Digest · Simulation and emulation · read after Chapter 7
導讀 · 模擬與模擬平台 · 第 7 章後閱讀
Exploring the "Internet from Space" with Hypatia用 Hypatia 探索「來自太空的網際網路」
Hypatia adds packets, queues and transport behavior to a geometric constellation model. Its experiments explain reactions to path changes under assumed link settings. Applying the results to Starlink requires comparison with the deployed network and its scheduler.Hypatia 在幾何星座模型中加入封包、佇列及傳輸行為，以指定鏈路設定解釋傳輸協定對路徑變化的反應。套用至 Starlink 時，須另行核對實際部署網路及排程器。
If you read nothing else
只讀這一段也可以
Hypatia predates Starlink service and awaits validation against it. Izhikevich et al. compared it with measured RTTs and found sustained user-observed spikes missing. Their measurement-driven model was 1.8 times more accurate on average. [35, §6.1.2, Fig. 6]這個工具早於 Starlink 開始服務，對照它的驗證仍待完成，而當 Izhikevich et al.[35, §6.1.2, Fig. 6]拿它和實測的 Starlink RTT 比較時，發現它產生的 RTT 曲線缺少真實使用者看到的持續性尖峰，而他們以量測驅動的模型平均準確 1.8 倍。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Queue behavior | Hypatia combines an ns-3 packet simulator with a Cesium visualizer. FCC filing parameters define the constellation and four static +Grid laser neighbors per satellite. Shortest paths refresh every 100 ms, and packets traverse 100-packet queues. Researchers can therefore observe TCP responses to moving topology alongside propagation delay. |
| Round-trip time | The results expose network dynamics. On Kuiper's first shell, Manila–Dalian RTT ranges from 25 to 48 ms within 200 s. Starlink and Kuiper city pairs have a median four path changes per 200 s. A 16 ms shortening reorders packets enough to halve NewReno's congestion window with zero packet loss. [6, Figs. 3, 4, 8] |
| Throughput | Transport experiments use 10 Mbit/s. Simulating one second of a 9.2 Gbit/s TCP flow requires about 555 s of wall-clock time. Throughput results therefore demonstrate path-change effects, with laser capacity outside their evidence scope. [6, Fig. 2] |
| 封包層模擬模型 | Hypatia 是一個以 ns-3 為核心的封包層模擬器，附帶 Cesium 視覺化工具，它把向 FCC 申請的星座參數讀進來，替每顆衛星接上四條靜態的 +Grid 雷射鄰居，每 100 ms 重算最短路徑，並讓真實封包穿過 100 個封包長的佇列，使研究者第一次能看見 TCP 如何對移動中的拓樸做出反應，將分析從傳播延遲曲線擴展到封包行為。 |
| 往返時間 | 它的主要發現都關於動態：在 Kuiper 第一層殼上，Manila–Dalian 的 RTT 在 200 s 內於 25 到 48 ms 之間擺盪，Starlink 與 Kuiper 的城市對每 200 s 中位數會換四次路徑，而一次 16 ms 的路徑縮短所造成的封包重排，足以在完整保留封包的情況下把 NewReno 的壅塞視窗砍半[6, Figs. 3, 4, 8]。 |
| 吞吐量 | 所有傳輸層實驗都跑在 10 Mbit/s，因為模擬一條 9.2 Gbit/s TCP 流的一秒鐘要花約 555 s 的實際時間[6, Fig. 2]，所以這些吞吐量數字說的是路徑變動，而適用於該模型的路徑動態，雷射容量屬於另一個問題。 |
1. Where this paper sits1. 這篇論文的位置
Hypatia placed orbital geometry inside ns-3, making queues, retransmissions and congestion-window dynamics observable. It shipped as open source with a visualizer for time-varying paths.Hypatia 把幾何放進 ns-3 裡，使佇列、重傳與壅塞視窗的動態變得可觀察，並以開放原始碼釋出，附上能把時變路徑畫清楚的視覺化工具。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Model configuration | Hypatia belongs to the 2018–2021 modeling era in Chapter 7. Before customer terminals shipped, researchers derived laser-linked LEO network behavior from regulatory filings. |
| Propagation delay | The ETH Zürich group developed the preceding HotNets 2018 papers and CoNEXT 2019 motif paper. These proposed that long-distance ISLs could beat fiber and improve +Grid wiring. Their calculators moved satellites and summed propagation paths, with packet behavior outside their scope. |
| Evaluation conditions | OpenAlex records 221 citations as of 29 October 2026, including 47 in 2026. This makes Hypatia the most-cited platform in the field. This is exactly why its assumptions deserve the critical reading in Section 6. |
| 評估條件 | Hypatia 屬於第 7 章所描述的建模時代，也就是 2018 到 2021 年之間，研究社群從監管申請文件，直接量測仍待補充資料建構出雷射互連 LEO 網路圖像的那幾年，因為當時正值使用者終端出貨之前。 |
| 傳播延遲 | 同一個 ETH Zürich 團隊早已在兩篇 HotNets 2018 論文與 CoNEXT 2019 的 motif 論文中主張，星間雷射在長距離上能勝過光纖，而 +Grid 佈線還能再改進，但那些結果全都來自純傳播延遲的計算器，它們移動衛星、加總路徑長度，封包行為則位於分析範圍之外。 |
| 評估條件 | 這篇論文是該領域引用最多的平台，截至 2026 年 10 月 29 日在 OpenAlex 上有 221 次引用，其中 47 次來自 2026 年，正因如此，它的假設才值得第 6 節那樣的批判性閱讀。 |
2. The problem, measured2. 問題，用量測說話
The question the paper sets out to answer is what an end-to-end connection experiences while the satellites beneath it keep moving. The answer it gives is that both the path and its length change constantly. A Rio de Janeiro–St. These are the numbers that Chapter 5 uses to motivate geography-based routing. They are entirely a product of the +Grid geometry under a static elevation mask, because failures, saturation and scheduling lie outside the model.這篇論文要回答的問題是，當腳下的衛星持續移動時，一條端到端連線會經歷什麼，而它給的答案是，路徑本身與路徑長度都持續變動。這些就是第 5 章用來引出地理式路由的數字，而它們完全是 +Grid 幾何在靜態仰角遮罩下的產物，因為模型只涵蓋幾何與固定鏈路，故障、飽和及排程屬於額外機制。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Round-trip time | On Kuiper's K1 shell with 1 ms pings over 200 s, the Manila–Dalian RTT varies between 25 and 48 ms. The authors describe this as nearly a factor of two, while Istanbul–Nairobi moves between 47 and 70 ms [6, §4.1, Fig. 3]. |
| Path length and hops | Petersburg RTT jumps from 96 to 111 ms at the t = 32.9 s path change. Around t = 155–165 s, connectivity fails for roughly ten seconds. The graph becomes disconnected under the 30° elevation mask. On Starlink's first shell, Paris–Luanda RTT ranges from 85 to 117 ms as routes alternate between six-hop and nine-hop zigzags. [6, Fig. 13] |
| Round-trip time | More than 80% of city pairs keep maximum RTT below twice the geodesic RTT. Starlink and Kuiper have a median four path changes per 200 s, compared with two for Telesat. One in ten Kuiper or Starlink pairs experiences seven or more changes. More than a third of Starlink connections temporarily exceed their minimum by two or more hops. [6, §5.1–5.2, Figs. 6–8] |
| 往返時間 | 在 Kuiper 的 K1 殼上以 1 ms 間隔 ping 200 s，Manila–Dalian 的 RTT 在 25 到 48 ms 之間變化，作者形容接近兩倍，而 Istanbul–Nairobi 則在 47 到 70 ms 之間移動[6, §4.1, Fig. 3]。 |
| 路徑長度與跳數 | 同一層殼上的 Rio de Janeiro–St. Petersburg 連線在 t = 32.9 s 路徑改變時從 96 跳到 111 ms，接著在 t = 155–165 s 左右有約十秒處於連線中斷狀態，因為在 30° 仰角遮罩下可行路徑集合為空，而在 Starlink 第一層殼上，Paris–Luanda 連線在 85 到 117 ms 之間擺盪，因為路由在六跳與九跳的鋸齒路徑之間交替[6, Fig. 13]。 |
| 往返時間 | 放到整個星座來看，超過 80% 的城市對能把最大 RTT 壓在測地線 RTT 的兩倍以內，Starlink 與 Kuiper 的城市對每 200 s 中位數換四次路徑而 Telesat 是兩次，每十對 Kuiper 或 Starlink 城市中有一對會換七次以上，而超過三分之一的 Starlink 連線在某個時刻會比最少跳數多出兩跳以上[6, §5.1–5.2, Figs. 6–8]。 |
3. The design3. 設計
Hypatia is a pipeline rather than a monolith, and Figure A follows it from left to right. The tool converts filed Keplerian elements into two-line element sets in the WGS72 frame and cross-checks them against pyephem. The same input file reproduces orbital positions at any instant [6, §3]. The inter-satellite graph is a +Grid with four lasers per satellite.
Two reach same-orbit neighbors, and two reach one neighbor in each adjacent orbit. It stays fixed throughout a run. This occurs because the paper explicitly leaves dynamic ISL reconfiguration out [6, §3.1]. The Cesium visualiser renders the same state as an animated globe, which is how the paper's path-change figures were produced.
Hypatia 由多個依序執行的階段組成，圖 A 由左到右跟著它走。星座以申請文件中的克卜勒根數進入，工具把它們轉成 WGS72 座標系下的兩行根數 (TLE)，並與 pyephem 交叉比對，使任何拿到同一份檔案的人都能重現任一時刻的軌道位置[6, §3]。星間圖是 +Grid，每顆衛星四條雷射，依序連到同軌道的前後衛星以及相鄰兩個軌道面各一顆鄰居，且在整個模擬過程中保持固定，因為論文明確把動態 ISL 重組排除在外[6, §3.1]。Cesium 視覺化工具把同一份狀態畫成動態地球，論文中的路徑變更圖就是這樣做出來的。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Ground infrastructure | Ground stations occupy the 100 most populous cities, each with a static parabolic antenna. Minimum elevation is 25° for Starlink, 30° for Kuiper and 10° for Telesat. Handoffs preserve packets by construction. [6, §2.2, §3.2, §5.1] |
| Ground infrastructure | Every 100 ms a Floyd–Warshall computation in networkx produces fresh forwarding state for the whole network. The state enters the ns-3 module. Each satellite and ground station becomes a node with 100-packet link queues and a uniform user-selected link rate. [6, §3.3, §4.2] |
| Latency comparison | Propagation delay follows orbital geometry continuously. Forwarding state updates every 100 ms, which lets Hypatia expose packet behavior beyond a static latency calculation. Packets can follow temporarily longer paths between forwarding updates. At a 1,000 ms step, missed changes affect 6% of Kuiper pairs. At 100 ms, the fraction is 0.4%. [6, §5.3, Fig. 9] |
| 地面設施 | 地面站是全球人口最多的 100 個城市，每個城市有一面靜態的拋物面天線，連到最低仰角以上的衛星，Starlink 為 25°、Kuiper 為 30°、Telesat 為 10°，而換手在設計上完整保留封包[6, §2.2, §3.2, §5.1]。 |
| 星座配置 | 每 100 ms，networkx 中的 Floyd–Warshall 計算為整個網路產生新的轉送狀態，載入 ns-3 模組，其中每顆衛星與地面站都是一個節點，鏈路上有 100 個封包的佇列，鏈路速率則由使用者選定一個均一值[6, §3.3, §4.2]。 |
| 傳播延遲 | Hypatia 進一步刻畫封包行為的關鍵在於，傳播延遲隨幾何連續變化，而轉送狀態只在 100 ms 的步長上更新，所以一個封包可能走在一條片刻之前還是最短、如今已繞路的路徑上，而步長的選擇確實有影響，1,000 ms 的步長會漏掉 6% Kuiper 城市對的路徑變更，100 ms 則只漏 0.4%[6, §5.3, Fig. 9]。 |
Original figure note原始圖說出處
Drawn for this primer from Kassing et al., IMC 2020, §3–§3.4, §5.3 and Figs. 2, 9.4. Evaluation setup4. 實驗設定
Hypatia is pure simulation, and the table below records the assumptions a reader must carry into every number it produces.Hypatia 是純模擬，下表記錄了讀者在解讀它產生的每一個數字時都必須記得的假設。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Plane configuration | Table 1 fixes the filed constellations. Starlink's first shell uses the August 2019 configuration: 72 planes of 22 satellites at 550 km and 53°. It also includes four Starlink shells at 1,110–1,325 km, three Kuiper shells and two Telesat shells. [6, Tab. 1] |
| Link or data rate | The default inter-satellite rate remains unspecified. The scalability sweep varies link rate from 1 Mbit/s to 10 Gbit/s as a computational-cost parameter. It leaves physical laser-terminal capacity open. The RTT and TCP experiments all use 10 Mbit/s [6, §3.4, §4.2, §5.4] . |
| Round-trip time | The validation column is the one that matters most for a 2026 reader. The paper appeared in October 2020, before independent Starlink measurements. Izhikevich et al. later found sustained customer RTT spikes missing from Hypatia. Their HitchHiking model was 1.8 times more accurate on average. Hypatia averaged 7.6 ms error with its second-nearest-ground-station configuration. [35, §6.1.2, Fig. 6] |
| 軌道面配置 | 星座來自論文的 Table 1，全是凍結的申請文件，Starlink 第一層殼是 2019 年 8 月申請的幾何，72 個軌道面各 22 顆衛星，位於 550 km、53°，另有四層 1,110–1,325 km 的 Starlink 殼、Kuiper 的三層與 Telesat 的兩層[6, Tab. 1]。 |
| 鏈路或資料速率 | 預設星間鏈路速率須從釋出的程式碼核對，因為論文把鏈路速率當成可擴展性掃描中從 1 Mbit/s 到 10 Gbit/s 的成本參數，雷射終端的實際容量屬於另一個問題，而 RTT 與 TCP 實驗全部使用 10 Mbit/s[6, §3.4, §4.2, §5.4]。 |
| 往返時間 | 對 2026 年的讀者而言最要緊的是驗證那一欄，因為論文發表於 2020 年 10 月，當時正值 Starlink 獨立量測出現之前，而後來唯一的比較來自 Izhikevich et al.，他們發現 Hypatia 的預測缺少真實使用者經歷的持續性 RTT 尖峰，他們的 HitchHiking 模型平均準確 1.8 倍，而 Hypatia 的次近地面站設定平均誤差為 7.6 ms[35, §6.1.2, Fig. 6]。 |
| Assumption假設 | Hypatia [6]Hypatia[6] |
|---|---|
| Method方法 | ns-3 packet-level simulation. Cesium visualiser [6, §3]ns-3 封包層模擬。Cesium 視覺化[6, §3] |
| Constellation星座 | Starlink S1 72×22 at 550 km, 53° (1,584 satellites), plus S2–S5 at 1,110–1,325 km. Kuiper K1 34×34 at 630 km, 51.9°. Telesat T1 27×13 at 1,015 km, 98.98° [6, Tab. 1]Starlink S1 於 550 km、53° 的 72×22（1,584 顆），加上 1,110–1,325 km 的 S2–S5。Kuiper K1 於 630 km、51.9° 的 34×34。Telesat T1 於 1,015 km、98.98° 的 27×13[6, Tab. 1] |
| ISL topologyISL 拓樸 | +Grid, four static lasers per satellite, dynamic reconfiguration excluded [6, §3.1]+Grid，每顆衛星四條靜態雷射，拓樸固定[6, §3.1] |
| Default ISL rate預設 ISL 速率 | None stated. Uniform rate swept 1 Mbit/s–10 Gbit/s for cost only. All RTT and TCP experiments at 10 Mbit/s [6, §3.4, §4.2, §5.4]數值待查。均一速率僅為成本掃描在 1 Mbit/s–10 Gbit/s 之間變動。所有 RTT 與 TCP 實驗皆為 10 Mbit/s[6, §3.4, §4.2, §5.4] |
| Latency model延遲模型 | Propagation at c along the current geometry, continuous in time. 100-packet queues add queueing delay. Per-hop node delay lies outside the model. GS attachment by elevation mask 25°/30°/10° with loss-free handoffs [6, §3.1–3.2, §4.2]沿當下幾何以光速傳播，時間上連續。100 封包佇列產生佇列延遲。每跳節點延遲位於範圍之外。地面站以 25°/30°/10° 仰角遮罩連線，換手完整保留封包[6, §3.1–3.2, §4.2] |
| Laser setup delay雷射建鏈延遲 | Outside model scope. Permanent ISLs remain acquired throughout execution [6, §3.1, §7]位於模型範圍之外。ISL 為永久鏈路，因此整段執行都維持擷取狀態[6, §3.1, §7] |
| 15-s schedule15 秒排程 | Outside model scope. The paper predates its discovery [13]位於模型範圍之外。論文早於它被發現[13] |
| Validation vs live Starlink對照真實 Starlink 的驗證 | Outside the paper's scope. Izhikevich et al. [35, §6.1.2, Fig. 6] found it misses sustained spikes and is 1.8× less accurate than HitchHiking論文的驗證範圍之外。Izhikevich et al.[35, §6.1.2, Fig. 6]發現它漏掉持續性尖峰，準確度比 HitchHiking 低 1.8 倍 |
5. What was measured5. 量到了什麼
The claims below are grouped by what they say about the network. The right-hand column records the setup each one depends on. The transport results deserve particular care. This occurs because they are the ones most often quoted out of context. Those conclusions are mechanism statements and remain valid. In comparison, the absolute throughputs are properties of a 10 Mbit/s link that nobody believes describes a laser terminal.下列宣稱依它們對網路的陳述分組，右欄記錄每一項所依賴的設定。傳輸層結果需要尤其小心，因為它們最常被脫離脈絡引用。這些結論是關於機制的陳述，至今仍然成立，而絕對吞吐量則是一條 10 Mbit/s 鏈路的性質，其解讀範圍限於設定的鏈路容量。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | NewReno repeatedly fills and drains a 100-packet buffer. Around t = 140 s, a roughly 16 ms shortening delivers later packets before earlier packets. Duplicate acknowledgements signal apparent loss, halving the congestion window while actual packet loss stays at zero. [6, §4.2, Fig. 4] |
| Delay comparison | Vegas interprets a t ≈ 33–35 s path-change delay increase as queue growth. Its throughput collapses and stays low. The authors conclude that loss and delay can both mislead LEO congestion control. [6, Fig. 5] |
| 評估條件 | NewReno 實驗顯示傳送端反覆填滿又排空 100 封包的緩衝區，接著在 t = 140 s 左右，一次約 16 ms 的路徑縮短讓後送的封包先於先送的封包抵達，由此產生的重複確認被解讀為遺失，於是壅塞視窗在完整保留封包的情況下砍半[6, §4.2, Fig. 4]。 |
| 延遲比較 | 從延遲上升推斷壅塞的 Vegas，則把 t ≈ 33–35 s 的路徑變更延遲增加誤讀為佇列累積，吞吐量崩落且持續低迷[6, Fig. 5]，作者因此得出結論，在 LEO 網路中遺失與延遲都可能是糟糕的壅塞訊號。 |
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Manila–Dalian RTT 25–48 ms, Istanbul–Nairobi 47–70 ms over 200 s [6, Fig. 3]200 s 內 Manila–Dalian RTT 25–48 ms、Istanbul–Nairobi 47–70 ms[6, Fig. 3] | Kuiper K1 shell, 1 ms pings, 10 Mbit/s links, propagation plus queueing only. Scheduling and PoP detours outside model scopeKuiper K1 殼，1 ms ping，10 Mbit/s 鏈路，只有傳播與佇列。排程器與 PoP 繞行位於範圍之外 |
| Median 4 path changes per 200 s on Starlink and Kuiper, 2 on Telesat. 10% of pairs ≥7 [6, Fig. 8]Starlink 與 Kuiper 每 200 s 中位數 4 次路徑變更，Telesat 2 次。10% 的城市對 ≥7 次[6, Fig. 8] | 100 most populous cities as static ground stations. +Grid with four ISLs. Result depends on the elevation mask and the 72×22 filing以 100 個人口最多城市為靜態地面站。四條 ISL 的 +Grid。結果取決於仰角遮罩與 72×22 的申請幾何 |
| >80% of pairs keep max RTT under 2× geodesic. >1/3 of Starlink pairs carry ≥2 extra hops. >10% see >50% hop-count change [6, Figs. 6–7]>80% 的城市對最大 RTT 低於測地線的 2 倍。>1/3 的 Starlink 城市對多出 ≥2 跳。>10% 跳數變化 >50%[6, Figs. 6–7] | Telesat scores lowest latency because of its 10° mask, so the comparison mixes constellation geometry with ground-segment policyTelesat 延遲最低是因為 10° 的遮罩，所以這個比較把星座幾何與地面段政策混在一起 |
| NewReno cwnd halved by reordering after a ~16 ms path shortening, zero loss [6, Fig. 4]. Vegas collapses after a latency rise at ~33–35 s [6, Fig. 5]約 16 ms 路徑縮短後的重排讓 NewReno cwnd 砍半，封包完整保留[6, Fig. 4]。Vegas 在 ~33–35 s 延遲上升後崩落[6, Fig. 5] | 10 Mbit/s, 100-packet queues, ns-3 stacks. A mechanism result, with Starlink throughput prediction requiring separate validation10 Mbit/s、100 封包佇列、ns-3 協定堆疊。是特定設定下的機制結果，Starlink 吞吐量預測需要實測容量 |
| Unused bandwidth exceeds 1/3 of capacity 31% of the time on Rio–St. Petersburg against 11% in a static network [6, Fig. 10]Rio–St. Petersburg 有 31% 的時間閒置頻寬超過容量的 1/3，靜態網路為 11%[6, Fig. 10] | NewReno with cross traffic at 10 Mbit/s. The static baseline is the t = 0 topology frozen in place10 Mbit/s 下帶交叉流量的 NewReno。靜態基準是凍結在 t = 0 的拓樸 |
| 100 ms routing step misses 0.4% of path changes against 6% at 1,000 ms [6, Fig. 9]100 ms 路由步長漏掉 0.4% 的路徑變更，1,000 ms 漏掉 6%[6, Fig. 9] | Kuiper K1. This is the justification for the default step and a cost knob for anyone extending the toolKuiper K1。這是預設步長的依據，也是任何擴充此工具者的成本旋鈕 |
| Ground relays instead of ISLs raise RTT by typically around 5 ms [6, App. A, Figs. 16–19]以地面中繼取代 ISL 使 RTT 通常高出約 5 ms[6, App. A, Figs. 16–19] | Bent-pipe appendix with the same 100 cities as relays. Compare Handley's HotNets 2019 relay study以同樣 100 個城市作為中繼的彎管附錄。可與 Handley 的 HotNets 2019 中繼研究對照 |
6. Limits and open questions6. 限制與待解的問題
- The tool still requires validation against the deployed constellation. Hypatia appeared before Starlink sold a terminal. The single independent comparison found that it misses the sustained RTT spikes that at least 70% of customers see at least once a day. Thus, the measurement-driven HitchHiking model came out 1.8 times more accurate on average [35, §6.1.2, §7.2, Fig. 6]. Any statement that Hypatia was "validated against Starlink" is wrong.這個工具對所建模星座的驗證仍待完成。Hypatia 問世時 Starlink 正值終端銷售開始之前，而唯一一次獨立比較發現其預測缺少至少 70% 使用者每天至少遇到一次的持續性 RTT 尖峰，因此以量測驅動的 HitchHiking 模型平均準確 1.8 倍[35, §6.1.2, §7.2, Fig. 6]。任何說 Hypatia「已對照 Starlink 驗證」的陳述都是錯的。
- The fixed four-laser +Grid follows a 2019 filing. Current Starlink satellites carry three terminals, and SpaceX keeps the topology proprietary. Thus, TinyLEO's baseline is a three-ISL grid [39, §6.1] . Ali and Zhang find a three-ISL grid disconnected on real ephemerides. Hypatia path-change statistics therefore describe a design-space baseline requiring separate deployment validation. This is a preprint result. [34, Tab. 2]每顆衛星四條雷射、固定的 +Grid 是 2019 年申請文件的假設，而目前的 Starlink 衛星只有三個終端，SpaceX 也拓樸仍屬專有資訊，這正是 TinyLEO 以三條 ISL 的網格作為基準的原因[39, §6.1]。Ali 與 Zhang 指出在真實星曆上三條 ISL 的網格含有互相隔離的連通分量[34, Tab. 2]preprint，所以 Hypatia 的路徑變更統計描述的是設計空間的基準，實際在軌網路需要量測驗證。
- Laser acquisition is absent. Real terminals require 2–30 s to acquire a new partner. Bhattacharjee et al. derive tolerable setup delays of only 2.3–80.63 ms. Above these thresholds, acquisition costs exceed path-change gains. Hypatia requires dynamic-link modeling to address the topology questions in Chapter 2. The latter source is a preprint. [17, Tab. 1] [32]雷射擷取位於模型範圍之外。真實終端需要 2–30 s 才能和新夥伴建鏈[17, Tab. 1]，Bhattacharjee 等人推導出的可容忍建鏈延遲只有 2.3–80.63 ms，超過就會讓換路徑額外延遲超過換路徑的收益[32]preprint，所以一個所有雷射永遠亮著的模擬器需要加入擷取機制才能回答第 2 章提出的拓樸問題。
- The 15-second reconfiguration at seconds :12, :27, :42 and :57 [13] lies outside the model, and ground-station handoffs are lossless by construction. Hypatia therefore produces RTT variability entirely from geometry. User measurements largely reflect the scheduler and the drop-front queue described in Chapter 6.在第 :12、:27、:42、:57 秒發生的 15 秒重組[13]位於模型範圍之外，地面站換手在設計上也完整保留封包，所以 Hypatia 顯示的 RTT 變動完全來自幾何，而使用者量到的變動主要來自排程器與第 6 章的前端丟棄佇列。
- Transport experiments use 10 Mbit/s, four orders of magnitude below the company claim of 100 Gbps sustained per laser. Per-hop switching delay lies outside Hypatia. Other papers select 1 ms or 10 ms. Cite Hypatia throughput as an illustration of network dynamics. [32] [33]所有傳輸層結果都在 10 Mbit/s，而公司宣稱每條雷射可持續 100 Gbps company，兩者差了四個數量級，而且將每跳交換延遲列於範圍之外，其他論文則選用 1 ms 或 10 ms[32][33]。來自 Hypatia 的吞吐量數字是動態的示意，引用時應該這樣說明。
- The radio ground-link model excludes antenna gain and interference. The ISL model excludes Doppler. Weather, mobile terminals, multipath and congestion-control hooks require additional implementation. Filings may be optimistic, and high-goodput packet simulation runs slowly. The xeoverse comparison records 266.76 s for 60 simulated seconds at 660 satellites. At 1,584 satellites, the cost is 698 s. [6, §7] Its London–San Francisco path grows from 18 to 21 hops, raising mean latency by 67.6%. Hypatia fails to reproduce that change in this comparison. [80, §V-A]作者列出的限制界定了模型範圍。射頻地面鏈路模型將天線增益與干擾列於範圍之外，ISL 模型將多普勒效應列於範圍之外。天氣、移動終端、多路徑與壅塞控制掛鉤都需要額外實作。申請文件可能過於樂觀，而逐封包模擬在高吞吐量下執行較慢[6, §7]。後來的工具量化了這種慢，xeoverse 記錄 Hypatia 在 660 顆衛星時模擬 60 s 要 266.76 s，1,584 顆時要 698 s[80, §V-A]，同一比較也發現 Hypatia 重現狀態仍待確認 xeoverse 產生的 London–San Francisco 跳數 18 變 21 所造成的 67.6% 平均延遲上升。
- Exact reproduction requires the released code to establish ISL range cutoff, atmospheric limits and Walker phasing. RTT uncertainty intervals also remain open.精確重現需要查閱釋出的程式碼，以補齊 ISL 距離上限、大氣限制與 Walker 相位參數。RTT 數字的誤差範圍也仍待補充。
7. How it connects7. 與其他內容的關聯
- Section 7.3 builds on this packet-simulation study. It covers the 100 ms step, 10 Mbit/s experiments and 555 s cost per simulated second.第 7.3 節：該章對封包層模擬的討論建立在這篇論文上，包括 100 ms 步長、10 Mbit/s 實驗與每模擬秒 555 s。
- Section 7.4: StarryNet, Celestial and xeoverse each position themselves against Hypatia on speed or fidelity. The capability table there lists what each models that outside Hypatia's scope.第 7.4 節：StarryNet、Celestial 與 xeoverse 各自在速度或擬真度上以 Hypatia 為對照，該節的能力表列出各平台相對於 Hypatia 增加的模型功能。
- Section 7.5: the shared blind spot, since Hypatia inherits the unknown laser rate, topology and scheduler like every other platform.第 7.5 節：共同的盲點，因為 Hypatia 和其他平台一樣承襲了仍待實測的雷射速率、拓樸與排程器。
- Section 5.4: the four path changes per 200 s are the churn figure that motivates SHORT's geographic hierarchy and TinyLEO's SRv6 anycast.第 5.4 節：每 200 s 四次路徑變更就是促成 SHORT 地理階層與 TinyLEO SRv6 anycast 的變動數字。
- Section 6.1 uses the reordering result to show that loss and delay can mislead LEO congestion control. Lai et al. later confirmed this on the live network.第 6.1 節：重排砍半 cwnd 的結果是最早指出遺失與延遲在 LEO 中是糟糕壅塞訊號的陳述，後來由 Lai 等人在真實網路上證實。
- Sibling digests: StarryNet runs real stacks in containers and shows Hypatia underestimating live RTT. Celestial trades network fidelity for a microVM per satellite. xeoverse claims 2.9× Hypatia's speed in real time. Upstream topology ideas appear in the motif paper. Validation evidence appears in HitchHiking. The 15-s schedule is modeled in SaTCP / LeoEM.同組導讀：StarryNet 在容器中跑真實協定堆疊並顯示 Hypatia 低估真實 RTT。Celestial 以每顆衛星一個微型虛擬機換取網路擬真度。xeoverse 宣稱以即時速度達到 Hypatia 的 2.9 倍。上游的拓樸想法在motif 論文，驗證落差在 HitchHiking，唯一內建 15 秒排程的模擬平台在 SaTCP / LeoEM。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 72×22, 550 km | Starlink S1 as filed in August 2019, 1,584 satellites at 53° [6, Tab. 1] | 2019 年 8 月申請的 Starlink S1，53° 的 1,584 顆衛星[6, Tab. 1] | |
| 4 | Static +Grid lasers per satellite. Starlink now flies 3 terminals with an undisclosed topology | 每顆衛星的靜態 +Grid 雷射數。Starlink 現在是 3 個終端且拓樸仍屬專有資訊 | |
| 100 ms | Forwarding-state recomputation interval. Misses 0.4% of path changes against 6% at 1,000 ms [6, Fig. 9] | 轉送狀態重算間隔。漏掉 0.4% 的路徑變更，1,000 ms 則漏 6%[6, Fig. 9] | |
| 10 Mbit/s | Link rate in every RTT and TCP experiment. The 1 Mbit/s–10 Gbit/s sweep measures simulation cost only | 所有 RTT 與 TCP 實驗的鏈路速率。1 Mbit/s–10 Gbit/s 的掃描只量模擬成本 | |
| 25–48 ms | Manila–Dalian RTT range over 200 s on Kuiper K1 [6, Fig. 3] | Kuiper K1 上 Manila–Dalian 200 s 內的 RTT 範圍[6, Fig. 3] | |
| 4 / 200 s | Median path changes per connection on Starlink and Kuiper. 2 on Telesat [6, Fig. 8] | Starlink 與 Kuiper 每條連線的路徑變更中位數。Telesat 為 2[6, Fig. 8] | |
| ~16 ms | Path shortening whose reordering halved NewReno's cwnd with zero loss [6, Fig. 4] | 其重排在完整保留封包時砍半 NewReno cwnd 的路徑縮短量[6, Fig. 4] | |
| ≈555 s | Wall-clock time to simulate 1 s of a 9.2 Gbit/s TCP flow. 269 s for 13.8 Gbit/s UDP [6, Fig. 2] | 模擬 9.2 Gbit/s TCP 流 1 s 所需的實際時間。13.8 Gbit/s UDP 為 269 s[6, Fig. 2] | |
| 1.8× | HitchHiking's accuracy advantage over Hypatia on measured Starlink RTTs [35, Fig. 6] | HitchHiking 在實測 Starlink RTT 上相對 Hypatia 的準確度優勢[35, Fig. 6] | |
| 0 | Laser setup delay, 15-s schedule events and live-Starlink validations in the paper | 論文中的雷射建鏈延遲、15 秒排程事件與對照真實 Starlink 的驗證次數 |