Digest · Computing in orbit · read after Chapter 8
導讀 · 在軌運算 · 第 8 章後閱讀
In-orbit Computing: An Outlandish Thought Experiment?在軌運算：一個異想天開的思想實驗？
This digest examines the mass, power and cost of orbital servers, then explains the handover trade behind their modeled latency advantage.本導讀檢視軌道伺服器的質量、電力與成本，再說明模型延遲優勢背後的換手取捨。
If you read nothing else
只讀這一段也可以
The thought experiment combines commodity-server budgets with geometric application latency. Its handover policy trades a small delay penalty for longer session continuity.
思想實驗結合商用伺服器預算與幾何應用延遲，換手策略以小幅延遲代價換較長會話連續性。
| # | Reported values and conditions |
|---|---|
| 1 | Bhattacherjee, Kassing, Licciardello and Singla analyze one commodity server on each satellite of a proposed 40,000-satellite constellation. |
| 2 | The 64-core HPE DL325 Gen10 weighs 15.6 kg, about 6% of satellite mass and 1% of volume. |
| 3 | Its 225–350 W draw uses 15–23% of an estimated 1.5 kW solar budget. |
| 4 | Launch costs about $42,000, implying $2,700/kg, and three-year orbital cost is about three times terrestrial ownership [2, §4]. |
| 5 | On the filed 4,409-satellite constellation, the nearest satellite lies within 11 ms propagation RTT everywhere. |
| 6 | A West African meeting server improves from 46 ms Azure to 16 ms orbital delay. |
| 7 | Sticky handover gives a 164 s median interval, about four times the minimum-latency policy [2, §3, §5, Figs. 1–7]. |
| 8 | This is a paper-based thought experiment with power identified as its largest impediment. |
| # | 數值、條件與證據 |
|---|---|
| 1 | Bhattacherjee、Kassing、Licciardello 與 Singla 分析提案中 40,000 顆星座每星搭載商用伺服器。 |
| 2 | 64 核 HPE DL325 Gen10 重 15.6 kg，約占衛星質量 6%、體積 1%，225–350 W 耗電占估計 1.5 kW 太陽能的 15–23%。 |
| 3 | 發射約 42,000 美元，隱含 2,700 美元/kg，三年成本約為地面三倍 [2, §4]。 |
| 4 | 4,409 顆申報模型中，全球最近衛星傳播 RTT 皆在 11 ms 內，西非會議從 Azure 46 ms 變為在軌 16 ms。 |
| 5 | Sticky 換手間隔中位數為 164 s，約是最低延遲策略四倍 [2, §3, §5, Figs. 1–7]。 |
| 6 | 這是紙上思想實驗，作者把電力列為最大障礙。 |
1. Where this paper sits1. 這篇論文的位置
The paper extends orbital computing from sensing CubeSats to communication megaconstellations through a networking question. After HotNets 2018, CoNEXT 2019 and Hypatia studied path latency, the group asks whether satellites could also host applications. Section 8.5 uses its paired mass, launch and ownership estimates as an orbital-server example.
本文從網路問題切入，把在軌運算從感測 CubeSat 擴展到通訊巨型星座。團隊繼 HotNets 2018、CoNEXT 2019 與 Hypatia 路徑延遲研究後，追問衛星承載應用的條件。第 8.5 節將質量、發射與持有成本估計並列，作為軌道伺服器範例。
The power scale differs from OEC. Denby and Lucia model a 7.5 W Jetson on a CubeSat, while this study budgets 350 W for a 64-core Starlink-hosted server. Later BUPT-1 measurements show 5–10 W boards saturating a 17 kg satellite’s thermal path, adding flight evidence to the comparison.
兩條路線有各自電力尺度：Denby 與 Lucia 為 CubeSat 的 Jetson 編列 7.5 W，本文則為 Starlink 的 64 核伺服器編列 350 W。後續北郵一號量到 5–10 W 板子使 17 kg 衛星傳導路徑飽和，為比較加入飛行證據。
2. The problem, measured2. 問題，用量測說話
Nearest-satellite propagation sets a geometric floor. Visibility determines which ground clients can use a satellite and how often access changes.
最近衛星傳播設定幾何下限，可見性則決定地面用戶可用衛星及存取更換頻率。
| # | Reported values and conditions |
|---|---|
| 1 | The latency analysis computes constellation geometry. |
| 2 | At publication, over 400 Starlink satellites were launched and 40,000–42,000 planned. |
| 3 | The authors give 550 km orbital speed as 27,306 km/h and period 95 min 39 s. |
| 4 | Propagation is roughly 65 times shorter than GEO at 35,786 km [2, §1–3]. |
| 5 | In the 4,409-satellite filing, nearest-satellite RTT is about 4 ms at most latitudes and at most 11 ms globally. |
| 6 | The farthest directly visible satellite is within 16 ms, while Kuiper Phase I coverage ends near 60° latitude [2, Fig. 1, §3.1]. |
| 7 | Starlink has at least 30 visible satellites, typically over 40. Kuiper has at least 10 [2, Fig. 2]. |
| # | 數值、條件與證據 |
|---|---|
| 1 | 延遲分析計算星座幾何。 |
| 2 | 發表時 Starlink 已發射逾 400 顆，規劃 40,000–42,000 顆。 |
| 3 | 作者給 550 km 速度 27,306 km/h、週期 95 分 39 秒，傳播延遲約比 35,786 km GEO 短 65 倍 [2, §1–3]。 |
| 4 | 4,409 顆申報模型的最近衛星 RTT 在多數緯度約 4 ms，全球最多 11 ms，最遠直接可見衛星在 16 ms 內，Kuiper 第一期涵蓋止於約 60° [2, Fig. 1, §3.1]。 |
| 5 | Starlink 至少可見 30 顆、通常超過 40 顆，Kuiper 至少可見 10 顆 [2, Fig. 2]。 |
Orbital placement can shorten meeting paths between dispersed users. Space-native workloads also need connectivity when their hosting satellite lies beyond ground-station view.
在軌配置可縮短分散使用者的會合路徑，太空原生工作也需處理主機衛星超出地面站視線的時段。
| # | Reported values and conditions |
|---|---|
| 1 | One server per 40,000 satellites gives an edge fleet about seven times smaller than Akamai’s footprint [2, §3.1]. |
| 2 | For three West African users, an orbital meeting server reduces propagation from Azure’s 46 to 16 ms. |
| 3 | A Kuiper scenario near Azure South Central US, Brazil South and Australia East reduces 97 to 66 ms [2, Fig. 3, §3.2]. |
| 4 | With ground stations in 1,000 cities, over one third of Starlink and one half of Kuiper satellites lie beyond population-center visibility at a given time. |
| 5 | This motivates ISLs for space-native data [2, Figs. 4–5, §3.3]. |
| # | 數值、條件與證據 |
|---|---|
| 1 | 40,000 顆各配一伺服器的邊緣規模約為 Akamai 七分之一 [2, §3.1]。 |
| 2 | 三位西非使用者的會議傳播延遲從 Azure 46 降為 16 ms，靠近 South Central US、Brazil South、Australia East 的 Kuiper 情境則從 97 降至 66 ms [2, Fig. 3, §3.2]。 |
| 3 | 即使在 1,000 城設站，任一時刻逾三分之一 Starlink、二分之一 Kuiper 衛星仍超出人口中心視線，形成太空原生資料使用 ISL 的動機 [2, Figs. 4–5, §3.3]。 |
3. The design3. 設計
The hardware budget asks whether a communication satellite can support an ordinary server. Power, cooling, radiation and replacement cycles each add conditions beyond processor capacity.
硬體預算檢驗通訊衛星支援普通伺服器的條件，電力、散熱、輻射與更新週期各自增加處理器之外的需求。
| # | Reported values and conditions |
|---|---|
| 1 | The hardware budget uses a single-socket HPE ProLiant DL325 Gen10 with 64 cores at 2.4–3.35 GHz and up to 2 TB RAM [2, §4]. |
| 2 | Its 15.6 kg mass occupies about 6% of satellite weight and 1% of volume. |
| 3 | Nominal 225 W and loaded 350 W use 15% and 23% of the authors’ rough 1.5 kW solar estimate, summarized as about 20%. |
| 4 | Radiation reasoning cites software hardening below an inner-belt boundary near 643 km and the ISS’s HPE computer at 408 km. |
| 5 | Thermal treatment proposes radiators or thermoelectric harvesting as possible added mass, with sizing left open. |
| 6 | Launch costs about $42,000, against terrestrial ownership near $5,000 annually. |
| 7 | Three-year orbital cost is about threefold, with five-year satellite life mismatched to a three-year server refresh. |
| # | 數值、條件與證據 |
|---|---|
| 1 | 硬體預算採單插槽 HPE ProLiant DL325 Gen10，64 核、2.4–3.35 GHz、最多 2 TB RAM [2, §4]。 |
| 2 | 15.6 kg 約占衛星重量 6%、體積 1%，225 W 標稱、350 W 負載占粗估 1.5 kW 太陽能 15%、23%，概括為約 20%。 |
| 3 | 輻射論證引用約 643 km 內帶界線以下的軟體強化，以及 408 km 國際太空站 HPE 電腦。 |
| 4 | 熱控提出散熱板、熱電回收作額外質量，尺寸仍待決定。 |
| 5 | 發射約 42,000 美元，地面持有每年約 5,000 美元，三年在軌約三倍，另有五年衛星壽命與三年伺服器更新週期的差異。 |
Sticky favors longer remaining visibility within a bounded latency band. It reduces state-migration frequency, while migration cost itself requires a separate model.
Sticky 在有界延遲帶中偏好較長剩餘可見時間，降低狀態遷移頻率，遷移成本本身則需要獨立模型。
| # | Reported values and conditions |
|---|---|
| 1 | MinMax follows the lowest-latency satellite, requiring migration whenever another becomes faster. |
| 2 | Sticky instead admits candidates within 10% of MinMax, keeps the five with longest remaining visibility, then chooses their lowest latency [2, §5]. |
| 3 | Median handover interval becomes 164 s, about fourfold, at roughly 1.4 ms added delay for West Africa [2, Figs. 6–7]. |
| 4 | It exchanges a bounded latency penalty for fewer state migrations. |
| 5 | Chapter 3’s later measured fifteen-second schedule addresses a related continuity trade under a separate policy. |
| # | 數值、條件與證據 |
|---|---|
| 1 | MinMax 跟隨最低延遲衛星，另一顆變快時就遷移。 |
| 2 | Sticky 改納入 MinMax 10% 內候選，保留剩餘可見最久五顆，再選最低延遲 [2, §5]。 |
| 3 | 換手中位數變為 164 s、約四倍，西非增加約 1.4 ms 延遲 [2, Figs. 6–7]，以有界懲罰換較少狀態遷移。 |
| 4 | 第 3 章後來量得的十五秒排程，則在另一策略下處理相關連續性取捨。 |
Original figure note原始圖說出處
Drawn for this primer from Bhattacherjee et al., HotNets 2020, §5 and Figs. 6–7.4. Evaluation setup4. 實驗設定
The evaluation combines filed geometry, a visibility policy and a paper-based hardware budget. Each supports a different part of the feasibility argument.
評估結合申報幾何、可見性策略與紙上硬體預算，各支持可行性論證的一部分。
| # | Reported values and conditions |
|---|---|
| 1 | Evaluation computes propagation on the filed 4,409-satellite Starlink Phase I and Kuiper Phase I orbits [2, §3]. |
| 2 | Queue, satellite processing and ground-segment costs are zero. [2, §3] |
| 3 | Access links are about 10 Gbps, while ISLs are “possibly faster”, with rate, node delay and topology outside the model. |
| 4 | Nearest-satellite and meeting-point paths use direct access. |
| 5 | Fixed-ground-point visibility drives the three-step Sticky simulation [2, §5]. |
| 6 | The hardware exercise combines a DL325 datasheet, rough solar estimate, secondary Falcon 9 price and literature ownership cost near $5,000/server-year [2, §4]. |
| # | 數值、條件與證據 |
|---|---|
| 1 | 評估在 4,409 顆 Starlink 第一期與 Kuiper 第一期申報軌道計傳播 [2, §3]，佇列、衛星處理與地面段成本為零。 |
| 2 | 存取約 10 Gbps，ISL 只稱「可能更快」，速率、節點延遲與拓樸在模型之外。 |
| 3 | 最近衛星與會合點採直接存取，固定地面可見性驅動三步 Sticky 模擬 [2, §5]。 |
| 4 | 硬體練習結合 DL325 規格、粗略太陽能、二手 Falcon 9 價與文獻約每伺服器年 5,000 美元持有成本 [2, §4]。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Nearest Starlink satellite within 11 ms RTT everywhere, about 4 ms at most latitudes, farthest visible 16 ms [2, Fig. 1, §3.1].任何地方到最近 Starlink 衛星的 RTT 在 11 ms 內，大多數緯度約 4 ms，最遠可見 16 ms[2, Fig. 1, §3.1]。 | The model counts propagation over the 4,409-satellite Phase I filing. Chapter 3’s measured bent-pipe medians of 25–50 ms also include PoP paths and scheduling delays.4,409 顆 Phase I 申報下的純傳播，第 3 章實測 25–50 ms 彎管中位數則包含 PoP 與排程。 |
| West Africa meeting server: 46 ms via Azure vs 16 ms in orbit, Kuiper scenario 97 vs 66 ms [2, Fig. 3, §3.2].西非會議伺服器：經 Azure 46 ms 對在軌 16 ms，Kuiper 情境 97 對 66 ms[2, Fig. 3, §3.2]。 | The terrestrial baseline uses Azure-region distance and the orbital baseline uses a mutually visible satellite. Processing cost is zero on both sides.地面基線採 Azure 區域距離，軌道基線採共同可見衛星，兩側處理成本都設零。 |
| DL325 server is 15.6 kg ≈ 6% of satellite mass, ≈1% of volume, 225–350 W = 15–23% of ≈1.5 kW solar [2, §4].DL325 伺服器 15.6 kg 約為衛星質量 6%、體積 1%，225–350 W 為約 1.5 kW 太陽能的 15–23%[2, §4]。 | Solar output is a rough estimate. Radiator mass, cooling power and eclipse batteries require additional budgets. Chapter 8 sizes radiator area for a 700 W payload at about 1.3–1.7 m².太陽能為粗估，散熱板質量、冷卻功率與地影電池需另編預算，第 8 章為 700 W 酬載估出約 1.3–1.7 m² 散熱板。 |
| Launch ≈ $42,000, orbital server ≈ 3× terrestrial TCO over three years [2, §4].發射約 $42,000，三年下來軌道伺服器約為地面 TCO 的 3 倍[2, §4]。 | Dividing $42,000 by 15.6 kg implies about $2,700/kg. The per-kilogram source value requires clarification, and Suncatcher later uses $3,600/kg for 2025–26.42,000 美元除以 15.6 kg 隱含約 2,700 美元/kg，原每公斤來源值需釐清，Suncatcher 後續對 2025–26 採 3,600 美元/kg。 |
| Sticky handoff: median 164 s between handoffs, ≈4× MinMax, +≈1.4 ms latency [2, Figs. 6–7, §5].Sticky 換手：換手間隔中位數 164 秒，約為 MinMax 的 4 倍，延遲 +約 1.4 ms[2, Figs. 6–7, §5]。 | Fixed-ground-point visibility gives migration frequency. Transfer size, time and energy remain separate questions.固定地面點可見性給出遷移頻率，傳輸大小、時間與能量則是另外的問題。 |
| With ground stations at 1,000 cities, >1/3 of Starlink and >1/2 of Kuiper satellites are invisible to population centres [2, Figs. 4–5, §3.3].地面站設於 1,000 座城市時，>1/3 的 Starlink 與 >1/2 的 Kuiper 衛星超出人口中心視線[2, Figs. 4–5, §3.3]。 | Geometric visibility motivates ISLs for space-native data, with their operation left outside the model.幾何可見性形成太空原生資料使用 ISL 的動機，其運作則在模型範圍之外。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| What per-kilogram launch price does the paper imply?論文隱含的每公斤發射價是多少？ | \( \dfrac{\$42{,}000}{15.6\ \text{kg}} \) | ≈ $2,700/kgcomputed from $42,000 and 15.6 kg in §4, yielding an implied per-kilogram price由 §4 的 42,000 美元與 15.6 kg 計得隱含每公斤價格 |
| What share of the solar budget does the server take?伺服器佔太陽能預算多少？ | \( \dfrac{225}{1500} = 0.15,\quad \dfrac{350}{1500} \approx 0.23 \) | 15% / 23%computed, reproduces §4, the 1.5 kW is the authors' rough estimate計算值，重現 §4，1.5 kW 為作者的粗略估計 |
| What is the implied MinMax handoff interval?隱含的 MinMax 換手間隔是多少？ | \( \dfrac{164\ \text{s}}{4} \) | ≈ 41 scomputed from the 164 s median and the ≈4× ratio in §5由 §5 的 164 秒中位數與約 4 倍比值計算 |
| How many Sticky handoffs per orbit does the median imply?中位數隱含每圈多少次 Sticky 換手？ | \( \dfrac{95 \times 60 + 39}{164} \) | ≈ 35 per orbitcomputed, 95 min 39 s period from §2, treats the median as a mean計算值，95 分 39 秒的週期來自 §2，把中位數當平均值處理 |
6. Limits and open questions6. 限制與待解的問題
- The paper is a thought experiment. Propagation calculations and datasheet-based hardware estimates establish model bounds, while flight predictions require further evidence. Chapter 5 applies the same distinction to path and end-to-end delay.本文屬思想實驗，傳播計算與規格表硬體估計建立模型界限，飛行預測則需要進一步證據，第 5 章也以相同區分處理路徑與端到端延遲。
- Later measurements give idle medians of 25–50 ms and a bent-pipe floor near 40 ms from PoPs and scheduling. An orbital server removes the PoP leg while retaining access scheduling, a later-observed mechanism outside the original model.後續量測給出 25–50 ms 閒置中位數與約 40 ms 彎管下限，來源包含 PoP 與排程。在軌伺服器移除 PoP 一段，仍保留存取排程，這是原模型以後才觀察到的機制。
- The 1.5 kW solar estimate is rough, and thermal sizing remains open. A 350 W server must reject the same heat. Later 5–10 W BUPT-1 boards reached thermal limits [85], making conduction and radiator budgets necessary.1.5 kW 太陽能為粗估，熱控尺寸仍待決定，350 W 伺服器須排相同熱量。後續 5–10 W 北郵一號板子已達熱限制 [85]，因此需編列傳導與散熱板預算。
- Radiation treatment uses an ISS-based argument. Chapter 8 contrasts Jetson TX2 proton-test interruptions with zero observed computational errors over six months at 490 km. Device and exposure conditions determine the comparison.輻射處理採國際太空站先例，第 8 章對照 Jetson TX2 質子中斷與 490 km 六個月的計算錯誤零觀察，元件與暴露條件決定比較範圍。
- The implied $2,700/kg uses a secondary historical source, while Suncatcher assumes $3,600/kg. Recompute the ownership ratio with an explicit launch-price basis.隱含 2,700 美元/kg 採歷史二手來源，Suncatcher 則假設 3,600 美元/kg，持有成本比應依明確的發射價格重算。
- The model uses the 2018 filing’s 4,409 satellites, with a 24×66 shell at 550 km. The later 72×22 layout changes visibility and handover intervals. ISL behavior remains an open model component.模型採 2018 申報 4,409 顆，其 550 km 殼層為 24×66，後續 72×22 改變可見性與換手間隔，ISL 行為仍是待建立的模型元件。
- Sticky reduces migration frequency to a 164 s median interval. Bytes and time required to move server state through ISLs or ground paths remain open, connecting to Chapter 12’s inference systems.Sticky 將遷移頻率降為 164 s 中位間隔，經 ISL 或地面搬狀態所需位元組與時間仍待建立，連結第 12 章推論系統。
7. How it connects7. 與其他內容的關聯
- Section 8.5 uses this paper’s mass, launch-cost and TCO calculations as a worked example. It extends launch costs to $3,600/kg today and a $200/kg projection.第 8.5 節：本文的質量、發射成本與 TCO 算術是該章的範例，並在那裡延伸到今日的 $3,600/kg 與 $200/kg 的預測。
- Section 8.2: the 15–23% power share is the input Panel 8.1 lets you re-derive with array area and eclipse storage included.第 8.2 節：15–23% 的電力佔比是面板 8.1 讓你把陣列面積與日蝕儲能一併納入後重新推導的輸入。
- Section 8.3: radiator sizing from Stefan–Boltzmann, extending the paper’s thermal budget.第 8.3 節：依 Stefan–Boltzmann 估散熱板尺寸，擴展本文熱控預算。
- Section 8.4: the below-the-belt radiation argument against the measured SEE rates.第 8.4 節：「內帶以下」的輻射論證對照實測的 SEE 率。
- Sections 3.6 and 11.4 cover the assumed 10 Gbps downlink and a later proposal at a thousand times the power. Its “specified versus open” table repeats the thermal and ISL questions left open here.第 3.6 節與第 11.4 節：本文假設 10 Gbps 下行，後續提案在千倍功率下仍需補足熱控與 ISL 細節。
- Handley [3] supplies the propagation-only latency method, while Hypatia [6] is the same group’s simulator. OEC [60] and Kodan [83] budget cubesat-class boards from a few watts to 15 W. This paper budgets hundreds of watts. Xing et al. [85] supplies flight evidence relevant to the power argument.Handley[3]提供只計傳播的延遲方法，Hypatia[6]是同一團隊的模擬器，OEC[60]與Kodan[83]以數瓦到 15 W 的立方衛星級板子編列預算，本文則以數百瓦，Xing 等人[85]是電力論證所需要的那個飛行測試。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 15.6 kg ≈ 6% · ≈1% | HPE DL325 Gen10 mass as share of Starlink satellite mass and volume [2, §4] | HPE DL325 Gen10 的質量佔 Starlink 衛星質量與體積的比例[2, §4] | |
| 225–350 W of ≈1.5 kW | Server draw vs estimated average solar power, 15–23% [2, §4] | 伺服器功耗對估計平均太陽能，15–23%[2, §4] | |
| ≈$42,000 · ≈$2,700/kg | Launch cost of the server at Falcon 9 per-kg pricing, per-kg value computed here | 以 Falcon 9 每公斤定價的伺服器發射成本，每公斤值為本文計算 | |
| ≈3× | Orbital vs terrestrial server cost over three years, against $5,000/server/yr DC TCO [2, §4] | 三年下來軌道對地面伺服器成本，對照每台每年 $5,000 的資料中心 TCO[2, §4] | |
| ≤11 ms · ≈4 ms | Propagation RTT to the nearest Starlink Phase I satellite, worst case and typical [2, Fig. 1] | 到最近 Starlink Phase I 衛星的傳播 RTT，最差與典型[2, Fig. 1] | |
| 46 → 16 ms | West Africa meeting server, Azure vs in-orbit, propagation-only [2, Fig. 3] | 西非會議伺服器，Azure 對在軌，只計傳播[2, Fig. 3] | |
| 164 s · ≈4× | Median time between Sticky handoffs and ratio to MinMax, +≈1.4 ms latency [2, Figs. 6–7] | Sticky 換手間隔中位數與對 MinMax 的比值，延遲 +約 1.4 ms[2, Figs. 6–7] | |
| ≈10 Gbps | Assumed up/downlink rate, ISLs "possibly faster", otherwise unmodelled [2, §3] | 假設上下行速率，ISL 稱可能更快，其他行為在模型範圍之外[2, §3] | |
| 4,409 · 40,000 | Starlink Phase I filing used for latency, planned constellation size used for edge scale [2, §2–3] | 用於延遲的 Starlink Phase I 申報數，用於邊緣規模的計畫星座數[2, §2–3] |