精细的处理。为了用精度换取执行时间,Kodan 通过调整图像帧的图块数量,以牺牲一定质量为代价来减少数据量。
Kodan increases downlinked data value density by mitigating the computational bottleneck. To show the value of Kodan, we implement seven end-to-end, deployment-ready, pixelsegmentation applications trained to filter low-value clouds using publicly-available, geospatial datasets. We quantify the improvement in valuable data downlinked with on-orbit computing using context-specialization to identify high-value observations; Kodan improves the data value density of the saturated downlink between 89 and 97 percent compared to the bent pipe.
Kodan 通过缓解计算瓶颈,提升了下传数据的价值密度。 为了展示 Kodan 的价值,我们利用公开的地理空间数据集,实施了七个端到端、可随时部署的像素分割应用,这些应用被训练用于滤除低价值的云层。我们量化了通过在轨计算,并利用上下文专业化 (context-specialization) 识别高价值观测,所带来的下传有价值数据的提升;与“弯管”模式相比,Kodan 将饱和下行链路的数据价值密度提升了 89% 至 97%
To summarize, the main contributions of this work are:
• We characterize the impact of orbital edge computing on both the downlink bottleneck and the computational bottleneck.
• We present Kodan, an OEC system that addresses the computational bottleneck using hardware-aware modifications of satellite applications.
• We characterize and evaluate context-specific models, frame tiling, and context-based elision to maximize data value density within the constraints of the computational bottleneck.
• We provide a comprehensive evaluation of Kodan across satellite data processing applications and hardware targets, resulting in improvements to the data value density of the saturated downlink between 89 and 97 percent.
总结而言,本文的主要贡献如下:
- 我们阐明了在轨边缘计算对下行链路瓶颈和计算瓶颈的双重影响
- 我们提出了 Kodan,一个利用硬件感知 (hardware-aware) 的方式修改卫星应用程序以解决计算瓶颈的在轨边缘计算系统
- 我们对上下文特定模型、图像帧分块以及基于上下文的省略策略 (context-based elision) 进行了描述和评估,这些技术旨在计算瓶颈的约束下最大化数据价值密度
- 我们对 Kodan 在多种卫星数据处理应用和硬件目标上进行了全面评估,结果表明,它能将饱和下行链路的数据价值密度提升 89% 至 97%