Small-scale LEO Satellite Networking for Global-scale Demands (TinyLEO)以小規模 LEO 衛星網路滿足全球規模需求（TinyLEO）
TinyLEO designs sparse constellations around geographic demand. Examine Earth-repeat orbits, model-predictive link control and geographic forwarding, with explicit limits on the Starlink baseline.TinyLEO 依地理需求設計稀疏星座。導讀說明地球重複軌道、模型預測鏈路控制與地理轉送，並界定 Starlink 比較基準的範圍。
TinyLEO matches orbit coverage to concentrated geographic demand. Matching pursuit selects Earth-repeat orbits, an orbital model-predictive controller favors long-lived ISLs through stable matching, and SRv6 forwards using geographic segment anycast. Synthesis results use fewer satellites than the dated Starlink fleet under the estimated demand [39, §6.1]. StarryNet emulation supplies control, repair and routing results [39, §6.2–6.3]. The model adopts company-stated link budgets and a standard three-ISL grid because Starlink's actual topology remains confidential. Tables distinguish synthesis assumptions from emulated network measurements.
TinyLEO 讓軌道覆蓋配合集中的地理需求。匹配追蹤選出地球重複軌道，軌道模型預測控制器以穩定配對偏好長壽鏈路，SRv6 則使用地理區段任播。在估計需求下，合成星座的衛星數低於指定日期的 Starlink 機隊[39, §6.1]，控制、修復及路由結果則來自 StarryNet 模擬平台[39, §6.2–6.3]。模型採公司聲明的鏈路預算，並因 Starlink 實際拓樸保密而使用標準三 ISL 網格。表格區分合成假設與模擬網路量測。
1. Where this paper sits1. 這篇論文的位置
TinyLEO illustrates the move from fixed link patterns to time-varying control in Chapter 5. The Tsinghua team's earlier work identified maneuver-induced link failures and unstable satellite-identity routing. This system combines orbit synthesis, geometry-aware topology control and forwarding keyed to Earth-fixed cells. Its stated per-satellite budget and explicit proxy-topology qualification also support Chapter 4's parameter sheet.
TinyLEO 呈現第 5 章從固定鏈路樣式走向時變控制的研究方向。清華團隊先前的工作指出軌道機動會中斷鏈路，衛星身分路由也會隨變動受到擾動。系統將軌道合成、依幾何調整的拓樸控制，以及地球固定格元轉送整合起來。明確的每衛星預算及代理拓樸限定，也提供第 4 章參數表的依據。
2. The problem, measured2. 問題，用量測說話
Uniform shells distribute coverage across oceans, deserts and cities, while the cited demand data concentrates users on a small land area [39, §1, §2.2]. TinyLEO tests a demand-matched design through synthesis. It assigns required capacity to Earth-fixed cells and checks spare access and ISL capacity at each time step. The baseline uses real Starlink positions from a dated fleet snapshot [39, §6.1]. Demand remains an acknowledged estimate, so satellite counts depend on those inputs.
均勻殼層將覆蓋分配至海洋、沙漠與城市，引用的需求資料則將使用者集中於少量陸地[39, §1, §2.2]。TinyLEO 透過合成測試依需求配置的設計，為地球固定格元指定所需容量，並在每個時間步檢查接取與 ISL 的剩餘容量。基準使用指定日期的實際 Starlink 衛星位置[39, §6.1]。需求仍是作者承認的估計，因此衛星數量取決於這些輸入。
| Item項目 | Value數值 | Conditions and interpretation條件與解讀 |
|---|---|---|
| Demand concentration需求集中程度 | >70% / ≈5% | Share of users and land, based on customer maps, Cloudflare Radar and ITU data.使用者與陸地比例，取自客戶地圖、Cloudflare Radar 及 ITU 資料。 |
| Uniform-shell reference均勻殼層參考 | 53° | Inclination used to illustrate demand mismatch.用於說明需求錯配的傾角。 |
| Demand grid需求網格 | 4,050 cells, 652 Tbps | A cell needs enough access and ISL capacity at every time step.每格在每個時間步都需具備足夠接取及 ISL 容量。 |
| Earth-repeat family地球重複軌道族 | 64,800, 423–1,873 km, 92.8–124.2 min | Candidate count, altitude and period ranges from Table 1.Table 1 的候選數、高度與週期範圍。 |
| Fleet baseline機隊基準 | 6,793, 5 shells, Jan 2025 | Actual Starlink positions.實際 Starlink 位置。 |
| ISL budgetISL 預算 | 3 × 200 Gbps/satellite | Company-stated peak rate. Sustained comparison is 100 Gbps in the fact base.公司聲明的峰值速率，事實庫的持續比較值為 100 Gbps。 |
| Access budget接取預算 | 96 Gbps, 100 Mbps/user → 960 users | Company specifications used as model inputs.公司規格作為模型輸入。 |
| Synthesis hardware合成硬體 | 2 CPUs, 64 threads, Gurobi 12.0 | Exact baseline remained unfinished after two months.精確基準在兩個月截止時仍持續求解。 |
| Emulated fleet模擬衛星群 | 1,741 satellites | StarryNet variant, Linux containers on an HPC cluster.StarryNet 變體，HPC 叢集上的 Linux 容器。 |
| Setup comparison建立時間比較 | 2–30 s | Chapter 2 gives current terminal timing. TinyLEO §2.3 identifies acquisition as a challenge, outside its topology-change budget.第 2 章現代終端的範圍，另於拓樸變更時間之外評估。 |
| User-link schedule scope使用者鏈路排程範圍 | 15 s | Measured Starlink reconfiguration requires an additional model.實測 Starlink 重新配置需要額外模型。 |
3. The design3. 設計
The synthesis layer selects Earth-repeat orbits whose ground tracks revisit the same demand cells. Matching pursuit greedily adds the orbit that reduces unmet demand most, while exact Gurobi runs remain unfinished at the comparison cutoff [39, §6.1]. The orbital model-predictive controller translates geographic intent into link schedules using predicted geometry and Gale–Shapley matching that favors long-lived partners [39, §4.2]. The data plane encodes cells as SRv6 segments, allowing any satellite covering a cell to forward toward the next cell [39, §4.3]. Covering satellites handle failures locally while the controller recompiles in the background. The tables separate local reroute time from controller repair and signaling.
合成層選擇地面軌跡會重訪相同需求格元的地球重複軌道。匹配追蹤每次加入最能減少需求缺口的軌道，精確 Gurobi 則在比較截止時仍持續求解[39, §6.1]。軌道模型預測控制器根據預測幾何，以偏好長壽對象的 Gale–Shapley 配對，將地理意圖轉成鏈路時程[39, §4.2]。資料平面將格元編碼為 SRv6 區段，讓當前覆蓋該格的任何衛星朝下一格轉送[39, §4.3]。覆蓋衛星先就地處理故障，控制器同時在背景重新編譯，表格區分局部重繞、控制器修復與訊令時間。
Original figure note原始圖說出處
Drawn for this primer from Li et al., SIGCOMM 2025, §4.1–4.3, §6.1–6.3, Figs. 15, 17, 19.4. Evaluation setup4. 實驗設定
The satellite-count study uses synthesis simulation with demand, orbit and position data from Space-Track, ITU and Cloudflare Radar. Gurobi supplies a time-limited exact baseline. Control and forwarding results come from a hardware-in-the-loop StarryNet variant using Linux containers, with configured link capacity and delay. Company specifications supply the access and ISL budgets.
Cell size bounds hop delay, while node delay, laser acquisition cost and user-link reconfiguration require additional models. Baselines include real Starlink positions, MegaReduce, Gurobi, TS-SDN with route aggregation enabled or disabled, and grid- or MPC-controlled Starlink variants. The stretch result uses the MPC-controlled variant.
衛星數研究採合成模擬，需求、軌道及位置資料來自 Space-Track、ITU 與 Cloudflare Radar，Gurobi 則提供設有截止時間的精確基準。控制與轉送結果來自硬體在環的 StarryNet 變體，以 Linux 容器執行，鏈路容量及延遲均由設定提供。接取與 ISL 預算取自公司規格，每跳延遲以格元大小為界。節點延遲、雷射擷取成本與使用者鏈路重組仍需額外模型。基準包含實際 Starlink 位置、MegaReduce、Gurobi、啟用或停用路由聚合的 TS-SDN，以及網格控制或 MPC 控制的 Starlink 變體，延展比採後者作比較。
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Aggregate ISL capacity per satellite under the §6.1 budget在 §6.1 預算下每顆衛星的 ISL 總容量 | \(3 \times 200\ \text{Gbps}\) | 600 Gbpscomputed. 3 ISLs × 200 Gbps from §6.1計算值。§6.1 的 3 條 ISL × 200 Gbps |
| Users one satellite's access link can carry一顆衛星的接取鏈路可承載的用戶數 | \(96\ \text{Gbps} / 100\ \text{Mbps}\) | 960computed. Matches the paper's 960 users per satellite [39, §6.1]計算值。與論文每衛星 960 名用戶相符[39, §6.1] |
| Reduction factor for the Starlink-customer caseStarlink 客戶情境的縮減倍數 | \(6{,}793 / 1{,}763\) | 3.85 ≈ 3.9×computed. Counts from §6.1計算值。數量取自 §6.1 |
| Share of the 83.8 ms repair that is signaling round trip83.8 ms 修復時間中訊令往返所佔比例 | \(83.5 / 83.8\) | 99.6%computed. Both from Fig. 17d計算值。兩者皆取自 Fig. 17d |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| 1,763 satellites meet Starlink-customer demand (3.9× fewer than 6,793), 3,344 meet backbone demand (2.0×), 1,066 meet Latin America (6.4×) [39, §6.1].1,763 顆衛星滿足 Starlink 客戶需求（比 6,793 少 3.9 倍）、3,344 顆滿足骨幹需求（2.0 倍）、1,066 顆滿足拉丁美洲（6.4 倍）[39, §6.1]。 | Synthesis uses estimated demand and company-stated link capacities. Full numerical assumptions appear in the parameter table.合成採估計需求與公司聲明的鏈路容量，完整數值假設集中列於參數表。 |
| At 99% availability the counts are 1,391 (4.9×), 3,184 (2.1×) and 865 (7.9×). Diurnal-aware design saves a further 271 satellites (18.5%), or 333 (26%) at 99% [39, §6.1].在 99% 可用性下為 1,391（4.9 倍）、3,184（2.1 倍）與 865（7.9 倍）。考慮日夜變化的設計再省 271 顆（18.5%），在 99% 下省 333 顆（26%）[39, §6.1]。 | Relaxing availability below 100% removes the satellites that only serve rare demand peaks.把可用性放寬到 100% 以下，就去掉了只為罕見需求尖峰服務的衛星。 |
| Control signaling 1–3 orders of magnitude below TS-SDN. Repair of 1,000 random link failures 83.8 ms on average [39, §6.2, Fig. 17].控制訊令比 TS-SDN 低 1–3 個數量級。1,000 次隨機鏈路失效的修復平均 83.8 ms[39, §6.2, Fig. 17]。 | StarryNet emulation. Signaling dominates average repair time, making controller placement the main timing constraint.StarryNet 模擬平台中，平均修復時間主要由訊令占據，因此控制器位置是主要時間約束。 |
| Routing stretch 1.29 at p90 and 1.63 maximum. Local ISL failover adds 13.6–44.3 ms [39, §6.3, Fig. 19].路由延展比 p90 1.29、最大 1.63。局部 ISL 故障切換增加 13.6–44.3 ms[39, §6.3, Fig. 19]。 | Stretch compares with Starlink under the same orbital MPC. Reacquisition requires a separate failover-time budget.延展比以相同 MPC 控制的 Starlink 為基準，雷射重新擷取需另外計入故障切換預算。 |
| Synthesis takes 6.5–7.7 h per case. Gurobi remains unfinished after two months [39, §6.1].合成每情境 6.5–7.7 小時，Gurobi 在兩個月截止時仍持續求解。[39, §6.1]。 | The cutoff leaves the matching-pursuit optimality gap open, so the exact baseline remains incomplete.截止時間使匹配追蹤與最佳解的差距仍待確認。 |
6. Limits and open questions6. 限制與待研究的問題
- Estimated demand sets the synthesis scope. Customer maps and Cloudflare Radar supply capacity and geographic distribution. Satellite counts scale with those inputs, as the paper acknowledges in its limitations.估計需求界定了合成範圍。容量與地理分布取自客戶地圖及 Cloudflare Radar，衛星數隨這些輸入縮放，論文也在限制中說明此條件。
- Company specifications set the link budget. TinyLEO uses the stated peak rate, while the fact base distinguishes peak and sustained figures relayed by journalists. Range-dependent performance and an optical link budget require separate analysis.公司規格設定了鏈路預算。TinyLEO 採用聲明的峰值速率，事實庫則區分記者轉述的峰值與持續數值。距離相關效能及光學鏈路預算仍需另外分析。
- Acquisition time requires an explicit budget. Lifetime-aware matching reduces repointing frequency, but each new link still requires setup. Daily acquisitions per satellite would help determine whether the compiled topologies remain feasible under real terminal timing.擷取時間需要明確預算。考量壽命的配對可減少重新指向頻率，每條新鏈路仍需建立。量化每顆衛星每日擷取次數，有助於判定編譯拓樸在真實終端時間下的可行性。
- The delay scope covers the ISL mesh. Cell size bounds per-hop delay, while node processing and user-link reconfiguration need separate models. Read failover and stretch results within that scope.延遲範圍涵蓋 ISL 網格，每跳延遲以格元大小為界，節點處理與使用者鏈路重組則需要另外建模。故障切換與延展比結果應依此範圍解讀。
- Earth-repeat orbits constrain coverage. The sparse design depends on repeated ground tracks and the discrete altitude family in Table 1. Drag, maneuvers and observed daily shell changes require further assessment of repeat-track stability.地球重複軌道限制了覆蓋設計。稀疏配置依賴重複地面軌跡及 Table 1 的離散高度集合，阻力、機動及實測每日殼層變化對重複軌跡穩定性的影響，仍需進一步評估。
- Centralized control makes repair depend mainly on signaling distance. The measured repair budget is dominated by the controller round trip. The paper identifies centralization as a limitation.集中控制使修復主要取決於訊令距離，量得的修復預算由控制器往返時間主導，論文也將集中化列為限制。
- The baselines answer separate questions. Real Starlink positions support satellite-count comparisons, while the authors' MPC supplies the routing comparison. The timed-out Gurobi run leaves an open optimality gap for synthesis.各項基準回答的問題各有範圍。實際 Starlink 位置支持衛星數比較，作者的 MPC 則提供路由比較。Gurobi 在截止時仍持續求解，使合成結果與最佳解的差距仍待確認。
7. How it connects7. 與其他內容的關聯
- Section 5.2: the 3-ISL baseline and the "undisclosed topology" statement are quoted there whenever +Grid is introduced.第 5.2 節：3-ISL 基準線與「拓樸仍屬保密」的聲明在那裡每次介紹 +Grid 時都會被引述。
- Section 5.4: SRv6 anycast supplies Earth-fixed forwarding. Compare SHORT's hierarchy and Hypatia's four path changes per 200 s.第 5.4 節：SRv6 上的地理區段任播是地球固定路由構想的資料平面一半，與 SHORT 的階層以及 Hypatia 每 200 秒 4 次路徑變更並列。
- Section 5.7: the assumptions table records three 200 Gbps terminals, with setup and scheduled reconfiguration modeled separately.第 5.7 節：TinyLEO 是假設表中已經採用三個終端與 200 Gbps、卻仍省略建立延遲與 15 秒排程的那一列。
- Section 4.2: the per-laser-rate row of the parameter sheet cites TinyLEO's 200 Gbps beside Ali and Zhang's 100 Gbps per direction.第 4.2 節：參數表每雷射速率的那一列把 TinyLEO 的 200 Gbps 與 Ali 和 Zhang 的每方向 100 Gbps 並列引用。
- Ali & Zhang digest: one study observes real shells, while the other designs sparse ones. Both use three terminals.Ali 與 Zhang 導讀：兩篇論文都放棄完整對稱殼層，一篇靠觀測真實殼層、一篇靠設計稀疏殼層，而兩者都採用三個終端。
- SHORT digest and MobiCom 2023 digest supply routing and maneuver findings used in TinyLEO's MPC and anycast design. CoNEXT 2019 motifs and Handley 2018 provide earlier fixed-link baselines.SHORT 導讀與MobiCom 2023 導讀提供 TinyLEO 的 MPC 及任播設計所依據的路由與機動研究，CoNEXT 2019 motif與Handley 2018則提供較早的固定鏈路基準。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| >70% on ~5% | Share of users on share of land, the demand mismatch motivating sparsity [39, §1, §2.2] | 用戶比例對陸地比例，促成稀疏化的需求錯配[39, §1, §2.2] | |
| 64,800 / 4,050 | Candidate Earth-repeat orbits (423–1,873 km, 92.8–124.2 min) and demand cells [39, Tab. 1, §4.1] | 候選地球重複軌道（423–1,873 km、92.8–124.2 分）與需求格子數[39, Tab. 1, §4.1] | |
| 3 × 200 Gbps + 96 Gbps | ISLs and radio access per satellite, from Starlink public statements [39, §6.1] | 每顆衛星的 ISL 與電波接取，取自 Starlink 公開聲明[39, §6.1] | |
| 1,763 vs 6,793 | Satellites for Starlink-customer demand vs Starlink's five shells in Jan 2025, 3.9× [39, §6.1] | 滿足 Starlink 客戶需求的衛星數對比 2025 年 1 月 Starlink 五殼層，3.9 倍[39, §6.1] | |
| 3,344 / 1,066 | Backbone (2.0×) and Latin America (6.4×) cases [39, §6.1] | 骨幹（2.0 倍）與拉丁美洲（6.4 倍）情境[39, §6.1] | |
| 1,391 / 3,184 / 865 | The same three cases at 99% availability, 4.9× / 2.1× / 7.9× [39, §6.1] | 同三個情境在 99% 可用性下，4.9 / 2.1 / 7.9 倍[39, §6.1] | |
| 6.5–7.7 h | Synthesis time per case. Gurobi unfinished after two months [39, §6.1] | 每情境合成時間，Gurobi 於兩個月截止時仍持續求解[39, §6.1] | |
| 1–3 orders | Control signaling reduction versus TS-SDN [39, §6.2, Fig. 17] | 相對 TS-SDN 的控制訊令縮減[39, §6.2, Fig. 17] | |
| 83.8 ms | Average repair over 1,000 random link failures, 83.5 ms of it signaling RTT [39] | 1,000 次隨機鏈路失效的平均修復時間，其中 83.5 ms 為訊令往返[39] | |
| 1.29 / 1.63 | Routing stretch at p90 and maximum vs Starlink under the same MPC [39] | 相對同一 MPC 下 Starlink 的路由延展比，p90 與最大值[39] | |
| 13.6–44.3 ms | Local ISL failover reroute cost [39] | 局部 ISL 故障切換的重繞成本[39] | |
| 1,741 | Satellites emulated as containers on the StarryNet variant [39, §6.2–6.3] | 在 StarryNet 變體上以容器模擬的衛星數[39, §6.2–6.3] |