Chapter 9 of 13 · Part II · The research
第 9 章（共 13 章）· 第二部 · 研究
Services in orbit: content delivery, storage, federated learning and on-board inference在軌服務：內容遞送、儲存、聯邦學習與機上推論
This chapter explains distant CDN exits, StarCDN’s response to orbital motion, and how visibility limits satellite federated learning. It identifies flown inference systems and compares the papers’ link assumptions with the rates that Part III requires.本章說明 Starlink 的 CDN 請求為何從遠處離開網路，以及 StarCDN 如何處理軌道運動。接著以計算說明可見性對衛星聯邦學習的限制，整理已在軌執行的推論，並比較各論文的鏈路假設與第三部所需的速率。
Before you start
開始之前
This chapter builds on Chapters 1, 2, 3 and 8. These terms recur throughout the chapter. Each link opens its definition.
New terms in this chapter (5)本章新名詞（5）
Each link opens the definition at the term’s first appearance.
各名詞在首次出現處都有定義，點選連結即可前往。
- Content delivery network (CDN)內容遞送網路（CDN） A set of caches placed near users so that popular objects are served from a short distance instead of from the origin server. 一組放在使用者附近的快取，讓熱門內容直接從近處送出。
- Consistent hashing一致性雜湊Each object hashes to one of N buckets, and each bucket maps to one cache. The group stores each object once and forwards requests to its holder. 以雜湊把每個物件對應到 N 個桶之一，再把每個桶對應到一台快取，使一組快取只保存每個物件的一份，請求則被轉送到持有者。
- Federated learning聯邦學習 Training in which each participant keeps its data locally, updates a shared model, and exchanges only model updates with an aggregator. 每個參與者保留本地資料，用這些資料更新共享模型，再把模型更新交給彙總者的訓練方式。
- Staleness陳舊度 The number of global rounds that have passed between a participant downloading a model and its update being aggregated. 參與者下載模型到它的更新被彙總之間，已經過去的全局回合數。
- On-board inference機上推論 A trained model runs on the satellite’s processor, which processes raw data locally and sends results down the link. 在衛星的處理器上執行訓練好的模型，就地處理原始資料，再透過下行鏈路傳送結果。
9.1Why services move up: PoP-exit geography服務為什麼上太空：PoP 出口的地理
Recall
回顧
Section 3.3 showed that a Starlink packet travels user → satellite → gateway → PoP before it reaches the Internet. The operator’s network ends at the PoP.
第 3.3 節說明了 Starlink 的封包要經過使用者 → 衛星 → 閘道 → PoP 才進入網際網路，營運商網路的終點位於 PoP。
CDNs use a request’s exit address to choose nearby caches. Starlink’s distant PoPs can therefore send users to distant CDN nodes.
CDN 依請求的出口位址選擇附近快取，Starlink 的遠端 PoP 因而可能把使用者送到遠處的 CDN 節點。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | A content delivery network (CDN) places caches close to users, so popular video segments and web objects travel only a few kilometers from a cache. |
| Result and scope | Qualitative scope | This design assumes that a request’s source address indicates the user’s location. |
| Result and scope | Qualitative scope | This assumption holds for a terrestrial ISP whose Internet exit shares its customers’ city or country. |
| Dataset and workload | Qualitative scope | In a LEO network, requests exit at a PoP determined by gateway and peering locations. |
| Result and scope | Qualitative scope | That PoP can be a continent away from the dish. |
| Dataset and workload | 22, 55, 800, 196 | Bose et al. compared Cloudflare AIM measurements: about 22K Starlink samples from 55 countries and about 800K terrestrial samples from 196 countries. |
| Latency comparison | 50 ms | Starlink users’ idle latency to the CDN was roughly 50 ms higher. |
| Result and scope | 120–150 ms | The difference reached 120–150 ms in Kenya, Mozambique and Zambia [89, §3.2, Fig. 2]. |
| Dataset and workload | 822, 1 | The methods list ~822K samples, while the introduction says "over 1M". |
| Result and scope | Qualitative scope | The figures here follow the methods count. |
| 結果與適用範圍 | 定性範圍 | 內容遞送網路（CDN）把快取放在使用者附近，讓熱門影片片段或網頁物件從幾公里外送出。 |
| 結果與適用範圍 | 定性範圍 | 這種設計假設，請求的來源位址能反映使用者的位置。 |
| 結果與適用範圍 | 定性範圍 | 在地面 ISP 上這個假設成立，因為 ISP 的網際網路出口與它的客戶在同一個城市或國家。 |
| 結果與適用範圍 | 定性範圍 | 在 LEO 網路上這個假設就失效了，因為請求是在 PoP 離開營運商網路，而 PoP 的位置由閘道與對等互連的所在決定，可能與天線相隔一個大陸。 |
| 結果與適用範圍 | 55, 22, 196, 800, 50 ms, 120–150 ms | Bose 等人用 Cloudflare 的 AIM 資料集量出了後果，他們比較了來自 55 個國家約 22K 筆 Starlink 樣本與來自 196 個國家約 800K 筆地面樣本，發現 Starlink 使用者到 CDN 的閒置延遲比地面使用者高出約 50 ms，在肯亞、莫三比克與尚比亞更高出 120–150 ms[89, §3.2, Fig. 2]。 |
| 結果與適用範圍 | 822, 1 | 論文的方法一節列出約 822K 筆樣本，導論則寫「超過 1M」，本章採用方法一節的計數。 |
The measured latency gap depends on both the terrestrial detour and the bent-pipe access path. A nearby CDN still leaves the satellite-access delay.
量測的延遲差距同時來自地面繞行與彎管接入路徑。即使 CDN 位於附近，衛星接入仍有其延遲成本。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Latency comparison | 1, 7.2 ms, 5 km | Table 1 reports a terrestrial median minimum RTT of 7.2 ms in Mozambique, to an optimal CDN node about 5 km away. |
| Latency comparison | 8,777 km, 138.7 ms | Starlink users in that country reached a node 8,777 km away with a 138.7 ms median minimum RTT. |
| Result and scope | 3, 160 ms, 250 ms | In Fig. 3’s Maputo case, Starlink traffic reached Frankfurt at roughly 160 ms, while the Cape Town node exceeded 250 ms [89, Tab. 1, Fig. 3]. |
| Result and scope | 4.895 µs | Distance explains most of this gap because light in fiber needs 4.895 µs per kilometer (the vacuum-versus-fiber figures of Section 2.1). |
| Orbital configuration | 8,777 km, 86 ms | A round trip over 8,777 km of fiber takes about 86 ms, before queueing and bent-pipe delays. |
| Dataset and workload | 13.4 km, 375.3 km, 33, 14.3 ms | Spanish and Japanese Starlink users reach closer CDN nodes yet still have higher latency. In Spain, distances are 13.4 versus 375.3 km and latencies are 33 versus 14.3 ms. |
| Latency comparison | Qualitative scope | This reflects Section 3.4’s bent-pipe latency floor. Two space legs and a gateway hop cost more than a few hundred kilometers of fiber. |
| 延遲比較 | 1, 5 km, 7.2 ms, 8,777 km, 138.7 ms, 3, 160 ms, 250 ms | 他們 Table 1 的莫三比克那一列把機制講得很具體：當地的地面使用者到約 5 km 外的最佳 CDN 節點，最小 RTT 中位數為 7.2 ms，而同一國家的 Starlink 使用者到 8,777 km 外的節點要 138.7 ms，Fig. 3 的馬布多案例則顯示 Starlink 流量被對應到法蘭克福，約 160 ms，而開普敦節點更超過 250 ms[89, Tab. 1, Fig. 3]。 |
| 軌道配置 | 4.895 µs, 8,777 km, 86 ms | 光是地理就能解釋大部分差距，因為光在光纖中每公里需要 4.895 µs（第 2.1 節的真空對光纖數字），所以在 8,777 km 光纖上跑一趟來回，在任何排隊或彎管跳數算進去之前就要約 86 ms，如算表所示。 |
| 軌道配置 | 13.4 km, 375.3 km, 33, 14.3 ms | 同一張表還記錄了一個更細微的案例：西班牙與日本的 Starlink 使用者被對應到比地面使用者更近的節點（西班牙 13.4 km 對 375.3 km），延遲卻仍然較高（33 對 14.3 ms），這反映了第 3.4 節的彎管延遲下限：兩段太空鏈路加一跳閘道，所需時間高於幾百公里光纖。 |
Original figure note原始圖說出處
Drawn for this primer from Bose et al., HotNets 2024, Table 1 and Fig. 3. Fiber constant from Section 2.1.SpaceCDN places caches on satellites to shorten those detours. Its simulation relates latency to ISL hop count, while storage power and cooling remain open.
SpaceCDN 將快取放上衛星，以縮短繞行路徑。模擬比較 ISL 跳數與延遲，儲存設備的電力與散熱則待評估。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Latency comparison | Qualitative scope | The idle-latency penalty also affected web performance. |
| Result and scope | 20–50 ms, 200 ms | Even Germany and the UK, which have local PoPs, showed 20–50 ms slower HTTP responses and about 200 ms higher First Contentful Paint [89, Figs. 4–5]. |
| Cache placement and fallback | Qualitative scope | The paper places the cache on the satellite overhead, where the user’s first hop already lands. A miss uses the nearest satellite holding the object. The ground supplies the final fallback. |
| Latency comparison | 1, 72, 22, 550 km, 2019, 5, 10 | Their xeoverse simulation uses the filed Starlink Shell 1: 72 planes of 22 satellites at 550 km, matching Section 1.4’s 2019 geometry. Fetching within 5 ISL hops competes with terrestrial CDN latency. Fetching within 10 hops halves current Starlink latency. About four copies per plane keep an object within 5 hops. |
| Cache replication | Qualitative scope | and that having only half the satellites act as caches is enough to stay competitive [89, §4–5, Figs. 7–8]. |
| Orbital configuration | 150 TB, 6,000, 900 | The storage estimate multiplies 150 TB of high-end server storage per satellite by 6,000 satellites. This gives about 900 PB in orbit [89, §5], as the calculation table reproduces. |
| Power and thermal conditions | Qualitative scope | What the paper leaves open is the power and heat of carrying that storage, the questions of Section 8.3. |
| 結果與適用範圍 | 20–50 ms, 200 ms | 網頁層的症狀跟著閒置延遲而來：Starlink 上的 HTTP 回應時間慢 20–50 ms，首次內容繪製（First Contentful Paint）即使在有本地 PoP 的德國與英國也高出約 200 ms[89, Figs. 4–5]。 |
| 結果與適用範圍 | 定性範圍 | 論文提出的補救是把快取放到使用者第一跳已經在的地方，也就是頭頂的衛星上，再以 ISL 退回到最近持有該物件的衛星，最後才退回地面。 |
| 軌道配置 | 1, 550 km, 72, 22, 2019, 5, 10 | 他們在 xeoverse 中模擬申報版的 Starlink Shell 1，即 550 km 高度 72 個軌道面各 22 顆（第 1.4 節的 2019 年申報幾何），發現從 5 跳 ISL 以內取得內容可與地面 CDN 延遲相當，10 跳以內則可把現在 Starlink 的延遲減半，每個軌道面約四份複本就能讓物件保持在 5 跳以內，而且只要一半的衛星擔任快取就足以維持競爭力[89, §4–5, Figs. 7–8]。 |
| 結果與適用範圍 | 150 TB, 6,000, 900 | 儲存論證只是一道乘法，每顆衛星 150 TB 的高階伺服器儲存乘上 6,000 顆，在軌約 900 PB[89, §5]，算表重算了這個數字。 |
| 結果與適用範圍 | 定性範圍 | 儲存設備所需的電力與散熱仍待評估，也就是第 8.3 節的問題。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Fibre round trip over the Mozambique detour莫三比克繞路的光纖來回時間 | \(2 \times 8{,}777\,\text{km} \times 4.895\,\mu\text{s/km}\) | 85.9 mscomputed. 8,777 km from Bose et al. Table 1, 4.895 µs/km from Section 2.1計算值。8,777 km 來自 Bose et al. Table 1，4.895 µs/km 來自第 2.1 節 |
| Share of the measured 138.7 ms explained by distance量到的 138.7 ms 中由距離解釋的比例 | \(85.9 / 138.7\) | 62%computed from the row above由上一列計算 |
| SpaceCDN's orbital storage estimateSpaceCDN 的在軌儲存估計 | \(150\,\text{TB} \times 6{,}000\) | 900 PBcomputed. Both inputs from [89, §5]計算值。兩個輸入皆來自 [89, §5] |
| Number數字 | Value數值 | Setup設定 | Source來源 |
|---|---|---|---|
| CDN idle-latency penalty, Starlink vs terrestrialCDN 閒置延遲懲罰，Starlink 對地面 | ≈50 ms. 120–150 ms (Kenya, Mozambique, Zambia) | Measurement, Cloudflare AIM, ~22K Starlink vs ~800K terrestrial samples量測，Cloudflare AIM，約 22K Starlink 對約 800K 地面樣本 | [89, §3.2, Fig. 2][89, §3.2, Fig. 2] |
| Mozambique median min RTT and distance to optimal CDN莫三比克最小 RTT 中位數與到最佳 CDN 的距離 | 7.2 ms / 5 km vs 138.7 ms / 8,777 km | Same measurement同一量測 | [89, Tab. 1][89, Tab. 1] |
| SpaceCDN latency vs ISL hop countSpaceCDN 延遲對 ISL 跳數 | ≤5 hops competitive. 10 hops ≈ half of Starlink | xeoverse simulation, Shell 1 72×22, hop-count latency, hop-count model onlyxeoverse 模擬，Shell 1 72×22，以跳數計延遲，模型僅計跳數 | [89, §4, Fig. 7][89, §4, Fig. 7] |
| StarCDN request hit rateStarCDN 請求命中率 | 60% (LRU) → 71% at 50 GB, N = 4. Headline 60 → 75% | Trace-driven simulation, 1,170 Gen-1 satellites, Akamai one-day trace, 15 s step軌跡驅動模擬，1,170 顆 Gen-1 衛星，Akamai 一日軌跡，15 s 步長 | [62, §5.2] . Abstract[62, §5.2] 。 摘要 |
| StarCDN uplink use as a fraction of cache-freeStarCDN 上行使用量佔零快取的比例 | LRU 30–35%. StarCDN 20–25% | Same simulation同一模擬 | [62, §5.2, Fig. 8][62, §5.2, Fig. 8] |
| StarCDN median latencyStarCDN 延遲中位數 | 22 ms vs 55 ms | Idle-latency values from prior measurement, with queueing excluded採先前量測的閒置延遲值，排除模擬排隊 | [62, §5.3, Fig. 10][62, §5.3, Fig. 10] |
9.2StarCDN: caching that moves against the orbitStarCDN：逆著軌道移動的快取
Recall
回顧
Section 2.5 gives each +Grid satellite two intra-plane and two inter-plane neighbors. Section 3.5 shows that serving satellites are reassigned on a 15-second schedule.
Orbital motion changes a satellite cache’s users. StarCDN distributes objects across nearby satellites through consistent hashing, trading a larger shared cache for ISL routing delay.
軌道運動讓衛星快取持續更換服務對象。StarCDN 以一致性雜湊在附近衛星間分配物件，以 ISL 路由延遲換取較大的共用快取。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | A satellite cache serves a changing group of users as it moves and a neighbor takes over the region. |
| Access-satellite changes | 15 s | Starlink users’ first-contact satellites change about every 15 s [62, §1]. |
| Cache performance | 60% | Zheng et al. estimate that a naive per-satellite LRU cache stays below a 60% hit rate. Simultaneous coverage also wastes storage because several satellites cache the same popular objects for one region [62, §1, §3.1.2]. |
| Result and scope | Qualitative scope | StarCDN attacks both with two mechanisms. |
| Result and scope | Qualitative scope | The first is consistent hashing made satellite-specific: every object is hashed into one of N buckets |
| Result and scope | 4, 9 | with N = 4 or 9 |
| Result and scope | Qualitative scope | and each bucket is mapped to a different satellite in a grid pattern |
| Result and scope | Qualitative scope | so that a satellite serves its own bucket directly and forwards any other request over ISLs to the neighbor holding that bucket |
| Result and scope | Qualitative scope | with every bucket reachable within two hops for both values of N [62, §3.2]. |
| Latency comparison | 40 ms, 9, 5 | The N co-covering satellites act as one cache N times larger, with one copy of each object. Routing latency grows with N. Beyond N = 9, worst-case latency reaches about 40 ms for at most about 5 extra hit-rate points [62, §5.3, Fig. 9]. |
| 結果與適用範圍 | 15 | 衛星快取的服務對象隨軌道運動而變動：隨著衛星前進、鄰居接管該區域，它的使用者每幾分鐘就換一批，而在 Starlink 的情況下，使用者的第一接觸衛星約每 15 秒就會改變[62, §1]。 |
| 結果與適用範圍 | 60% | Zheng 等人估計，各衛星獨立運作的基本 LRU 快取因此命中率會停在 60% 以下，他們還指出第二種浪費，也就是同時覆蓋同一區域的幾顆衛星，全都用相同的熱門物件填滿各自的快取[62, §1, §3.1.2]。 |
| 結果與適用範圍 | 定性範圍 | StarCDN 以兩個機制同時對付這兩個問題。 |
| 結果與適用範圍 | 4, 9 | 第一個是為衛星量身打造的一致性雜湊：每個物件被雜湊到 N 個桶之一，N = 4 或 9，每個桶再對應到網格圖樣中各顆衛星，於是衛星直接服務自己的桶，其餘請求則透過 ISL 轉送給持有該桶的鄰居，而 N 取這兩個值時，每個桶都能在兩跳內到達[62, §3.2]。 |
| 結果與適用範圍 | 9, 40 ms, 5 | N 顆共同覆蓋的衛星因此成為一個 N 倍大、每個物件各存一份的快取，代價是路由延遲，它隨 N 增加，在 N 超過 9 之後最差情況達到約 40 ms，卻只換來至多約 5 個百分點的額外命中率[62, §5.3, Fig. 9]。 |
The second mechanism, relayed fetch, uses orbital motion to predict where content is cached. On a cache miss, a satellite first asks the nearest inter-orbit neighbor with the same bucket. That neighbor recently covered the same region and already holds its users’ requested content. Content therefore moves against orbital motion to follow the users. The authors use only inter-orbit links for this because intra-orbit ISL latency is higher in their geometry.
They report that proactive prefetching was rejected as less efficient [62, §3.3, Tab. 1]. Failure handling has two tiers. A transient failure sends the miss to the ground. A long-term failure remaps the bucket to the next available satellite [62, §3.4]. Figure 9.2 draws both mechanisms on a +Grid.
第二個機制是中繼擷取（relayed fetch），利用軌道運動預測內容所在。遇到快取遺漏時，衛星先向持有相同桶的最近跨軌道面鄰居查詢，那顆衛星片刻之前才覆蓋過同一批使用者的區域，因此已經持有這些使用者要過的內容，於是內容沿著與衛星運動相反的方向流動，跟隨使用者的需求。作者只用跨軌道面鏈路做這件事，因為在他們的幾何中同軌道面 ISL 的延遲較高，他們也報告主動預取因效率較差而被捨棄[62, §3.3, Tab. 1]。故障分兩層處理：暫時性故障變成一次向地面的快取遺漏，長期故障則把桶重新對應到下一顆可用的衛星[62, §3.4]。圖 9.2 把兩個機制畫在 +Grid 上。
Original figure note原始圖說出處
Drawn for this primer from Zheng et al., SIGCOMM 2025, §3.2–3.3 and Table 1.The trace-driven evaluation separates gains from object distribution and relayed fetch. Reported uplink savings depend on the baseline, and failures reduce the hit rate.
軌跡驅動評估區分了物件分工與中繼擷取的貢獻。上行節省幅度隨比較基準而變，故障則會降低命中率。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Dataset and workload | 1%, 423, 512 TB, 24, 1,170, 72, 53°, 1 | The evaluation replays a one-day Akamai trace from nine clusters. The sites are Mexico City, Dallas, Atlanta, Washington DC, New York, London, Frankfurt, Vienna and Istanbul. A 1% sample gives 423 M video requests, 512 TB and 24 M unique objects. The Starlink Gen-1 TLE configuration has 1,170 satellites in 72 orbits at 53°. |
| Result and scope | 15, 100, 8.03 ms, 2.15 ms | Microsoft’s CosmicBeats generates orbital motion and link schedules in 15-second steps. ISLs run at 100 Gbps, with mean delays of 8.03 ms intra-orbit and 2.15 ms inter-orbit. |
| Capacity and rate | 20 | and 20 Gbps ground–satellite links |
| Result and scope | 10, 100 GB | with cache sizes swept from 10 to 100 GB and scaled for the sampling [62, §5.1, Tab. 1]. |
| Cache performance | 50 GB, 4, 60%, 71%, 15, 60 GB, 9, 75% | At 50 GB and N = 4, request hit rate rises from 60% for per-satellite LRU to 71%. The largest gap is 15 points at 60 GB and N = 9. The abstract instead headlines 60% to 75% [62, §5.2]. |
| Cache performance | 6, 4, 9.7, 9, 4.8, 4.1 | Consistent hashing adds 6 hit-rate points at N = 4 and 9.7 points at N = 9. Relayed fetch adds another 4.8 and 4.1 points, respectively [62, §5.2.1–5.2.2]. |
| Result and scope | Qualitative scope | The resource that matters most is the uplink from ground to satellite |
| Result and scope | Qualitative scope | which the paper treats as the scarce link |
| Result and scope | 30–35%, 20–25% | and there LRU uses 30–35% of the cache-free volume while StarCDN uses 20–25% |
| Result and scope | 80%, 30% | These levels give the abstract’s “up to 80%” reduction against an uncached baseline. Against LRU, the reduction is roughly 30%, as the calculation table shows [62, §5.2, Fig. 8]. |
| Cache performance | 126, 1,296, 438, 7, 74% | The failure scenario has 126 of 1,296 active satellites out of slot and 438 broken ISLs. Hit rate falls by at most 7 points, while uplink saving stays at 74% [62, §5.4, Fig. 11]. |
| Result and scope | 1, 5.4, 1,170 | The satellite count differs between Table 1 and §5.4, and we cite 1,170 as the evaluated constellation. |
| 結果與適用範圍 | 1%, 423, 512 TB, 24, 1, 1,170, 72, 53°, 15, 100, 8.03 ms, 2.15 ms, 20, 10, 100 GB | 評估重播了來自九個叢集（墨西哥城、達拉斯、亞特蘭大、華盛頓特區、紐約、倫敦、法蘭克福、維也納與伊斯坦堡）的一日 Akamai 軌跡，取樣 1%，得到 423 M 個影片請求、512 TB 與 24 M 個相異物件，跑在取自 Starlink Gen-1 TLE 的 1,170 顆衛星、72 條軌道、53° 傾角上，軌道運動與鏈路排程由 Microsoft 的 CosmicBeats 以 15 秒步長產生，ISL 為 100 Gbps，同軌道面平均延遲 8.03 ms、跨軌道面 2.15 ms，地面–衛星鏈路為 20 Gbps，快取大小從 10 掃到 100 GB 並依取樣比例縮放[62, §5.1, Tab. 1]。 |
| 結果與適用範圍 | 50 GB, 4, 60%, 71%, 60 GB, 9, 15, 75% | 在 50 GB、N = 4 時請求命中率從每衛星 LRU 的 60% 升到 71%，最大差距是 60 GB、N = 9 時的 15 個百分點，摘要的標題數字則是 60% 到 75%[62, §5.2]。 |
| 結果與適用範圍 | 4, 6, 9, 9.7, 4.8, 4.1 | 消融實驗把兩個機制分開：單靠一致性雜湊在 N = 4 時多 6 個百分點、N = 9 時多 9.7 個百分點的請求命中率，中繼擷取再加 4.8 與 4.1[62, §5.2.1–5.2.2]。 |
| 結果與適用範圍 | 30–35%, 20–25%, 80%, 30% | 最要緊的資源是從地面到衛星的上行鏈路，論文把它視為稀缺鏈路，而 LRU 使用零快取流量的 30–35%，StarCDN 使用 20–25%，以完全零快取為基準這就是摘要所說的「最多 80%」減少，以 LRU 為基準則約減少 30%，算表把這一點講明[62, §5.2, Fig. 8]。 |
| 結果與適用範圍 | 1,296, 126, 438, 7, 74% | 在 1,296 顆活躍衛星中有 126 顆偏離軌位、438 條 ISL 中斷的故障情境下，命中率最多下降 7 個百分點，上行節省仍維持 74%[62, §5.4, Fig. 11]。 |
| 結果與適用範圍 | 1, 5.4, 1,170 | Table 1 與 §5.4 列出的衛星數存在差異，我們以 1,170 作為受評估的星座。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Uplink reduction against a cache-free baseline相對於零快取的上行減少 | \(1 - (0.20 \ldots 0.25)\) | 75–80%computed. StarCDN uplink 20–25% of cache-free from Fig. 8計算值。StarCDN 上行為零快取的 20–25%，取自 Fig. 8 |
| Uplink reduction against per-satellite LRU相對於每衛星 LRU 的上行減少 | \(1 - 22.5/32.5\) | ≈31%computed from the midpoints of 20–25% and 30–35%由 20–25% 與 30–35% 的中點計算 |
| Mean object size in the Akamai traceAkamai 軌跡的平均物件大小 | \(512\,\text{TB} / 24\,\text{M}\) | ≈21 MBcomputed. Trace totals from Zheng et al. §5.1計算值。軌跡總量來自 Zheng et al. §5.1 |
| Objects held by one 50 GB satellite cache一顆 50 GB 衛星快取能放的物件數 | \(50\,\text{GB} / 21\,\text{MB}\) | ≈2,400computed from the row above. ×4 or ×9 when pooled by hashing由上一列計算。以雜湊共用時再乘 4 或 9 |
Stop and think停下來想一想
Set α to 0.8 and compare N = 1 with N = 9 at 50 GB. Repeat at α = 1.2. Explain why pooling helps more under flatter popularity, then identify which video, web or download workloads gain most. Consider relayed fetch separately, because its extra 4–5 hit-rate points come entirely from orbital motion outside this model.
把 α 設為 0.8，在 50 GB 比較 N = 1 與 N = 9，再在 α = 1.2 重做一次。為什麼熱門度分布越平，共用買到的命中率越多？這對哪些工作負載（影片、網頁、下載）最能從 StarCDN 的雜湊獲益有什麼含意？接著問模型範圍之外的效果：StarCDN 的中繼擷取在雜湊之上又加了 4–5 個百分點，而這份增益完全來自模型省略掉的運動。
StarCDN’s latency comparison uses earlier idle measurements. Its simulation leaves mid-transfer disconnection, handover failure and hardware validation for further work.
StarCDN 的延遲比較採用先前的閒置量測。傳輸中斷、換手失敗與硬體驗證則留待後續研究。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | Three caveats belong beside these numbers. |
| Latency comparison | 22 ms, 55 ms, 1 | StarCDN compares idle latencies of 22 ms and 55 ms from prior measurements, with queueing outside that comparison. Its discrete-time simulator excludes mid-transfer disconnection and handover failure. The Gen-1 TLE satellites carry ground links only, so their geometry serves as a topology proxy. The four-link grid is Section 5.7’s modeling baseline, with actual Starlink wiring requiring separate evidence [62, §5.1, §5.3, §7]. |
| Result and scope | Qualitative scope | Both CDN papers therefore share the assumptions the rest of Part II has learned to flag, and both remain simulation studies. |
| 結果與適用範圍 | 定性範圍 | 這些數字旁邊應該放三個但書。 |
| 結果與適用範圍 | 22 ms, 55 ms, 1 | StarCDN 22 ms 對 55 ms 的延遲比較採用先前量測得到的閒置延遲值，排隊則在模擬範圍之外，它的離散時間模擬器忽略傳輸中途的斷線與換手失敗，而提供幾何的 Gen-1 衛星根本僅有地面鏈路，所以 Gen-1 只是拓樸代理，四鏈路網格是第 5.7 節所說的建模基準，其與實際 Starlink 拓樸的關係仍待驗證[62, §5.1, §5.3, §7]。 |
| 結果與適用範圍 | 定性範圍 | 兩篇 CDN 論文因此都帶著第二部其餘章節學會標記的那些假設，而且都停留在模擬階段。 |
9.3Federated learning governed by contact visibility受接觸可見性主導的聯邦學習
Recall
回顧
Section 1.2 computes a pass lasting a few minutes at 550 km. Section 8.1 shows the downlink limit when a satellite meets a ground station only a few times daily.
Federated learning keeps training data on each satellite and exchanges model updates. FedSat uses predictable contacts to update the global model asynchronously.
聯邦學習將訓練資料留在各衛星，只交換模型更新。FedSat 利用可預測的接觸時程，以異步方式更新全局模型。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Model and compute | Qualitative scope | Federated learning trains a shared model while participants retain their data. Each downloads the global model, runs local gradient steps and returns an update. The aggregator averages those updates into the next global model. |
| Satellite learning workload | Qualitative scope | Earth-observation satellites are a natural fit, since the images they hold are too large to downlink and too valuable to discard |
| Result and scope | Qualitative scope | but the constellation imposes a constraint that specific to satellite federated systems, as Razmi et al. put it |
| Result and scope | Qualitative scope | which is that "the time between contacts is much longer than the actual online time" |
| Latency comparison | Qualitative scope | with the delay from assigning a task to receiving its result ranging from one orbital period to half a day [90, §I]. |
| Model and compute | Qualitative scope | FedSat uses predictable contacts to turn synchronous FedAvg into an asynchronous scheme. Each reconnection corrects the global model by the weighted difference between that satellite’s previous and new updates. |
| Orbital configuration | 500, 2,000 km, 80°, 10° | The Walker-Delta evaluation has two shells at 500 and 2,000 km, with five satellites each. It uses 80° inclination and a 10° elevation mask. |
| Capacity and rate | 96.2%, 99.4–99.6% | The single ground station is at the North Pole or Bremen. Connectivity is ground-only, with link rate unspecified. FedSat reaches 96.2% of centralized MNIST accuracy with IID data and 99.4–99.6% with non-IID data. |
| Result and scope | 127 | and the authors state that it cuts FedAvg's training time "by several hours" against a maximum orbital period of about 127 min [90, §II–IV, Figs. 1–3]. |
| 結果與適用範圍 | 定性範圍 | 聯邦學習在保有各自資料的參與者之間訓練一個共享模型，每個參與者下載目前的全局模型，在自己的資料上跑本地梯度步驟，只把更新送回去，由彙總者平均成下一個全局模型。 |
| 結果與適用範圍 | 定性範圍 | 地球觀測衛星是天然的適用對象，因為它們持有的影像下行成本高、又值得保留，但星座帶來一個任何地面聯邦系統都特有的限制，用 Razmi 等人的話說，就是「接觸之間的時間遠長於實際在線時間」，從指派任務到收到結果的延遲從一個軌道週期到半天之間[90, §I]。 |
| 結果與適用範圍 | 定性範圍 | 他們的 FedSat 把同步的 FedAvg 展開成利用接觸可預測性的異步方案，在每次重新連線時以衛星前一次與這一次更新的加權差值修正全局模型。 |
| 結果與適用範圍 | 500, 2,000 km, 80°, 10°, 96.2%, 99.4–99.6%, 127 | 在 500 與 2,000 km 兩層 Walker-Delta 玩具星座上評估，每層五顆衛星、80° 傾角、10° 仰角遮罩、單一地面站在北極或不來梅、僅有地面鏈路、鏈路速率則待說明，它達到集中式 MNIST 準確率的 96.2%（IID 資料）與 99.4–99.6%（non-IID 資料），作者表示 FedAvg 的訓練時間「縮短了數小時」，評估時間以約 127 min 的最大軌道週期為單位[90, §II–IV, Figs. 1–3]。 |
FedSpace schedules aggregation by trading idle time against update staleness. Its connectivity assumptions describe the dated fleet used in the study.
FedSpace 在等待時間與更新陳舊度之間取捨，以安排彙總。連線假設應連同研究所用機隊及日期一起解讀。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 191, 12, 15, 68, 4, 5–19 | FedSpace schedules aggregation across Planet’s 191 satellites and 12 ground stations. Orbital mechanics predicts connectivity in 15-minute slots. Connected-satellite counts range from 68 to 4, while each satellite sees a station 5–19 times daily [91, §2.1–2.2, Figs. 1–2] preprint. |
| Model and compute | 1 | The scheduler trades idleness against staleness. Idleness is time spent waiting for a contact. Staleness counts global rounds between model download and update aggregation. FedSpace applies the weight \((s+1)^{-\alpha}\) to stale updates [91, §2.3–2.4]. |
| Reliability and accuracy | 1.7, 38.6%, 43.1, 16.5, 45.8 | Training takes 1.7 days (38.6%) less than FedBuff and 43.1 days (16.5×) less than synchronous FL. Synchronous FL needs 45.8 days on this fleet. Plain asynchronous FL fails to reach target accuracy because of staleness [91, §1]. |
| Orbital configuration | 2022, 5 | FedSpace assumes ground-only communication in LEO. This 2022 assumption is dated in the fact base. Starlink has carried lasers since v1.5. Four years later, OrbitalBrain gives the same Planet-class fleet ISLs, with overlapping authorship. |
| 結果與適用範圍 | 191, 12, 15, 68, 4, 5–19 | FedSpace 把同一個觀察放大到真實的機隊，在 Planet 的 191 顆衛星與 12 個地面站之間排程彙總，連線性由軌道力學以 15 分鐘的時槽預測，同時連上的衛星數在 68 與 4 之間擺動，每顆衛星一天看到地面站 5–19 次[91, §2.1–2.2, Figs. 1–2]preprint。 |
| 結果與適用範圍 | 1 | 它的排程器在兩種成本之間取捨：閒置，即衛星等待接觸卻處於等待狀態，以及陳舊度，即衛星下載模型到它的更新被彙總之間已過去的全局回合數，FedSpace 以 \((s+1)^{-\alpha}\) 把它降權[91, §2.3–2.4]。 |
| 結果與適用範圍 | 1.7, 38.6%, 43.1, 16.5, 45.8 | 結果是訓練時間比 FedBuff 短 1.7 天（38.6%），比同步聯邦學習短 43.1 天（16.5×），後者在這個機隊上需要 45.8 天，而單純的異步 FL 則因陳舊度而準確率低於目標[91, §1]。 |
| 結果與適用範圍 | 2022, 5 | FedSpace 以 LEO 衛星僅與地面通訊作為前提，事實庫把這個 2022 年的假設標為過時，因為 Starlink 自 v1.5 起就帶著雷射，而四年後由部分相同作者撰寫的 OrbitalBrain，就為同一類 Planet 級機隊配上了 ISL。 |
Original figure note原始圖說出處
Drawn for this primer from Razmi et al., IEEE WCL 2022 (T_o,max ≈ 127 min). Elmahallawy & Luo, IEEE BigData 2022, Table I (16 Mb/s). Chabra et al., NINeS 2026, §1 (passes ≤ 7 min). 22.5 s computed.The ISL-based learning designs improve access to aggregators. Their comparisons keep link capacity fixed and show how contact frequency changes time to accuracy.
使用 ISL 的學習設計增加了接觸彙總者的機會。比較在鏈路容量固定時，顯示接觸頻率如何改變達標時間。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | The papers that add ISLs make the visibility argument explicit by removing it. |
| Result and scope | 40, 5, 2,000 km, 80°, 16 Mb/s | AsyncFLEO models 40 satellites in 5 orbits at 2,000 km and 80° inclination with ISLs only between adjacent satellites in the same orbit. This occurs because satellites in different orbits "have very high relative velocity". It models every link as RF at a single rate of 16 Mb/s with delay equal to transmission plus propagation plus processing [92, §III, Tab. I, Eqs. 7–8]. |
| Orbital configuration | 72, 30, 12, 6, 5, 3, 20 | Table II reports convergence times under identical link assumptions. FedISL with an arbitrary ground station and FedSpace each need 72 h. FedHAP needs 30 h and FedSat 12 h. AsyncFLEO needs 6 h with a ground station, 5 h with one high-altitude platform and 3 h 20 min with two. The range exceeds an order of magnitude. |
| Result and scope | 22 | which is the "22×" the paper claims [92, Tab. II]. |
| Result and scope | 72, 3 | That spread is the chapter's central lesson, since the 72-hour and 3-hour rows use identical links. |
| Result and scope | Qualitative scope | The aggregator’s visibility frequency changes. |
| Result and scope | 6, 15, 550 km, 65°, 85%, 50, 121 | FedRings uses intra-plane rings in a Walker Star constellation: 6 planes × 15 satellites at 550 km and 65°. Inter-plane links activate within visibility windows. Rates use symbolic Shannon expressions, with numeric bandwidth unspecified. DenseNet-121 exceeds 85% EuroSAT accuracy within 50 rounds [93, §3–4, Tab. 1] preprint. |
| Result and scope | 7.5 W, 100, 0.9, 10 s | OrbitalBrain is the peer-reviewed design here with numeric ISL rates. Each satellite uses a Jetson Orin Nano drawing about 7.5 W. ISLs are capped near 100 Mbps, with a 0.9 efficiency factor when setup exceeds 10 s. Updates follow a shortest-path aggregation tree rooted at the best-connected satellite. |
| Orbital configuration | 1.52–12.4, 207, 117, 12 | and reports 1.52–12.4× faster time-to-accuracy on 207 Planet and 117 Spire satellites with 12 Planet ground stations |
| Reliability and accuracy | 30% | with accuracy collapsing when participation falls to 30% [94, §2.1, §3.2, §5, Fig. 1]. |
| 結果與適用範圍 | 定性範圍 | 加入 ISL 的論文反而透過移除可見性限制，把這個論證講得最清楚。 |
| 結果與適用範圍 | 2,000 km, 80°, 5, 40, 16 Mb/s | AsyncFLEO 建模 2,000 km、80° 傾角、5 條軌道共 40 顆衛星，ISL 只存在於同一軌道相鄰衛星之間，因為相異軌道的衛星相對速度很高，而且它把每條鏈路都建模為單一速率 16 Mb/s 的 RF 鏈路，延遲等於傳輸加傳播加處理[92, §III, Tab. I, Eqs. 7–8]。 |
| 結果與適用範圍 | 72, 30, 12, 6, 5, 3, 20, 22 | 它的 Table II 接著列出在相同鏈路假設下相差超過一個數量級的收斂時間：任意地面站的 FedISL 與 FedSpace 都是 72 h，FedHAP 30 h，FedSat 12 h，AsyncFLEO 配地面站 6 h、配一座高空平台 5 h、配兩座 3 h 20 min，這就是論文宣稱的「22×」[92, Tab. II]。 |
| 結果與適用範圍 | 72, 3 | 這個差距就是本章的核心教訓，因為從 72 小時那一列到 3 小時那一列，鏈路維持固定，彙總者的可見頻率則隨方案而變。 |
| 結果與適用範圍 | 550 km, 65°, 6, 15, 121, 50, 85% | FedRings 在 550 km、65° 的 Walker Star 6 個軌道面 × 15 顆衛星上保留同樣的同軌道面環，只在可見窗口內啟用跨軌道面鏈路，鏈路速率只以 Shannon 形式書寫，數值頻寬待說明，並以 DenseNet-121 在 50 回合內於 EuroSAT 上達到 85% 以上的準確率[93, §3–4, Tab. 1]preprint。 |
| 結果與適用範圍 | 7.5 W, 100, 10 s, 0.9, 207, 117, 12, 1.52–12.4, 30% | OrbitalBrain 是唯一給出 ISL 數值速率、且經同儕審查的設計，它給每顆衛星一塊約 7.5 W 的 Jetson Orin Nano，把 ISL 上限設在約 100 Mbps，對超過 10 s 的 ISL 建立時間套用 0.9 的效率因子，沿著以連線最好的衛星為根的最短路徑樹彙總，並在 207 顆 Planet 與 117 顆 Spire 衛星、12 個 Planet 地面站上報告 1.52–12.4× 的達標時間加速，而參與率降到 30% 時準確率會崩落[94, §2.1, §3.2, §5, Fig. 1]。 |
| Paper論文 | Constellation星座 | ISL assumptionISL 假設 | Result結果 |
|---|---|---|---|
| FedSat, Razmi et al., IEEE WCL 2022FedSat，Razmi et al., IEEE WCL 2022 | 2 Walker-Delta shells (500 / 2,000 km), 5 satellites each, 80°, 10° mask, one GS2 層 Walker-Delta（500 / 2,000 km），每層 5 顆，80°，10° 遮罩，單一地面站 | Ground links only. Link rate unspecified僅地面鏈路，速率待說明 | 96.2% (IID) and 99.4–99.6% (non-IID) of centralized MNIST accuracy. "several hours" saved vs FedAvg. To,max ≈ 127 min集中式 MNIST 準確率的 96.2%（IID）與 99.4–99.6%（non-IID）。比 FedAvg 省「數小時」。To,max ≈ 127 min |
| [91] preprint[91] preprint | 191 Planet satellites, 12 GSs, 15-min slots191 顆 Planet 衛星，12 個地面站，15 分鐘時槽 | Ground-only model (2022 assumption)僅地面鏈路（2022 年假設） | −1.7 days (38.6%) vs FedBuff. 16.5× vs synchronous FL (45.8 days)比 FedBuff 少 1.7 天（38.6%）。比同步 FL（45.8 天）快 16.5× |
| AsyncFLEO, Elmahallawy & Luo, IEEE BigData 2022AsyncFLEO，Elmahallawy & Luo, IEEE BigData 2022 | 40 satellites, 5 orbits, 2,000 km, 80°. GS or 1–2 HAPs40 顆衛星，5 條軌道，2,000 km，80°。地面站或 1–2 座 HAP | Intra-plane only. RF 16 Mb/s僅同軌道面。RF 16 Mb/s | Convergence 72 h (FedISL, arbitrary GS) → 3 h 20 min (two HAPs), Table II收斂時間 72 h（FedISL，任意地面站）→ 3 h 20 min（兩座 HAP），Table II |
| [93] preprint[93] preprint | Walker Star 6 × 15 at 550 km, 65°Walker Star 6 × 15，550 km，65° | Intra-plane ring. Inter-plane only in windows. Shannon-form rate, numeric value unspecified同軌道面環。跨軌道面僅限窗口內。Shannon 形式速率，數值待說明 | >85% on EuroSAT within 50 rounds (DenseNet-121)EuroSAT 50 回合內 >85%（DenseNet-121） |
| [94][94] | 207 Planet / 117 Spire satellites, 12 Planet GSs. Jetson Orin Nano ~7.5 W207 顆 Planet / 117 顆 Spire 衛星，12 個 Planet 地面站。Jetson Orin Nano 約 7.5 W | ~100 Mbps cap. Setup >10 s → 0.9 efficiency上限約 100 Mbps。建立 >10 s → 效率 0.9 | 1.52–12.4× time-to-accuracy. Collapse at 30% participation達標時間 1.52–12.4×。參與率 30% 時崩落 |
Model exchange occupies a small part of each orbital period under these assumptions. Contact timing therefore dominates the learning schedule.
在這些假設下，模型交換只佔軌道週期的一小部分，因此接觸時程主導了學習排程。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | The arithmetic that unifies the table is in the calc table below. |
| Capacity and rate | 18, 45 MB, 16 Mb/s, 22.5 s, 0.3%, 127 | A ResNet-18 of about 45 MB crosses a 16 Mb/s link in 22.5 s, which is 0.3% of a 127-minute orbital period. |
| Model and compute | 100, 7, 5 GB | Also, OrbitalBrain's 100 Mbps link inside a 7-minute pass can move more than 5 GB, far more than any model these papers train. Thus, the communication-to-computation ratio in this literature depends on contact timing, with ample capacity during each contact. |
| Result and scope | 45.8, 2.7 | FedSpace provides another visibility comparison: synchronous training takes 45.8 days, falling to about 2.7 days while link capacity stays fixed. |
| Capacity and rate | Qualitative scope | These papers study rates below Chapter 11’s Tbps regime. Their ISL designs use intra-plane rings because of relative-velocity constraints. The formation in Section 11.2 reduces relative velocity to meters per second and targets removal of that constraint. |
| 結果與適用範圍 | 定性範圍 | 把這張表統一起來的算術在下面的算表裡。 |
| 容量與速率 | 45 MB, 18, 16 Mb/s, 22.5 s, 127, 0.3%, 100, 7, 5 GB | 約 45 MB 的 ResNet-18 穿過 16 Mb/s 的鏈路需要 22.5 s，是 127 分鐘軌道週期的 0.3%，而即使 OrbitalBrain 的 100 Mbps 鏈路在 7 分鐘的過頂內也能傳超過 5 GB，遠大於這些論文訓練的任何模型，所以這套文獻中的通訊對運算比例是由鏈路何時存在決定，可見窗口中的容量則綽綽有餘。 |
| 結果與適用範圍 | 45.8, 2.7 | FedSpace 自己的數字從另一面說了同一件事：同步訓練的 45.8 天在鏈路容量維持固定的情況下縮到約 2.7 天。 |
| 容量與速率 | 定性範圍 | 這也是為什麼這些論文的研究速率均低於第 11 章的 Tbps 等級，以及為什麼每一篇用到 ISL 的論文都以相對速度為由把 ISL 侷限在同軌道面環內，而第 11.2 節所談、相對速度只有每秒幾公尺的編隊飛行，正是為了移除這個限制而設計的。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Time to move a 45 MB ResNet-18 over 16 Mb/s以 16 Mb/s 傳送 45 MB ResNet-18 的時間 | \(45 \times 8\,\text{Mb} / 16\,\text{Mb/s}\) | 22.5 scomputed. 16 Mb/s from AsyncFLEO Table I, ~45 MB fp32 ResNet-18 as in the fact base計算值。16 Mb/s 來自 AsyncFLEO Table I，約 45 MB 的 fp32 ResNet-18 依事實庫 |
| Transfer as a share of the longest period傳輸佔最長週期的比例 | \(22.5\,\text{s} / (127 \times 60\,\text{s})\) | 0.3%computed. 127 min from Razmi et al. §IV計算值。127 min 來自 Razmi et al. §IV |
| Bytes one OrbitalBrain pass can carry一次 OrbitalBrain 過頂能傳的位元組 | \(100\,\text{Mb/s} \times 420\,\text{s} / 8\) | 5.25 GBcomputed. 100 Mbps and ≤7 min passes from Chabra et al. §1–2.1計算值。100 Mbps 與 ≤7 min 過頂來自 Chabra et al. §1–2.1 |
| FedSpace training time implied by its savings由 FedSpace 的節省推得的訓練時間 | \(45.8 - 43.1\ \text{days}\) | ≈2.7 dayscomputed. 45.8 and 43.1 days from So et al. §1計算值。45.8 與 43.1 天來自 So et al. §1 |
| AsyncFLEO speed-up as a ratioAsyncFLEO 加速倍數 | \(72\,\text{h} / 3.33\,\text{h}\) | 21.6×computed from Table II. The paper rounds to 22×由 Table II 計算。論文四捨五入為 22× |
Stop and think停下來想一想
Keep the link at 16 Mb/s and increase contacts from 1 to 3 per orbit. Then restore 1 contact and raise rate to 1 Gb/s. Compare the training-time reductions. Find the model-size threshold where one exchange exceeds a 7-minute pass and doubles round time at one contact per orbit. Explain the scale gap between FedRings’ 8 Mb DenseNet-121 and a language model’s gigabytes.
把鏈路留在 16 Mb/s，把接觸次數從每軌道 1 次改成 3 次，再把接觸改回 1 次，把鏈路提高到 1 Gb/s。哪個改變更能縮短訓練，縮短多少？接著找出在每軌道一次接觸時，交換超過 7 分鐘過頂窗口、回合時間翻倍的模型大小。這個門檻對 8 Mb 的 DenseNet-121（FedRings）與語言模型的數 GB 之間的差距說了什麼？
9.4On-board inference that flew真正飛過的機上推論
Recall
回顧
Section 8.1 introduced orbital edge computing and the downlink bottleneck, and Section 8.6 tabulated the compute hardware that has reached orbit.
Flown inference processes observations locally and downlinks results. Published evidence ranges from ground tests for flight payloads to in-orbit thermal measurements.
已飛行的推論在本地處理觀測資料，再下傳結果。公開證據涵蓋飛行酬載的地面測試，以及在軌熱量測。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Model and compute | Qualitative scope | On-board inference runs a trained model on the satellite’s processor, processes raw data locally, and sends results down the link. |
| Result and scope | Qualitative scope | This flown service uses only the downlink, which local processing helps conserve. |
| Result and scope | 1, 2.1 MB, 2, 325 ms, 1.8 W, 92% | ESA’s Φ-Sat-1 supplied the first civil demonstration. Its CloudScout cloud detector uses a 2.1 MB CNN on hyperspectral tiles. The Intel Movidius Myriad 2 runs each inference in 325 ms at 1.8 W, with 92% accuracy. |
| Orbital configuration | Qualitative scope | The published results are ground tests, with in-orbit validation left to the mission [87, Tabs. 3, 5]. |
| Orbital configuration | 4.8, 4.8 km, 0.110 s, 129, 0.091 s | RaVAEn on D-Orbit’s ION SCV004 encodes 4.8 × 4.8 km² tiles with a Myriad X variational autoencoder in 0.110 s per tile. Its 129-parameter few-shot classifier trains on board in 0.091 s per epoch. The authors present this as the first fast on-board training, while timing provenance still requires documentation [95] preprint. |
| Result and scope | 1, 4B, 200, 80, 50, 10% | BUPT-1 supplies the peer-reviewed in-orbit commodity-compute measurement, using Raspberry Pi 4B and Huawei Atlas 200 DK boards. Under high load, the Pi enters the >80 °C throttling zone within about 50 min. YOLO latency rises about 10%. Section 8.3 explains this thermal behavior [85, §4]. |
| 結果與適用範圍 | 定性範圍 | 機上推論在衛星自己的處理器上執行訓練好的模型，就地處理原始資料，再透過下行鏈路傳送結果。 |
| 結果與適用範圍 | 定性範圍 | 這種服務已有在軌示範，所需的網路連線僅限於它要節省的下行鏈路。 |
| 結果與適用範圍 | 1, 2.1 MB, 2, 325 ms, 1.8 W, 92% | 第一次民用示範是 ESA 的 Φ-Sat-1，它的 CloudScout 網路是一個 2.1 MB、用於高光譜切片雲偵測的 CNN，在 Intel Movidius Myriad 2 上以每次推論 325 ms、1.8 W 執行，準確率 92%。 |
| 結果與適用範圍 | 定性範圍 | 已發表的結果是地面測試，在軌驗證留給任務本身[87, Tabs. 3, 5]。 |
| 軌道配置 | 4.8, 4.8 km, 0.110 s, 0.091 s, 129 | D-Orbit ION SCV004 上的 RaVAEn 以 Myriad X 上的變分自編碼器編碼 4.8 × 4.8 km² 的切片，每片 0.110 s，並在機上以每個 epoch 0.091 s 訓練一個 129 個參數的少樣本分類器，作者把它呈現為第一次快速的機上訓練，雖然計時數據的出處待補充[95]preprint。 |
| 結果與適用範圍 | 1, 4B, 200, 50, 80, 10% | 唯一經同儕審查的商用運算在軌量測，是 BUPT-1 上的 Raspberry Pi 4B 與華為 Atlas 200 DK 板，Pi 在高負載下約 50 min 內進入 >80 °C 的降頻區，YOLO 延遲增加約 10%，這就是第 8.3 節的散熱故事[85, §4]。 |
Larger payload demonstrations come from company announcements and press reports. Distributed inference and training over ISLs still await flight demonstration.
較大型酬載的示範來自公司公告與媒體報導，跨 ISL 的分散式推論及訓練仍待在軌示範。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | The larger payloads are company announcements. |
| Result and scope | 1, 60 kg, 2025, 325 km | Starcloud-1 launched in November 2025 to 325 km, with a 60 kg satellite carrying one NVIDIA H100. It demonstrated Gemma-family inference and Karpathy’s nanoGPT training. Throughput, power, temperature and duty-cycle figures await publication. The company page says “Gemini”, while the quoted statements say “Gemma” (NVIDIA blog. |
| Result and scope | 12, 2025 | DCD, 12 Dec 2025) company. |
| Dataset and workload | 12, 14, 2025, 744 TOPS, 8, 1 kW, 15, 1, 2026 | China’s Three-Body Computing Constellation launched 12 satellites on 14 May 2025. Each reportedly has 744 TOPS and an 8-billion-parameter model (Live Science) company via press. Suncatcher’s four-TPU MVP reached orbit on 1 Oct 2026, with roughly 1 kW and 15-minute duty cycles. It uses ground connectivity only, as Section 11.1 records. |
| Latency comparison | 2.7, 81.9%, 2505.14427 | SkyMemory proposes LEO KV caching with a Jetson Nano testbed. Constella separates processor and communicator roles, reporting 2.7× lower latency and at least 81.9% simulated success. Both remain simulation-stage preprints, with ISL rates unspecified in their abstracts (arXiv 2505.14427. |
| Result and scope | 2609.05427 | ArXiv 2609.05427) preprint. |
| Model and compute | 2026 | The flown systems use single-spacecraft inference or tiny on-board fine-tuning. Model footprints range from a few megabytes to one H100’s memory. As of October 2026, multi-satellite distributed inference and training over ISLs still await flight demonstration. |
| 結果與適用範圍 | 定性範圍 | 較大的酬載都是公司公告。 |
| 軌道配置 | 1, 2025, 11, 325 km, 60 kg | Starcloud-1 是 2025 年 11 月發射到 325 km、重 60 kg、搭載單一顆 NVIDIA H100 的衛星，它以 Gemma 系列模型做推論示範、以 Karpathy 的 nanoGPT 做訓練示範，但吞吐量、功率、溫度與工作週期數字待公開，而公司自己的網頁寫「Gemini」，引述的發言卻說「Gemma」（NVIDIA 部落格。 |
| 結果與適用範圍 | 2025, 12 | DCD，2025 年 12 月 12 日）company。 |
| 電力與熱條件 | 2025, 5, 14, 12, 744 TOPS, 80, 1 kW, 15, 2026, 10, 1 | 中國的三體計算星座在 2025 年 5 月 14 日發射 12 顆衛星，宣稱每顆 744 TOPS 並搭載一個 80 億參數模型（Live Science）company via press，而搭載四顆 TPU、約 1 kW、15 分鐘工作週期的 Suncatcher MVP 於 2026 年 10 月 1 日進入軌道且僅有地面鏈路，第 11.1 節有記錄。 |
| 結果與適用範圍 | 2.7, 81.9%, 2505.14427 | SkyMemory 提議把 KV 快取放進 LEO 並以 Jetson Nano 做測試平台，Constella 則把衛星分成處理者與通訊者兩種角色，在模擬中延遲低 2.7×、成功率至少 81.9%，兩者都停留在模擬階段的預印本，摘要的 ISL 數值速率待補充（arXiv 2505.14427。 |
| 結果與適用範圍 | 2609.05427 | arXiv 2609.05427）preprint。 |
| 結果與適用範圍 | 2026, 10 | 表中的模式是一致的：所有飛過的都是單一太空船的推論或微小的機上微調，模型佔用介於幾 MB 與一顆 H100 的記憶體容量之間，而截至 2026 年 10 月，跨 ISL 的多衛星分散式推論與訓練仍待在軌示範。 |
| Mission任務 | Hardware and model硬體與模型 | What was measured量到什麼 | Evidence證據 |
|---|---|---|---|
| Φ-Sat-1 / CloudScout (2020) | Movidius Myriad 2. 2.1 MB cloud-detection CNNMovidius Myriad 2。2.1 MB 雲偵測 CNN | 325 ms and 1.8 W per inference, 92% accuracy (ground tests. In-orbit validation pending at publication)每次推論 325 ms、1.8 W，準確率 92%（地面測試。發表時在軌驗證待完成） | peer-reviewed Giuffrida et al., Remote Sensing 2020 |
| D-Orbit ION SCV004 / RaVAEn (2023) | Myriad X. VAE encoder + 129-parameter classifierMyriad X。VAE 編碼器＋129 參數分類器 | 0.110 s per tile. 0.091 s per training epoch on board每片 0.110 s。機上每個 epoch 0.091 s | preprint Růžička et al., arXiv 2307.08700 |
| BUPT-1 / Tiansuan (2023–24) | Raspberry Pi 4B, Atlas 200 DK. YOLORaspberry Pi 4B、Atlas 200 DK。YOLO | Throttling >80 °C within ~50 min. +~10% latency. Zero SEE errors in six months約 50 min 內進入 >80 °C 降頻。延遲增加約 10%。六個月 SEE 錯誤計數為零 | peer-reviewed [85] |
| Starcloud-1 (Nov 2025) | One NVIDIA H100, 60 kg, 325 km. Gemma-family model, nanoGPT單顆 NVIDIA H100，60 kg，325 km。Gemma 系列模型、nanoGPT | Inference and a training-loop demonstration. Telemetry awaiting publication推論與訓練迴圈示範。遙測待公開 | company NVIDIA blog. DCD |
| Three-Body Computing Constellation (May 2025)三體計算星座（2025 年 5 月） | 12 satellites. Claimed 744 TOPS each. 8B-parameter model12 顆衛星。宣稱每顆 744 TOPS。80 億參數模型 | Independent measurement pending獨立量測待完成 | company via press Live Science |
| Suncatcher MVP (Oct 2026) | Four TPUs, ≈1 kW. Ground links only四顆 TPU，≈1 kW。僅地面鏈路 | 15-minute duty cycles reported. In contact, "operating as expected"報導稱 15 分鐘工作週期。已建立聯繫，「運作如預期」 | company. Press Google blog. Payload |
9.5What these services assume about links這些服務對鏈路的假設
Service and inference studies assume widely separated link rates. The evidence table distinguishes those assumptions from flown demonstrations, company claims and training-cluster targets.
服務與推論研究的鏈路假設相差很大，證據表將這些假設與在軌示範、公司宣稱及訓練叢集目標分開列出。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | 9.4 | Figure 9.4 compares assumed rates across five service papers and Chapter 12’s inference preprints. Their link-rate range spans six orders of magnitude. |
| Capacity and rate | 16 Mb/s, 100 | FedSat and FedSpace use ground links only. AsyncFLEO assumes 16 Mb/s RF links, FedRings specifies symbolic Shannon rates, and OrbitalBrain uses about 100 Mbps [92, Tab. I][94, §2.1]. |
| Link-rate assumptions | 100, 20 | SpaceCDN reports hop counts with link rate unspecified. StarCDN assumes 100 Gbps ISLs and 20 Gbps ground links [62, Tab. 1]. These CDN studies occupy the ladder’s middle. |
| Capacity and rate | 0.5 Gbit/s, 100, 200, 12 | The inference preprints spread across the top, from CollabLLM-LEO's 0.5 Gbit/s through SpaceMoE's "≥100 Gbps" to LLMSpace's 200 Gbps terminals (Chapter 12). |
| Result and scope | Qualitative scope | Against these stand the numbers that have been demonstrated or claimed |
| Capacity and rate | 200, 100 | with TBIRD's 200 Gbps optical downlink flown [37], Starlink's 100 Gbps sustained and 200 Gbps peak per laser as company claims via journalists company |
| Capacity and rate | 400, 3, 6.4 | Section 10.2 supplies terrestrial references above the service rates. Meta’s Llama 3 cluster provides 400 Gbps per GPU preprint report. TPU v6e provides 6.4 Tbps ICI per chip company. |
| Capacity and rate | 9.6–12.8 | and the 9.6–12.8 Tbps per aperture that Suncatcher targets [36, p. 8]. |
| 結果與適用範圍 | 9.4 | 把五篇服務論文與第 12 章的推論預印本並排閱讀，會得到一把橫跨六個數量級的假設鏈路速率之梯，畫在圖 9.4。 |
| 容量與速率 | 16 Mb/s, 100 | 最底層是聯邦學習論文：FedSat 與 FedSpace 僅使用地面鏈路，AsyncFLEO 是 16 Mb/s 的 RF，FedRings 是符號式的 Shannon 速率，OrbitalBrain 約 100 Mbps[92, Tab. I][94, §2.1]。 |
| 容量與速率 | 100, 20 | CDN 論文在中段，SpaceCDN 只報告跳數，速率待說明，StarCDN 則假設 100 Gbps 的 ISL 與 20 Gbps 的地面鏈路[62, Tab. 1]。 |
| 容量與速率 | 0.5 Gbit/s, 100, 200, 12 | 推論預印本散布在上層，從 CollabLLM-LEO 的 0.5 Gbit/s、SpaceMoE 的「≥100 Gbps」到 LLMSpace 的 200 Gbps 終端（第 12 章）。 |
| 容量與速率 | 200, 100, 3, 400, 6.4, 9.6–12.8 | 與它們相對的是已示範或已宣稱的數字：TBIRD 的 200 Gbps 光學下行已飛過[37]，Starlink 每具雷射 100 Gbps 持續、200 Gbps 峰值是透過記者轉述的公司宣稱company，而在所有東西之上的，是第 10.2 節的地面參考點：Meta Llama 3 叢集中單顆 GPU 擁有的 400 Gbpspreprint report、每顆 TPU v6e 晶片 6.4 Tbps 的 ICIcompany，以及 Suncatcher 目標的每個光學孔徑 9.6–12.8 Tbps[36, p. 8]。 |
Original figure note原始圖說出處
Drawn for this primer from the fact base: Elmahallawy & Luo 2022 Table I. Chabra et al. 2026 §2.1. Zheng et al. 2025 Table 1. CollabLLM-LEO, SpaceMoE, LLMSpace (Chapter 12). Schieler et al. ICSOS 2023. Llama 3 §3.3. TPU v6e docs. Agüera y Arcas et al., Joule 2026, p. 8.| Paper論文 | ISL rate assumed假設的 ISL 速率 | Latency model延遲模型 | Traffic unit流量單位 |
|---|---|---|---|
| SpaceCDN [89] | Rate unspecified. Hop count only速率待說明，僅有跳數 | xeoverse hop-count latencyxeoverse 跳數延遲 | One web or video object一個網頁或影片物件 |
| StarCDN [62] | 100 Gbps ISL. 20 Gbps GSL | Mean per-link delay 8.03 / 2.15 / 2.94 ms, propagation only. Idle latency for the user comparison每鏈路平均延遲 8.03 / 2.15 / 2.94 ms，僅傳播。使用者比較採閒置延遲 | ≈21 MB object≈21 MB 物件 |
| FedSat [90], FedSpace (2022) | Ground links only零 | Contact windows only只有接觸窗口 | One model update per contact每次接觸一個模型更新 |
| AsyncFLEO [92] | 16 Mb/s RF, intra-plane | Transmission + propagation + processing傳輸＋傳播＋處理 | One CNN/MLP model一個 CNN/MLP 模型 |
| FedRings [93] | Shannon form, numeric value unspecifiedShannon 形式，數值待說明 | Σ(propagation + processing), values unspecifiedΣ（傳播＋處理），數值待說明 | Sparsified DenseNet-121 (~8 Mb)稀疏化的 DenseNet-121（約 8 Mb） |
| OrbitalBrain [94] | ~100 Mbps cap | Setup >10 s → 0.9 efficiency建立 >10 s → 效率 0.9 | One model update per tree edge樹的每條邊一個模型更新 |
CDNs exchange objects, while federated learning exchanges model updates between contacts. Synchronous training adds repeated collective communication whose completion depends on every participant.
CDN 交換物件，聯邦學習則利用接觸交換模型更新。同步訓練還需要反覆執行集體通訊，完成時間取決於每個參與者。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | Why does Part III need a different regime rather than a faster version of this one? |
| Result and scope | Qualitative scope | The answer lies in the traffic unit column of the table. |
| Model and compute | 21 MB | A CDN moves about 21 MB per request and tolerates tens of milliseconds. Federated learning exchanges megabyte-scale models each round and tolerates minutes to hours of staleness. Both use satellite networks as wide-area services, where geography and visibility matter more than link rate under the stated assumptions. |
| Latency comparison | 400 | Section 10.1 derives synchronous training’s different requirements. Each step sends roughly twice the model’s bytes per worker through all-reduce and repeats every few seconds. All participants wait for the slowest link. Terrestrial clusters therefore provide 400 Gbps per GPU and latency of tens of microseconds Section 10.2. |
| Capacity and rate | 16 Mb/s, 200 | The service papers study different traffic patterns under link assumptions from 16 Mb/s to 200 Gbps. |
| Capacity and rate | Qualitative scope | Orbital training requires a separate assessment of cluster bisection bandwidth and synchronization, beyond these overhead-access models. |
| Scheduling implications | 9.3 | Section 9.3’s scheduling lesson carries over: predictable geometry turns waiting into scheduling. Chapters 11 and 13 apply it to a formation with continuously visible links. |
| 結果與適用範圍 | 定性範圍 | 為什麼第三部需要另一個具有獨立需求的等級？ |
| 結果與適用範圍 | 定性範圍 | 答案在表的「流量單位」那一欄。 |
| 結果與適用範圍 | 21 MB | CDN 每個請求搬一個約 21 MB 的物件，容忍數十毫秒。 |
| 結果與適用範圍 | 定性範圍 | 聯邦學習每回合搬一個數 MB 的模型，容忍數分鐘到數小時的陳舊度。 |
| 結果與適用範圍 | 定性範圍 | 而在這兩種情況下，衛星網路都是一種廣域服務，如前幾節所示，它的地理與可見性比鏈路速率更重要。 |
| 容量與速率 | 400 | 同步訓練把這些性質全部反轉，正如第 10.1 節所推導，因為每一步要透過 all-reduce 搬動約兩倍模型位元組的每工作者流量，每隔幾秒重複一次，並且讓每個參與者停住等到最慢的鏈路完成，這就是第 10.2 節的地面叢集給單顆 GPU 一條 400 Gbps 鏈路、並把延遲壓在數十微秒的原因。 |
| 容量與速率 | 16 Mb/s, 200 | 本章的服務論文在 16 Mb/s 到 200 Gbps 的鏈路假設下研究各自的流量模式。 |
| 結果與適用範圍 | 定性範圍 | 評估在軌訓練還需要另外分析叢集的對切頻寬與同步，才能涵蓋頭頂衛星接入模型之外的需求。 |
| 結果與適用範圍 | 9.3 | 唯一能延續的性質是第 9.3 節的教訓：可預測的幾何能把等待變成排程，第 11 章與第 13 章會把它用在鏈路永遠可見的編隊上。 |
Chapter summary
本章摘要
Orbital caching reduces terrestrial detours, while satellite learning benefits from more frequent aggregation contacts. Training requires a separate assessment of collective traffic, cluster connectivity and reliability.
軌道快取縮短地面繞行，衛星學習則受益於更頻繁的彙總接觸。訓練仍需另外評估集體通訊流量、叢集連線與可靠度。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 8,777 km, 138.7 ms, 7.2 ms, 50 ms | Orbital services can shorten terrestrial detours. A Starlink PoP can map users to CDN nodes 8,777 km away through its peering. Bose et al. measured 138.7 versus 7.2 ms in Mozambique and roughly 50 ms of worldwide CDN penalty. |
| Dataset and workload | 4, 9, 60%, 71%, 50 GB, 20–25%, 100, 1 | StarCDN distributes content through satellite-specific hashing into N = 4 or 9 buckets, reachable within two ISL hops. Relayed fetch retrieves content against satellite motion. At 50 GB, hit rate rises from 60% to 71%, while ground uplink falls to 20–25% of uncached volume. The trace-driven simulation assumes 100 Gbps ISLs and a four-link Gen-1 topology proxy. |
| Capacity and rate | 45 MB, 16 Mb/s, 22.5 s, 90–127 | Visibility governs the satellite-learning examples. A 45 MB model crosses a 16 Mb/s link in 22.5 s, while aggregators appear for minutes within 90–127-minute orbits. FedSat, FedSpace, AsyncFLEO and OrbitalBrain gain through contact scheduling and intra-plane ISLs, with bandwidth held fixed. |
| Power and thermal conditions | 1, 2.1 MB, 325 ms, 1.8 W | The flown inference record uses one satellite per system. Examples range from Φ-Sat-1’s 2.1 MB CNN, at 325 ms and 1.8 W, to the company-announced Starcloud-1 H100. Multi-satellite inference and training over ISLs still await flight demonstration. |
| Capacity and rate | 16 Mb/s, 200, 400 | These papers assume 16 Mb/s to 200 Gbps. These rates remain below a terrestrial GPU’s 400 Gbps and Part III’s Tbps-per-aperture regime. |
| 軌道配置 | 8,777 km, 138.7 ms, 7.2 ms, 50 ms | 服務移到軌道上，可以縮短地面路徑的繞行：Starlink 使用者的流量在 PoP 離開，而 PoP 的對等互連可能把它對應到 8,777 km 外的 CDN 節點，Bose 等人在莫三比克量到 138.7 ms 對 7.2 ms，全球則約有 50 ms 的 CDN 懲罰。 |
| 快取表現 | 4, 9, 100, 1, 50 GB, 60%, 71%, 20–25% | StarCDN 以衛星專屬的一致性雜湊回應軌道運動，把物件分到兩跳 ISL 內可達的 N = 4 或 9 個桶，再以逆著衛星運動拉取內容的中繼擷取補上，在以 100 Gbps ISL 與四鏈路 Gen-1 代理拓樸進行的軌跡驅動模擬中，把 50 GB 時的命中率從 60% 提到 71%，並把稀缺的地面到衛星上行壓到零快取流量的 20–25%。 |
| 容量與速率 | 45 MB, 16 Mb/s, 22.5 s, 90–127 | 衛星聯邦學習受可見性限制，因為 45 MB 的模型穿過 16 Mb/s 鏈路只要 22.5 s，而彙總者在 90–127 分鐘的軌道中只有幾分鐘可見，所以 FedSat、FedSpace、AsyncFLEO 與 OrbitalBrain 的勝利都來自圍繞接觸排程與加入同軌道面 ISL，實驗中的頻寬則維持固定。 |
| 電力與熱條件 | 1, 325 ms, 1.8 W, 2.1 MB | 真正飛過的機上推論都是單衛星的，從 Φ-Sat-1 以 325 ms、1.8 W 執行的 2.1 MB CNN 到 Starcloud-1 由公司公告的 H100，跨 ISL 的多衛星推論與訓練仍待在軌示範。 |
| 容量與速率 | 16 Mb/s, 200, 400 | 這些論文假設的鏈路速率從 16 Mb/s 到 200 Gbps，全都低於地面單顆 GPU 擁有的 400 Gbps，更遠低於第三部所檢視的每孔徑 Tbps 等級。 |
Where these ideas go next
這些概念接下來用在哪裡
Chapter 10 opens Part III with AI training’s network requirements and derives the all-reduce traffic that distinguishes training steps from service requests. It also establishes Figure 9.4’s terrestrial references. Chapter 11 returns to Section 9.3’s visibility lesson through continuously visible formation links. Chapter 12 examines the gap between flown single-satellite inference and distributed inference preprints.
第 10 章以「AI 訓練對網路有什麼要求」開啟第三部，推導出讓訓練步驟與本章請求具有相異需求的 all-reduce 流量，並確立圖 9.4 的鏈路之梯只是點到為止的地面參考數字。第 9.3 節的可見性教訓會在第 11 章回來，那裡的編隊讓每條鏈路永遠可見，而已飛過的單衛星推論與分散式推論預印本之間的差距，則是第 12 章的主題。
Papers behind this chapter本章所依據的論文
HotNets 2024It's a Bird? It's a Plane? It's CDN! (SpaceCDN)
SIGCOMM 2025StarCDN: Moving Content Delivery Networks to Space
IEEE WCL 2022Ground-Assisted Federated Learning in LEO Satellite Constellations
arXiv 2202.01267FedSpace: Federated Learning at Satellites and Ground Stations
NINeS 2026OrbitalBrain: Federated Learning over Inter-Satellite Links