Research digests
研究導讀
48 papers, one page each48 篇論文，每篇一頁
Papers are grouped by topic, with a prerequisite chapter for each group. Each digest explains constellation, link-rate, latency and failure assumptions, measurement coverage and links to the chapters. Venue and evidence tags follow the policy in How to use.論文依主題分組，各組列出先備章節。每篇導讀說明星座、鏈路速率、延遲及故障假設、量測與證據範圍，以及相關章節。發表場域與證據標記依使用說明中的規則。
Link physics鏈路物理 5 · read after Chapter 25 篇 · 第 2 章後
IEEE VTM 2021 peer-reviewedLaser Intersatellite Links in a Starlink Constellation: A Classification and AnalysisLaser range determines how many neighbors a Starlink satellite can reach. Most contacts last minutes, so setup time limits the use of temporary links.雷射通訊距離決定 Starlink 衛星可連接的鄰居數量。大多數接觸僅維持數分鐘，因此鏈路建立時間限制了暫時性鏈路的使用。Chaudhry & Yanikomeroglu · Ch. 2
IEEE OJCOMS 2022 peer-reviewedTemporary Laser Inter-Satellite Links in Free-Space Optical Satellite NetworksMillisecond repointing could add links, especially near the poles and at medium laser ranges. Assess the intercontinental latency gains and assumed per-hop node delay.毫秒級重新指向能增加鏈路，極區附近及中等雷射距離的增益尤其明顯。導讀分析洲際端到端延遲的改善，並檢視論文假設的每跳節點延遲。Chaudhry & Yanikomeroglu · Ch. 2
arXiv 2301.05285 (WCNC) preprintLaser Inter-Satellite Link Setup Delay: Quantification, Impact, and Tolerable ValueSetup delay applies once per path change. A one-second setup raises a 26 ms path to a 124 ms average. Tolerable setup delays range from a few milliseconds on short routes to tens of milliseconds on long routes.鏈路建立延遲按每次路徑切換計算。一秒的建立時間可使原本 26 ms 的路徑達到平均 124 ms。可容忍的建立延遲隨路線長度增加，從短路線的幾毫秒，增至長路線的數十毫秒。Bhattacharjee, Chaudhry, Yanikomeroglu, Hu, Lamontagne · Ch. 2
ICSOS 2023 / NASA NTRS peer-reviewedRecent On-Orbit Results and ARQ Performance Analysis for the TBIRD 200-Gbps MissionA 6U CubeSat sent 4.8 TB during a five-minute pass using commercial transceivers. Selective-repeat ARQ delivered error-free data. The receiver's Doppler window and one-meter adaptive-optics telescope define this ground-link demonstration's scope.一顆 6U 立方衛星使用現成電信收發器，在五分鐘過頂期間下傳 4.8 TB，並以選擇性重傳 ARQ 達成零錯誤傳輸。接收器的多普勒窗口與一公尺自適應光學望遠鏡界定了這項地面鏈路展示的適用範圍。Schieler et al. · Ch. 2
IEEE TAES 2023 peer-reviewedSignal Structure of the Starlink Ku-Band DownlinkThree published parameters yield a 4.4 µs OFDM symbol and a 1.333 ms Starlink downlink frame. The primer calculates a 456 Mbps per-channel ceiling from them.三個公開參數可推導出 Starlink 下行的 4.4 µs OFDM 符號與 1.333 ms 幀。本教材再據此計算出每頻道 456 Mbps 的上限。Humphreys, Iannucci, Komodromos, Graff · Ch. 3
Measuring Starlink量測 Starlink 8 · read after Chapter 38 篇 · 第 3 章後
ACM IMC 2022 peer-reviewedA First Look at Starlink PerformanceThe first academic Starlink study measured one Belgian dish. Its packet-loss results mainly describe the access link. The digest also identifies a frequently misattributed finding.第一篇學術 Starlink 量測研究使用一支比利時終端，封包遺失結果主要反映存取鏈路。導讀也釐清一項常被誤植至此論文的結論。Michel, Trevisan, Giordano, Bonaventure · Ch. 3
ACM WWW 2024 peer-reviewedA Multifaceted Look at Starlink PerformancePoP placement helps explain a 40 ms bent-pipe floor and 100 ms island latency where gateways are absent. Loaded RTT rises two to four times. Two controlled dishes identify the 15-second boundaries as a global schedule.PoP 位置有助於解釋 40 ms 的彎管延遲下限，以及閘道缺席的島嶼為何達到 100 ms。負載下的 RTT 升為兩到四倍，而兩支受控終端的量測將 15 秒邊界定位為全球排程。Mohan et al. · Ch. 3
Proc. ACM MACS / SIGMETRICS 2024 peer-reviewedDemocratizing LEO Satellite Network MeasurementTwo TTL-limited ping streams isolate a remote connection's satellite hop, with validation against a real dish. They reveal inter-satellite routing in Nigeria, Brazil and the Seychelles beyond the reach of earlier dish-based studies.兩道限制 TTL 的 ping 串流可分離遠端 Starlink 連線中的衛星跳段，研究並以實際終端驗證方法。它涵蓋先前終端量測研究較難觸及的奈及利亞、巴西與塞席爾，揭露當地的星間路由。Izhikevich, Tran, Izhikevich, Akiwate, Durumeric · Ch. 3
CoNEXT 2023 Companion peer-reviewedMaking Sense of Constellations: Methodologies for Understanding Starlink's Scheduling AlgorithmsObstruction maps and public orbital elements identify the serving satellite. Latency changes at seconds 12, 27, 42 and 57 reveal scheduler timing and satellite preferences.讀完後，你能說明遮蔽圖與公開軌道根數如何結合來辨識終端正在使用哪顆衛星、為什麼延遲狀態在每分鐘的第 12、27、42 與 57 秒改變，以及 Starlink 排程器偏好哪些衛星及其原因。Tanveer, Puchol, Singh, Bianchi, Nithyanand · Ch. 3
IEEE PIMRC 2023 peer-reviewedMeasuring a Low-Earth-Orbit Satellite NetworkThe 2023 ground-segment map connects 24 km service cells, nine-dish gateways and 21 PoPs. The digest distinguishes measured values from rules of thumb.讀完後，你能按論文 2023 年所描繪的畫出 Starlink 的地面段，從 24 km 的服務細胞經過九支天線的閘道到 21 個接入點之一，並說出其中哪些被廣泛引用的數字是量測、哪些是經驗法則。Pan, Zhao, Cai · Ch. 3
IEEE INFOCOM 2023 peer-reviewedNetwork Characteristics of LEO Satellite Constellations: A Starlink-Based Measurement from End UsersA Vancouver-area study measured rain, geomagnetic storms and vehicle motion. These effects act on separate system components. Later work reinterprets several observations from 2022.溫哥華地區的一項研究量測雨、地磁風暴及車輛移動對 Starlink 鏈路的影響。這些因素各自作用於系統的特定部位，後續研究也重新詮釋了部分 2022 年觀察。Ma et al. · Ch. 3
USENIX NSDI 2024 peer-reviewedDemocratizing Direct-to-Cell Low Earth Orbit Satellite NetworksMOSAIC replaces ground-built cellular sessions with one-time tokens that satellites verify locally. Assess its speedups against the assumed inter-satellite links outside the paper's model.MOSAIC 以衛星可在本地驗證的一次性權杖，取代依地面逐跳建立的蜂巢式會談。導讀說明高速移動基地台帶來的會談問題，並檢視模型範圍之外的星間鏈路假設如何影響加速結果。Liu et al. · Ch. 3
ACM SIGCOMM 2021 peer-reviewedL2D2: Low Latency Distributed Downlink for LEO SatellitesMany small receive-only stations and a learned link predictor reduce hours-long waits for Earth-observation data by about four times. Separate real-link measurements from trace-driven simulation.讀完後，你能說明為什麼地球觀測資料要等上數小時才能抵達地面，為什麼以許多小型僅接收站與一個學習而來的鏈路預測器取代少數大型地面站，能把等待時間縮短約四倍，以及論文中哪些數字來自真實鏈路、哪些來自軌跡驅動的模擬。Vasisht, Shenoy, Chandra · Ch. 3
Topology and routing拓樸與路由 10 · read after Chapter 510 篇 · 第 5 章後
ACM HotNets 2018 peer-reviewedDelay is Not an Option: Low Latency Routing in SpaceThe paper reports propagation-only results for a proposed five-laser constellation. Trace the zigzag paths behind “47% faster than fiber” and “beats fiber beyond 3,000 km”.這篇論文針對擬議的五雷射星座計算純傳播延遲。導讀追溯「比光纖快 47%」與「超過 3,000 公里就勝過光纖」的出處，並說明網格中的鋸齒狀路徑。Handley · Ch. 5
ACM HotNets 2019 peer-reviewedUsing Ground Relays for Low-Latency Wide-Area Routing in MegaconstellationsGround relays let a 550 km shell using ground links beat a great-circle fiber path. Examine the one-relay-per-degree saturation and the hybrid relay–ISL design.一個 550 公里殼層可藉地面中繼轉送，勝過大圓光纖路徑。導讀說明「每一度一個中繼」的飽和點，以及結合地面中繼與星間鏈路的設計。Handley · Ch. 5
ACM HotNets 2018 peer-reviewedGearing up for the 21st Century Space RaceCompare four Frankfurt–Washington latency cases and polar +Grid link shutdowns. Examine the capacity and cost estimates, then trace the research agenda into later work.讀完後，你能說明法蘭克福–華盛頓的四階延遲梯、為什麼極軌道 +Grid 在接近極區時會關閉跨軌道鏈路、論文的容量與成本估計是如何建立的，以及它的議程中有哪些項目後來的文獻真的接手了。Bhattacherjee et al. · Ch. 5
ACM CoNEXT 2019 peer-reviewedNetwork Topology Design at 27,000 km/hourA motif repeats links among co-moving satellites to keep topology fixed over an orbit. Examine the stretch-plus-hop objective and the 54% gain over +Grid. At some link ranges, gains fall to 1–4%, making this a design baseline with specific assumptions.Motif 在同向運動的衛星間重複配置鏈路，使拓樸在整個軌道週期內保持固定。導讀說明「伸長比加跳數」的計分方式，以及相對 +Grid 的 54% 改善如何隨鏈路距離縮至 1–4%。這些結果適合作為具有明確假設的設計空間基準。Bhattacherjee & Singla · Ch. 5
ACM SIGCOMM CCR 2020 peer-reviewedInternet Backbones in SpaceA transit constellation resembles an autonomous system whose peerings change every few minutes. BGP-compatible filtering reduces connectivity, and ground infrastructure may cost three times the satellites. Compare the four integration designs and the paper's deployment assessment.讀完後，你能說明為什麼一個販售中轉服務的星座看起來像一個對等連線每隔幾分鐘就出現又消失的自治系統、對 BGP 友善的過濾器會丟掉多少連線能力、為什麼地面段的成本可能是衛星的三倍，以及論文判斷四種整合設計中哪一種可部署。Giuliari, Klenze, Legner, Basin, Perrig, Singla · Ch. 5
arXiv 2602.07756 preprintLEO Topology Design Under Real-World Deployment ConstraintsLinks feasible over a full orbit reduce average delay from 89.0 ms to 52.9 ms and hops from 23.4 to 8.2. The evaluation uses 90 days of Starlink ephemerides and examines +Grid's failures on partially deployed shells.讀完後，你能說明為什麼在 90 天的真實 Starlink 星曆上，只用整個軌道週期都保持可行的鏈路所建的拓樸，能把衛星對衛星的平均延遲從 89.0 ms 降到 52.9 ms、跳數從 23.4 降到 8.2，以及為什麼熟悉的 +Grid 在真正在飛的部分部署殼層上會失效。Ali & Zhang · Ch. 5
ACM SIGCOMM 2025 peer-reviewedSmall-scale LEO Satellite Networking for Global-scale DemandsEarth-repeat orbits serve modeled Starlink demand with 1,763 satellites versus 6,793. A model-predictive controller converts operator intent into changing topologies using three 200 Gbps lasers per satellite. Starlink's actual topology remains confidential, which limits comparisons with +Grid.地球重複軌道可在模型中以 1,763 顆衛星服務原本需 6,793 顆的 Starlink 客戶需求。模型預測控制器利用每顆衛星的三條 200 Gbps 雷射，將營運意圖轉成時變 ISL 拓樸。Starlink 的實際拓樸仍屬保密資訊，這也界定了第 5 章各項 +Grid 結果的比較範圍。Li et al. · Ch. 5
ACM MobiCom 2023 peer-reviewedA Networking Perspective on Starlink's Self-Driving LEO Mega-ConstellationOne collision-avoidance maneuver can trigger dozens of neighboring maneuvers and break ISLs. Examine how these movements challenge topology models that hold links fixed.一次碰撞閃避機動可能引發數十次鄰近衛星機動，並中斷星間鏈路。導讀說明這些連鎖變化如何影響網路，以及固定 ISL 拓樸模型所依賴的運動假設。Li et al. · Ch. 5
ACM MobiCom 2024 peer-reviewedStable Hierarchical Routing for Operational LEO NetworksChanging links, ground attachments and membership disrupt satellite-identity routing. Earth-fixed geographic cells give coordinate distances a stable reference, while a three-tier hierarchy handles remaining changes.鏈路、地面連接與成員持續變動，會擾動依衛星身分建立的路由。地球固定格元提供穩定的座標距離參考，三層階層則處理其餘變動。Li et al. · Ch. 5
ACM SIGCOMM 2025 peer-reviewedSaTE: Low-Latency Traffic Engineering for Satellite NetworksA topology holds for about 70 ms. A graph-attention model trained on 512 representative topologies allocates traffic for 4,236 satellites in 17 ms, versus Gurobi's 46 s. Examine the ISL, bandwidth and 15-second scheduling assumptions behind these results.讀完後，你能說明為什麼在 LEO 星座上做流量工程是一場與只靜止約 70 ms 的拓樸賽跑，一個在 512 個代表性拓樸上訓練的圖注意力模型如何在 17 ms 內算出 4,236 顆衛星的配置而 Gurobi 需要 46 秒，以及這個結果建立在哪些關於 ISL、頻寬與 15 秒排程的假設上。Wu, Han, Rajpal, Zhang, Wang · Ch. 5
Transport and congestion control傳輸層與壅塞控制 6 · read after Chapter 66 篇 · 第 6 章後
ACM IMC 2022 peer-reviewedA Browser-side View of Starlink ConnectivityA browser extension and three Raspberry Pis reveal clustered Starlink loss and doubled page transit time in rain. BBR leads five controllers but reaches about half the UDP-measured capacity.讀完後，你能說明為什麼一個瀏覽器擴充套件加三台 Raspberry Pi 就足以證明 Starlink 的封包遺失是成叢出現的、雨天會讓網頁傳輸時間翻倍，以及 BBR 雖是五種壅塞控制器中最好的，卻仍只達到 UDP 量測容量的一半左右。Kassem, Raman, Perino, Sastry · Ch. 6
IEEE INFOCOM 2023 peer-reviewedSaTCP: Link-Layer Informed TCP Adaptation for Highly Dynamic LEO Satellite NetworksHandover timing lets senders retain their congestion windows across predicted disruptions. Emulated CUBIC throughput rises from 34.3 to 138.1 Mbps, while competing standard CUBIC flows face starvation.預知換手時點可讓傳送端在預期中斷期間保留壅塞視窗，模擬中的 CUBIC 吞吐量因而從 34.3 升至 138.1 Mbps。同一機制也會使共用鏈路的原版 CUBIC 流遭遇頻寬飢餓。Cao & Zhang · Ch. 6
ACM HotNets 2024 peer-reviewedMind the Misleading Effects of LEO Mobility on End-to-End Congestion ControlNine controllers overfill or underuse a Starlink link whose capacity changes by tens of Mbps every 15 seconds. In trace replay, discarding samples from before reconfiguration reduces BBRv1 RTT by 30% and raises Copa throughput by 74%.九種控制器在每 15 秒容量變動的 Starlink 鏈路上，呈現過度傳送與利用偏低兩群。丟棄重組前樣本的軌跡重播結果，集中列於導讀的比較表。Lai et al. · Ch. 6
ACM SIGCOMM 2025 peer-reviewedLeoCC: Making Internet Congestion Control Robust to LEO Satellite DynamicsLeoCC infers reconfiguration from ACK arrival patterns and discards older bottleneck samples. Examine the scope of its 85–494% headline gains and independent LeoEM comparisons with CUBIC, SaTCP and BBRv3.LeoCC 從 ACK 到達模式推斷重組，並丟棄事件之前的瓶頸樣本。導讀解釋 85–494% 主要增益的適用範圍，以及獨立 LeoEM 評估中與 CUBIC、SaTCP、BBRv3 的比較。Lai et al. · Ch. 6
ACM IMC 2026 peer-reviewedCharacterizing the Configuration of Starlink QueuingOver ten million hardware-timestamped probes identify a single drop-front queue holding about 1,400–1,600 packets. Flows share capacity unevenly, with passive queue management. Examine how front drops reduce delay and deliver earlier congestion signals to loss-based senders.超過一千萬個硬體時間戳探測封包，揭露約可容納 1,400–1,600 個封包的單一前端丟棄佇列。各流取得的容量分配失衡，佇列採被動管理。導讀說明前端丟棄如何降低延遲，同時提早向基於遺失的傳送端提供壅塞訊號。Garcia, Sundberg, Brunstrom · Ch. 6
ACM IMC 2026 peer-reviewedLooking at Stars: Analyzing Real-Time Video Conferencing over StarlinkA 2.5 Mbps Teams uplink can fall below target despite ample terminal capacity. WebRTC interprets 15-second reconfiguration delays as congestion. The widely cited handover statistics come from this video-conferencing study.讀完後，你能說明為什麼一個輕鬆容納在 Starlink 終端容量內的 2.5 Mbps Teams 上行仍會掉到目標以下、WebRTC 基於延遲的壅塞估計器如何把 15 秒重組誤認為壅塞，以及為什麼這篇論文常被單獨引用的換手對齊統計，其實是一項視訊會議研究的副產品。Fries, Matezki, Bose, Mohan · Ch. 6
Simulation and emulation模擬與模擬平台 4 · read after Chapter 74 篇 · 第 7 章後
ACM IMC 2020 peer-reviewedExploring the "Internet from Space" with HypatiaHypatia adds packet-level protocols to propagation models and exposes path dynamics. Its transport results depend on modeled capacity, while validation against deployed Starlink remains pending.Hypatia 在純傳播延遲模型之外加入封包層協定，呈現路徑動態。傳輸層結果依賴模型設定的容量，與實際 Starlink 網路的對照驗證仍待補充。Kassing, Bhattacherjee, Águas, Saethre, Singla · Ch. 7
USENIX NSDI 2023 peer-reviewedStarryNet: Empowering Researchers to Evaluate Futuristic Integrated Space and Terrestrial NetworksStarryNet runs routing software on 4,408 emulated satellites. Setup takes minutes and roughly one-third of seven servers' resources. Users supply the ISL capacity as a model parameter.StarryNet 在 4,408 顆模擬衛星上執行真實路由程式，建置耗時數分鐘，約使用七台伺服器總資源的三分之一。星間鏈路容量由使用者設定為模型參數。Lai et al. · Ch. 7
ACM/IFIP Middleware 2022 peer-reviewedCelestial: Virtual Software System Testbeds for the LEO EdgeCelestial gives each satellite a virtual machine and suspends idle microVMs to support 4,409 instances. Its accuracy comparison uses its own simulation. ISL conclusions therefore require additional validation.Celestial 為每顆衛星提供虛擬機，並暫停閒置實例，以支援 4,409 台 microVM。其準確度比對採用自身模擬，星間鏈路相關結論仍需額外驗證。Pfandzelter & Bermbach · Ch. 7
IFIP Networking 2024 peer-reviewedxeoverse: A Real-time Simulation Platform for Large LEO Satellite Mega-ConstellationsStandard Linux software needs constellation emulation to keep pace with real time. xeoverse precomputes topology and updates changed links. Its reported performance depends on user-selected model parameters.原版 Linux 軟體要求星座模擬跟上實際時鐘。xeoverse 預先計算可預測的拓樸，再只更新發生變化的鏈路。主要效能數值取決於使用者設定的模型參數。Kassem & Sastry · Ch. 7
Computing in orbit在軌運算 4 · read after Chapter 84 篇 · 第 8 章後
ASPLOS 2020 peer-reviewedOrbital Edge Computing: Nanosatellite Constellations as a New Class of Computer SystemDownlink capacity limits Earth-observation data delivery. A computational nanosatellite pipeline distributes one frame's inference across satellites. The reported results come from simulation.下行容量限制地球觀測資料的遞送速度。運算奈米衛星管線將單一影像的推論分散至多顆衛星，論文報告的結果均來自模擬。Denby & Lucia · Ch. 8
ASPLOS 2023 peer-reviewedKodan: Addressing the Computational Bottleneck in SpaceOn-board inference shifts the constraint to compute deadlines. Geospatial model specialization helps a slow space processor match a fast camera. The digest identifies Kodan's documented results and source limits.機上推論將限制轉為運算期限。依地理情境特化模型，有助於讓較慢的太空處理器跟上高速相機。導讀說明 Kodan 各項數值的來源與證據範圍。Denby, Chintalapudi, Chandra, Lucia, Noghabi · Ch. 8
ACM HotNets 2020 peer-reviewedIn-orbit Computing: An Outlandish Thought Experiment?Reproduce estimates for one commercial server per Starlink satellite: mass, power and cost. Handoff is the main network obstacle. Distinguish these estimates from propagation-only latency calculations.導讀重現在每顆 Starlink 衛星配置一台商用伺服器的質量、電力與成本估算。網路方面的主要障礙在於換手，文中也區分估計值與純傳播延遲。Bhattacherjee, Kassing, Licciardello, Singla · Ch. 8
ACM MobiCom 2024 peer-reviewedDeciphering the Enigma of Satellite Computing with COTS Devices: Measurement and AnalysisA 5 W single-board computer can overheat inside a satellite despite cold surroundings. Two thermal curves set its duty cycle. Six months of error-free operation defines a specific observation period.即使周圍太空寒冷，衛星內部的 5 W 單板電腦仍可能過熱。兩條熱曲線界定其工作週期，而六個月零單事件錯誤的結果，具有明確的觀測期間與推論範圍。Xing, Xu, Zhou, Zhang, Wang, Li, Qian · Ch. 8
Services in orbit在軌服務 5 · read after Chapter 95 篇 · 第 9 章後
ACM HotNets 2024 peer-reviewedIt's a Bird? It's a Plane? It's CDN! Investigating Content Delivery Networks in the LEO Satellite Networks EraDistant PoPs weaken CDN locality for Starlink users. The study measures this effect by country and simulates satellite-hosted SpaceCDN caches. Assess its gains within that simulation's scope.遠方 PoP 會削弱 Starlink 使用者取得 CDN 鄰近服務的優勢。研究逐國量測這項影響，並模擬在衛星部署快取的 SpaceCDN，導讀據此說明增益的適用範圍。Bose, Fadaei, Mohan, Kassem, Sastry, Ott · Ch. 9
ACM SIGCOMM 2025 peer-reviewedStarCDN: Moving Content Delivery Networks to SpaceMoving caches require coordination beyond local LRU. Satellite-specific consistent hashing and relayed fetch combine small caches into a larger service. Assess the Gen-1 proxy topology, 15-second steps and measured idle-latency inputs.移動快取需要跨衛星協調。衛星專屬的一致性雜湊與接力擷取，將多個小快取整合為較大的服務。導讀檢視 Gen-1 代理拓樸、15 秒步長，以及取自量測的閒置延遲。Zheng, Taneja, Masood, Sabnis, Sitaraman, Vasisht · Ch. 9
IEEE WCL 2022 peer-reviewedGround-Assisted Federated Learning in LEO Satellite ConstellationsGround-contact timing limits satellite federated learning. FedSat makes FedAvg asynchronous using predictable contacts, while AsyncFLEO and FedRings add intra-plane ISLs to reduce waiting.地面接觸時點限制了跨衛星聯邦學習。FedSat 利用接觸的可預測性，將同步 FedAvg 改為非同步方案，後續的 AsyncFLEO 與 FedRings 則加入同軌道面 ISL，減少等待。Razmi, Matthiesen, Dekorsy, Popovski · Ch. 9
arXiv 2202.01267 preprintFedSpace: An Efficient Federated Learning Framework at Satellites and Ground StationsIntermittent ground contacts create a trade-off between idleness and staleness. FedSpace schedules aggregation using the orbits of a real 191-satellite Planet fleet. Its ground-only communication assumption needs a date-specific scope.間歇性的地面接觸造成閒置與陳舊度之間的取捨。FedSpace 利用真實的 191 顆 Planet 衛星軌道排程彙總，其僅經地面通訊的核心假設需要附上日期範圍。So, Hsieh, Arzani, Noghabi, Avestimehr, Chandra · Ch. 9
NINeS 2026 (OASIcs) peer-reviewedOrbitalBrain: Federated Learning over Inter-Satellite LinksISLs shift federated learning from ground-pass waiting to routing over changing graphs. OrbitalBrain aggregates along a shortest-path tree. Its ~100 Mbps ISL cap is five orders of magnitude below Part III's formation clusters.加入 ISL 後，聯邦學習的問題從等待地面過頂轉為在時變圖上路由。OrbitalBrain 沿最短路徑樹彙總更新，其約 100 Mbps 的 ISL 上限，比第三部編隊叢集低五個數量級。Chabra, Li, Hsieh, Arzani, Olsen, Chandra, Segarra · Ch. 9
AI clusters and inference in orbit在軌 AI 叢集與推論 6 · read after Chapter 116 篇 · 第 11 章後
Joule 2026 / arXiv 2511.19468 peer-reviewedToward a Future Space-Based, Highly Scalable AI Infrastructure System DesignGoogle proposes satellites 100–200 m apart and about 10 Tbps from a 5 W transmitter. Examine disputed link-budget constants and the additional details needed to analyze the cluster network.Google 提議將衛星配置在相距 100–200 m 的位置，以 5 W 發射器承載約 10 Tbps。導讀檢視兩次閱讀間存在爭議的鏈路預算常數，以及分析叢集內部網路所需的補充資訊。Agüera y Arcas et al. (Google) · Ch. 11
arXiv 2605.15335 preprintDesigning Dense Satellite Clusters for Distributed Space-based DatacentersCircular relative orbits fit over four times as many satellites as Suncatcher's 2:1 ellipses. Compare planar and three-dimensional clusters, including compute positions consumed by Clos switching satellites.讀完後，你能說明為什麼圓形相對軌道能比 Suncatcher 的 2:1 橢圓多排進四倍以上的衛星、三維叢集何時勝過平面叢集，以及在視線雷射鏈路上運行 Clos 網路架構要付出多少運算衛星的代價。Pénot & Balakrishnan · Ch. 11
arXiv 2607.14172 preprintThe Cost and Network Limits of Space-Based AI ComputeAt 8,000 nodes, a laser torus has three to four orders of magnitude less bisection bandwidth than a terrestrial Clos. Assess training implications, single-satellite inference scope and the formulas behind the comparison.在 8,000 個節點下，雷射環面的對分頻寬比地面 Clos 少三到四個數量級。導讀說明對訓練的影響、單衛星推論的適用範圍，以及引用比較結果前應檢查的公式與假設。van Berkel · Ch. 10
arXiv 2605.00515 preprintSpaceMoE: Realizing Distributed Mixture-of-Experts Inference over Space NetworksSpaceMoE distributes experts across 1,056 satellites. A radiation-hardened processor sets its 1.02–1.07 s per-token time despite “≥ 100 Gbps” links. Examine the modeling choices behind that result.SpaceMoE 將混合專家模型分布至 1,056 顆衛星。即使假設「≥ 100 Gbps」雷射鏈路，每 token 1.02–1.07 s 的耗時仍由抗輻射處理器決定。導讀說明引用前應檢查的建模選擇。Wang, Yang, Huang, Sheng, Letaief · Ch. 12
arXiv 2604.04654 preprintCommunication-Efficient Collaborative LLM Inference over LEO Satellite NetworksCollabLLM-LEO splits a transformer across satellites. Learned masks, 8-bit quantization and entropy coding shrink layer-boundary tensors by over 20×. Its 42% delay reduction compares in-orbit processing with downloading data before processing.CollabLLM-LEO 將 transformer 拆分至多顆衛星，以學習式遮罩、8 位元量化與熵編碼，將層邊界張量縮小超過 20 倍。42% 的延遲降低比較的是在軌處理與先下傳再處理，服務端位於軌道或地面的比較則需要另外評估。Zhang, Wu, Li, Wang, Shen · Ch. 12
arXiv 2605.05615 preprintLLMSpace: Carbon Footprint Modeling for Large Language Model Inference on LEO SatellitesA bottom-up model derives embodied carbon from payload mass, launch intensity and hardware lifetime. Orbital operational carbon approaches zero, so embodied carbon dominates. Examine assumptions about mass, links and radiation lifetime.讀完後，你能說明一個由下而上的碳模型如何把酬載質量、發射碳強度與硬體壽命轉換成隱含碳、為什麼在軌的營運碳接近零而隱含碳成為主體，以及網路讀者應該質疑模型中哪些關於質量、鏈路與輻射壽命的假設。Jiang, Ildefonso, Loveless, Chen · Ch. 12