Gearing up for the 21st Century Space Race為 21 世紀太空競賽做準備

ACM HotNets 2018Bhattacherjee, Aqeel, Bozkurt, Aguirre, Chandrasekaran, Godfrey, Laughlin, Maggs, Singla · paper論文 peer-reviewed

This digest explains the Frankfurt–Washington latency comparison, polar +Grid link shutdowns, capacity and cost estimates, and the later work that develops these research questions.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。

≈ 8 min · 5-question quiz≈ 8 分鐘 · 5 題測驗Chapter 5 ↗第 5 章 ↗Jump to the quiz ↓跳到測驗 ↓

If you read nothing else

只讀這一段也可以

The paper connects topology, routing, transport and economics through a geometric LEO model. Its quantitative evidence belongs beside the corresponding assumptions.

本文透過幾何 LEO 模型連結拓樸、路由、傳輸與經濟學，量化證據須連同相應假設。

Research agenda and quantitative assumptions
#	Reported values and conditions
1	This ETH Zürich and collaborators’ HotNets paper sets a research agenda for LEO topology, routing, congestion control and ISP economics.
2	Frankfurt–Washington latency is 46.4 ms over the contemporary Internet, 35.8 ms over Hibernia Express, 32.6 ms over great-circle fiber and 21.7 ms in vacuum.
3	Dense LEO therefore projects about 35% below Internet latency and 16% below Hibernia ^{[68, §3.3]}.
4	Simulations use uniformly spaced polar orbits at 1,160 km, four lasers per satellite and polar cross-link shutdowns.
5	The size sweep is 102–502 satellites, compared with the filing’s 4,425 ^{[68, §3.1–3.2]}.
6	Rough capacity estimates multiply about 12,000 satellites by 20 Gbps to obtain 240 Tbps, against 295 Tbps of fiber.
7	Estimated transit cost is $0.06/GB against terrestrial $0.003–0.03 ^{[68, §2]}.
8	These are propagation calculations on synthetic geometry and rough economic assumptions.

研究議程與數值假設
#	數值、條件與證據
1	ETH Zürich 與合作者這篇 HotNets 論文提出 LEO 拓樸、路由、壅塞控制與 ISP 經濟學的研究方向。
2	法蘭克福–華盛頓延遲為當時網際網路 46.4 ms、Hibernia Express 35.8 ms、大圓光纖 32.6 ms、真空 21.7 ms，因此稠密 LEO 推估比網際網路低約 35%、Hibernia 低約 16% ^{[68, §3.3]}。
3	模擬採 1,160 km 均勻極軌道，每星四雷射，極區關閉跨面鏈路，規模掃描為 102–502 顆，申報則為 4,425 顆 ^{[68, §3.1–3.2]}。
4	粗估容量以約 12,000 顆乘 20 Gbps 得 240 Tbps，對照光纖 295 Tbps，中轉成本為 0.06 美元/GB，地面為 0.003–0.03 ^{[68, §2]}。
5	這些結果採合成幾何的傳播計算與粗略經濟假設。

1. Where this paper sits1. 這篇論文的位置

The paper develops a community agenda beyond a single constellation’s latency comparison. Its group later produced CoNEXT 2019 motifs, IMC 2020 Hypatia and the HotNets 2020 orbital-computing study. Chapter 5 follows its sequence from wiring through moving-route behavior and congestion control to capacity economics. Section 5.1 pairs its Frankfurt–Washington comparison with Handley’s New York–London work. Chapter 3 later contrasts the whole-constellation estimate with measurement-era uncertainty in per-satellite capacity.

本文把單一星座的延遲比較擴展為研究議程，作者群後續提出 CoNEXT 2019 motif、IMC 2020 Hypatia 與 HotNets 2020 在軌運算研究。第 5 章沿其順序，從接線、移動路由、壅塞控制談到容量經濟學。第 5.1 節將法蘭克福–華盛頓比較與 Handley 的紐約–倫敦工作並列，第 3 章則以量測時代的每衛星容量資料缺口對照整星座估計。

2. The problem, measured2. 問題，用量測說話

The latency ladder separates medium speed, cable detour and Internet effects. Its distance conversion also exposes a source convention requiring clarification.

延遲梯分開介質速度、纜線繞行與網際網路效應，距離換算也揭露需要來源釐清的口徑。

Latency ladder and convention check
#	Reported values and conditions
1	The Frankfurt–Washington ladder separates propagation media, cable detour and Internet routing effects.
2	Its four values are 46.4 ms Internet, 35.8 ms Hibernia, 32.6 ms great-circle fiber and 21.7 ms vacuum c-latency ^{[68, §3.3]}.
3	Relative to vacuum, the ratios are 2.14, 1.65 and 1.50 respectively.
4	The last ratio is the assumed glass index.
5	The original narrative associates 21.7 ms with about 6,500 km, beyond Handley’s 3,000 km crossover.
6	The calculation table separately treats 21.7 ms as RTT and gives about 3,250 km one way, exposing a convention that needs source clarification.
7	A satellite path replaces fiber propagation with vacuum and changes route detour, while adding ascent and grid distance.
8	The two percentage gains both imply about 30 ms, providing an internal consistency check.

延遲梯與口徑檢查
#	數值、條件與證據
1	法蘭克福–華盛頓延遲梯分開傳播介質、纜線繞行與網際網路路由效果，數字各為 46.4、35.8、32.6 與 21.7 ms ^{[68, §3.3]}。
2	相對真空，比值依序為 2.14、1.65、1.50，最後一項為玻璃折射率假設。
3	原正文把 21.7 ms 連至約 6,500 km、超過 Handley 3,000 km 交叉點，計算表則以它作 RTT、得到約 3,250 km 單程，顯示口徑需向來源釐清。
4	衛星以真空取代光纖傳播並改變繞路，同時增加上下行與網格距離。
5	兩項百分比增益都推得約 30 ms，可作內部一致性檢查。

Question	問題	Formula
Each rung of the ladder as a multiple of the c-latency延遲梯每一階相對 c 延遲的倍數	$ r=\tfrac{t}{t_c},\ t_c=21.7\ \text{ms} $	fibre 1.50 · Hibernia 1.65 · Internet 2.14computed from 32.6, 35.8 and 46.4 ms ^{[68, §3.3]}, 1.50 is the implied refractive index由 32.6、35.8 與 46.4 ms 計算^{[68, §3.3]}，1.50 是隱含的折射率
Great-circle distance implied by the c-latencyc 延遲隱含的大圓距離	$ L=\tfrac{t_c}{2}\,c $	≈ 3,250 km one way, ≈ 6,500 km round-trip pathcomputed, c = 299,792,458 m/s, 21.7 ms is a round-trip figure, so the one-way great-circle distance is ≈ 3,250 km計算值，c = 299,792,458 m/s，21.7 ms 為往返值，所以單程大圓距離約 3,250 公里
Dense-LEO RTT implied by the two percentage claims兩個百分比宣稱隱含的密集 LEO RTT	$ 0.65\times46.4,\ 0.84\times35.8 $	30.2 ms and 30.1 mscomputed consistency check of "≈35% below the Internet" and "≈16% below Hibernia" ^{[68, §3.3]}, the absolute value remains unspecified對「約低於網際網路 35%」與「約低於 Hibernia 16%」^{[68, §3.3]}的計算一致性檢查，絕對值仍待列出
Aggregate capacity from the per-satellite assumption由每衛星假設得到的總容量	$ C=N\times c_{sat} $	12,000 × 20 Gbps = 240 Tbpsthe paper's own arithmetic ^{[68, §2]}, 20 Gbps per satellite is a model assumption, compared with ≈295 Tbps of fiber論文自己的算術^{[68, §2]}，每顆 20 Gbps 為模型假設，對照約 295 Tbps 光纖

3. The design3. 設計

Polar geometry forces a trade between short cross-plane links and link lifetime. The density sweep shows how that topology changes routes and delay.

極軌幾何形成短跨面鏈路與壽命的取捨，密度掃描則顯示拓樸如何改變路由與延遲。

Topology and density-sweep results
#	Reported values and conditions
1	The paper proposes research directions grounded in geometric simulation.
2	Four lasers connect fore, aft and adjacent-plane neighbors.
3	Polar convergence shortens cross-plane distance and raises relative speed, so those links shut down near the poles, leaving isolated rings ^{[68, §3.1]}.
4	Topology stays static within each simulated minute, exposing the trade between short links and long lifetimes later addressed by motifs.
5	In a 302-satellite polar model, shortest paths nearly always beat great-circle fiber with a median twelve hops ^{[68, Fig. 3]}.
6	Frankfurt–São Paulo median RTT falls from 121 ms at 102 satellites to 98 ms at 502 ^{[68, §3.2]}.
7	With 252 satellites, one pair varies about 5 ms around a 32 ms median ^{[68, Fig. 7]}.
8	The authors use this variation to motivate congestion-control work.
9	Their economics combines the rough 240 Tbps estimate with manufacturing and launch assumptions to obtain $0.06/GB.

拓樸與密度掃描結果
#	數值、條件與證據
1	本文依幾何模擬提出研究方向。
2	四具雷射連接前後與鄰面衛星，極區匯聚使跨面距離縮短、相對速度上升，因此關閉跨面鏈路，留下各自軌道環 ^{[68, §3.1]}。
3	每個模擬分鐘內拓樸固定，揭露後續 motif 處理的短鏈路與長壽命取捨。
4	302 顆極軌模型的最短路徑幾乎總勝大圓光纖，中位十二跳（Fig. 3）。
5	法蘭克福–聖保羅 RTT 中位數從 102 顆的 121 ms 降至 502 顆的 98 ms ^{[68, §3.2]}，252 顆時某城市對在 32 ms 中位附近變化約 5 ms（Fig. 7）。
6	作者以此動機討論壅塞控制，並結合粗估 240 Tbps、製造與發射假設，得到 0.06 美元/GB 的經濟估計。

Figure A. Left: the paper’s topology uses uniformly spaced polar planes with fore-and-aft links throughout. Cross-orbit links form the +Grid. These links switch off within the polar cap, where neighboring planes converge and have high relative speeds. Right, the Frankfurt–Washington ladder that anchors the latency argument, from the 2018 Internet down to the vacuum-speed floor.圖 A。左：論文的拓樸——均勻分布的極軌道面在各處都有前後鏈路，而構成 +Grid 的跨軌道鏈路在相鄰軌道面匯聚且相對速度快的極冠內關閉。右：支撐延遲論證的法蘭克福–華盛頓梯，從 2018 年的網際網路一路降到真空光速下限。Sources來源: ^{[68, §3.1, §3.3]}

Original figure note原始圖說出處

Drawn for this primer from Bhattacherjee et al., HotNets 2018, §3.1 and §3.3.

4. Evaluation setup4. 實驗設定

The evaluation isolates geometric propagation and varies synthetic constellation density. Capacity and cost estimates add separate aggregate-rate and price assumptions.

評估分離幾何傳播並改變合成星座密度，容量與成本則加入各自的總速率及價格假設。

Simulation configuration and economic inputs
#	Reported values and conditions
1	Evaluation uses the authors’ geometric simulation.
2	The filing supplies context, while experiments sweep synthetic polar constellations of 102–502 satellites at 1,160 km, then extrapolate dense-LEO behavior ^{[68, §3.1–3.3]}.
3	Four lasers per satellite form a +Grid away from polar shutdowns.
4	Topology updates each simulated minute.
5	Delay includes geometric propagation, with queueing, node processing and acquisition costs set to zero.
6	Fiber references are real Hibernia and ideal great-circle fiber at index 1.5.
7	Capacity assumes about 12,000 satellites at 20 Gbps each, against an assembled 295 Tbps terrestrial estimate.
8	Cost combines those values with rough launch and manufacturing assumptions.

模擬設定與經濟輸入
#	數值、條件與證據
1	評估使用作者的幾何模擬，申報提供背景，實驗則掃描 1,160 km、102–502 顆合成極軌星座，再外推稠密 LEO 行為 ^{[68, §3.1–3.3]}。
2	每星四具雷射在極區關閉區以外形成 +Grid，拓樸每模擬分鐘更新。
3	延遲計幾何傳播，佇列、節點處理與擷取成本設零。
4	光纖參考為實際 Hibernia 與折射率 1.5 的理想大圓光纖。
5	容量假設約 12,000 顆各 20 Gbps，對照彙整的 295 Tbps 地面估計，成本再結合粗略發射與製造假設。

5. What was measured5. 量到了什麼

Claim宣稱	Setup and caveat設定與附註
Frankfurt–Washington: 46.4 ms Internet, 35.8 ms Hibernia Express, 32.6 ms great-circle fibre, 21.7 ms c-latency ^{[68, §3.3]}.法蘭克福–華盛頓：網際網路 46.4 ms、Hibernia Express 35.8 ms、大圓光纖 32.6 ms、c 延遲 21.7 ms^{[68, §3.3]}。	The Internet and Hibernia values are 2018 measurements or published figures, the other two are computed bounds. The ladder is for one city pair.網際網路與 Hibernia 的值是 2018 年的量測或公開數字，另兩個是計算的下限。這條梯子只針對一對城市。
A dense LEO network sits ≈ 35% below today's Internet and ≈ 16% below Hibernia ^{[68, §3.3]}.密集 LEO 網路約低於今天的網際網路 35%、低於 Hibernia 約 16%^{[68, §3.3]}。	Propagation-only extrapolation from the polar sweep, both percentages imply ≈ 30 ms, which is below the measured ~40 ms Starlink bent-pipe floor.由極軌道掃描外推的純傳播結果，兩個百分比都隱含約 30 ms，低於實測約 40 ms 的 Starlink 彎管下限。
With 302 polar satellites, paths almost always beat great-circle fibre, with a median of 12 hops ^{[68, Fig. 3]}. Frankfurt–São Paulo median RTT is 121 ms at 102 satellites and 98 ms at 502 ^{[68, §3.2]}.302 顆極軌道衛星幾乎總是勝過大圓光纖，中位數 12 跳（圖 3），法蘭克福–聖保羅中位數在 102 顆時 121 ms、502 顆時 98 ms^{[68, §3.2]}。	The synthetic polar constellations use 1,160 km altitude. Their geometry produces high hop counts, while per-hop delay is zero.合成極軌星座採 1,160 km 高度，幾何使跳數較高，每跳延遲則設為零。
With 252 satellites one pair's latency varies ≈ 5 ms around a 32 ms median ^{[68, Fig. 7]}.在 252 顆衛星下某一對城市的延遲在 32 ms 中位數附近變動約 5 ms（圖 7）。	The variability argument for congestion control, later quantified by Hypatia as 25–48 ms swings on Manila–Dalian.壅塞控制的變動性論證，後來由 Hypatia 量化為馬尼拉–大連 25–48 ms 的擺動。
Aggregate capacity is ≈ 240 Tbps from ≈ 12,000 satellites at 20 Gbps each, against ≈ 295 Tbps of fibre. Cost is ≈ $0.06 per GB, against $0.003–0.03 for transit ^{[68, §2]}.約 12,000 顆衛星各 20 Gbps 合計約 240 Tbps，對比約 295 Tbps 的光纖，每 GB 約 0.06 美元，對比中轉的 0.003–0.03 美元^{[68, §2]}。	These estimates give orders of magnitude and assume the per-satellite rate. Later sources report contested Gen1 capacity estimates from 6.5 to 24 Gbps (fact base Part C, item 22).數量級估計，每衛星速率是假設，而 Gen1 每衛星容量在後來的來源中仍在 6.5 到 24 Gbps 之間有爭議（事實庫 C 部第 22 項）。

6. Limits and open questions6. 限制與待解的問題

The synthetic polar model has a specific scope. Its 102–502 satellites at 1,160 km follow polar geometry, while Starlink’s 53° shell retains inter-plane distance near 0.60 of the equatorial value. Polar shutdown behavior therefore informs Iridium-style and Telesat-polar designs (fact base Chapter 2).合成極軌模型有其適用範圍：1,160 km、102–502 顆採極軌幾何，Starlink 的 53° 殼層則保留約赤道值 0.60 的面間距離。因此極區關閉機制適用於 Iridium 式與 Telesat 極軌設計（事實庫第 2 章）。
Propagation delay describes geometric paths with zer o queue, node and acquisition costs. Ideal fiber at index 1.5 provides a lower bound. The implied ≈30 ms result measures a separate quantity from measured Starlink delays. These include a bent-pipe floor around 40 ms and idle medians of 25–50 ms (Part C, item 58).傳播延遲描述幾何路徑，佇列、節點與擷取成本為零，折射率 1.5 的理想光纖提供下限。推得約 30 ms 的結果與實測 Starlink 約 40 ms 彎管下限、25–50 ms 閒置中位數，各有量測範圍（C 部第 58 項）。
Capacity and cost depend on assumed per-satellite output and terrestrial aggregate capacity. The authors label estimates rough, and later sources still span a wide capacity range (Part C, items 21–22).容量與成本依賴每衛星輸出及地面總容量假設，作者將估計標為粗略，後續來源的容量範圍也仍很廣（C 部第 21–22 項）。
Topology updates once per minute, with four static ISLs and zero acquisition cost. Current satellites carry three terminals in an undisclosed arrangement (Part C, items 13, 55 and 56). The lifetime issue motivates later motif work.拓樸每分鐘更新，採四條靜態 ISL 與零擷取成本。目前衛星則有三具終端，排列仍待公開（C 部第 13、55、56 項），鏈路壽命問題促成後續 motif 研究。
Congestion-control discussion is a mechanism argument based on RTT variation. Transport testing arrives later in Hypatia, and measured 15-second scheduling adds a disturbance outside this model (Part C, item 57).壅塞控制討論是依 RTT 變化提出的機制論證，Hypatia 後續加入傳輸測試，實測15 秒排程則補上本模型範圍外的干擾（C 部第 57 項）。
Interpret each comparison under its definition. Dense-LEO delay uses percentages, city-pair medians use small synthetic constellations, and capacity compares downlink output with backbone capacity under separate definitions.各項比較須依其定義解讀：稠密 LEO 延遲採百分比，城市中位數採小型合成星座，容量則以各自定義對照下行輸出與骨幹容量。

7. How it connects7. 與其他內容的關聯

Section 5.1 uses the Frankfurt–Washington ladder as the second pillar of the fibre comparison and separates it from the measured floor.第 5.1 節以法蘭克福–華盛頓梯作為光纖比較的第二根支柱，並將它與實測下限分開。
Section 5.2 starts from this paper's four-link +Grid and its per-minute static topology before introducing motifs.第 5.2 節從這篇論文的四鏈路 +Grid 與每分鐘靜態拓樸出發，再引入 motif。
Section 5.7 records the 1,160 km polar stand-in and the four-ISL assumption among the superseded parameters.第 5.7 節把 1,160 公里極軌道替代品與四 ISL 假設記入已被取代的參數。
Section 2.5 explains the polar seam and the cos(latitude) shrinkage that make the cross-orbit switch-off necessary, and Section 3.2 takes up the contested per-satellite capacity.第 2.5 節說明極區縫隙與 cos(緯度) 收縮如何使跨軌道關閉成為必要，第 3.2 節接手爭議中的每衛星容量。
Related pages: Handley 2018 covers routing in the same session. CoNEXT 2019, Hypatia and In-orbit Computing extend this group’s topology, transport and economics work. CCR 2020 develops the ISP economics question. Ali & Zhang 2026 and TinyLEO revisit topology with three terminals and real data.相關導讀：Handley 2018 是同一場次的伴隨路由論文，CoNEXT 2019、Hypatia 與在軌運算是同一群人對拓樸、傳輸與經濟學項目的後續，CCR 2020 把 ISP 經濟學項目推得最遠，Ali 與 Zhang 2026 和 TinyLEO 在真實資料上以三個終端重做拓樸。

Key numbers關鍵數字

Number	數字	Meaning (setup)
46.4 / 35.8 / 32.6 / 21.7 ms	Frankfurt–Washington RTT: Internet 2018 / Hibernia Express / great-circle fibre / vacuum c-latency ^{[68, §3.3]}.	法蘭克福–華盛頓 RTT：2018 年網際網路／Hibernia Express／大圓光纖／真空 c 延遲^{[68, §3.3]}。
≈ 35% / ≈ 16%	Dense-LEO latency below the Internet / below Hibernia ^{[68, §3.3]}, both imply ≈ 30 ms [computed].	密集 LEO 延遲低於網際網路／低於 Hibernia 的幅度^{[68, §3.3]}，兩者都隱含約 30 ms［計算值］。
4,425 at ≈ 1,160 km	The filed constellation named in the paper, experiments instead use 102–502 satellites in uniform polar orbits ^{[68, §3.1]}.	論文點名的申請星座，實驗改用 102–502 顆均勻極軌道衛星^{[68, §3.1]}。
4	Laser ISLs per satellite, two in-orbit and two cross-orbit, the latter off near the poles, topology static per simulated minute ^{[68, §3.1]}.	每顆衛星的雷射 ISL 數，兩條同軌、兩條跨軌，後者在極區關閉，拓樸每模擬分鐘靜態^{[68, §3.1]}。
12 hops	Median path on a 302-satellite polar constellation that almost always beats great-circle fibre ^{[68, Fig. 3]}.	302 顆極軌道衛星星座上幾乎總能勝過大圓光纖的路徑中位跳數（圖 3）。
121 → 98 ms	Frankfurt–São Paulo median RTT from 102 to 502 satellites ^{[68, §3.2]}.	法蘭克福–聖保羅 RTT 中位數從 102 顆到 502 顆衛星的變化^{[68, §3.2]}。
≈ 5 ms around 32 ms	Latency wander of one pair at 252 satellites ^{[68, Fig. 7]}, the paper's motivation for new congestion control.	252 顆衛星下某一對城市的延遲游走（圖 7），論文提出新壅塞控制的動機。
≈ 240 vs ≈ 295 Tbps	Aggregate LEO capacity at 20 Gbps × 12,000 satellites versus estimated terrestrial fibre ^{[68, §2]}, model assumptions.	20 Gbps × 12,000 顆的 LEO 總容量對比估計的地面光纖^{[68, §2]}，屬模型假設。
≈ $0.06/GB	Estimated cost of LEO transit against $0.003–0.03 terrestrial ^{[68, §2]}, rough by the authors' own description.	LEO 中轉的估計成本，對比地面 0.003–0.03 美元^{[68, §2]}，作者自述為粗略。
~ 40 ms	Measured Starlink bent-pipe RTT floor, for contrast, external comparison (fact base Part C, item 58).	實測 Starlink 彎管 RTT 下限，作外部對照（事實庫 C 部第 58 項）。

Question	問題	Formula
Each rung of the ladder as a multiple of the c-latency延遲梯每一階相對 c 延遲的倍數	\( r=\tfrac{t}{t_c},\ t_c=21.7\ \text{ms} \)	fibre 1.50 · Hibernia 1.65 · Internet 2.14computed from 32.6, 35.8 and 46.4 ms ^{[68, §3.3]}, 1.50 is the implied refractive index由 32.6、35.8 與 46.4 ms 計算^{[68, §3.3]}，1.50 是隱含的折射率
Great-circle distance implied by the c-latencyc 延遲隱含的大圓距離	\( L=\tfrac{t_c}{2}\,c \)	≈ 3,250 km one way, ≈ 6,500 km round-trip pathcomputed, c = 299,792,458 m/s, 21.7 ms is a round-trip figure, so the one-way great-circle distance is ≈ 3,250 km計算值，c = 299,792,458 m/s，21.7 ms 為往返值，所以單程大圓距離約 3,250 公里
Dense-LEO RTT implied by the two percentage claims兩個百分比宣稱隱含的密集 LEO RTT	\( 0.65\times46.4,\ 0.84\times35.8 \)	30.2 ms and 30.1 mscomputed consistency check of "≈35% below the Internet" and "≈16% below Hibernia" ^{[68, §3.3]}, the absolute value remains unspecified對「約低於網際網路 35%」與「約低於 Hibernia 16%」^{[68, §3.3]}的計算一致性檢查，絕對值仍待列出
Aggregate capacity from the per-satellite assumption由每衛星假設得到的總容量	\( C=N\times c_{sat} \)	12,000 × 20 Gbps = 240 Tbpsthe paper's own arithmetic ^{[68, §2]}, 20 Gbps per satellite is a model assumption, compared with ≈295 Tbps of fiber論文自己的算術^{[68, §2]}，每顆 20 Gbps 為模型假設，對照約 295 Tbps 光纖