跳转至

Related Work

Satellite Networking: Our work follows recent work [5, 6, 11, 21–23, 43, 44] in the satellite networking domain, including: edge computing on the satellites [6, 11], ground station architectures [22, 43, 44], security of satellite networks [19], inter-satellite links [21, 23], network benchmarking [21], etc. For satellite-ground station traffic, past work [11, 43, 44] treats the satellite-ground station contact as the bottleneck and schedules traffic greedily, i.e., transfers as much data as possible in every contact. Unlike past work, Umbra takes a withhold scheduling approach, where all or part of the data can be withheld for subsequent contacts between satellite and Earth. We are also the first ones to focus on the ground station-cloud bandwidth as an emerging bottleneck given the rapid advances in satellite-ground station radio speeds [13].

卫星网络 (Satellite Networking)

本研究遵循了卫星网络领域的近期工作 [5, 6, 11, 21–23, 43, 44]，这些工作涵盖了：卫星边缘计算 [6, 11]、地面站架构 [22, 43, 44]、卫星网络安全 [19]、星间链路 [21, 23]、网络基准测试 [21] 等。在星地通信流量方面，过去的工作 [11, 43, 44] 通常将星地接触本身视为瓶颈，并采用贪心策略 (greedily) 进行流量调度，即在每次接触中尽可能多地传输数据。与以往工作不同，Umbra 采用了一种暂缓调度 (withhold scheduling) 方法，允许将全部或部分数据保留至后续的星地接触窗口。鉴于星地无线电速率的飞速发展 [13]，我们也是首个将地面站到云端的带宽视为一个新出现的瓶颈并加以关注的研究。

Time Expanded Networks: Scheduling dynamic network flows is well-studied [16, 17, 40]. Flow scheduling using time expanded networks has been explored in the context of scheduling traffic in the internet [16] and sneakernets [7]. Recently, some research has looked to formulate time expanded networks in the satellite context [39, 45, 47]. This work focuses on the task of relaying traffic through a network of interconnected satellites and models it from an energy [39], compute [45], and network perspective [47]. Our modeling of this problem is unique because we are the first to model the end-to-end data transfer from large scale satellite constellations to the cloud as a time expanded network. This modelling is challenging in its scale – hundreds of satellites, tens of ground station antennas, and time varying links. In addition, we are the first to leverage time expanded networks for load balancing. Finally, our work reveals new insights like how ground stations can suffer from load imbalance and how we can frame a new withhold scheduling approach by performing analysis on this time expanded graph.

时间扩展网络 (Time Expanded Networks)

动态网络流的调度是一个被深入研究的领域 [16, 17, 40]。利用时间扩展网络 (TEN) 进行流调度的方法已在互联网流量调度 [16] 和潜行网 (sneakernets) [7] 的场景中被探索过。近期，一些研究尝试在卫星通信的背景下构建时间扩展网络 [39, 45, 47]。这些工作主要聚焦于通过互联卫星网络中继流量的任务，并分别从能量 [39]、计算 [45] 和网络 [47] 的角度对其进行建模。我们的建模方法是独特的，因为我们首次将从大规模卫星星座到云端的端到端数据传输问题建模为一个时间扩展网络。该建模在规模上极具挑战性——涉及数百颗卫星、数十个地面站天线以及时变链路 (time varying links)。此外，我们也是首次利用时间扩展网络来实现负载均衡 (load balancing)。最后，我们的工作揭示了新的见解，例如地面站如何遭受负载不均 (load imbalance) 的问题，以及我们如何通过分析这个时间扩展图来构建一种全新的暂缓调度方法。

Traffic Scheduling Algorithms: There has been a large body of work on scheduling algorithms in other contexts such as data centers. Our observations and results in satellite traffic are analogous to the delay scheduling work for cluster scheduling [50]. Delay scheduling observed that instead of scheduling jobs to the first available node on a cluster, it is advantageous to wait for a small amount of time and find a node that has favorable features (e.g., data locality). This improves overall system performance. In spite of similarity, Umbra’s setting and techniques are different— network traffic instead of cluster scheduling. Furthermore, the scale of our problem is enormous and requires network flow formulations & optimizations.

流量调度算法 (Traffic Scheduling Algorithms)

在其他领域（如数据中心）已有大量关于调度算法的研究。我们在卫星流量中观察到的现象和结果， 与集群调度中的延迟调度 (delay scheduling) 工作 [50] 有相似之处。 延迟调度发现，与其将作业调度到集群中第一个可用的节点，不如等待一小段时间，去寻找一个具有有利特征（例如，数据局部性 (data locality)）的节点，这样做反而能提升整体系统性能。尽管存在相似性，Umbra 的应用场景和技术是不同的——我们处理的是网络流量而非集群调度。此外，我们问题的规模极其庞大，需要依赖网络流的公式化建模与优化。

Disruption Tolerant Networks (DTN): DTN is a class of networks experiencing intermittent connections between endpoints. There is a rich literature in routing and traffic engineering in DTN [30], and in applications of TEN in DTN such as running shortest path algorithm for routing [35] or maximum flow algorithm for optimizing throughput [46]. Traffic in DTNs flow in a "Store, Carry and Forward" manner [34] similar to LEO satellite systems, but Umbra is different from the works on DTN in terms of the scale of the system and the objective being optimized. First, Umbra deals with earth imaging satellites which transmit massive images rather than short messages, where the bottleneck is not only intermittent connections but also limited networking bandwidth in the system. Second, while previous work on DTN mostly deals with a multi-agent network and focuses on optimizing for one device in the network, Umbra works in a centralized network where all the ground stations and satellites are owned by the same entity and they can collaborate to optimize for a global objective. Finally, while the DTN works try to optimize either latency or throughput using TENs, Umbra is, to our best knowledge, the first algorithm that can optimize the 2 objectives simultaneously.

容断网络 (DTN) 是一类端点之间连接呈间歇性的网络。在DTN的路由和流量工程 [30] 方面，以及TEN在DTN中的应用（例如使用最短路径算法进行路由 [35] 或使用最大流算法优化吞吐量 [46]）方面，已有丰富的研究文献。DTN中的流量以“存储-携带-转发” (Store, Carry and Forward) 的方式流动 [34]，这与低轨（LEO）卫星系统相似。

然而，Umbra在以下几个方面与传统的DTN研究不同：

  1. 系统规模与瓶颈：Umbra处理的是地球成像卫星传输的海量图像而非短消息，其瓶颈不仅在于间歇性连接，还包括整个系统有限的网络带宽

  2. 优化目标：以往的DTN研究大多处理多智能体网络，并侧重于为网络中的单个设备进行优化。而Umbra工作在一个集中式网络中，所有地面站和卫星都由同一实体拥有，它们可以协同工作以实现一个全局优化目标

  3. 多目标优化：虽然之前的DTN工作尝试使用TEN来优化延迟或吞吐量，但据我们所知，Umbra是首个能够同时优化这两个目标的算法