Delay is Not an Option: Low Latency Routing in Space延遲不是選項：太空中的低延遲路由

ACM HotNets 2018Mark Handley (University College London) · paper論文 peer-reviewed

This digest traces the vacuum-speed and long-distance latency claims to Handley’s five-laser simulation. It explains the lattice, geometric detours and historical constellation assumptions.本導讀追溯真空速度與長距延遲論斷的五雷射模擬依據，說明格狀接線、幾何繞路與歷史星座假設。

≈ 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

Handley quantifies when vacuum propagation can outweigh orbital detours. The conclusion follows a historical filing and an inferred laser layout.

Handley 量化真空傳播何時可補償軌道繞路，結論依據歷史申報與推論的雷射布局。

Founding latency argument and historical inputs
#	Reported values and conditions
1	Handley argues that vacuum light travels about 47% faster than light in glass. Long satellite paths can beat terrestrial fiber beyond roughly 3,000 km ^{[3, §4]}.
2	His custom simulator uses the 2016 FCC filing: 1,600 satellites, 32 planes of 50, 1,150 km altitude and 53° inclination.
3	He infers five laser links and continually recomputes shortest paths.
4	New York–London RTT is about 57–66 ms, against 55 ms great-circle fiber and 76 ms measured Internet.
5	London–Johannesburg approaches half the best Internet path’s 182 ms ^{[3, Figs. 7, 9]}.
6	The model includes geometric propagation with queueing and acquisition set to zero, and capacity outside its scope.
7	Later deployment moved to 550 km and three terminals, making the paper a historical mechanism argument.

奠基延遲論證與歷史輸入
#	數值、條件與證據
1	Handley 指出真空光比玻璃快約 47%，長距衛星路徑因此可在約 3,000 km 之外勝過地面光纖 ^{[3, §4]}。
2	自製模擬器採 2016 FCC 申報：1,600 顆、32 面各 50 顆、1,150 km 與 53°，推得五條雷射並持續重算最短路徑。
3	紐約–倫敦 RTT 約 57–66 ms，對照 55 ms 大圓光纖與 76 ms 實測網際網路，倫敦–約翰尼斯堡接近最佳網際網路 182 ms 的一半（Figs. 7、9）。
4	模型計幾何傳播，佇列、擷取設零，容量在範圍之外。
5	後續部署改為 550 km、三終端，因此本文是歷史機制論證。

1. Where this paper sits1. 這篇論文的位置

Public FCC orbital parameters were the basis for Handley’s 2018 routing study. He inferred a plausible laser layout and tested whether ascent, inter-satellite hops and descent could beat ground fiber. The long-distance crossover became a reference for Bhattacherjee’s agenda paper, CoNEXT motifs, Hypatia and CCR’s transit economics. Section 5.1 develops this quantitative starting point.

公開 FCC 軌道參數是 Handley 2018 路由研究的基礎。他推得一種可行雷射布局，測試上下行與星間跳躍的路徑在何種條件下勝過地面光纖。長距交叉點成為 Bhattacherjee 議程、CoNEXT motif、Hypatia 與 CCR 經濟學的參考，第 5.1 節從這個量化起點發展。

2. The problem, measured2. 問題，用量測說話

The crossover balances faster propagation against ascent, descent and grid stretch. Orbit height fixes both the geometric cost and satellite motion.

交叉點平衡較快傳播與上下行、網格伸長，軌道高度同時決定幾何成本與衛星運動。

Propagation constants and filed orbit
#	Reported values and conditions
1	Vacuum propagation takes 3.3356 µs/km, while single-mode fiber at index 1.4675 takes 4.895 µs/km.
2	Handley rounds the 46.75% speed difference to about 47% ^{[3, §4]} ^{[32, §II]}.
3	Satellite geometry adds ascent, descent and lattice stretch, so distance savings must first cover these costs.
4	This produces the modeled 3,000 km crossover.
5	At the filed 1,150 km altitude, Handley gives about 7.3 km/s and 107 min per orbit.
6	Kepler calculations give 7.28 km/s and 108.3 min.
7	Roughly thirty satellites remain above London within the 40°-from-vertical visibility cone ^{[3, §2–3]}, with computed values from the fact base.

傳播常數與申報軌道
#	數值、條件與證據
1	真空傳播需 3.3356 µs/km，折射率 1.4675 的單模光纖需 4.895 µs/km，46.75% 速度差由 Handley 取整為約 47% ^{[3, §4]} ^{[32, §II]}。
2	衛星幾何增加上下行與格網伸長，距離節省須先補償這些成本，形成模型的 3,000 km 交叉點。
3	申報 1,150 km 高度下，Handley 給出約 7.3 km/s、107 分鐘，克卜勒計算則為 7.28 km/s、108.3 分鐘。
4	倫敦在距天頂 40° 視錐內約有三十顆衛星 ^{[3, §2–3]}，計算值來自事實庫。

Question	問題	Formula	公式	Result	結果
Propagation cost per kilometre, vacuum versus fibre每公里傳播成本，真空對光纖	\( \tau_{vac}=\tfrac{1}{c},\quad \tau_{fib}=\tfrac{n}{c} \)	3.3356 µs/km, 4.895 µs/km, ratio 1.4675computed, c = 299,792,458 m/s (exact), n = 1.4675 ^{[32, §II]}, Handley quotes "≈ 47%"計算值，c = 299,792,458 m/s（精確值），n = 1.4675^{[32, §II]}，Handley 引為「≈ 47%」
Speed and period of the 1,150 km shell1,150 公里殼層的速度與週期	\( v=\sqrt{\mu/a},\quad T=2\pi\sqrt{a^{3}/\mu} \)	7.28 km/s, 108.3 mincomputed, a = 6,378.137 + 1,150 km, μ = 398,600.4418 km³/s² ^[5], Handley writes ≈ 7.3 km/s and ≈ 107 min ^{[3, §3]}計算值，a = 6,378.137 + 1,150 km，μ = 398,600.4418 km³/s²^[5]，Handley 寫 ≈ 7.3 km/s 與 ≈ 107 分鐘^{[3, §3]}

3. The design3. 設計

Co-moving links form two meshes, and a repeatedly realigned bridge joins them. Source routing uses predicted geometry to choose global shortest paths.

同向鏈路形成兩網格，以反覆對準的橋接相連，來源路由再用預測幾何選全域最短路徑。

Five-link design and route computation
#	Reported values and conditions
1	Handley infers five lasers from the filing’s five “communication components” and calls them “effectively the minimum” for his mesh ^{[3, §2–3]}.
2	Links 1–2 join fore and aft neighbors, while 3–4 join same-index satellites in adjacent planes.
3	They form separate co-moving north-east-bound and south-east-bound meshes in the 53° constellation.
4	Link 5 bridges these meshes and repeatedly realigns as its partners close rapidly.
5	It is “often down”, with EDRS acquisition below one minute cited for context ^{[3, §3]}.
6	Latency-weighted Dijkstra recomputes every 50 ms for the topology 200 ms ahead.
7	Source-installed paths remove satellite routing-table requirements, and the author reports a feasible 10 ms laptop interval ^{[3, §4]}.
8	Routes alternate intra-plane and cross-plane hops.
9	Selecting a lower-elevation satellite toward the destination shortens paths at about 3 dB radio cost. Greedy geographic routing has a longer latency tail ^{[3, §4]}.
10	Of twenty link-disjoint New York–London paths, five beat the fiber bound and all beat measured Internet RTT. This supports load spreading with the modeled latency advantage ^{[3, Fig. 11]}.

五鏈路設計與求路
#	數值、條件與證據
1	Handley 從五個「通訊元件」推得五具雷射，稱為其網格「實際上的最低需求」^{[3, §2–3]}。
2	鏈路 1–2 連前後，3–4 連鄰面同編號衛星，在 53° 星座形成各自同向的東北、東南網格。
3	鏈路 5 橋接兩網格，夥伴快速接近時反覆對準，因此「經常斷線」，文中以 EDRS 小於一分鐘擷取作參考 ^{[3, §3]}。
4	延遲權重 Dijkstra 每 50 ms 計算 200 ms 後拓樸，來源端安裝路徑使衛星可直接依來源路由轉送，作者也報告筆電可用 10 ms 間隔 ^{[3, §4]}。
5	路徑交替面內與跨面跳躍，朝目的地選低仰角衛星可縮路、代價約 3 dB，貪婪地理路由則有較長延遲尾端 ^{[3, §4]}。
6	二十條鏈路分離的紐約–倫敦路徑中，五條勝光纖下限，全數勝實測網際網路，支持保留模型延遲優勢的分流（Fig. 11）。

Figure A. Left: Handley inferred this five-link topology from the 2016 filing. Links 1–4 form a lattice within each co-moving mesh, while link 5 joins the meshes. The shortest path between ground points zig-zags across the lattice. Right, the headline New York–London comparison, in which the satellite RTT oscillates between the fibre bound and the measured Internet.圖 A。左：Handley 從 2016 年申請文件推斷的五鏈路拓樸，鏈路 1–4 在每個同向網格內構成格狀結構，鏈路 5 橋接兩張網格，而兩個地面點之間的最短路徑在格狀結構上呈鋸齒狀前進。右：紐約–倫敦的主要比較，衛星 RTT 在光纖下限與實測網際網路之間擺動。Sources來源: ^{[3, §3–4, Fig. 7]}

Original figure note原始圖說出處

Drawn for this primer from Handley, HotNets 2018, §3–4 and Fig. 7.

4. Evaluation setup4. 實驗設定

The simulation computes geometric propagation on the filed shells. Admission control, unassigned rates and simplified acquisition define the resulting latency scope.

模擬在申報殼層上計幾何傳播，准入控制、待指定速率與簡化擷取共同界定延遲範圍。

Simulation scope and constellation parameters
#	Reported values and conditions
1	The purpose-built simulator propagates FCC orbital elements and calculates geometric routes.
2	Phase 1 has 1,600 satellites, 32×50, at 1,150 km and 53°, with Handley’s collision-based phase offset 5/32.
3	Phase 2 adds 1,600 at 53.8°, about 40 km lower, with offset 17/32.
4	Higher shells are 8×50 at 74°, 5×75 at 81°, and 6×75 at 70°, between 1,130 and 1,325 km.
5	The modeled population totals 4,425 satellites, excluding VLEO ^{[3, §2]}.
6	Visibility within 40° of vertical means a 50° elevation mask, stricter than the later 25° mask.
7	Five lasers use Figure A’s layout.
8	Rates remain outside the model as “too speculative”, with 100 Gb/s anticipated as possible.
9	Admission control sets queueing to zero, and only link 5 has intermittent availability ^{[3, §2–3, §5]}.
10	Fiber is an ideal great-circle propagation bound, supplemented by measured 2018 Internet RTT.

模擬範圍與星座參數
#	數值、條件與證據
1	專用模擬器推算 FCC 軌道元素並計幾何路徑。
2	第一階段為 1,600 顆、32×50、1,150 km、53°，碰撞分析選相位 5/32。
3	第二階段再加 1,600 顆、53.8°、低約 40 km、偏移 17/32。
4	較高殼層為 74° 的 8×50、81° 的 5×75、70° 的 6×75，高度 1,130–1,325 km，合計 4,425 顆，排除 VLEO ^{[3, §2]}。
5	距天頂 40° 可見相當於 50° 仰角，比後來 25° 更嚴。
6	五雷射採圖 A 接線，速率被視為「太過推測性」而留在模型之外，並預期 100 Gb/s 可行。
7	准入控制使佇列成本設零，僅鏈路 5 採間歇可用 ^{[3, §2–3, §5]}。
8	光纖是理想大圓傳播下限，另以 2018 實測網際網路 RTT 對照。

5. What was measured5. 量到了什麼

Claim宣稱	Setup and caveat設定與附註
New York–London RTT is ≈ 57–66 ms over 200 s. The comparison values are 55 ms for great-circle fibre and 76 ms for measured Internet ^{[3, Fig. 7, §4]}.紐約–倫敦 RTT 在 200 秒視窗內約 57–66 ms，對比 55 ms 大圓光纖與 76 ms 實測網際網路（圖 7，§4）。	Phase 1 only, propagation only, values read from the figure. The RTT spike at 70–95 s occurs during a crossing between meshes. At that moment, the required link 5 is down.僅第一階段、純傳播，數值由圖讀出。70–95 秒處的 RTT 尖峰是路徑必須經由一條斷線的鏈路 5 從一張網格跨到另一張的時刻。
Satellite RTT falls below the fibre bound at times for NY–London, SF–London and London–Singapore ^{[3, Fig. 8]}.紐約–倫敦、舊金山–倫敦與倫敦–新加坡的衛星 RTT 有時低於光纖下限（圖 8）。	Plotted as a ratio to the fibre bound, roughly 0.6–1.6 across the window, the ratio is below 1 only part of the time.以相對光纖下限的比值繪製，視窗內約 0.6–1.6，比值只有部分時間低於 1。
London–Johannesburg RTT "almost half" of the 182 ms best Internet path, Phase 2 improves it by about 20% ^{[3, Fig. 9]}.倫敦–約翰尼斯堡 RTT「幾乎只有」182 ms 最佳網際網路路徑的一半，第二階段再改善約 20%（圖 9）。	The figure axis spans 60–100 ms, with the exact value unspecified. The north–south route is compared with the Internet, under its long lattice detour.圖軸為 60–100 ms，精確值待列出。南北向路徑有較長格網繞路，此處以網際網路作對照。
Beats any terrestrial fibre network for distances greater than about 3,000 km (Abstract).在距離超過約 3,000 公里時勝過任何地面光纖網路（摘要）。	The abstract gives a crossover summary. Its altitude and great-circle fiber assumptions define that claim, while a full crossover curve requires further evidence.摘要給出交叉點概述，其高度與大圓光纖假設界定主張，完整交叉曲線則需要進一步證據。
Among 20 link-disjoint NY–London paths, five beat the fibre bound and all 20 beat 76 ms. Path 20 takes about 74 ms ^{[3, Fig. 11]}.20 條鏈路不相交的紐約–倫敦路徑：五條勝過光纖下限、全部 20 條勝過 76 ms，第 20 條約 74 ms（圖 11）。	RF and laser links have equal assumed capacity. Path 20’s one-way delay varies by about 10% ^{[3, Fig. 12]}. A reorder buffer absorbs this variation while retaining an RTT of about 74 ms ^{[3, §5]}.假設無線電與雷射鏈路容量相等，第 20 條路徑的單向延遲變動約 10%（圖 12），而吸收該變動的重排緩衝仍給出約 74 ms 的 RTT^{[3, §5]}。

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

The 55 ms fiber value is an ideal great-circle bound, and 76 ms is measured Internet RTT. London–Johannesburg’s comparison with 182 ms also uses the Internet as baseline.55 ms 光纖值是理想大圓下限，76 ms 是實測網際網路 RTT。倫敦–約翰尼斯堡與 182 ms 的比較也採網際網路作基準。
Model delay includes propagation, with queueing, switching and capacity outside its scope. Multipath assumes equal radio and laser rates. Measured bent-pipe RTT includes gateways and PoPs, with about 40 ms floor and 25–50 ms idle medians (Part C, item 58). Seychelles ISL routing reached 181 ms against a direct-ISL estimate near 40 ms ^{[35, §7.2]}. Each comparison consequently needs its measurement definition.模型計傳播，佇列、交換與容量在其範圍之外，多路徑則假設無線電與雷射等速。實測彎管 RTT 包含閘道、PoP，約有 40 ms 下限與 25–50 ms 閒置中位數（C 部第 58 項）。塞席爾 ISL 路由曾達 181 ms，直接 ISL 估計則約 40 ms ^{[35, §7.2]}，因此各項比較須附量測定義。
The filed layout was superseded. November 2018 moved Phase 1 to 550 km, 24×66, then August 2019 changed it to 72×22 (Part C, item 12). Shorter ground legs and changed hop geometry make crossover altitude-dependent.申報布局後來更改：2018 年 11 月第一階段改為 550 km、24×66，2019 年 8 月再改為 72×22（C 部第 12 項）。地面腿縮短、跳躍幾何改變，使交叉點依賴高度。
Five lasers are Handley’s inference. The 2018 filing implies four, and the current technology page states three with an undisclosed topology (Part C, items 13 and 55). The two-mesh bridge remains a design-space baseline.五雷射來自 Handley 推論，2018 申報隱含四具，目前技術頁列三具，拓樸仍待公開（C 部第 13、55 項）。雙網格橋接因此屬設計空間的基準。
Link-5 downtime acknowledges acquisition qualitatively. Modern terminals require 2–30 s (Part C, item 18), and later work finds even 100 ms setup raises mean delay ^{[32, §IV-B]}. Fifteen-second scheduling belongs to later measured behavior outside this historical model (Part C, item 57).鏈路 5 停用以定性方式反映擷取。現代終端需 2–30 s（C 部第 18 項），後續研究顯示 100 ms 建立也會提高平均延遲 ^{[32, §IV-B]}。十五秒排程是後來實測行為，位於本歷史模型範圍之外（C 部第 57 項）。
The text describes 50 ms recomputation with 200 ms look-ahead and separately reports 10 ms as feasible. These are computational workload statements. Reordering, BBR-style control, failures and load-dependent routing remain research questions ^{[3, §5]}.正文描述 50 ms 重算與 200 ms 前瞻，另報告 10 ms 可行，兩者都是計算負載陳述。重排、BBR 式控制、故障與負載相依路由仍屬研究問題 ^{[3, §5]}。

7. How it connects7. 與其他內容的關聯

Section 5.1 builds the latency-versus-distance argument from this paper and separates it from the measured 40 ms floor.第 5.1 節從這篇論文建立延遲對距離的論證，並將它與實測的 40 ms 下限分開。
Section 5.2 treats the lattice of links 1–4 as the +Grid baseline that motifs and stable-link designs are measured against.第 5.2 節把鏈路 1–4 的格狀結構當作 +Grid 基準，motif 與穩定鏈路設計都以它為對照。
Section 5.7 tabulates the assumptions that have changed since 2018: altitude, terminal count, setup delay and the 15-second schedule.第 5.7 節列表整理 2018 年以來已改變的假設：高度、終端數、建鏈延遲與 15 秒排程。
Section 2.5 explains why the inter-mesh link 5 is the inclined-constellation analogue of a polar seam, and Section 2.1 gives the vacuum-versus-fibre arithmetic.第 2.5 節說明為什麼跨網格的鏈路 5 是傾斜星座中極區縫隙的對應物，第 2.1 節給出真空對光纖的算術。
Related pages: Handley 2019 recovers the latency advantage through ground relays. Gearing Up 2018 compares Frankfurt–Washington against fibre on polar orbits. CoNEXT 2019 optimizes the lattice, while CCR 2020 examines transit costs. Ali & Zhang 2026 and TinyLEO revisit wiring with three terminals and real ephemerides.相關導讀：Handley 2019 拿掉雷射，用地面中繼找回優勢，Gearing Up 2018 在極軌道上對法蘭克福–華盛頓做同樣的光纖比較，CoNEXT 2019 最佳化格狀結構，CCR 2020 問把這些路徑當作中轉服務販售要花多少成本，Ali 與 Zhang 2026 和 TinyLEO 在真實星曆上以三個終端重做接線。

Key numbers關鍵數字

Number	數字	Meaning (setup)
≈ 47%	Speed advantage of light in vacuum over light in glass ^{[3, §4]}, exact ratio 1.4675 for n = 1.4675 [computed].	真空中光速相對玻璃中光速的優勢^{[3, §4]}，n = 1.4675 時精確比值 1.4675［計算值］。
≈ 3,000 km	Ground distance beyond which the satellite path beats any fibre path (Abstract), specific to 1,150 km altitude and a great-circle fibre.	衛星路徑開始勝過任何光纖路徑的地面距離（摘要），針對 1,150 公里高度與大圓光纖。
32 × 50, 1,150 km, 53°	Phase 1 of the 2016 filing, 1,600 satellites, phase offset 5/32 ^{[3, §2]}, superseded by 550 km / 72 × 22.	2016 年申請文件的第一階段，1,600 顆，相位偏移 5/32^{[3, §2]}，已被 550 公里／72×22 取代。
5	Laser terminals per satellite, inferred from five "communication components" ^{[3, §2–3]}, Starlink now states three.	每顆衛星的雷射終端數，由五個「通訊元件」推斷^{[3, §2–3]}，Starlink 目前聲明為三個。
57–66 / 55 / 76 ms	NY–London RTT: satellite path over 200 s / great-circle fibre bound / measured Internet ^{[3, Fig. 7]}.	紐約–倫敦 RTT：200 秒內衛星路徑／大圓光纖下限／實測網際網路（圖 7）。
≈ ½ of 182 ms	London–Johannesburg satellite RTT against the best Internet path, Phase 1, Phase 2 adds ≈ 20% ^{[3, Fig. 9]}.	倫敦–約翰尼斯堡衛星 RTT 對比最佳網際網路路徑，第一階段，第二階段再改善約 20%（圖 9）。
20 / 5 / ≈ 74 ms	Link-disjoint NY–London paths, count below the fibre bound, path 20 RTT, and reorder-buffered aggregate RTT ^{[3, Fig. 11, §5]}.	鏈路不相交的紐約–倫敦路徑數／勝過光纖下限的條數／第 20 條路徑與重排緩衝合流的 RTT（圖 11，§5）。
50 ms / 200 ms	Route recomputation interval and look-ahead horizon for source routing ^{[3, §4]}, 10 ms stated feasible.	來源路由的重算間隔與前瞻時間^{[3, §4]}，另稱 10 ms 亦可行。
~ 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）。