Chapter 3 of 13 · Part I · The physics
第 3 章（共 13 章）· 第一部 · 物理
Ground links: terminals, gateways, beams, handovers and what Starlink measures like地面鏈路：終端、閘道、波束、換手，以及實測的 Starlink
Use Starlink's Ku-band waveform to calculate a rate budget. Compare the user, satellite, gateway and PoP contributions to latency. Trace per-satellite capacity estimates, from 6.5 Gbps to above 1 Tbps, to their definitions. Distinguish 15-second scheduling, dish handovers and drop-front queue effects in measured traces.本章從 Starlink 的 Ku 頻段波形計算速率預算，並比較使用者、衛星、閘道及 PoP 對延遲的貢獻。每顆衛星 6.5 Gbps 至超過 1 Tbps 的容量估計，須對照各自的定義。解讀實測軌跡時，也須區分 15 秒排程、天線換手與前端丟棄佇列的影響。
Before you start
開始之前
This chapter builds on Chapters 1 and 2. The geometry of a pass sets how long a terminal can hold a satellite and how far the signal travels. The propagation and node-delay vocabulary of the ISL chapter reappears when a path leaves the bent pipe. Each chip links to its definition.
本章建立在第 1 章與第 2 章之上。過頂的幾何決定終端能握住一顆衛星多久、訊號要走多遠，而星間鏈路那一章的傳播延遲與節點延遲詞彙，會在路徑離開彎管時再次出現。點選每個標籤即可回到它的定義。
New terms in this chapter (14)本章新名詞（14）
Each term is defined in the text where it first appears. The link jumps there.
每個名詞都在它第一次出現的段落中定義，點連結可跳過去。
- User link使用者鏈路 The radio link between a subscriber terminal and the satellite, as opposed to the gateway (feeder) link. 使用者終端與衛星之間的無線電鏈路，相對於閘道（饋線）鏈路。
- Phased array相位陣列 A flat antenna of many elements whose relative phases steer the beam electronically, using electronic beam steering. 由許多元件組成的平板天線，以元件間的相對相位電子化地轉向波束，以電子方式控制波束。
- OFDM frameOFDM 幀 The repeating 1.333 ms unit of the Starlink downlink: sync symbols, header, then data symbols over 1,024 subcarriers. Starlink 下行每 1.333 ms 重複一次的單位：同步符號、標頭，然後是分佈在 1,024 個子載波上的資料符號。
- Rain fade雨衰 Extra attenuation of a Ku- or Ka-band link while rain lies on the path, seen in Starlink as a 17–27% throughput drop. 當路徑上有雨時 Ku 或 Ka 頻段鏈路的額外衰減，在 Starlink 上表現為 17–27% 的吞吐量下降。
- Beam波束 One steerable spot of radio illumination formed by the satellite's array. A satellite forms many at once. 衛星陣列形成的一個可轉向的照射光斑。一顆衛星同時形成許多個。
- Cell服務格 The ground area whose users are served together by a beam, roughly 24 km across in Starlink. 由同一波束一起服務的地面區域，在 Starlink 中直徑約 24 km。
- Gateway閘道站 A ground station of several parabolic dishes that carries the satellite's traffic to and from the terrestrial network. 由數個拋物面天線組成的地面站，把衛星的流量送往地面網路並接回。
- Point of presence (PoP)網路接入點（PoP） The site where the operator's network hands traffic to the public Internet. Every Starlink flow is tunnelled to one. 業者網路把流量交給公共網際網路的地點。每一條 Starlink 流量都被隧道送到某一個 PoP。
- Bent pipe彎管轉送 The single-satellite path terminal → satellite → gateway in the same footprint, with ISLs excluded. 終端 → 衛星 → 同一覆蓋區內閘道的單衛星路徑，路徑排除 ISL。
- Bufferbloat緩衝區膨脹 Latency inflation caused by a large standing queue, 2–4× the idle RTT on a loaded Starlink link. 由大型常駐佇列造成的延遲膨脹，在負載下的 Starlink 鏈路上達到閒置 RTT 的 2–4 倍。
- Drop-front queue前端丟棄佇列 A queue that discards the oldest packet when full, rather than the arriving one. 滿載時丟棄最舊封包，保留新到封包的佇列。
- Reconfiguration interval重新配置週期 The 15-second period on which Starlink's global scheduler reassigns satellites and resources, at seconds :12, :27, :42 and :57. Starlink 全域排程器重新指派衛星與資源的 15 秒週期，落在每分鐘的第 12、27、42、57 秒。
- Handover換手 The moment a terminal switches from one serving satellite to another. 終端從一顆服務衛星切換到另一顆的時刻。
- Direct-to-cell (D2C)直連手機（D2C） Service from a satellite to an unmodified mobile phone in a terrestrial cellular band. 衛星以地面行動通訊頻段直接服務原有規格的手機。
3.1The user link: bands, phased arrays and the reverse-engineered waveform使用者鏈路：頻段、相位陣列與被逆向工程的波形
Recall
回顧
Section 1.2 gives 550 km satellite slant ranges of 550 km at zenith and 1,123 km at the 25° mask. Section 2.1 gives vacuum propagation cost as 3.336 µs per kilometer. Both numbers return below as the two ends of the user link's path-loss and delay range.
Every subscriber path starts with the user link. This radio connection joins the roof terminal to its assigned satellite. This is distinct from the feeder link that carries the same traffic between the satellite and a gateway on the ground. Starlink's uplink waveform, terminal transmit power and antenna element count still require source evidence. This chapter therefore describes the downlink in detail and uses measured throughput for the uplink.用戶感受到的一切，都先經過使用者鏈路，也就是屋頂上的終端與網路指派給它的那顆衛星之間的無線電鏈路，這對照於衛星與地面閘道之間承載同一份流量的饋線鏈路。Starlink 的上行波形、終端發射功率與天線元件數目仍待論文或申請文件提供證據，所以本章詳細描述下行，而上行只透過實測吞吐量來談。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Frequency | Starlink's user downlink occupies the Ku band between 10.7 and 12.7 GHz. This is the band Humphreys, Iannucci, Komodromos and Graff captured and decoded when they reverse-engineered the signal [42, §V–VII]. |
| Frequency | OneWeb lists 10.7–12.7 GHz for satellite-to-user transmission and 14.0–14.5 GHz for return traffic. Each beam has one 250 MHz forward carrier, with 32 user beams per satellite [11, pp. 4–9] FCC filing. |
| 頻率 | Starlink 的使用者下行佔用 10.7 到 12.7 GHz 的 Ku 頻段，Humphreys、Iannucci、Komodromos 與 Graff 逆向工程訊號時擷取並解碼的就是這個頻段[42, §V–VII]。 |
| 頻譜與波束配置 | OneWeb 為自家系統提交的文件列出同樣的 10.7–12.7 GHz 作為衛星到使用者的傳輸、14.0–14.5 GHz 作為使用者到衛星的回程，每個波束一個 250 MHz 的前向載波，每顆衛星 32 個使用者波束[11, pp. 4–9]FCC filing。 |
The terminal is a phased array, a flat panel containing many radiating elements. Electronic phase control steers its beam toward the satellite. This is what lets a terminal follow one satellite across the sky and then jump to the next within a fraction of a second. The gain of an ordinary subscriber panel remains proprietary. Thus, the 15 dB figure is a property of their receiver rather than of the service.終端本身是一面相位陣列，由許多小型輻射元件組成的平板，各元件的相對相位以電子方式設定，使面板的波束指向衛星而以電子方式轉向，這正是讓終端能跟著一顆衛星劃過天空、再於一秒內切換到下一顆的原因。一般用戶面板的增益仍待公開，所以 15 dB 這個數字適用於他們的接收機條件，服務本身的面板增益則仍待公開。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Terminal field of view | A Starlink dish connects within its forward field of view. Including the panel’s −22° pointing offset, Izhikevich et al. model usable azimuths below 158° or above 338° (Izhikevich et al., Proc. ACM MACS 8(1) / SIGMETRICS 2024, App. E). |
| Receiver hardware | Humphreys and colleagues used a 90 cm offset parabolic dish with roughly 3° beamwidth and nominal 40 dBi gain at 12.5 GHz. Its low-noise block had a 0.8 dB noise figure. They observed clean 4QAM clusters near 15 dB SNR. Synchronization sequences remained detectable below −6 dB [42, §II, Fig. 4, §VII.C]. |
| 評估條件 | 這面陣列的視野有限：Starlink 的天線通訊範圍限於天線前方視野，Izhikevich 等人在納入面板 −22° 的指向偏置後，把可用方位角建模為小於 158° 或大於 338°[35, App. E]。 |
| 擷取 | Humphreys 等人擷取訊號時使用一面 90 cm 的偏置拋物面天線，波束寬約 3°、在 12.5 GHz 名目增益 40 dBi，後接雜訊指數 0.8 dB 的低雜訊降頻器，憑這樣的增益他們在約 15 dB 的訊雜比下看到乾淨的 4QAM 星座群，而同步序列在 −6 dB 以下仍可偵測[42, §II, Fig. 4, §VII.C]。 |
The reverse-engineered waveform is what gives the chapter its first hard numbers. Figure 3.1 draws this structure.被逆向工程的波形給了本章第一批確實的數字。圖 3.1 畫出這個結構。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Frequency | The 2 GHz downlink spectrum contains eight 240 MHz channels spaced 250 MHz apart. The 10 MHz guard bands give spectral occupancy of 24/25. |
| Channel center frequencies | with center frequencies following \(F_{c,i} = 10.7\,\text{GHz} + F/2 + 0.25(i - \tfrac12)\,\text{GHz}\) |
| Vacant channels | and channels 1 and 2 were vacant in every observation. The authors attribute this to protecting the 10.6–10.7 GHz radio-astronomy band next door [42, §VII.A, Tab. II]. |
| Frequency | Each channel uses OFDM with N = 1,024 subcarriers spaced 234.375 kHz apart. Useful symbol duration is 4.267 µs. A 32-sample, 0.133 µs cyclic prefix gives total symbol duration of 4.4 µs. The four central subcarriers are left empty as a gutter so at most 1,020 of 1,024 carry anything [42, §VII.A, VII.D, Tab. II] . |
| Evaluation conditions | Symbols are grouped into an OFDM frame of 1.333 ms, so that exactly 750 frames pass each second. Each frame holds 302 non-zero symbols plus one vacant guard interval of 4.533 µs, of which 298 symbols carry data [42, §VI, VII.B, Tab. II]. |
| Evaluation conditions | The first symbol is a primary synchronization sequence built from eight repetitions of a 128-sample subsequence derived from a length-127 m-sequence |
| Secondary synchronization | the second is a secondary synchronization sequence |
| Frame header | symbols 2–5 form a header |
| Data modulation | symbols roughly 6 to 12 are modulated with 16QAM while almost everything else uses 4QAM |
| Coda synchronization | and two coda sync symbols close the frame at positions 300–301 |
| Synchronization density | which puts four synchronization symbols into every 1.33 ms where LTE sends two every 5 ms [42, §V.G, VI, VII.B–C]. |
| 頻率 | 2 GHz 的下行頻譜分成八個 240 MHz 的頻道，間距 250 MHz，留下 10 MHz 的護頻帶、頻譜佔用率 24/25，中心頻率遵循 \(F_{c,i} = 10.7\,\text{GHz} + F/2 + 0.25(i - \tfrac12)\,\text{GHz}\)，而第 1 與第 2 頻道在每次觀測中都是空的，作者把這歸因於保護隔鄰的 10.6–10.7 GHz 電波天文頻段[42, §VII.A, Tab. II]。 |
| 頻率 | 在一個頻道內，訊號是 N = 1,024 個子載波、間距 234.375 kHz 的正交分頻多工，有效符號 4.267 µs，循環前綴 32 個樣本即 0.133 µs，因此符號長 4.4 µs，而中央四個子載波留空作為溝槽，所以 1,024 個中最多 1,020 個承載內容[42, §VII.A, VII.D, Tab. II]。 |
| 評估條件 | 符號被組成 1.333 ms 的 OFDM 幀，每秒恰好 750 幀，每幀有 302 個非零符號加上一個 4.533 µs 的空置護衛間隔，其中 298 個符號承載資料[42, §VI, VII.B, Tab. II]。 |
| 評估條件 | 第一個符號是由長度 127 的 m 序列衍生出的 128 樣本子序列重複八次而成的主同步序列，第二個是次同步序列，第 2–5 個符號構成標頭，大約第 6 到 12 個符號用 16QAM 調變而其餘幾乎全部用 4QAM，兩個尾端同步符號在第 300–301 位置結束這一幀，使得每 1.33 ms 就有四個同步符號，而 LTE 是每 5 ms 送兩個[42, §V.G, VI, VII.B–C]。 |
Original figure note原始圖說出處
Drawn for this primer from Humphreys et al., IEEE TAES 2023, Table II and Secs. V–VII. Measured PHY rates from Garcia, Sundberg & Brunstrom, IFIP WONS 2024, Secs. III–IV.Both of those are inferences in the research notes rather than statements by any paper.這兩點都是研究筆記中的推論，其歸因屬於本教材的分析。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | The frame parameters are enough to compute what one channel can carry before error-correction coding. The arithmetic in the first calc table gives 456 Mbps for 4QAM throughout and 912 Mbps if every data symbol were 16QAM. |
| Link or data rate | Garcia, Sundberg and Brunstrom independently recovered this frame timing from packet bursts. They analyzed 422 million packets across 2,195 speed tests and found sub-bursts 1.333 ms apart. Burst sizes indicated physical-layer rates of 373.3 and 430.5 Mbps [43, §III–IV, Tab. I]. |
| Link or data rate | Those two rates sit at 82% and 94% of the raw 4QAM figure. This is the range one would expect after forward-error-correction overhead, so the decoded waveform and the traffic trace agree. |
| Capacity comparison | Raw spectral efficiency is 1.9 bit/s/Hz. Shannon’s bound permits more at the 15 dB SNR Humphreys measured with a 90 cm dish. This is about 5.0 bit/s/Hz. The gap supports two possibilities: robustness trades away capacity, or subscriber panels receive lower SNR than the 40 dBi research dish. |
| 頻道原始容量 | 幀參數已足以算出一個頻道在錯誤更正編碼之前能承載多少，第一張計算表的算式對全程 4QAM 給出 456 Mbps，若每個資料符號都是 16QAM 則為 912 Mbps。 |
| 鏈路或資料速率 | Garcia、Sundberg 與 Brunstrom 從完全相異的方向得到同樣的幀時序，他們觀察 2,195 次速度測試中 4.22 億個封包的叢發間距，發現連續子叢發相隔 1.333 ms，並從叢發大小辨識出 373.3 與 430.5 Mbps 的實體層速率[43, §III–IV, Tab. I]。 |
| 鏈路或資料速率 | 這兩個速率落在 4QAM 原始值的 82% 與 94%，正是前向錯誤更正開銷之後會預期的範圍，因此解碼出的波形與流量軌跡彼此吻合。 |
| 頻譜效率 | 原始頻譜效率 1.9 bit/s/Hz 遠低於 Shannon 界限在 Humphreys 以 90 cm 天線測得的 15 dB 下所允許的約 5.0 bit/s/Hz，這個差距意味著波形以容量換取穩健性，或者一般用戶面板看到的 SNR 低於 40 dBi 的天線。 |
The second thing the geometry of Chapter 1 fixes is the path loss. That is the physical reason the scheduler's preference for high-elevation satellites in Section 3.5 matters for capacity as well as obstruction. Rain fade is the extra Ku-band attenuation caused by rain along the path. It adds to geometric loss.第 1 章的幾何固定的第二件事是路徑損耗。這就是第 3.5 節中排程器偏好高仰角衛星之所以關乎容量也關乎遮蔽的物理原因。天氣在幾何損耗之上再加一層損耗，本教材稱之為雨衰，也就是當路徑上有雨時 Ku 頻段路徑承受的額外衰減。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Free-space path loss | The textbook equation is \(\text{FSPL}_{\text{dB}} = 20\log_{10} d_{\text{km}} + 20\log_{10} f_{\text{GHz}} + 92.45\) textbook. At 12 GHz, loss is 168.8 dB over the 550 km zenith path. The 1,123 km path at 25° gives 175.0 dB loss. Mask-edge reception is therefore 6.2 dB weaker. This is a factor of about 4.2 in power, than one looking straight up. |
| Model configuration | The ITU-R P.618 margin model requires further source review in this fact base. This chapter therefore uses measured Starlink effects. Section 3.4 reports 17–27% throughput loss and doubled web transit time in moderate rain. |
| 星座配置 | 以教科書形式 \(\text{FSPL}_{\text{dB}} = 20\log_{10} d_{\text{km}} + 20\log_{10} f_{\text{GHz}} + 92.45\) textbook 計算，12 GHz 的訊號在到天頂衛星的 550 km 上損失 168.8 dB，在 25° 遮罩處 1,123 km 的斜距上損失 175.0 dB，所以位於遮罩邊緣的終端比直視上方的終端少收到 6.2 dB，相當於功率約 4.2 倍的差距。 |
| 模型配置 | 工程師用來估算餘裕的 ITU-R P.618 模型仍待這份事實庫補充，所以本章只引用在 Starlink 上實測到的效應，第 3.4 節把它們列為 17–27% 的吞吐量下降與中雨時網頁傳輸時間加倍。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Raw bits in one frame at 4QAM4QAM 下一幀的原始位元數 | \(1{,}020 \times 298 \times 2\) | 607,920 bitcomputed. Subcarriers, data symbols and modulation from [42, Tab. II]計算值。子載波數、資料符號數與調變來自 [42, Tab. II] |
| Raw rate per 240 MHz channel, 4QAM每個 240 MHz 頻道的原始速率，4QAM | \(607{,}920 \times 750\ \text{s}^{-1}\) | 456 Mbpscomputed. 750 frames/s from the 1.333 ms frame計算值。每秒 750 幀來自 1.333 ms 的幀長 |
| Raw rate if every data symbol were 16QAM若每個資料符號皆為 16QAM 的原始速率 | \(1{,}020 \times 298 \times 4 \times 750\) | 912 Mbpscomputed. Upper bound, the observed mix is mostly 4QAM計算值。上界，實際觀測主要是 4QAM |
| Measured PHY rates as a share of raw 4QAM實測 PHY 速率佔 4QAM 原始值的比例 | \(373.3/456,\ 430.5/456\) | 82%, 94%computed. Rates from [43, Tab. I]. Remainder is FEC and control overhead計算值。速率來自 [43, Tab. I]。差額為 FEC 與控制開銷 |
| Raw spectral efficiency原始頻譜效率 | \(456\ \text{Mbps} / 240\ \text{MHz}\) | 1.9 bit/s/Hzcomputed. Shannon at 15 dB would allow log₂(1+31.6) ≈ 5.0計算值。15 dB 下 Shannon 界限允許 log₂(1+31.6) ≈ 5.0 |
| Free-space path loss at 12 GHz, nadir, 550 km12 GHz、天頂、550 km 的自由空間路徑損耗 | \(20\log_{10}550 + 20\log_{10}12 + 92.45\) | 168.8 dBcomputed. Textbook FSPL form, altitude from Chapter 1計算值。教科書 FSPL 形式，高度來自第 1 章 |
| Free-space path loss at the 25° mask, 1,123 km25° 遮罩、1,123 km 的自由空間路徑損耗 | \(20\log_{10}1{,}123 + 20\log_{10}12 + 92.45\) | 175.0 dBcomputed. Slant range from Section 1.2計算值。斜距來自第 1.2 節 |
| Power ratio between zenith and mask edge天頂與遮罩邊緣之間的功率比 | \(10^{(175.0-168.8)/10}\) | 4.2×computed. 6.2 dB計算值。6.2 dB |
| One-way propagation, zenith and mask edge單程傳播，天頂與遮罩邊緣 | \(550/c,\ 1{,}123/c\) | 1.83 ms, 3.75 mscomputed. Brackets SpaceX's stated 1.8–3.6 ms per leg (company)計算值。涵蓋 SpaceX 所述每段 1.8–3.6 ms（公司說法） |
3.2Beams, cells and per-satellite capacity波束、服務格與每顆衛星的容量
A satellite divides its coverage into simultaneously steered spots. The satellite’s phased array forms simultaneous beams. Each beam illuminates a steerable ground patch as the satellite passes, then hands that patch to another satellite. The ground area whose subscribers are served together by a beam is a cell .
The smaller hexagonal cell size often quoted elsewhere requires peer-reviewed or regulatory evidence. This primer uses the documented cell size.
一顆衛星會把覆蓋區分成多個可同時控制的照射區。它自己的相位陣列同時形成許多光斑，每個光斑是一道波束，也就是衛星在飛越時固定照在一塊地面、然後交給下一顆衛星的可轉向無線電照射區。由同一道波束一起服務的地面區域是一個服務格。
經常被轉述的更小六角形服務格尺寸，仍需同儕審查或監管來源佐證。本教材採用已有記錄的服務格尺寸。
| Measure | Values, conditions and evidence |
|---|---|
| Cell size and access | Pan, Zhao and Cai provide the peer-reviewed cell description from a year-long campaign in Victoria, British Columbia. Their rules of thumb use roughly 24 km cell diameter and 100 users per 300 km². Normally, four users share a channel. The satellite polls dishes to grant uplink access [44, §III-A]. |
| Packet-level sharing | Garcia’s study followed Pan’s by three years. Garcia, Sundberg and Brunstrom built a queue simulator from over ten million hardware-timestamped probes. Reproducing loss placement required one accessible frame followed by three inaccessible frames, each lasting 1.33 ms. This is the same four-way sharing seen from the packet side [45, §2, 6] . |
| Density and scheduling | Kassem et al. relay supportable-density estimates as low as 6 users per km² [46, §5]. SpaceX names user count per beam as a driver of fronthaul scheduling delay [47] company. |
| 項目 | 數值、條件與證據 |
|---|---|
| 服務格與存取 | 關於 Starlink 服務格唯一經同儕審查的描述來自 Pan、Zhao 與 Cai 在加拿大卑詩省維多利亞為期一年的量測，他們以經驗法則的方式陳述：使用者被分到直徑約 24 km 的服務格，規劃數字是每 300 km² 100 個使用者，通常四個使用者共用一個通訊頻道，而衛星以輪詢方式授予使用者天線上行存取[44, §III-A]。 |
| 封包層共用 | Garcia、Sundberg 與 Brunstrom 三年後根據超過一千萬個硬體時間戳探測封包建立的佇列模擬器，只有在 1.33 ms 幀上採用「一幀可存取、三幀暫停存取」的模式時才重現了觀測到的丟包位置，這與從封包側看到的四方共用是同一件事[45, §2, 6]。 |
| 密度與排程 | Kassem 等人轉述可支撐密度低至每 km² 6 個使用者的估計[46, §5]，而 Starlink 自己的工程筆記把每個波束的使用者數列為所謂前傳排程延遲的驅動因素之一[47]company。 |
Those two rules of thumb combine into a planning number. The capacity itself is the most contested number in this chapter. The table lists every claim with its tag. The panel lets you choose which one to believe.這兩條經驗法則合起來給出一個規劃數字。容量本身則是本章最具爭議的數字。下表把每項說法連同標籤列出，面板則讓你選擇要相信哪一個。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Capacity comparison | A circle 24 km across covers about 452 km², and at 100 users per 300 km² that cell holds about 150 subscribers. Thus, the per-user share of whatever capacity the satellite devotes to the cell is that capacity divided by roughly 150. This is the default Panel 3.1 starts from. |
| Model configuration | Pachler, del Portillo, Crawley and Cameron model four constellations with physical constraints. Their SpaceX inputs include usable spectrum of 1.3 GHz down and 2.1 GHz up, gateway placement and masks. SpaceX capacity is 6.5 Gbps per satellite initially and 4.76 Gbps finally. Kuiper gives 15.5–19.6 Gbps, Telesat 25.2 Gbps and OneWeb 2.01 Gbps [31]. |
| Capacity comparison | Wang et al.'s IPv6 census states a V2 Mini bandwidth of 96 Gbps per satellite against 24 Gbps for V1.5 [26, §IV-D] preprint. The 96 Gbps figure is a third party's back-calculation from SpaceX's statement that V2 Mini has more than ten times the capacity of earlier satellites. |
| Capacity comparison | SpaceX says second-generation capacity is about four times earlier models. It reported roughly 450 Tbps deployed capacity in July 2025. V3 targets exceed 1 Tbps downlink and 200 Gbps uplink per satellite [23] company. |
| Capacity accounting | Dividing 450 Tbps by roughly 7,000 operational satellites gives about 64 Gbps each. This lies between Wang’s estimates and is ten times Pachler’s result. The company may sum beam capacity with full frequency reuse. The model instead counts delivery constrained by gateways and spectrum. |
| 星座配置 | 直徑 24 km 的圓約覆蓋 452 km²，依每 300 km² 100 個使用者計算，這個服務格容納約 150 個用戶，所以衛星分配給該服務格的任何容量，每個使用者能分到的份額就是那份容量除以大約 150，這也是面板 3.1 預設的起點。 |
| 模型配置 | Pachler、del Portillo、Crawley 與 Cameron 對四個星座的物理模型，計入 SpaceX 可用頻譜下行 1.3 GHz、上行 2.1 GHz、閘道位置與仰角遮罩，得出 SpaceX 初始星座每顆衛星 6.5 Gbps、最終星座 4.76 Gbps，相對於 Kuiper 的 15.5–19.6 Gbps、Telesat 的 25.2 Gbps 與 OneWeb 的 2.01 Gbps[31]。 |
| 單星容量估算 | Wang 等人的 IPv6 普查陳述 V2 Mini 每顆衛星頻寬 96 Gbps、V1.5 為 24 Gbps[26, §IV-D]preprint，而 96 Gbps 這個數字是第三方從 SpaceX「V2 Mini 容量超過早期衛星十倍」的說法反推出來的。 |
| 公司容量聲明 | SpaceX 自己說第二代衛星容量約為早期型號的四倍，2025 年 7 月報告累計部署容量約 450 Tbps，並預期第三代 V3 衛星每顆下行超過 1 Tbps、上行 200 Gbps[23]company。 |
| 容量計算範圍 | 把 450 Tbps 除以當時約 7,000 顆運作中的衛星，得到每顆約 64 Gbps，這個數字介於 Wang 的兩個值之間、是 Pachler 模型的十倍，顯示公司是在完全頻率重用下加總波束容量，而模型計算的是閘道與頻譜允許一顆衛星實際交付的量。 |
| Per-satellite capacity claim每顆衛星的容量說法 | Value數值 | Source and tag來源與標籤 | What it counts計算的是什麼 |
|---|---|---|---|
| SpaceX Gen1, physics model, initial constellationSpaceX 第一代，物理模型，初始星座 | 6.5 Gbps | [31] peer-reviewed[31] peer-reviewed | Deliverable throughput given spectrum, gateways and 25° mask. 4.76 Gbps for the final constellation在頻譜、閘道與 25° 遮罩限制下可交付的吞吐量。最終星座為 4.76 Gbps |
| V1.5 satelliteV1.5 衛星 | 24 Gbps | [26, §IV-D] preprint[26, §IV-D] preprint | Stated as an estimate. Its derivation and an FCC or SpaceX Gen1 value remain to be established估計值。推導過程與 FCC 或 SpaceX 的第一代數值仍待補充 |
| V2 Mini satelliteV2 Mini 衛星 | 96 Gbps | [26, §IV-D] preprint[26, §IV-D] preprint | Back-calculated from SpaceX's "more than 10×" statement從 SpaceX「超過 10 倍」的說法反推 |
| V2 Mini relative to earlier modelsV2 Mini 相對於早期型號 | ≈4× | SpaceX via Via Satellite, 27 Feb 2023 companySpaceX，經 Via Satellite 2023 年 2 月 27 日轉述 company | Attributed to E-band backhaul and more powerful arrays. Absolute value pending歸因於 E 頻段回程與更強的陣列。絕對數值待補 |
| Fleet average, July 2025全艦隊平均，2025 年 7 月 | ≈64 Gbps | 450 Tbps ÷ ~7,000 operational satellites, computed from SpaceX via Converge Digest company450 Tbps ÷ 約 7,000 顆運作中衛星，由 SpaceX 經 Converge Digest 轉述的數字算出 company | Cumulative deployed capacity, definition pending累計部署容量，定義待補 |
| V3 satelliteV3 衛星 | >1 Tbps down, 200 Gbps up下行 >1 Tbps，上行 200 Gbps | SpaceX via ISPreview, 4 Jan 2025 and Converge Digest, 15 Jul 2025 companySpaceX，經 ISPreview 2025 年 1 月 4 日與 Converge Digest 2025 年 7 月 15 日轉述 company | Design target for a planned satellite awaiting service entry at the time陳述時仍待服役的衛星的設計目標 |
| OneWeb Gen1 satelliteOneWeb 第一代衛星 | 2.01 Gbps | [31] . 32 beams × one 250 MHz carrier per the filing FCC filing FCC filing[31] 。 依申請文件為 32 個波束 × 一個 250 MHz 載波 FCC filing FCC filing | Bent-pipe system, gateway-limited彎管系統，受閘道限制 |
| Modelling assumption in recent systems papers近期系統論文中的建模假設 | 96 Gbps access link. 20 Gbps ground link96 Gbps 接取鏈路。20 Gbps 對地鏈路 | [39, §6.1] . [62, Tab. 1] peer-reviewed[39, §6.1] 。 [62, Tab. 1] peer-reviewed | Assumptions chosen by the papers, with measurements treated separately論文自行選擇的假設，量測結果須另行核對 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Area of a 24 km cell直徑 24 km 服務格的面積 | \(\pi \times 12^2\) | 452 km²computed. Diameter is Pan et al.'s rule of thumb, PIMRC 2023, Sec. III-A計算值。直徑為 Pan et al. 的經驗法則，PIMRC 2023, Sec. III-A |
| Subscribers per cell at the planning density以規劃密度計的每格用戶數 | \(452 \times 100/300\) | ≈150computed. Both inputs are Pan et al.'s rules of thumb計算值。兩個輸入皆為 Pan et al. 的經驗法則 |
| Fleet-average capacity, July 2025全艦隊平均容量，2025 年 7 月 | \(450\ \text{Tbps} / 7{,}000\) | ≈64 Gbpscomputed. 450 Tbps is a company figure via Converge Digest. Satellite count approximate計算值。450 Tbps 為公司數字，經 Converge Digest 轉述。衛星數為約數 |
| Per-user share, 6.5 Gbps model, 8 cells, 150 users每使用者份額，6.5 Gbps 模型、8 格、150 人 | \(6.5\ \text{Gbps}/(8 \times 150)\) | 5.4 Mbpscomputed. The panel's starting point計算值。面板的起點 |
Stop and think停下來想一想
Which of the four numbers would you be willing to put in a paper, and with what tag?在經同儕審查的 6.5 Gbps 模型下，衛星能服務幾個格且需求維持在容量以內？四個數字中你願意把哪一個寫進論文，帶什麼標籤？
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Capacity comparison | Set 150 users per cell and 20 Mbps each, then switch between the four capacity scenarios. |
| Capacity comparison | Under the 6.5 Gbps model, how many cells can a satellite serve before demand exceeds capacity? Compare with 24 km cells and Tanveer et al.’s 35–44 visible satellites in Section 3.5. |
| 容量比較 | 把每格使用者設為 150 人、每人 20 Mbps，然後在四個容量情境之間切換。 |
| 星座配置 | 這與約 24 km 的格大小以及第 3.5 節引用 Tanveer 等人的 35–44 顆可見衛星相比如何？ |
3.3Gateways, PoPs and the bent pipe閘道、PoP 與彎管轉送
Recall
回顧
Section 2.1 gives the 72×22 shell’s intra-plane and equatorial inter-plane hops as 1,969 km and 604 km. Their one-way propagation costs are 6.57 ms and 2.01 ms. Public per-satellite switching measurements remain necessary. Bhattacharjee et al. assume 1 ms, while Chaudhry and Yanikomeroglu assume 10 ms. Those are the ingredients of the ISL leg of the budget below.
第 2.1 節算出 72×22 殼層的 ISL 跳距沿軌道面為 1,969 km、至赤道處相鄰軌道面為 604 km，各自耗費 6.57 ms 與 2.01 ms 的單程傳播，並把每顆衛星的交換延遲留為 1 ms（Bhattacharjee 等人）或 10 ms（Chaudhry 與 Yanikomeroglu）的假設，其公開量測資料仍待補充。這些就是下方預算中 ISL 段的材料。
On the ground side of the satellite sit two kinds of site. The global gateway count still requires source evidence. A point of presence, or PoP, is the site where the operator's private network hands traffic to the public Internet. Every traceroute in every study shows user packets tunnelled to the PoP assigned to the terminal before they emerge. Thus, the PoP rather than the gateway is the user's logical exit.在衛星的地面一側有兩種站點。全球閘道總數仍待來源證據。網路接入點（PoP）是業者的私有網路把流量交給公共網際網路的地點，每一份研究中的每一條 traceroute 都顯示使用者封包先被隧道送到指派給該終端的 PoP 再浮出，所以使用者在邏輯上的出口位於 PoP，閘道則負責衛星與地面的接續。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Ground infrastructure | A gateway is a ground station with several large dishes, often nine under protective domes according to Pan, Zhao and Cai. Feeder links exchange traffic with overhead satellites. Starlink addresses gateways within 172.16/12 [44, §III-B]. |
| Evaluation conditions | SpaceX reports over 1,500 gateway antennas across over 100 US sites (Converge Digest, 15 Jul 2025) company. Nine dishes at 100 sites would give about 900. Newer sites may have more dishes, or the earlier observation may be dated. |
| PoP coverage | Pan et al. listed 21 PoPs in mid-2023. These comprised seven in the US, three each in Europe and Oceania, two in Asia, and four in South America. Africa and Mexico had one each. |
| 2025 IPv6 census | Wang et al.’s IPv6 census found 49 PoPs and 98 backbone links on 29 December 2025. They served 5,975,440 active routers in 165 countries. Dallas alone served 321,065 [44, §III-C] [26, §III] preprint. |
| PoP coverage | Kassem et al. observed PoPs near Google Cloud sites in early 2022. London users moved from Google AS36492 to SpaceX AS14593 during 16–24 February 2022 [46, §3]. |
| 地面設施 | 閘道站是由數個大型拋物面天線組成的地面站，依 Pan、Zhao 與 Cai 的觀察常是九個裝在保護罩內，它透過饋線鏈路與上方衛星交換流量，在 Starlink 網路內以 172.16/12 範圍定址[44, §III-B]。 |
| 評估條件 | SpaceX 報告僅在美國就有超過 100 個站點、超過 1,500 面閘道天線（Starlink 進度報告，經 Converge Digest 2025 年 7 月 15 日轉述）company，而一百個站點各九面天線只會得到約 900 面，所以要麼較新的閘道有更多天線，要麼「常是九個」的觀察已經過時。 |
| PoP 覆蓋 | Pan 等人在 2023 年中列出 21 個 PoP（美國七個、歐洲三個、大洋洲三個、亞洲兩個、非洲與墨西哥各一個、南美洲四個） |
| 2025 年 IPv6 普查 | 而 Wang 等人的 IPv6 普查在 2025 年 12 月 29 日辨識出由 98 條骨幹鏈路連接的 49 個 PoP，服務 165 個國家的 5,975,440 台活躍使用者路由器，其中達拉斯 PoP 就服務 321,065 台[44, §III-C][26, §III]preprint。 |
| PoP 覆蓋 | Kassem 等人在 2022 年初觀察到 PoP 通常位於鄰近的 Google Cloud 機房，而倫敦使用者在 2022 年 2 月 16 至 24 日間從 Google 的 AS36492 移到 SpaceX 自己的 AS14593[46, §3]。 |
The usual bent-pipe path sends terminal traffic to a satellite, then directly to a gateway within its footprint. Terrestrial fiber carries it onward to the PoP. The space segment uses groundward links. Every independent measurement finds this to be the dominant mode.
Michel et al.’s 2022 Belgian traceroutes to San Francisco, Singapore and European anchors shared European exits. This indicated groundward routing at that terminal and time [48, §3.1]. The fraction of global traffic that crosses an ISL still requires quantitative study. The Nigerian one-third is the only estimate, and it describes a region with two ground stations.
在一般情形下把這些元件串起來的路徑是彎管轉送：終端向上發射到一顆衛星，同一顆衛星直接向下發射到其覆蓋區內的閘道，閘道再經地面光纖轉送到 PoP，整條路徑經由對地鏈路完成。每一項獨立量測都發現這是主要模式。Michel 等人 2022 年從比利時到舊金山與新加坡的 traceroute 和到歐洲錨點的流量經由同樣的歐洲出口離開，顯示當時該終端採用對地轉送路徑[48, §3.1]。全球流量中跨越 ISL 的比例仍待論文量化。奈及利亞的三分之一是唯一的估計，而它描述的是一個只有兩座地面站的地區。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| PoP coverage | In October 2023, Lagos was Africa’s sole PoP, within gateway-sparse Nigeria. Izhikevich et al. found over 70% of RTTs below a two-satellite ISL path’s theoretical minimum. This supports single-hop relay paths. During roughly one third of observations, RTT rose to 2–5× the median. Starlink engineers attributed these spikes to ISL routing (Izhikevich et al., Proc. ACM MACS 8(1) / SIGMETRICS 2024, §7.2). |
| Ground infrastructure | Mohan et al. place most terminals within about 1,200 km of their ground station. Gateway–PoP latency is about 5 ms generally and 6–8 ms in North America [49, §6.1, App. D]. |
| Propagation delay | SpaceX uses lasers for remote locations and congestion routing. It gives 1.8–3.6 ms propagation per leg and usually under 10 ms RTT. Indirect laser routes to the ground can add latency [47] company. |
| PoP 覆蓋 | Izhikevich 等人發現在閘道稀疏的奈及利亞，其拉哥斯 PoP 在 2023 年 10 月是非洲唯一的 PoP，超過 70% 的往返時間低於兩衛星 ISL 路徑的理論最小值，這意味著單跳中繼，而約三分之一的時間 RTT 以尖峰上升到中位數的 2–5 倍，Starlink 工程師確認那是 ISL 路由[35, §7.2]。 |
| 地面設施 | Mohan 等人顯示多數終端與其地面站相距約 1,200 km 以內，閘道到 PoP 的延遲幾乎到處都約 5 ms，北美為 6–8 ms[49, §6.1, App. D]。 |
| 傳播延遲 | SpaceX 把雷射描述為用於偏遠地點與壅塞繞路，把傳播定為每段 1.8–3.6 ms、往返通常低於 10 ms，並承認當流量沿間接路徑落地時雷射路徑可能增加延遲[47]company。 |
A satellite whose footprint lies outside gateway coverage must route traffic beyond the bent pipe. The measured penalty mainly reflects the landing point chosen by routing policy, beyond Chapter 2's few milliseconds of extra propagation. PoP geography accounts for much of the measured latency beyond the propagation floor.當衛星的覆蓋區位於閘道服務範圍之外，流量便須離開彎管路徑。實測延遲主要反映路由政策選定的落地位置，超出第 2 章預測的幾毫秒額外傳播。Starlink 使用者的延遲在傳播下限之外，主要反映 PoP 的地理位置。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Latency comparison | Mohan et al. measure about 100 ms in the US Virgin Islands, Réunion and Falklands, each beyond local gateway and PoP coverage. A Réunion probe reaches Frankfurt roughly 9,000 km away through long ISL chains. It still beats submarine cable by about 60 ms [49, §6.1, Figs. 13–14]. |
| PoP coverage | The Seychelles yacht’s Nigeria-PoP path had 181 ms minimum RTT. Direct ISL propagation would give about 40 ms, and zig-zag propagation at most about 80 ms. A marine customer off Fortaleza averaged 96 ms, against 14 ms theoretical delay [35, §7.2–7.3]. |
| Ground infrastructure | LENS’s Seychelles dish was over 5,000 km from its nearest gateway. Its PoP moved from Lagos to Frankfurt on 8 December 2023. Minimum RTT rose while variability fell [50, §3.2, 4.4]. Antarctica’s McMurdo station uses 97.6°-shell ISLs to Sydney’s PoP [51, §3]. |
| PoP coverage | The Manila PoP opened in May 2023, reducing Philippine domestic latency by about 90%. Earlier traffic detoured through Japan, adding 50–70 ms [49, Fig. 19]. |
| 延遲比較 | Mohan 等人在美屬維京群島、留尼旺與福克蘭群島量到約 100 ms 的延遲，三地的地面站與 PoP 均位於外地，並顯示一個留尼旺探針經由只可能是長 ISL 鏈的路徑到達約 9,000 km 外的法蘭克福 PoP，卻仍比海底電纜快約 60 ms[49, §6.1, Figs. 13–14]。 |
| PoP 覆蓋 | Izhikevich 等人在塞席爾的一艘遊艇落地於奈及利亞 PoP，最小 RTT 為 181 ms，而直接 ISL 路徑約為 40 ms、即使之字形路徑最多也只約 80 ms，另一位在福塔雷薩外海的海上用戶平均 96 ms，理論值卻是 14 ms[35, §7.2–7.3]。 |
| 地面設施 | LENS 資料集的塞席爾天線與最近地面站相距超過 5,000 km，當其 PoP 於 2023 年 12 月 8 日從拉哥斯改為法蘭克福時，最小 RTT 上升但變異性下降[50, §3.2, 4.4]，而南極麥克默多站則經由 97.6° 殼層透過 ISL 連到雪梨 PoP[51, §3]。 |
| PoP 覆蓋 | 最清楚的自然實驗是 2023 年 5 月啟用的馬尼拉 PoP，它讓菲律賓境內延遲下降約 90%，因為此前流量要繞經日本多走 50–70 ms[49, Fig. 19]。 |
Original figure note原始圖說出處
Drawn for this primer from the Starlink latency note (March 2024, company), Mohan et al., WWW 2024, §6.1 and App. D, Izhikevich et al., SIGMETRICS 2024, §7.2, and the hop lengths of Section 2.1.The budget behind both pictures can be written down. Pan et al. put the floor of the access RTT at about 20 ms and attribute it to propagation plus the MAC polling delay [44, §III-A].兩張圖背後的預算可以寫下來。Pan 等人把接取 RTT 的下限定在約 20 ms，歸因於傳播加 MAC 輪詢延遲[44, §III-A]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| PoP coverage | A bent pipe includes user and feeder legs plus gateway–PoP fiber. At 45° user elevation, 40° gateway elevation and 500 km fiber, one-way propagation is about 7.7 ms. RTT is about 15 ms. This is below every measured idle median in Section 3.4. |
| Propagation delay | SpaceX’s note identifies four latency contributors. Propagation takes 1.8–3.6 ms per leg, followed by gateway–PoP ground transport. Fronthaul scheduling depends on topology and users per beam. Processing, buffer sizing and excess drops complete the budget. In early 2024, SpaceX adjusted buffers and introduced fq_codel. Reported US peak-hour median latency fell from 48.5 to 33 ms. The 99th percentile fell from above 150 to below 65 ms [47] company. |
| Path length and hops | Each ISL adds propagation, and each satellite adds a node delay that needs public measurement. Panel 3.2 exposes that assumption. A ten-hop chain at 10 ms per node approaches the Seychelles yacht’s measured 181 ms. |
| PoP 覆蓋 | 在彎管轉送中，封包跨越使用者段、饋線段與閘道到 PoP 的光纖，所以在使用者仰角 45°、閘道仰角 40°、500 km 光纖下，單程傳播約 7.7 ms、往返約 15 ms，低於第 3.4 節中每一個實測的閒置中位數。 |
| 傳播延遲 | 其餘部分就是 SpaceX 工程筆記拆解出的四個貢獻者：每段 1.8–3.6 ms 的傳播、閘道與 PoP 之間的地面網路、由拓樸與每波束使用者數決定的前傳排程延遲，以及處理、緩衝區大小與可減少的丟棄，公司說在 2024 年初以適當大小的緩衝區與路由器上的 fq_codel 處理了後兩者，並報告美國尖峰時段中位數延遲從 48.5 降到 33 ms、第 99 百分位從超過 150 降到低於 65 ms[47]company。 |
| 路徑長度與跳數 | 當路徑離開彎管時，每一跳 ISL 加上它的傳播、每顆衛星加上仍待量測的節點延遲，面板 3.2 讓你選擇假設，並看看一條十跳、每節點 10 ms 的鏈多快就達到塞席爾遊艇實際看到的 181 ms。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Bent-pipe one-way propagation, user 45°, gateway 40°, 500 km fiber彎管單程傳播，使用者 45°、閘道 40°、500 km 光纖 | \(\tfrac{749}{c} + \tfrac{812}{c} + 500 \times 4.895\,\mu\text{s}\) | 7.7 mscomputed. Slant ranges from the Chapter 1 formula at 550 km, fiber index from Section 2.1計算值。斜距由第 1 章公式在 550 km 算出，光纖折射率來自第 2.1 節 |
| Bent-pipe round trip, same geometry同一幾何下的彎管往返 | \(2 \times 7.7\) | 15.3 mscomputed. Compare Pan et al.'s ~20 ms access floor and SpaceX's 25.7 ms US median (company)計算值。對照 Pan et al. 約 20 ms 的接取下限與 SpaceX 25.7 ms 的美國中位數（公司） |
| Three intra-plane ISL hops plus four satellites at 1 ms, one way三跳軌道面內 ISL 加四顆衛星各 1 ms，單程 | \(3 \times 6.57 + 4 \times 1\) | 23.7 mscomputed. 6.57 ms per 1,969 km hop from Section 2.1. 1 ms is Bhattacharjee et al.'s assumption計算值。每跳 1,969 km 的 6.57 ms 來自第 2.1 節。1 ms 為 Bhattacharjee et al. 的假設 |
| Same chain with 10 ms per satellite同一條鏈、每顆衛星 10 ms | \(3 \times 6.57 + 4 \times 10\) | 59.7 mscomputed. 10 ms is Chaudhry & Yanikomeroglu's assumption, OJCOMS 2022, Sec. IV計算值。10 ms 為 Chaudhry 與 Yanikomeroglu 的假設，OJCOMS 2022, Sec. IV |
Stop and think停下來想一想
What does that tell you about where the latency lives?這告訴你延遲住在哪裡？
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Path length and hops | Leave the hop count at zero and move the two elevation sliders from 25° to 90°. |
| Evaluation conditions | Geometric changes affect RTT by only a few milliseconds. Measured medians fell from 46–52 ms in 2022 to SpaceX’s 25.7 ms in 2025. |
| Path length and hops | Set 10 hops at 1,970 km and 10 ms node delay. Explain how the chain and PoP location contribute to the Seychelles 181 ms. |
| 路徑長度與跳數 | 把跳數留在零，把兩個仰角滑桿從 25° 移到 90°。 |
| 評估條件 | 往返最多只變化幾毫秒，而實測中位數卻從 2022 年的 46–52 ms 降到 SpaceX 2025 年數字的 25.7 ms。 |
| 路徑長度與跳數 | 接著設定 10 跳、每跳 1,970 km、每節點 10 ms：光是這條鏈就能解釋塞席爾的 181 ms，還是結果仍取決於 PoP 在哪裡？ |
3.4Measured performance 2022–20252022–2025 年的實測效能
Starlink measurements began in late 2021 and now span a dozen campaigns. Each result requires its observation point, year, target and dish because these conditions define the quantity measured. A Belgian dish's 2021 RTT to a German RIPE Atlas anchor describes an end-to-end path. SpaceX's 2025 median for millions of US terminals describes router-to-PoP paths. Their average would combine separate measurement scopes. The study-level results support the trends discussed below.
Starlink 的量測文獻始於 2021 年底，如今涵蓋十多個量測活動。每筆結果均須附上觀測點、年份、目標與天線，這些條件共同定義所量測的數量。2021 年比利時天線到德國 RIPE Atlas 錨點的 RTT 中位數，描述一條端到端路徑。SpaceX 2025 年數百萬台美國終端的中位數，描述路由器到 PoP 的路徑。將兩者平均會混合各異的量測範圍，因此下文以各研究的結果解讀趨勢。
| Study (venue, year of data)研究（場合、資料年份） | Setup量測設定 | Idle RTT median閒置 RTT 中位數 | Down / up下行 / 上行 | Loss丟包 |
|---|---|---|---|---|
| [48] (Dec 2021–Apr 2022)[48]（2021 年 12 月–2022 年 4 月） | One dish, Louvain-la-Neuve, Belgium. 11 Atlas anchors, Ookla, QUIC H3 to campus server比利時新魯汶一台天線。11 個 Atlas 錨點、Ookla、QUIC H3 至校園伺服器 | 46–52 ms to Belgian anchors (min 24–28 ms). ~42 ms Germany. Under H3 load 95 ms down, 104 ms up至比利時錨點 46–52 ms（最小 24–28 ms）。德國約 42 ms。H3 負載下下行 95 ms、上行 104 ms | 178 / 17 Mbit/s (Ookla medians)178 / 17 Mbit/s（Ookla 中位數） | 1.56% down, 1.96% up in bulk. 0.40% light load (Table 2)大量傳輸下下行 1.56%、上行 1.96%。輕負載 0.40%（Table 2） |
| [46] (Dec 2021–mid 2022)[46]（2021 年 12 月–2022 年中） | Browser extension, 18 users in 10 cities. 3 Raspberry Pi nodes (North Carolina, Barcelona, Wiltshire)瀏覽器擴充套件，10 個城市 18 位使用者。3 台 Raspberry Pi 節點（北卡羅來納、巴塞隆納、威爾特郡） | Wireless-link queueing delay medians 48.3 / 24.3 / 16.5 ms (NC / London / Barcelona, Table 2)無線鏈路排隊延遲中位數 48.3 / 24.3 / 16.5 ms（北卡 / 倫敦 / 巴塞隆納，Table 2） | London 123.2 / 11.3. Seattle 90.3 / 6.6. Toronto 65.8 / 6.9 Mbps (Table 3)倫敦 123.2 / 11.3。西雅圖 90.3 / 6.6。多倫多 65.8 / 6.9 Mbps（Table 3） | Clusters up to 50%. ~12% of iperf tests above 5% (Figs. 6c, 7)叢集最高達 50%。約 12% 的 iperf 測試超過 5%（Figs. 6c, 7） |
| [53] (2022)[53] （2022 年） | Dishes in the Vancouver area against a cable line. Stationary and in-motion溫哥華地區天線對照有線寬頻。靜止與移動中 | ~50 ms stationary, ~10% above cable (Sec. IV-A)靜止約 50 ms，比有線高約 10%（Sec. IV-A） | TCP ~80 / 7. UDP ~200 / 10 Mb/s (Table II)TCP 約 80 / 7。UDP 約 200 / 10 Mb/s（Table II） | 0.24% down, 1.24% up. Storm cut download 100 → 5 Mb/s (Sec. IV-B, IV-D)下行 0.24%、上行 1.24%。地磁風暴使下載從 100 降到 5 Mb/s（Sec. IV-B, IV-D） |
| [44] (2022–23)[44] （2022–23 年） | Several owned dishes plus about a dozen user dishes, Victoria BC, almost a year數台自有天線加約十二台使用者天線，卑詩省維多利亞，近一年 | Minimum access RTT ~20 ms (Sec. III-A). Median values pending最小接取 RTT 約 20 ms（Sec. III-A）。中位數待補 | Outside the reported table表列資料以外 | Reporting pending報告資料待補 |
| [35] (May–Jun 2023)[35] （2023 年 5–6 月） | ~2,400 customer IPs with exposed services in 13 countries, 20 PoP cities. Validated against one dish13 個國家、20 個 PoP 城市約 2,400 個有暴露服務的用戶 IP。以一台天線驗證 | Means: Mexico 28 ms, Georgia US 58 ms, Brazil 104 ms, Nigeria 149 ms. Seychelles minimum 181 ms (Fig. 11, §7.2)平均：墨西哥 28 ms、美國喬治亞 58 ms、巴西 104 ms、奈及利亞 149 ms。塞席爾最小 181 ms（Fig. 11, §7.2） | Outside the measurement scope (latency only)量測範圍以外（僅延遲） | Zero observed drops during RTT spikes on the ground-truth dish (§6.1.1)基準天線在 RTT 尖峰期間觀測到的丟包為零（§6.1.1） |
| [49] (Dec 2022–Sep 2023)[49]（2022 年 12 月–2023 年 9 月） | 98 RIPE Atlas probes in 21 countries. 19.2M M-Lab tests from 34 countries. Two controlled dishes21 個國家 98 個 RIPE Atlas 探針。34 個國家 1,920 萬次 M-Lab 測試。兩台受控天線 | minRTT 40–50 ms in most countries. Bent-pipe 36–48 ms, median ~40 ms. ~100 ms at ISL islands. 2–4× under load (Figs. 5–8, 12)多數國家 minRTT 40–50 ms。彎管 36–48 ms、中位數約 40 ms。ISL 島嶼約 100 ms。負載下 2–4 倍（Figs. 5–8, 12） | p75 ~50–100 / 4–12 Mbps, M-Lab (Fig. 9)p75 約 50–100 / 4–12 Mbps，M-Lab（Fig. 9） | p75 TCP loss ~4–8% (Fig. 21)p75 TCP 丟包約 4–8%（Fig. 21） |
| [13] (2023)[13] （2023 年） | Four dishes (W. Europe, NE/Midwest/NW US). IRTT at 20 ms to PoP-colocated servers四台天線（西歐、美國東北/中西/西北）。每 20 ms 一次 iRTT 至 PoP 同址伺服器 | Outside the study focus. Regime shifts every 15 s at :12/:27/:42/:57 (Fig. 2)研究焦點以外。每 15 s 在 :12/:27/:42/:57 出現狀態轉移（Fig. 2） | Reporting pending報告資料待補 | Reporting pending報告資料待補 |
| WetLinks, [54] (Sep 2023–Mar 2024) preprintWetLinks, [54]（2023 年 9 月–2024 年 3 月）preprint | Two dishes, Osnabrück and Enschede, ~140,000 measurements with weather stations兩台天線，奧斯納布魯克與恩斯赫德，約 14 萬次量測並配合氣象站 | 62 ± 3 ms end to end. First hop 31–35 ms mean (Sec. V-C)端到端 62 ± 3 ms。第一跳平均 31–35 ms（Sec. V-C） | 180–250 / 13–17 Mbit/s hourly medians (Sec. V-A)每小時中位數 180–250 / 13–17 Mbit/s（Sec. V-A） | Mean 1.0%, median 0.4%. ~65% of loss events single packets (Table II)平均 1.0%、中位數 0.4%。約 65% 的丟包事件為單一封包（Table II） |
| [63] company[63] company | Router telemetry every 15 s from millions of US terminals, peak hours 6–9 PM數百萬台美國終端每 15 s 一次的路由器遙測，尖峰時段晚間 6–9 點 | 48.5 → 33 ms. P99 >150 → <65 ms after buffer changes48.5 → 33 ms。緩衝區調整後 p99 從 >150 降到 <65 ms | Specification pending說明資料待補 | Specification pending說明資料待補 |
| SpaceX progress report, July 2025 companySpaceX 進度報告，2025 年 7 月 company | US fleet-wide telemetry, peak hours美國全網遙測，尖峰時段 | 25.7 ms. <1% of samples above 55 ms25.7 ms。<1% 的樣本超過 55 ms | ~200 Mbps median download下載中位數約 200 Mbps | Specification pending說明資料待補 |
| [64] (2025) preprint[64] （2025 年） preprint | Residential terminal in Pisa. In-flight terminals over the Baltic and the Pacific比薩的住宅終端。波羅的海與太平洋上空的機上終端 | 22 ms Pisa. 23 ms in flight to Frankfurt. 62 ms Pacific to Tokyo比薩 22 ms。機上至法蘭克福 23 ms。太平洋至東京 62 ms | 188 / ~30 Mbps Pisa. 64 / 24 in flight比薩 188 / 約 30 Mbps。機上 64 / 24 | Packet delivery ratio >99.4%封包送達率 >99.4% |
| [45] (May and Oct 2025)[45] （2025 年 5 月與 10 月） | >10M hardware-timestamped UDP probes in 120 burst patterns, Karlstad. Check site Málaga. Downlink only卡爾斯塔德超過一千萬個硬體時間戳 UDP 探測封包、120 種叢發模式。馬拉加為對照站。僅下行 | Outside the study focus. Queueing plateau 30–60 ms. 1,500-packet queue at 124 Mbps: 85 ms mean delay drop-front vs 122 ms drop-tail (§6–7)研究焦點以外。排隊高原 30–60 ms。124 Mbps 下 1,500 封包佇列：前端丟棄平均延遲 85 ms 對尾端丟棄 122 ms（§6–7） | Fitted drain rates 33–218 Mbps. Zero loss at 48 and 96 Mbps sends擬合排空速率 33–218 Mbps。以 48 與 96 Mbps 發送時零丟包 | 6,000-packet burst at 500 Mbps: 55% loss, 87 ms one-way delay (§5)500 Mbps 的 6,000 封包叢發：55% 丟包、單程延遲 87 ms（§5） |
| [51] (Nov 2025–Apr 2026)[51]（2025 年 11 月–2026 年 4 月） | Teams and instrumented WebRTC from Munich, Delft, Oslo and McMurdo. Uplink only從慕尼黑、台夫特、奧斯陸與麥克默多執行 Teams 與儀器化 WebRTC。僅上行 | Outside the study focus. ≥89% of handovers within ±1 s of a reconfiguration at the European sites (§4.2)研究焦點以外。歐洲站點 ≥89% 的換手落在重新配置的 ±1 s 內（§4.2） | Saturating uplink flows reach ~30 Mbps飽和的上行流量達約 30 Mbps | Loss in 1.0% of 100 ms intervals at 2.5 Mbps, 3.5% at 10 Mbps (App. B.5)2.5 Mbps 時 1.0% 的 100 ms 區間有丟包，10 Mbps 時為 3.5%（App. B.5） |
The latency results show a trend at the level of individual studies.沿延遲欄往下讀，各研究結果本身即可顯示趨勢。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Path length and hops | Belgian idle medians were 46–52 ms in 2021–22. Most of Mohan et al.’s 21 countries had 40–50 ms medians in 2023. |
| WetLinks latency | WetLinks measures 62 ms end to end but only 31–35 ms to the first hop in 2023–24 |
| SpaceX latency statements | SpaceX reports 33 ms for US peak hours in March 2024 and 25.7 ms in July 2025 |
| Pisa latency | and Ullah et al. see 22 ms from Pisa in 2025. |
| PoP coverage | The independent academic medians lag SpaceX's self-reported figures by roughly 10–20 ms throughout. Measurement endpoints explain part of this difference. Academic RIPE and CDN targets lie beyond the PoP, while SpaceX measures router-to-PoP latency. Company-described buffer and scheduling changes explain another part. |
| Evaluation conditions | Bose et al.’s 2025 content-target comparison gives 19–20 ms median for terrestrial ISPs and about 50 ms for Starlink. Chapter 9 uses this endpoint definition [52] preprint. |
| 路徑長度與跳數 | 2021–22 年比利時 46–52 ms 的閒置中位數，到 2023 年在 Mohan 等人的 21 個國家中大多成為 40–50 ms |
| WetLinks 延遲 | WetLinks 在 2023–24 年量到端到端 62 ms 但到第一跳只有 31–35 ms |
| SpaceX 延遲聲明 | SpaceX 報告 2024 年 3 月美國尖峰時段 33 ms、2025 年 7 月 25.7 ms |
| 比薩延遲 | 而 Ullah 等人 2025 年從比薩看到 22 ms。 |
| PoP 覆蓋 | 獨立的學術中位數始終比 SpaceX 自報的數字落後約 10–20 ms，這有一部分是量測上的人為差異，兩者的量測條件各有定義，因為 RIPE 錨點與 CDN 邊緣等學術目標位於 PoP 之外，而 SpaceX 量的是從路由器到 PoP，另一部分則是公司所述緩衝區與排程器變更的真實效果。 |
| 評估條件 | Bose 等人 2025 年的比較發現地面 ISP 到內容的中位數為 19–20 ms、Starlink 約 50 ms，用的又是另一種目標，屬於第 9 章的內容[52]preprint。 |
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Throughput | The throughput column shows an asymmetry that every study reproduces at close to ten to one: |
| Belgium: downlink and uplink | 178 against 17 Mbit/s in Belgium |
| London: downlink and uplink | 123.2 against 11.3 in London |
| Seattle: downlink and uplink | and 90.3 against 6.6 in Seattle |
| Vancouver: TCP and UDP | about 80 against 7 Mb/s over TCP and 200 against 10 over UDP in Vancouver |
| Germany and Netherlands | 180–250 against 13–17 in Germany and the Netherlands |
| Pisa: downlink and uplink | and 188 against about 30 in Pisa in 2025 |
| Fries et al.: saturated uplink | while Fries et al. find uplink flows that saturate their allocation at about 30 Mbps in 2025–26. |
| Evaluation conditions | Uploads roughly doubled, narrowing the ratio, while design still favors downlink. Satellite polling grants uplink access through a shared, request-driven channel. Downlink uses continuously scheduled frames. SpaceX’s V3 target retains roughly five-to-one asymmetry: over 1 Tbps down and 200 Gbps up (Converge Digest, 15 Jul 2025) company. |
| Evaluation conditions | Michel et al. and WetLinks report idle or light-load loss below 0.5%. Bulk transfer raises this to 1.5–2%, often in clusters. Michel et al. recorded 244,008 events, with 49 µs median duration and 7.5 ms at p99. A wired Amsterdam control lost 10 packets among over 5.8 million. This locates loss within Starlink [48, §3.2, Tab. 2, Fig. 4]. Kassem et al. observed clusters reaching 50% near the serving satellite’s departure from view [46]. Section 3.5 revisits that interpretation. |
| 吞吐量 | 吞吐量欄顯示每份研究都重現、接近十比一的上下行速率差距： |
| 比利時：下行與上行 | 比利時 178 對 17 Mbit/s |
| 倫敦：下行與上行 | 倫敦 123.2 對 11.3、 |
| 西雅圖：下行與上行 | 西雅圖 90.3 對 6.6 |
| 溫哥華：TCP 與 UDP | 溫哥華 TCP 約 80 對 7 Mb/s、UDP 200 對 10 |
| 德國與荷蘭 | 德國與荷蘭 180–250 對 13–17 |
| 比薩：下行與上行 | 比薩 2025 年 188 對約 30 |
| Fries 等人：飽和上行 | 而 Fries 等人在 2025–26 年發現飽和其配額的上行流量約為 30 Mbps。 |
| 上行與下行容量 | 隨著上傳大致翻倍，這個比例已經縮小，但它內建在設計中，因為 Pan 等人描述的衛星輪詢天線以授予上行存取是一條分時、請求驅動的頻道，而下行是連續排程的幀，SpaceX 自己 V3 下行超過 1 Tbps 對上行 200 Gbps 的目標也在衛星端維持五比一（Converge Digest, 2025 年 7 月 15 日）company。 |
| 封包遺失統計 | 閒置或輕負載狀態下的丟包在 Michel 等人與 WetLinks 中低於 0.5%，大量傳輸時升到 1.5–2%，而且成叢出現：Michel 等人記錄 244,008 次丟包事件，中位持續時間 49 µs、第 99 百分位 7.5 ms，而經有線路徑到阿姆斯特丹的對照在超過 580 萬個封包中只丟 10 個，這把丟包定位在 Starlink 內部[48, §3.2, Tab. 2, Fig. 4]，Kassem 等人則看到高達 50% 的叢集與服務衛星離開視線同時發生[46]，第 3.5 節會重新檢視這個詮釋。 |
Garcia, Sundberg and Brunstrom characterize a single shared downlink bottleneck queue per user. Sharing is determined at user level.
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Latency comparison | Bufferbloat is latency growth caused by a large standing bottleneck queue. Mohan et al. measure download RTT rising two- to fourfold, reaching about 400–500 ms. The 60th percentile stays near 100 ms under upload and reaches about 200 ms under download [49, §4, Fig. 8]. |
| Round-trip time | Michel et al.’s QUIC bulk-download RTT has median 95 ms, p95 175 ms and p99 210 ms. At a light 3 Mbit/s message load, these values are 50, 71 and 87 ms [48, Fig. 3]. |
| Capacity comparison | Ten CUBIC and ten BBRv1 flows retain unequal shares. Queue capacity is about 1,400–1,600 packets and is packet-limited. Drops follow queue length rather than delay. This supports a length-based policy outside CoDel-like active queue management. The discipline is drop-front , which discards the oldest packet when the queue is full rather than the arriving one [45, §4, §6–7] . |
| Queue behavior | At 124 Mbps, their 1,500-packet simulator gives 85 ms mean delay with drop-front and 122 ms with drop-tail. Drop-front reduces delivered-packet delay and sends loss signals earlier. Chapter 6 examines the congestion-control implications. |
| 佇列行為 | 在負載下鏈路表現出緩衝區膨脹，也就是瓶頸前方的大型常駐佇列把延遲推高，Mohan 等人量到下載期間 RTT 上升兩到四倍、達到約 400–500 ms，上傳負載下第 60 百分位維持在約 100 ms，而下載負載下則達到約 200 ms[49, §4, Fig. 8]。 |
| 往返時間 | Michel 等人在 QUIC 大量下載下的每封包 RTT 中位數為 95 ms、第 95 百分位 175 ms、第 99 百分位 210 ms，相對於 3 Mbit/s 輕量訊息負載下的 50、71 與 87 ms[48, Fig. 3]。 |
| 佇列行為 | Garcia、Sundberg 與 Brunstrom 接著刻畫了佇列本身：下行瓶頸表現為每個使用者一條共用佇列、公平份額的分配單位為使用者，十條 CUBIC 與十條 BBRv1 流量仍維持相異份額，其容量以封包為計數單位、約 1,400–1,600 個封包，丟棄決策依佇列長度決定，屬於長度導向策略，與類似 CoDel 的延遲導向主動佇列管理各有機制，而其規則是前端丟棄，即佇列滿時丟棄最舊的封包，新到封包則進入佇列[45, §4, §6–7]。 |
| 佇列行為 | 在他們的模擬器中，同一條 1,500 封包的佇列以 124 Mbps 排空時，前端丟棄的平均延遲為 85 ms、尾端丟棄為 122 ms，所以這個選擇縮短了發送端感受到的佇列，卻更早送出丟包訊號，第 6 章把它當作壅塞控制問題來處理。 |
Weather closes the list. Richter et al. report apparently negligible geomagnetic effects on latency [55]. This measures a separate quantity from Ma et al.’s throughput result and remains compatible with it.天氣結束這份清單。Richter 等人報告地磁風暴強度與延遲的關聯似乎很弱[55]，此結果與 Ma 等人的發現相容，兩者各自量測一種數量。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Throughput | Ma et al. report about 27% mean throughput loss during precipitation. UDP download falls from about 215 Mb/s dry to 120 Mb/s at 4.1–5.2 mm rain. During the 3–4 February 2022 geomagnetic storm, download fell from about 100 to 5 Mb/s. Upload remained largely stable [53, §IV-D]. |
| Ground infrastructure | WetLinks places weather stations beside both dishes. Rain-window median download is 181.5 Mbit/s, against 218.1 dry, roughly 17% lower. The 4–5 mm bucket gives roughly 30% loss, while upload correlation is insignificant [54, §V-E] preprint. Kassem et al.’s London median page transit time rises from 470.5 ms clear to 931.5 ms in moderate rain [46, Fig. 4]. |
| 吞吐量 | Ma 等人量到降水期間吞吐量平均下降約 27%，UDP 下載從乾燥時約 215 Mb/s 降到 4.1–5.2 mm 雨量時約 120 Mb/s，並記錄 2022 年 2 月 3–4 日的一次地磁風暴使下載從約 100 降到 5 Mb/s，而上傳則大致維持原水準[53, §IV-D]。 |
| 評估條件 | WetLinks 在兩台天線旁都設有氣象站，發現雨窗內的下載中位數為 181.5 Mbit/s、乾燥時為 218.1，約低 17%，在 4–5 mm 的區間約低 30%，而上傳的相關性則低於顯著門檻[54, §V-E]preprint，Kassem 等人來自倫敦的瀏覽器資料則顯示網頁傳輸時間中位數從晴空的 470.5 ms 加倍到中雨的 931.5 ms[46, Fig. 4]。 |
3.5The 15-second schedule and handovers15 秒排程與換手
Measurements show that Starlink's latency changes in steps on a fixed clock, alongside the effects of orbital geometry. This primer calls the 15 s period the reconfiguration interval , the cadence on which a global controller reassigns satellites, beams and resources to terminals. Mohan et al. observed simultaneous boundaries in Edinburgh and Munich, establishing a global clock. They also saw shifts with the dish restricted to one usable satellite. Thus, the boundary effect occurs even while the serving satellite stays fixed [49, §6.2, Fig. 16].量測顯示，Starlink 延遲除了受到軌道幾何影響，也會依固定時鐘呈現階躍變化。本教材把這 15 s 的週期稱為重新配置週期，即全域控制器把衛星、波束與資源重新指派給終端的節拍。Mohan 等人在愛丁堡與慕尼黑同時觀察到同樣的邊界，確立了這個時鐘是全球性的，並用一台視野被遮蔽到只剩一顆可用衛星的天線看到同樣的轉移，這排除了衛星更換是邊界效應本身的原因[49, §6.2, Fig. 16]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| PoP coverage | Tanveer, Puchol, Singh, Bianchi and Nithyanand used four terminals in Western Europe and the northeastern, midwestern and northwestern United States. They sent probes every 20 ms to servers at the serving PoP. Dish obstruction maps were polled every 15 s. XOR differences and dynamic time warping matched tracks to CelesTrak ephemerides. Latency regimes changed every 15 s at :12, :27, :42 and :57 across all sites. Consecutive windows differed under a Mann–Whitney test at p < .05. Parallel latency bands within each window were a few milliseconds apart. The authors attribute these bands to satellite round-robin frame allocation [13, §3–4, Fig. 2]. |
| Capacity comparison | Lai et al. measure capacity changes exceeding 10 Mbps at over 80% of reconfigurations. Roughly one fifth change capacity by at least 40 Mbps [56, Fig. 7]. |
| Constellation configuration | Terminals see 35–44 satellites above the 25° mask in any slot [13, §3], so the scheduler has real choice at every tick. |
| PoP 覆蓋 | Tanveer、Puchol、Singh、Bianchi 與 Nithyanand 運行四台終端，一台在西歐、三台在美國東北、中西與西北部，每 20 ms 向服務 PoP 同址的伺服器送出一個探測封包，同時每 15 s 輪詢每台天線的遮蔽圖，並以連續遮蔽圖的 XOR 差分與動態時間規整把掃過的軌跡對應到 CelesTrak 星曆，他們發現延遲在每分鐘第 12、27、42 與 57 秒、每 15 s 於每個觀測點進入新狀態，連續視窗在 Mann–Whitney 檢定下於 p < .05 統計上相異，而視窗內的延遲形成相隔幾毫秒的平行帶，他們歸因於衛星上的輪詢式幀分配[13, §3–4, Fig. 2]。 |
| 容量比較 | Lai 等人接著量化邊界上什麼在變：超過 80% 的重新配置讓可用容量改變超過 10 Mbps，約五分之一改變 40 Mbps 以上[56, Fig. 7]。 |
| 星座配置 | 終端在任一時槽都能看到 25° 遮罩以上的 35–44 顆衛星[13, §3]，所以排程器在每一個節拍都有真實的選擇。 |
Once the boundary regions are excluded, period-level models predict the 99th-percentile latency to within a few milliseconds after about a second of samples. This is the observation that makes the schedule exploitable rather than merely annoying.一旦排除邊界區域，週期層級的模型在約一秒的樣本後就能把第 99 百分位延遲預測到幾毫秒之內，正是這個觀察讓排程成為可用於預測與調適的條件。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| PoP coverage | Casparsen, Jakobsen, Nielsen, Popovski and Leyva-Mayorga measured high-resolution boundary behavior in Aalborg. Client and server shared a university, allowing submillisecond clock synchronization. One-way probes ran every 2 ms, or 500 Hz. Higher rates caused probe loss. |
| Scheduling boundaries and delay | Across months of 15 s periods, the initial spike averages about 74 ms above period mean and decays within 140 ms. A secondary rise occurs during the final 75 ms. Uplink tails exceed downlink tails, and phase alignment persists across months [57, Fig. 3] preprint. |
| Evaluation conditions | The abstract calls these spikes "about 140 ms at the start and 75 ms at the end". These are window durations, distinct from spike height. |
| Evaluation conditions | The figure supports the reading used here, a ~74 ms excursion confined to the first 140 ms. |
| PoP 覆蓋 | 邊界本身在高時間解析度下的樣子來自 Casparsen、Jakobsen、Nielsen、Popovski 與 Leyva-Mayorga，他們把用戶端與伺服器放在奧爾堡同一所大學，使時鐘能同步到遠低於一毫秒，並每 2 ms 探測一次單程延遲，這是 500 Hz 的速率，再高鏈路就開始丟失探測封包。 |
| 排程邊界與延遲 | 把數月的軌跡切成 15 s 的週期後，他們量到週期開始處一個平均比該週期平均延遲高約 74 ms、在最初 140 ms 內衰減的尖銳尖峰，下一個邊界前最後 75 ms 的一次次要上升，上行比下行更重的尾部，以及跨越數月持續的相位對齊[57, Fig. 3]preprint。 |
| 軌道高度 | 他們的摘要把同一結果寫成「開頭約 140 ms、結尾 75 ms」的尖峰，把視窗長度與尖峰高度混在一起。 |
| 評估條件 | 圖支持這裡採用的讀法，即限於最初 140 ms 內約 74 ms 的偏移。 |
A handover is the moment a terminal switches from one serving satellite to another. It is a different event from a reconfiguration even though the two are synchronised. Figure 3.3 lays one minute of all three on a single timeline.換手是終端從一顆服務衛星切換到另一顆的時刻，它與重新配置是相異的事件，雖然兩者同步。圖 3.3 把這三者的一分鐘放在同一條時間軸上。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Handover and scheduling | Fries, Matezki, Bose and Mohan detect handovers through angular shifts between gRPC obstruction maps, polled once per second. Sites are Munich, Delft, Oslo and McMurdo. Regular handovers fall within ±1 s of :12/:27/:42/:57 boundaries. At least 89% are regular at the European sites. McMurdo uses the 97.6° shell and ISLs to Sydney. There, 208 of 678 handovers, or 30.7%, are irregular [51, §4.2] . |
| Handover effects | Their satellites remain visible for 5–10 min before handover. Ma et al. similarly observe roughly 10 min dwell per spot. Latency rises as the satellite recedes, then resets at handover [53, §III-A, IV-A]. Teams uplink bitrate targets fall about 0.8 Mbps within 1–2 s after reconfiguration. Recovery over the next 10 s is partial (Fries et al., §4.1). |
| Handover and scheduling | Kassem et al. in 2022 had attributed their 50% loss clusters to handovers because the scheduler remained to be identified. Three mechanisms shape these events. Scheduled reconfiguration creates fifteen-second spikes, with satellite changes as a separate event. Dish handovers align with boundaries at mid-latitudes and often occur between boundaries in Antarctica. Section 3.4’s drop-front queue converts either event into loss when full. |
| 換手與排程 | Fries、Matezki、Bose 與 Mohan 從慕尼黑、台夫特、奧斯陸與麥克默多站每秒輪詢一次的連續 gRPC 遮蔽圖之間的角度位移偵測換手，並把落在 :12/:27/:42/:57 邊界 ±1 s 內的換手歸類為規律：三個歐洲站點至少 89% 的換手是規律的，而由 97.6° 殼層經 ISL 連到雪梨服務的麥克默多，678 次換手中有 208 次即 30.7% 是偏離排程邊界的[51, §4.2]。 |
| 換手與排程 | 在他們的資料中衛星在換手前保持可見 5–10 分鐘，與 Ma 等人觀察到每個位置約 10 分鐘的駐留一致，期間延遲隨衛星遠去逐漸上升直到換手把它重設[53, §III-A, IV-A]，而一通 Teams 通話的上行位元率目標在重新配置後 1–2 s 內下降約 0.8 Mbps，並在接下來 10 s 內只部分恢復（Fries et al., §4.1）。 |
| 換手與排程 | Kassem 等人在 2022 年把他們 50% 的丟包叢集歸因於換手，因為排程器當時仍待辨識，而如今三種現象可以分開：衛星更換則視為獨立事件，每 15 s 都有一次排定的重新配置尖峰、在中緯度與邊界對齊而在南極常偏離邊界的天線換手，以及第 3.4 節那條在佇列滿時把前述任何一者轉成丟包的前端丟棄佇列。 |
Original figure note原始圖說出處
Drawn for this primer from Tanveer et al., CoNEXT 2023 Companion, Fig. 2. Casparsen et al., arXiv 2601.08439, Fig. 3 (preprint). Fries et al., ACM IMC 2026, §4.2.The scheduler's choices follow measurable preferences, and Tanveer et al.'s matching of trajectories to ephemerides exposes its preferences.排程器的選擇呈現可量測的偏好，Tanveer 等人把軌跡對應到星曆的做法揭露了它的偏好。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Constellation configuration | Terminals connect only to satellites above 25° of elevation |
| Selected elevation | the satellites it selects sit at a median 22.9° higher elevation than the available ones it passes over |
| High-elevation selection bias | and although satellites between 45° and 90° make up only about 30% of the candidates they receive 80% of the selections. The authors attribute this to the exclusion zone around the geostationary arc that terminals above about 40°N must respect. |
| Northern selection bias | 58% of available satellites lie to the north of the terminals yet 82% of selections do |
| Sunlight selection bias | sunlit satellites are chosen 72.3% of the time when both sunlit and eclipsed candidates are available |
| Satellite prediction accuracy | A random forest uses local time and satellite counts per cluster. It places the chosen satellite within its top five 65% of the time, against a 22% baseline [13, §5.1–6, Figs. 4–7]. |
| Slant range | The 25° mask is a license condition, reduced from 40° in the 2018 modification [7, §2]. The table compares operator masks because they set pass length, slant range and remaining link margin through Section 3.1’s path loss. |
| 星座配置 | 終端只連接仰角 25° 以上的衛星，它選中的衛星比它略過的可用衛星仰角中位數高 22.9°，而且雖然 45° 到 90° 之間的衛星只佔候選者約 30%，它們卻獲得 80% 的選擇，作者把這歸因於約 40°N 以北的終端必須迴避的地球同步弧排除區。 |
| 選星偏好 | 58% 的可用衛星位於終端北側，選擇卻有 82% 在北側，當日照與蝕影候選者都存在時日照衛星被選中 72.3% 的時間，而以當地時間與各群衛星數為特徵的隨機森林能在前五名內預測被選衛星 65% 的時間，相對於 22% 的基準[13, §5.1–6, Figs. 4–7]。 |
| 斜距 | 25° 的遮罩本身是執照條件，如 Handley 所記錄在 2018 年的修改中從 40° 降低[7, §2]，下表彙整主要業者的遮罩，因為它們決定過頂長度、斜距，並透過第 3.1 節的路徑損耗算式決定鏈路預算還剩多少給資料。 |
| Operator業者 | Minimum elevation最低仰角 | Source來源 | Note備註 |
|---|---|---|---|
| StarlinkStarlink | 25° user terminals and gateways. 5° polar gateways使用者終端與閘道 25°。極區閘道 5° | FCC April 2021 modification via ISPreview, 28 Apr 2021 FCC filing. [13] measuredFCC 2021 年 4 月修改案，經 ISPreview 2021 年 4 月 28 日 FCC filing。[13] 實測 | Lowered from 40° in the 2018 filing [7, §2]. Scheduler picks 45–90° for 80% of selections在 2018 年申請中從 40° 降低[7, §2]。排程器 80% 的選擇落在 45–90° |
| Amazon KuiperAmazon Kuiper | 35° user terminals. 20° gateways使用者終端 35°。閘道 20° | [9, Tab. 7] FCC filing[9, Tab. 7] FCC filing | Hypatia models Kuiper at 30°, slightly overstating pass lengthHypatia 以 30° 建模 Kuiper，略高估過頂長度 |
| Telesat LightspeedTelesat Lightspeed | 10° | [10, p. 15] FCC filing[10, p. 15] FCC filing | Pachler et al. estimate 40° as the operational minimum for full coveragePachler et al. 估計完整覆蓋的運作最低值為 40° |
| OneWeb Gen1OneWeb 第一代 | ≥55° guaranteed visibility. Beams may point above 25°保證可見 ≥55°。波束可指向 25° 以上 | [11, pp. 11–12] FCC filing[11, pp. 11–12] FCC filing | 55° is a coverage-geometry figure at 1,200 km, with the pointing floor defined separately55° 是 1,200 km 高度的覆蓋幾何數字，指向下限則須另行定義 |
| Pachler et al. operational estimatesPachler et al. 的運作估計 | SpaceX 35°, Amazon 35°, Telesat 40°, OneWeb 70°SpaceX 35°、Amazon 35°、Telesat 40°、OneWeb 70° | [31] peer-reviewed[31] peer-reviewed | Minimum angle that guarantees continuous coverage in their model, above the licensed floor其模型中保證連續覆蓋的最低角度，高於執照下限 |
3.6Direct-to-cell and Earth-observation downlink直連手機與地球觀測下行
Recall
回顧
Section 3.1 calculates 456 Mbps raw rate for one 240 MHz Ku channel and a dedicated terminal. Section 3.3 defines bent pipe as a single-satellite path to a gateway within its footprint. This section changes the user end of the first and reverses the direction of the second.
Everything so far assumed a terminal built for the satellite, and the last two topics of the chapter relax that assumption from opposite ends. Direct-to-cell service, abbreviated D2C, connects a satellite to an unmodified mobile phone in a terrestrial cellular band. The handset lacks a phased-array dish. The satellite supplies most of the link gain against a small, nearly omnidirectional phone antenna.到目前為止的一切都假設終端是專為衛星打造的，本章最後兩個主題則從相反的兩端放寬這個假設。直連手機（direct-to-cell，簡稱 D2C）服務讓衛星以地面行動通訊頻段直接連接原有規格的手機，於是相位陣列天線從鏈路的使用者端消失，衛星幾乎要獨力面對一支小而近乎全向的手機天線來完成鏈路預算。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Constellation configuration | Garcia-Cabeza et al. analyze over one million US D2C measurements from October 2024–April 2025. Starlink uses T-Mobile’s PCS G block: 1910–1915 MHz uplink and 1990–1995 MHz downlink, or 2×5 MHz. Around 400 D2C satellites completed deployment in December 2024, roughly 6% of the fleet. This grew above 650 by 2026 ( [58] 64(6), 2026). |
| Constellation configuration | The same authors report that the FCC allows up to 7,500 D2C-capable satellites while it denied a request for 30,000. This makes the size of this layer a regulatory quantity as much as an engineering one. |
| 頻率 | Garcia-Cabeza 等人在 2024 年 10 月到 2025 年 4 月之間透過美國各地超過一百萬筆群眾外包量測研究了 Starlink 的 D2C 服務，它使用授權給 T-Mobile 的 PCS G 區塊中 2×5 MHz 的頻譜，上行 1910–1915 MHz、下行 1990–1995 MHz，由 2024 年 12 月完成、約 400 顆具 D2C 能力的衛星組成的星座提供，約佔整個機隊的 6%，到 2026 年已增加到超過 650 顆[58]。 |
| 星座配置 | 同一批作者指出，FCC 允許最多 7,500 顆具 D2C 能力的衛星，同時駁回了 30,000 顆的申請，這使得這一層的規模既是工程量，也同樣是一個監理量。 |
The measured radio quality shows what removing the dish costs. Published D2C latency measurements remain to be established. This primer therefore leaves its latency value open and limits Panel 3.2 to documented settings.實測的無線電品質顯示了拿掉天線的代價。D2C 的延遲仍待已發表量測確認，因此本教材保留此數值，面板 3.2 的選項範圍則限於已有依據的設定。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Power budget | The median reference signal received power (RSRP). The phone reports power on serving-cell reference symbols. Satellite power was −121 dBm, against −97 dBm from terrestrial cells, a 24 dB deficit. The median signal-to-interference-plus-noise ratio (SINR) was 0 dB against 5 dB. Thus, the median phone sits at the point where the wanted signal and the noise plus interference are equal in power. |
| Evaluation conditions | At that SINR the link adaptation has to choose robust modulation and heavy coding. Spectral efficiency has median 0.64 bit/s/Hz and mean 0.79 bit/s/Hz. The authors estimate roughly 4 Mbps per beam for outdoor users [58]. |
| Link or data rate | Multiplying the 5 MHz downlink by median efficiency gives 3.2 Mbps. Mean efficiency gives 4.0 Mbps, as the calculation table shows. Thus, the published figure corresponds to the mean efficiency and is shared by every phone in the beam. This is less than a hundredth of the raw rate of one Ku channel. |
| 功率預算 | 參考訊號接收功率（RSRP）是手機回報的服務基地台參考符號功率，其中位數來自衛星時為 −121 dBm，來自地面基地台時為 −97 dBm，差了 24 dB，而訊號對干擾加雜訊比（SINR）的中位數為 0 dB 對 5 dB，所以中位數的手機正好處在想要的訊號與雜訊加干擾功率相等的位置。 |
| 頻譜效率 | 在這個 SINR 下，鏈路調適只能選擇穩健的調變與重度編碼，實測頻譜效率的中位數為 0.64 bit/s/Hz、平均值為 0.79 bit/s/Hz，作者據此估計戶外使用者每個波束約 4 Mbps[58]。 |
| 波束共享容量 | 下方的計算表顯示，5 MHz 下行載波乘以中位效率得到 3.2 Mbps，乘以平均效率得到 4.0 Mbps，所以公布的數字對應的是平均效率，而且由波束內所有手機共用，這低於一個 Ku 頻道原始速率的百分之一。 |
The systems problem in D2C lies less in the radio than in the cellular core behind it. Liu et al. describe cellular session state at each hop between phone, radio access network and mobile core. A satellite base station leaves within minutes. Repeated registration and handover then couple the satellite operator, mobile operator and user.
MOSAIC runs radio and user planes onboard each satellite. It replaces session state with policy-embedded one-time tokens issued by the mobile operator and enforced by the SIM. In-band tokens support operation with either available or absent ISLs. This suits LEO satellites with unreliable ISLs [59, §3–6].
D2C 的系統設計重點，在於無線電背後的行動核心網路。Liu 等人指出，一個行動通訊會話以狀態的形式保存在手機、無線接取網路與行動業者核心網路之間的每一跳上，因此當基地台是一顆幾分鐘內就離開的衛星時，衛星業者、行動業者與使用者會因為一連串的重新註冊與換手而緊密耦合。他們的設計 MOSAIC 讓每顆衛星自成一體：在星上執行無線電與使用者平面，並以行動業者簽發、由 SIM 卡強制執行、內嵌政策的一次性權杖取代逐會話的狀態，權杖以頻內方式傳遞，所以這個設計在具備或缺少 ISL 的情境下皆能運作，這點很重要，因為許多 LEO 衛星的 ISL 可靠度有限[59, §3–6]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Handover and scheduling | MOSAIC uses data-driven emulation and an Amarisoft 3GPP NTN prototype with commercial phones. Serviceable area increases up to 116%. Setup is up to 5.19 times faster than the 3GPP NTN Starlink baseline. Inter-operator handover signaling falls by 850–7,640 times. Baseline signaling storms can consume up to 15.75% of ISL bandwidth. The estimate assumes a typical 20 Gbps ISL [59, §3.2, §7, Figs. 7, 15, 17, 21]. |
| 鏈路建立 | 在資料驅動的模擬，以及以 Amarisoft 3GPP NTN 協定堆疊搭配商用手機打造的原型中，MOSAIC 把可服務面積擴大最多 116%，在 Starlink 星座上的服務建立比 3GPP NTN 基準快最多 5.19 倍，並把業者之間的換手信令減少 850–7,640 倍，而基準設計的信令風暴可能吃掉最多 15.75% 的 ISL 頻寬，這個估計以作者引用的典型 ISL 容量 20 Gbps 為基礎[59, §3.2, §7, Figs. 7, 15, 17, 21]。 |
Earth-observation traffic starts in orbit as imagery. A visible ground station must receive it, reversing the subscriber bent-pipe direction.地球觀測衛星把彎管反了過來，因為它們的流量以影像的形式在軌道上產生，必須經由當下剛好可見的地面站送到地面。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Ground-contact capacity | Denby and Lucia study nanosatellites using groundward links. A station receives 200 Mbit/s during 5–10 minute passes, carrying up to 15 GB per pass. One station serves nine satellites per revolution. The data delay from capture to ground reaches 5.5 hours. |
| Constellation configuration | Serving 1,000 satellites ideally requires 112 stations. Onboard imagery filtering reduces this to six by sending useful results [60, §3.3, §6–7]. |
| Constellation configuration | A 10-minute pass at 200 Mbit/s carries 15 GB. Serving 1,000 satellites at nine per station requires about 112 stations. These calculations reproduce the paper’s figures. The gap between 112 and 6 stations is the founding argument for the orbital edge computing that Chapter 8 develops. |
| 星座配置 | Denby 與 Lucia 為僅採對地鏈路的奈米衛星星座量化了由此而來的下行之牆：地面站在 5–10 分鐘的過頂期間以 200 Mbit/s 接收，所以一次過頂最多帶下 15 GB，一個站每圈只能服務 9 顆衛星，從拍攝到落地的資料延遲達 5.5 小時。 |
| 星座配置 | 服務 1,000 顆衛星理想上需要 112 個站，而如果衛星在星上過濾影像、只送下有用的部分，則只需 6 個站[60, §3.3, §6–7]。 |
| 星座配置 | 計算表確認 10 分鐘過頂傳 15 GB 恰好是 200 Mbit/s，而每站 9 顆時 1,000 顆衛星約需 112 個站，所以這篇論文的數字彼此一致，而 112 對 6 個站的差距正是第 8 章所發展的軌道邊緣運算的奠基論證。 |
Vasisht, Shenoy and Chandra attack the same wall from the ground. L2D2 uses many inexpensive receive-only stations with roughly 1 m dishes. A learned model predicts link quality, and a matching algorithm assigns satellites to stations. The learned model replaces ITU propagation models. Satellites can send data to whichever inexpensive station they pass. This reduces waiting for one of a few large stations [61, §4–5].Vasisht、Shenoy 與 Chandra 從地面攻擊同一道牆。L2D2 以許多約 1 m 天線的小型僅接收站取代少數昂貴的大型站，用學習模型取代 ITU 傳播模型預測每條衛星到地面站鏈路的品質，並把衛星指派到地面站當作一個匹配問題來解，使衛星能下傳到它正經過的任何一個便宜的站，使資料可提早落地，縮短等待大型站進入視野的時間[61, §4–5]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Ground infrastructure | Trace-driven evaluation uses 173 SatNOGS stations and 259 satellites. Median downlink latency falls from 90 to 21 minutes, and p90 from 323 to 71 minutes. The learned predictor's median SNR error was 0.39 dB against 2.39 dB for the ITU models [61, §6, Tabs. 2–3] . |
| Throughput | The introduction and link-budget section give conflicting weather-attenuation ranges. Their rain-fade interpretation requires source clarification. This primer therefore uses Section 3.4’s measured Starlink throughput reductions of 17–27%. |
| 星座配置 | 在以 173 個 SatNOGS 站與 259 顆衛星進行的軌跡驅動評估中，下行延遲中位數從 90 分鐘降到 21 分鐘，第 90 百分位從 323 分鐘降到 71 分鐘，而學習預測器的 SNR 誤差中位數為 0.39 dB，ITU 模型則為 2.39 dB[61, §6, Tabs. 2–3]。 |
| 吞吐量 | 這篇論文在引言與鏈路預算兩節給出兩個相異的天候衰減範圍，這些數值作為雨衰依據前，須先釐清定義及差異，第 3.4 節中 Starlink 實測吞吐量下降 17–27% 仍是應該使用的數字。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| D2C signal deficit against a terrestrial cell (median RSRP)D2C 相對地面基地台的訊號差距（RSRP 中位數） | \(-97 - (-121)\) | 24 dBcomputed. RSRP medians from [58]計算值。RSRP 中位數來自 [58] |
| D2C downlink rate per beam at the median spectral efficiency以頻譜效率中位數計的每波束 D2C 下行速率 | \(5\,\text{MHz} \times 0.64\,\text{bit/s/Hz}\) | 3.2 Mbpscomputed. 5 MHz downlink of the PCS G block and 0.64 median from Garcia-Cabeza et al.計算值。PCS G 區塊的 5 MHz 下行與 0.64 中位數來自 Garcia-Cabeza et al. |
| Same at the mean spectral efficiency以平均頻譜效率計 | \(5\,\text{MHz} \times 0.79\,\text{bit/s/Hz}\) | 4.0 Mbpscomputed. Matches the authors' ~4 Mbps per-beam estimate for outdoor users計算值。與作者對戶外使用者每波束約 4 Mbps 的估計相符 |
| One Ku channel against one D2C beam一個 Ku 頻道對一個 D2C 波束 | \(456 / 4\) | ≈ 114×computed. 456 Mbps raw from Section 3.1 (before FEC), 4 Mbps from Garcia-Cabeza et al.計算值。456 Mbps 原始速率來自第 3.1 節（FEC 前），4 Mbps 來自 Garcia-Cabeza et al. |
| Station rate implied by 15 GB in a 10-minute pass10 分鐘過頂傳 15 GB 所隱含的地面站速率 | \(15 \times 8\,\text{Gbit} / 600\,\text{s}\) | 200 Mbit/scomputed. 15 GB per pass and 5–10 min passes from [60, §3.3]計算值。每次過頂 15 GB 與 5–10 分鐘過頂來自 [60, §3.3] |
| Stations for 1,000 satellites at 9 satellites per station per revolution每站每圈 9 顆衛星時 1,000 顆衛星所需站數 | \(\lceil 1000 / 9 \rceil\) | 112computed. Matches Denby & Lucia's 112 ideal stations (against 6 with on-board filtering)計算值。與 Denby & Lucia 的 112 個理想站數相符（星上過濾時為 6 個） |
| L2D2 latency reduction, median and 90th percentileL2D2 的延遲降幅，中位數與第 90 百分位 | \(90/21,\; 323/71\) | 4.3×, 4.5×computed. Latencies from [61, §6]計算值。延遲來自 [61, §6] |
Original figure note原始圖說出處
Drawn for this primer from Humphreys et al. (TAES 2023) via the Section 3.1 calculation, Garcia-Cabeza et al. (IEEE Commun. Mag. 2026), Denby & Lucia (ASPLOS 2020, §3.3) and Vasisht et al. (SIGCOMM 2021, §6). 本入門依 Humphreys et al.（TAES 2023，經第 3.1 節計算）、Garcia-Cabeza et al.（IEEE Commun. Mag. 2026）、Denby & Lucia（ASPLOS 2020, §3.3）與 Vasisht et al.（SIGCOMM 2021, §6）繪製。The two downlinks close the chapter on the same lesson as its first sections. For a dish, phone or orbital camera, spectrum and SINR on one radio hop set the available rate. A phone shares 5 MHz near 0 dB and obtains a few Mbps. A dish uses a 240 MHz channel and obtains hundreds.
Ground infrastructure and routing shape latency alongside propagation. Subscribers depend on their assigned PoP. Imaging satellites depend on the next visible ground station.
這兩種下行以本章開頭幾節同樣的教訓作結。端點涵蓋 Starlink 天線、手機及軌道上的相機時，速率都由單一無線電跳上的頻譜與 SINR 決定，這就是為什麼在 0 dB 下共用 5 MHz 的手機只得到每秒幾 Mb，而在 240 MHz 頻道上的天線能得到數百 Mb。而延遲主要受最近地面設施的位置及導流方式影響，對用戶而言那是 PoP，對成像衛星而言則是下一個可見的地面站。
Chapter summary
本章摘要
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | Humphreys et al. decoded the Ku-band OFDM waveform externally. Its eight 240 MHz channels, 1,024 subcarriers and 1.333 ms frames yield 456 Mbps raw rate per channel before coding. |
| Capacity comparison | Beams serve cells roughly 24 km across. Per-satellite capacity estimates differ: 6.5 Gbps in a peer-reviewed model, 24 and 96 Gbps in preprints. SpaceX claims over 1 Tbps for V3. Thus, a paper must name the figure it uses together with its tag. |
| PoP coverage | Traffic usually follows bent pipe to a gateway, then a tunnel to its PoP. Measured latency is 40–50 ms in most countries and 25.7 ms under SpaceX’s US definition. Islands beyond local PoP coverage see 100–181 ms. These values show PoP geography’s influence. |
| Round-trip time | Measured downlink/uplink rates differ roughly tenfold. Load raises RTT two- to fourfold within a shared 1,400–1,600-packet drop-front queue. Rain reduces throughput 17–27%. |
| Constellation configuration | The global scheduler reconfigures every 15 s at :12, :27, :42 and :57. Initial spikes average about 74 ms and decay within 140 ms. At mid-latitudes, at least 89% of handovers align with boundaries. |
| Ground infrastructure | D2C shares a 5 MHz carrier at about 4 Mbps per beam. Earth-observation delivery depends on visible stations. L2D2 adds inexpensive stations, while onboard filtering reduces data volume. |
| 波形與頻道容量 | Starlink 的使用者鏈路是 Humphreys 等人從外部解碼的 Ku 頻段 OFDM 波形，八個 240 MHz 頻道、1,024 個子載波與 1.333 ms 的幀合起來給出每頻道編碼前 456 Mbps 的原始速率。 |
| 星座配置 | 在其之上，波束服務直徑約 24 km 的服務格，而每顆衛星的容量是一個有爭議的數字，從同儕審查的 6.5 Gbps 模型，經過 24 與 96 Gbps 的預印本估計，到 V3 超過 1 Tbps 的公司說法，所以論文必須寫明所用的數字及其標籤。 |
| PoP 覆蓋 | 流量通常經彎管到閘道，再以隧道送到 PoP，而多數國家 40–50 ms、SpaceX 自報的美國 25.7 ms、位於 PoP 覆蓋範圍之外的島嶼 100–181 ms 的實測延遲顯示，主宰用戶體驗的是 PoP 的地理位置，傳播則提供延遲下限。 |
| 佇列行為 | 實測鏈路約有十比一的上下行速率差距，在負載下於一個 1,400–1,600 封包的前端丟棄單一佇列中使 RTT 膨脹二到四倍，並在雨中損失 17–27% 的吞吐量。 |
| 星座配置 | 全域排程器每 15 秒在 :12、:27、:42、:57 重新指派衛星，每個邊界帶來一個約 74 ms、在 140 ms 內結束的尖峰，而在中緯度至少 89% 的換手對齊邊界。 |
| 地面設施 | 直連手機服務以每波束約 4 Mbps 共用一個 5 MHz 載波，而地球觀測下行主要受可見地面站數目限制，L2D2 以許多便宜的地面站改善此限制，星上過濾則以少送資料改善此限制。 |
Where these ideas go next
這些概念接下來用在哪裡
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Capacity comparison | Chapter 4 gathers cell sizes, capacity scenarios, PoP counts, scheduling intervals and measured medians into an evidence-tagged parameter sheet. It then calculates Panel 3.2’s bent-pipe and ISL-chain budgets end to end. |
| Queue behavior | Chapter 6 examines drop-front queues and periodic boundaries as congestion-control problems. Chapter 8 starts orbital computing from Denby and Lucia’s downlink wall. |
| 容量比較 | 第 4 章把本章的數字，從服務格大小與容量情境到 PoP 數目、15 秒週期與實測中位數，收進一張每列都保留證據標籤的參數表，然後把面板 3.2 的彎管與 ISL 鏈預算從頭到尾算一遍。 |
| 佇列行為 | 第 6 章把前端丟棄佇列與 15 秒邊界當作壅塞控制的問題再次討論，而第 8 章以 Denby 與 Lucia 的下行之牆作為軌道運算的起點。 |
Papers behind this chapter本章所依據的論文
IEEE TAES 2023Signal Structure of the Starlink Ku-Band Downlink
ACM IMC 2022A First Look at Starlink Performance
ACM WWW 2024A Multifaceted Look at Starlink Performance
SIGMETRICS 2024Democratizing LEO Satellite Network Measurement
CoNEXT 2023 CompanionMaking Sense of Constellations: Methodologies for Understanding Starlink's Scheduling Algorithms
IEEE PIMRC 2023Measuring a Low-Earth-Orbit Satellite Network
USENIX NSDI 2024Democratizing Direct-to-Cell Low Earth Orbit Satellite Networks
ACM SIGCOMM 2021L2D2: Low Latency Distributed Downlink for LEO Satellites