It's a Bird? It's a Plane? It's CDN! Investigating Content Delivery Networks in the LEO Satellite Networks Era是鳥？是飛機？是 CDN！探討 LEO 衛星網路時代的內容遞送網路
SpaceCDN examines how distant Starlink exit points weaken CDN locality. Measurements establish the latency gap, while a simulation explores satellite caches. The cache proposal requires further evaluation of link performance, replacement policy and hardware.SpaceCDN 研究 Starlink 的遠端網路出口如何削弱 CDN 的地理鄰近優勢。量測建立延遲差距，模擬則探索衛星快取的可行性。其鏈路性能、替換策略與硬體需求仍須進一步評估。
The traffic exit point explains the detour beyond the satellite hop. Peering at that point of presence determines which CDN node answers. That simulation counts hops on a four-link +Grid with ISL rate unspecified and laser setup and the 15-second schedule excluded. Its model excludes cache behavior, so the paper establishes the latency gap firmly and the remedy only as a plausibility argument.衛星跳段之外的繞路，主要由流量離開網路的位置決定，因為 Starlink 把流量送到某個 PoP，再由該 PoP 的對等互連決定由哪個 CDN 節點回應。論文提出的對策 SpaceCDN 把內容快取在衛星上，先透過星間鏈路向鄰近衛星取用，最後才回到地面。這個模擬是在四鏈路 +Grid 上數跳數，ISL 速率待定，雷射建鏈時間、15 秒排程及快取模型皆位於範圍之外，所以論文對延遲差距的證據紮實，對解方則只是可行性論證。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Latency comparison | Bose et al. combine about 22K Starlink and about 800K terrestrial Cloudflare AIM measurements, roughly 822K total. They add crowdsourced browser tests and LEOScope probes. Starlink CDN idle latency exceeds terrestrial latency by about 50 ms. The gap reaches 120–150 ms in Kenya, Mozambique and Zambia. In Mozambique, the best Starlink CDN node is 8,777 km away, with median minimum RTT of 138.7 ms. The terrestrial comparison is 5 km and 7.2 ms. [89, §3.2, Fig. 2, Tab. 1] |
| Constellation configuration | The proposed remedy, SpaceCDN, caches content on the satellites and falls back over inter-satellite links before going to the ground. The xeoverse simulation uses Starlink Shell 1 with 72 planes of 22 satellites at 550 km. Fetching within 5 ISL hops competes with terrestrial CDN latency. At 10 hops, latency is half the current Starlink result. [89, §4, Fig. 7] |
| 延遲比較 | Bose 等人結合約 22K 筆 Starlink 與約 800K 筆地面 Cloudflare AIM 量測（合計約 822K 筆）、群眾外包的瀏覽器測試以及 LEOScope 探針，說明 Starlink 使用者連到 CDN 的閒置延遲比地面使用者高約 50 ms，在肯亞、莫三比克與尚比亞更高出 120–150 ms。 |
| 往返時間 | 莫三比克的 Starlink 使用者，最佳 CDN 節點遠在 8,777 km 外，最小 RTT 中位數 138.7 ms，而地面上的鄰居只要 5 km、7.2 ms[89, §3.2, Fig. 2, Tab. 1]。 |
| 軌道面配置 | 以 xeoverse 模擬 Starlink Shell 1（550 km、72 個軌道面各 22 顆）的結果顯示，在 5 個 ISL 跳數內取得內容即可與地面 CDN 延遲競爭，10 跳則能把現今 Starlink 的延遲減半[89, §4, Fig. 7]。 |
1. Where this paper sits1. 這篇論文的位置
This workshop paper opens the services discussion in Chapter 9. It asks how satellite access affects content delivery networks, a widely used latency optimization. A content delivery network places copies of popular objects close to users. This benefit depends on a nearby access-network exit and a nearby CDN node.
Chapter 3 traces Starlink traffic through a bent pipe, a gateway and a point of presence (PoP). Starlink hands traffic to the Internet at this facility. The Internet therefore sees the PoP's location. Bose et al. measure the consequence of that fact for content delivery.
They are the first to propose moving the cache onto the satellite. StarCDN [62], comes from an independent group a year later. It develops cache placement and evaluates hit rates with request traces. Thus, the two papers are best read as a measurement-plus-vision paper followed by a systems paper.
這篇工作坊論文為第 9 章的服務主線開場，它問的是：當使用者搬到衛星網路上，網際網路最成功的延遲最佳化會發生什麼事。內容遞送網路之所以有效，是因為它把熱門物件的副本放在使用者附近，而這預設了使用者的網路路徑會在住家附近離開接取網路、並在附近遇到 CDN 節點。第 3 章已說明 Starlink 使用者的流量經由彎管送到閘道，再送到 PoP（Starlink 把流量交給網際網路其他部分的設施），而網際網路看到的是 PoP 的位置，使用者位置則須另行定位。Bose 等人量測的正是這個事實對內容遞送造成的後果，而且他們是第一個提議把快取搬到衛星上的團隊。一年後由另一個獨立團隊發表的 StarCDN[62]把這個提議發展成完整設計，包含快取放置演算法與以真實軌跡驅動的命中率，所以兩篇最好連起來讀：先是一篇量測加願景的論文，接著是一篇系統論文。
2. The problem, measured2. 問題，用量測說話
Table 1 shows why, because it lists both the distance to the optimal CDN node and the median minimum RTT.Table 1 說明了原因，因為它同時列出到最佳 CDN 節點的距離與最小 RTT 中位數。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | Cloudflare Aggregated Internet Measurement (AIM) records Starlink as AS14593 from March 2024. The dataset contributes about 22K Starlink measurements from 55 countries. The terrestrial comparison contains about 800K measurements from 196 countries. The authors add about 5K NetMet Chromium-extension measurements. LEOScope probes in the UK, Germany, Canada and Nigeria fetch top-20 Tranco sites on Cloudflare and CloudFront. [89, §3.1] |
| Latency comparison | Terrestrial CDN idle latency is about 50 ms below Starlink latency. The gap reaches 120–150 ms in Kenya, Mozambique and Zambia. [89, §3.2, Fig. 2] |
| Evaluation conditions | In Mozambique the terrestrial user's optimal node is 5.0 km away at 7.2 ms |
| Mozambique: Starlink | while the Starlink user's is 8,776.5 km away at 138.7 ms |
| Kenya comparison | and Kenya (197.5 km and 16 ms against 6,310.8 km and 110.9 ms) |
| Zambia comparison | and Zambia (1,202.64 km and 44 ms against 7,545.9 km and 143.5 ms) follow the same pattern. |
| Evaluation conditions | Figure 3 makes it concrete for Maputo, whose Starlink users are served from Frankfurt at roughly 150–180 ms. A local cache would answer in under 20 ms [89, Fig. 3]. |
| 評估條件 | 量測的基礎是 Cloudflare 的 Aggregated Internet Measurement（AIM）資料集，Starlink 自 2024 年 3 月起以 AS14593 出現在其中，貢獻約 22K 筆來自 55 國的 Starlink 量測，對照約 800K 筆來自 196 國的地面量測。 |
| 評估條件 | 另外補上作者 NetMet Chromium 擴充套件收集的約 5K 筆群眾外包量測，以及在英國、德國、加拿大與奈及利亞的 LEOScope 探針，抓取託管在 Cloudflare 與 CloudFront 上的 Tranco 前 20 名網站[89, §3.1]。 |
| 延遲比較 | 主要結果是地面使用者到 CDN 的閒置延遲比 Starlink 使用者低約 50 ms，而在肯亞、莫三比克與尚比亞差距擴大到 120–150 ms[89, §3.2, Fig. 2]。 |
| 評估條件 | 莫三比克的地面使用者最佳節點在 5.0 km 外、7.2 ms，Starlink 使用者則在 8,776.5 km 外、138.7 ms。 |
| 評估條件 | 肯亞（197.5 km、16 ms 對 6,310.8 km、110.9 ms）與尚比亞（1,202.64 km、44 ms 對 7,545.9 km、143.5 ms）呈現同樣的模式。 |
| 評估條件 | Fig. 3 以馬布多為例讓這件事具體化：當地 Starlink 使用者由法蘭克福以約 150–180 ms 服務，而本地快取低於 20 ms 就能回應[89, Fig. 3]。 |
Two rows of the same table separate the two causes that the African numbers combine. A local PoP, as in Spain, leaves only the bent-pipe latency floor. Southern African traffic exits in Europe, adding thousands of kilometers of fiber.同一張表中的另外兩列，把非洲數字裡混在一起的兩個成因拆開了。PoP 在本地時（如西班牙），Starlink 使用者只損失彎管的那一截底線。在南部非洲，流量從歐洲出口離開網路，使用者損失的是底線再加上數千公里的光纖。即使在德國與英國，First Contentful Paint 的中位數也高出約 200 ms。而哥斯大黎加、保加利亞與肯亞等經由 ISL 服務的國家，在下載期間出現超過 200 ms 的緩衝膨脹[89, Figs. 4–5]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Round-trip time | In Spain, the optimal Starlink node is 13.4 km away, with median minimum RTT of 33 ms. Terrestrial RTT is 14.3 ms. Japan gives 57.0 km and 34 ms for Starlink, compared with 9 ms terrestrially. These local nodes isolate the bent-pipe cost of about 20–25 ms from PoP geography. [89, Tab. 1] |
| Evaluation conditions | Starlink HTTP responses were 20–50 ms slower, occasionally reaching an additional 100 ms. Median First Contentful Paint increased by about 200 ms even in Germany and the UK. ISL-served Costa Rica, Bulgaria and Kenya showed download bufferbloat exceeding 200 ms. [89, Figs. 4–5] |
| 往返時間 | 在西班牙，Starlink 使用者的最佳節點只在 13.4 km 外，最小 RTT 中位數卻是 33 ms，地面使用者則是 14.3 ms。 |
| 評估條件 | 在日本則是 57.0 km、34 ms 對 9 ms[89, Tab. 1]。 |
| PoP 覆蓋 | 這把彎管本身的成本（約 20–25 ms）與 PoP 地理的成本分離開來。 |
| 評估條件 | 網頁層面的症狀隨之而來：經由 Starlink 的 HTTP 回應時間慢了 20–50 ms，偶爾慢到 100 ms。 |
3. The design3. 設計
SpaceCDN puts the cache where the user's first hop already is. A request first reaches the satellite overhead. A local copy answers immediately. Otherwise, ISLs carry it to the nearest copy.
The final fallback uses a gateway, a PoP and a terrestrial CDN [89, §4]. Because every satellite overhead is replaced within minutes, the paper adds two ideas that exploit motion rather than fight it. The first distributes DASH video segments across successive satellites. Each holds the segments that users will request during its overhead pass.
The second is duty-cycled caching. A subset of satellites stores objects. The rest relay requests over ISLs. Extra hops reduce storage, power and thermal requirements [89, §4–5].
Further ideas, namely machine-learned "content bubbles", replicated "space VMs" and MetaCDN-style economics, are described as future work with evaluation left open. The storage estimate assigns 150 TB of high-end server storage to each of 6,000 satellites. The total exceeds 900 PB. The authors equate this to over 300 million two-hour 1080p videos [89, §5].
SpaceCDN 把快取放在使用者第一跳已經在的地方。請求先送到頭頂的衛星，若它持有該物件就直接回應，其餘情況下請求經由星間鏈路送到最近一顆持有副本的衛星，閘道、PoP 與地面 CDN 則是最後手段[89, §4]。由於頭頂上的每顆衛星幾分鐘內就會被取代，論文加入兩個利用運動的方式配合衛星移動的想法。第一個是把 DASH 影片分段條帶化到即將飛過使用者上空的一連串衛星上，讓每顆衛星持有它在頭頂期間會被請求的片段。第二個是工作週期式快取，只有一部分衛星擔任快取，其餘衛星透過 ISL 轉送，用跳數換取儲存、電力與散熱[89, §4–5]。其他想法，包括以機器學習產生的「內容泡泡」、複製的「太空虛擬機」與 MetaCDN 式的經濟模型，都被列為後續工作，評估仍待完成。儲存論證只是一個乘法：每顆衛星 150 TB 的高階伺服器儲存，乘以 6,000 顆衛星，得到超過 900 PB，作者把它換算成超過 3 億部兩小時的 1080p 影片[89, §5]。
Original figure note原始圖說出處
Drawn for this primer from Bose et al., HotNets 2024, Table 1, Fig. 3, §4–5 and Figs. 7–8.4. Evaluation setup4. 實驗設定
The paper has two halves with different kinds of evidence. The latency gap is a measurement, taken from production data (Cloudflare AIM), from the authors' own browser extension and from LEOScope probes. Its main metric is idle latency to the CDN. Thus, it excludes queueing under load except where the bufferbloat figures report it separately [89, §3].
The evaluation measures distance to a copy in hops at each replication level. It selects duty-cycled caches randomly. Replacement policy, request traces and hit rates lie outside its scope [89, §5, Fig. 8].
這篇論文的兩半有各種性質的證據。延遲差距是量測，資料來自營運資料（Cloudflare AIM）、作者自己的瀏覽器擴充套件以及 LEOScope 探針，主要指標是到 CDN 的閒置延遲，所以除了另外報告的緩衝膨脹數字之外，將負載下排隊列於範圍之外[89, §3]。論文中快取替換策略、請求軌跡及命中率皆位於評估範圍之外，因為評估只問在給定複本數下副本離使用者幾跳。工作週期式快取的評估則是隨機挑選擔任快取的衛星子集[89, §5, Fig. 8]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Plane configuration | The SpaceCDN simulation uses xeoverse with the August 2019 Starlink Shell 1 filing. Its 72 planes of 22 satellites total 1,584 satellites at 550 km. [89, §4] |
| Latency comparison | Latency in that simulation is computed from ISL hop counts on the four-link +Grid that xeoverse assumes. The paper leaves ISL data rate and per-hop node delay unspecified and excludes link setup time. The model excludes the 15-second global reconfiguration described in Chapter 3. Thus, the overhead satellite in the simulation is a geometric choice rather than the one Starlink's scheduler would assign. |
| 軌道面配置 | SpaceCDN 那一半則是 xeoverse 中的模擬，設定為 2019 年 8 月申請文件幾何下的 Starlink Shell 1，也就是 550 km、72 個軌道面各 22 顆，共 1,584 顆衛星[89, §4]。 |
| 延遲比較 | 模擬中的延遲是由 xeoverse 假設的四鏈路 +Grid 上的 ISL 跳數算出，論文將 ISL 資料速率、每跳節點延遲及建鏈時間列為待定，第 3 章描述的 15 秒全域重新配置亦位於模型範圍之外，所以模擬中的「頭頂衛星」是幾何上的選擇，Starlink 排程器的實際指派則須另行量測。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Starlink users see CDN idle latency about 50 ms above terrestrial users, 120–150 ms above in Kenya, Mozambique and Zambia [89, §3.2, Fig. 2].Starlink 使用者的 CDN 閒置延遲比地面使用者高約 50 ms，在肯亞、莫三比克與尚比亞高 120–150 ms[89, §3.2, Fig. 2]。 | Measurement, about 22K Starlink against about 800K terrestrial AIM samples. Cloudflare only, and AIM combines wired and wireless terrestrial access.量測，約 22K 筆 Starlink 對約 800K 筆地面 AIM 樣本。只涵蓋 Cloudflare，且 AIM 合併呈現有線及無線地面接取。 |
| Mozambique's optimal CDN is 8,776.5 km away at 138.7 ms over Starlink against 5.0 km and 7.2 ms terrestrially [89, Tab. 1].莫三比克經 Starlink 的最佳 CDN 在 8,776.5 km 外、138.7 ms，地面則是 5.0 km、7.2 ms[89, Tab. 1]。 | Median minimum RTT. The calc table shows that fiber distance alone explains about 62% of the Starlink figure.最小 RTT 中位數。計算表顯示單是光纖距離就解釋了 Starlink 數字的約 62%。 |
| First Contentful Paint is about 200 ms higher at the median in Germany and the UK [89, Fig. 5].在德國與英國，First Contentful Paint 中位數高約 200 ms[89, Fig. 5]。 | NetMet and LEOScope browser measurements. Read from a figure, and present even where PoPs are local.NetMet 與 LEOScope 瀏覽器量測。數值讀自圖表，即使 PoP 在本地也存在。 |
| Fetching within 5 ISL hops is competitive with terrestrial CDN latency, and within 10 hops halves current Starlink latency [89, §4, Fig. 7].在 5 跳 ISL 內取得內容可與地面 CDN 延遲競爭，10 跳內可把現今 Starlink 延遲減半[89, §4, Fig. 7]。 | xeoverse simulation, Shell 1 72×22, hop-count latency on a +Grid. ISL rate, setup delay, scheduling and cache behavior lie outside its scope.xeoverse 模擬，Shell 1 72×22，+Grid 上以跳數計延遲。ISL 速率待定，建鏈延遲、排程器及快取模型排除。 |
| About four copies per plane keep an object within 5 hops. Caching on 50% of satellites gives results competitive with the terrestrial median [89, §4–5, Fig. 8].每軌道面約四份副本就能讓物件保持在 5 跳內，50% 的衛星擔任快取即可與地面中位數競爭[89, §4–5, Fig. 8]。 | Random choice of caching satellites. Values read from figures.隨機選擇擔任快取的衛星。數值讀自圖表。 |
| 150 TB per satellite across 6,000 satellites gives more than 900 PB of orbital storage [89, §5].每顆衛星 150 TB、共 6,000 顆，可提供超過 900 PB 的在軌儲存[89, §5]。 | An arithmetic estimate, with mass, power and thermal budgets outside its scope for the storage.算術估計，儲存設備的質量、電力及散熱預算位於範圍之外。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| What does the Mozambique detour cost in fiber round trip alone?莫三比克的繞路單是光纖來回要多少時間？ | \( 2 \times 8{,}777\,\text{km} \times 4.895\,\mu\text{s/km} \) | ≈ 85.9 mscomputed. 8,777 km from Table 1, 4.895 µs/km fiber constant from Chapter 2 (fact base Part A)計算值。8,777 km 來自 Table 1，光纖常數 4.895 µs/km 來自第 2 章（事實庫 Part A） |
| What share of the measured 138.7 ms is distance?量到的 138.7 ms 中有多少是距離造成的？ | \( 85.9 / 138.7 \) | ≈ 62%computed from the row above and Table 1由上一列與 Table 1 計算 |
| What is the bent-pipe floor where the PoP is local?PoP 在本地時，彎管的底線是多少？ | \( 33 - 14.3,\quad 34 - 9 \) | ≈ 19–25 mscomputed. Spain and Japan rows of Table 1, where the Starlink user's CDN is closer than the terrestrial user's計算值。取自 Table 1 的西班牙與日本兩列，Starlink 使用者的 CDN 甚至比地面使用者的更近 |
| How many copies of one object does "four per plane" imply in Shell 1?「每軌道面四份」在 Shell 1 中代表一個物件有幾份副本？ | \( 4 \times 72 = 288,\quad 288/1{,}584 \) | 288 copies ≈ 18% of satellitescomputed. 4 per plane from §4, 72×22 = 1,584 from §4計算值。每軌道面 4 份來自 §4，72×22 = 1,584 來自 §4 |
| Orbital storage from the paper's inputs由論文輸入算出的在軌儲存 | \( 150\,\text{TB} \times 6{,}000 \) | 900 PBcomputed. Both inputs from §5計算值。兩個輸入皆來自 §5 |
6. Limits and open questions6. 限制與待解的問題
- The measurement count is inconsistent. The introduction claims more than 1M measurements. The methods list about 22K Starlink and about 800K terrestrial samples, roughly 822K. The PoP count appears as both 22 and about 20. This digest uses about 822K (fact base Part C, item 43).量測筆數的兩處記載存在差異。引言宣稱超過 100 萬筆量測，方法一節列出的卻是約 22K 筆 Starlink 與約 800K 筆地面樣本，合計約 822K。PoP 數量也同時出現 22 與約 20。本導讀採用約 822K（事實庫 Part C 第 43 項）。
- The hop axis requires further specification. The text lists hop limits of 1, 2, 3, 5 and 10. The Fig. 7 legend shows 1, 3, 5 and 10. The reported figure therefore leaves the 2-hop result open.跳數軸仍須補充說明。內文列出 1、2、3、5、10 跳的上限，Fig. 7 圖例卻只有 1、3、5、10，所以仍待論文提供 2 跳的結果。
- Four ISLs on a filed geometry. The simulation uses the 2019 72×22 filing with a four-link +Grid. Starlink states three laser terminals per satellite and keeps its topology proprietary. Real-network hop counts would therefore differ (fact base Part C, items 12, 13 and 55).以申請文件幾何加四條 ISL。模擬用的是 2019 年 72×22 的申請幾何與四鏈路 +Grid，而 Starlink 表示每顆衛星有三具雷射終端，且拓樸仍屬專有資訊，所以在真實網路上跳數會各種（事實庫 Part C 第 12、13、55 項）。
- Link rate, setup delay and scheduling remain open. Latency follows hop count, with ISL rate and node delay unspecified. Links become available instantly, while modern terminals require 2–30 s for setup. The 15-second reassignment of users to satellites is absent (Part C, items 18, 19, 56 and 57).鏈路速率、建鏈延遲及排程仍待補充。延遲只是跳數，ISL 速率與節點延遲皆待定。鏈路被視為即時可用，但現代終端的雷射建鏈需要 2–30 s。使用者每 15 秒被重新指派給衛星的機制亦位於模型範圍之外（Part C 第 18、19、56、57 項）。
- Cache behavior and hardware remain open. Replacement policy, traces and hit rates lie outside the model. The 900 PB estimate also leaves mass, power and thermal budgets open. The paper itself cites a sub-30 °C limit for COTS hardware. The authors note that better ground infrastructure alone could bring latency to 20–30 ms. This is the baseline a SpaceCDN must beat.快取行為與硬體需求仍待補充。替換策略、軌跡及命中率皆位於模型範圍之外。900 PB 這個數字的質量、電力及散熱預算仍待補充，儘管論文自己引用了 COTS 硬體低於 30 °C 的限制。作者也指出單靠更好的地面基礎設施就能把延遲降到 20–30 ms，這才是 SpaceCDN 必須勝過的基準。
7. How it connects7. 與其他內容的關聯
- Section 9.1 opens with PoP-exit geography. Figure 9.1 redraws the Maputo path, and its calculation table gives the 86 ms fiber detour.第 9.1 節：PoP 出口的地理是該章的開場論證，馬布多路徑重繪為圖 9.1，86 ms 的光纖繞路列在其計算表中。
- Section 9.2: StarCDN, the full design that follows from this proposal.第 9.2 節：StarCDN，也就是由這個提議發展出的完整設計。
- Section 9.3 and Section 9.4: the other services in orbit, federated learning and on-board inference, which face the same visibility and power limits.第 9.3 節與第 9.4 節：其他在軌服務，即聯邦學習與機上推論，它們面對同樣的可見性與電力限制。
- Section 9.5: SpaceCDN sits on the link-rate ladder as "hop counts only", the one service paper with rate left unspecified.第 9.5 節：SpaceCDN 在鏈路速率之梯上標為「只數跳數」，是唯一速率仍待說明的服務論文。
- Section 3.3: gateways, PoPs and the bent pipe, the mechanism this paper measures the cost of.第 3.3 節：閘道、PoP 與彎管，也就是這篇論文量測其成本的機制。
- Section 10.6: the constellation-wide regime of Gbps links and millisecond hops, where caching services live.第 10.6 節：Gbps 鏈路與毫秒級跳數的星座尺度區間，快取服務就在這個區間。
- StarCDN [62] supplies the cache algorithm and hit rates this paper lacks. xeoverse is the simulator used here. Mohan et al. [49] and Kassem et al. [46], which share authors, measure the PoP and bent-pipe latencies underneath. FedSat and OrbitalBrain are the federated-learning services of the same chapter.StarCDN[62]補上本文缺少的快取演算法與命中率。xeoverse 是本文使用的模擬器。共享作者的 Mohan 等人[49]與 Kassem 等人[46]量測了底層的 PoP 與彎管延遲。FedSat 與 OrbitalBrain 則是同一章的聯邦學習服務。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| ~822K | AIM measurements, ~22K Starlink (55 countries) and ~800K terrestrial (196 countries). The intro's ">1M" requires further evidence [89, §3.1] | AIM 量測筆數，約 22K 筆 Starlink（55 國）與約 800K 筆地面（196 國）。引言的「>1M」仍待證據[89, §3.1] | |
| ≈50 ms · 120–150 ms | CDN idle-latency penalty over Starlink, global and in Kenya, Mozambique, Zambia [89, Fig. 2] | 經 Starlink 的 CDN 閒置延遲懲罰，全球與肯亞、莫三比克、尚比亞[89, Fig. 2] | |
| 138.7 vs 7.2 ms | Mozambique median min RTT, Starlink vs terrestrial. 8,776.5 vs 5.0 km to the optimal CDN [89, Tab. 1] | 莫三比克最小 RTT 中位數，Starlink 對地面。到最佳 CDN 8,776.5 對 5.0 km[89, Tab. 1] | |
| ≈200 ms | Higher median First Contentful Paint in Germany and the UK [89, Fig. 5] | 德國與英國 First Contentful Paint 中位數高出的量[89, Fig. 5] | |
| 5 · 10 hops | ISL hops at which SpaceCDN matches terrestrial CDN and halves Starlink latency. Xeoverse, 72×22 [89, Fig. 7] | SpaceCDN 追上地面 CDN 與把 Starlink 延遲減半所需的 ISL 跳數。xeoverse，72×22[89, Fig. 7] | |
| ≈4 per plane · 50% | Copies per plane to stay within 5 hops. Fraction of satellites caching that stays competitive [89, Figs. 7–8] | 維持 5 跳內所需的每軌道面副本數。仍具競爭力的快取衛星比例[89, Figs. 7–8] | |
| 150 TB × 6,000 ≈ 900 PB | Orbital storage estimate, power and thermal budgets outside its scope [89, §5] | 在軌儲存估計，電力及散熱預算位於範圍之外[89, §5] | |
| 20–30 ms | Best-case Starlink latency with better ground infrastructure, the baseline a SpaceCDN must beat [89, §5] | 地面基礎設施改善後的 Starlink 最佳延遲，也就是 SpaceCDN 必須勝過的基準[89, §5] |