Digest · Measuring Starlink · read after Chapter 3
導讀 · 量測 Starlink · 第 3 章後閱讀
Making Sense of Constellations: Methodologies for Understanding Starlink's Scheduling Algorithms理解星座：解讀 Starlink 排程演算法的方法論
This digest explains how obstruction maps and public orbital elements identify a serving satellite. It examines the fifteen-second latency schedule and the satellite-selection preferences inferred from it.讀完後，你能說明遮蔽圖與公開軌道根數如何結合來辨識終端正在使用哪顆衛星、為什麼延遲狀態在每分鐘的第 12、27、42 與 57 秒改變，以及 Starlink 排程器偏好哪些衛星及其原因。
If you read nothing else
只讀這一段也可以
Obstruction-map matching connects a serving satellite to its latency trace. This reveals a common reassignment cadence and preferences among visible candidates.
遮蔽圖比對將服務衛星連到延遲軌跡，揭露共同重新指派節律與候選偏好。
| # | Reported values and conditions |
|---|---|
| 1 | Tanveer et al. measured four terminals in Western Europe and the north-eastern, midwestern and north-western United States. |
| 2 | They probed serving PoPs every 20 ms and collected gRPC obstruction maps every 15 s. |
| 3 | XOR differences exposed sky tracks, which dynamic time warping matched to SGP4-propagated CelesTrak elements. |
| 4 | Agreement exceeded 99% across 500 manual checks. |
| 5 | Latency regimes changed at :12/:27/:42/:57 across sites, with statistically distinct adjacent windows and parallel bands interpreted as round-robin frame allocation [13, §3–4, Fig. 2]. |
| 6 | Selected satellites had a 22.9° median elevation advantage. |
| 7 | The 45–90° band held 30% of candidates and 80% of selections. |
| 8 | Northern candidates accounted for 58% of availability and 82% of selections, while sunlit satellites drew 72.3%. |
| 9 | A random forest achieved 65% top-five prediction against a 22% baseline [13, §5–6]. |
| 10 | The venue is the CoNEXT 2023 Companion short-paper track, with collection duration and sample counts awaiting clarification. |
| # | 數值、條件與證據 |
|---|---|
| 1 | Tanveer 等人量測西歐及美國東北、中西部、西北部四支終端，每 20 ms 探測服務 PoP，每 15 s 經 gRPC 收集遮蔽圖。 |
| 2 | XOR 差分分離天空軌跡，再以動態時間規整比對 CelesTrak 根數的 SGP4 軌跡，500 次人工核對的一致率超過 99%。 |
| 3 | 各站延遲都在 :12/:27/:42/:57 切換，鄰接窗口分布有統計差異，平行延遲帶則被解讀為輪詢訊框分配 [13, §3–4, Fig. 2]。 |
| 4 | 選中衛星的仰角中位優勢為 22.9°，45–90° 帶占候選 30%、選擇 80%。 |
| 5 | 北方占可用候選 58%、選擇 82%，受日照衛星占選擇 72.3%。 |
| 6 | 隨機森林前五名準確率為 65%，基準為 22% [13, §5–6]。 |
| 7 | 本文發表於 CoNEXT 2023 Companion 短文軌，蒐集時長與樣本數仍待說明。 |
1. Where this paper sits1. 這篇論文的位置
Earlier studies described Starlink latency distributions. Kassem et al. attributed their 2022 loss clusters to handover. Tanveer et al. connected those observations to a global fifteen-second reassignment cadence. Mohan, Lai, Izhikevich, Casparsen and Fries later used its :12/:27/:42/:57 boundaries. The arXiv copy retains a stale conference header, which can mislead venue citations.
Crossref places DOI 10.1145/3624354.3630586 in the CoNEXT 2023 Companion short-paper track. Claims therefore need that publication scope. Chapter 3 uses the scheduling result, Chapter 1 the elevation preference beside visible-satellite counts, and Chapter 7 the contrast with simulator handover policies.
早期研究描述 Starlink 延遲分布，Kassem 等人則把 2022 年的遺失叢集歸因於換手。Tanveer 等人把觀察連結到全球十五秒重配節律，Mohan、Lai、Izhikevich、Casparsen 與 Fries 後來也使用 :12/:27/:42/:57 邊界。arXiv 保留過期會議頁首，容易影響場合引用，Crossref 則把 DOI 10.1145/3624354.3630586 列於 CoNEXT 2023 Companion 短文軌，主張須連同此發表範圍解讀。第 3 章使用排程結果，第 1 章把仰角偏好與可見衛星數並列，第 7 章則對照模擬器的換手策略。
2. The problem, measured2. 問題，用量測說話
Track reconstruction combines consecutive map differences with orbital predictions. Clock alignment lets the study examine satellite choice and path delay together.
軌跡重建結合連續地圖差分與軌道預測，時鐘對齊則讓研究聯合檢視衛星選擇與路徑延遲。
| # | Reported values and conditions |
|---|---|
| 1 | Starlink’s serving-satellite identity and selection policy are hidden from ordinary terminal diagnostics. |
| 2 | The study reconstructs identity from the gRPC obstruction map: a 123 × 123 polar image centered at pixel (62, 62). |
| 3 | A 45-pixel radius spans 25° elevation at the rim to 90° at the center, with north at zero azimuth. |
| 4 | App alignment and a two-day fully colored-map experiment establish these parameters [13, §3]. |
| 5 | XOR between maps 15 s apart isolates the current track segment. |
| 6 | Ten-minute resets keep successive tracks separate. |
| 7 | Dynamic time warping matches each observed track to SGP4 predictions, with over 99% agreement in 500 manual checks. iRTT probes every 20 ms to PoP-colocated servers. iPerf3 holds 50% upstream load, and NTP aligns the clocks for joint latency–identity analysis. |
| # | 數值、條件與證據 |
|---|---|
| 1 | 一般終端診斷會隱藏服務衛星身分與選擇策略。 |
| 2 | 研究從 gRPC 遮蔽圖重建身分：123 × 123 極座標影像，以像素 (62, 62) 為中心，45 像素半徑涵蓋邊緣 25° 至中央 90° 仰角，方位零為北。 |
| 3 | 作者透過手機視圖對齊與兩天塗滿地圖實驗建立參數 [13, §3]。 |
| 4 | 相隔 15 s 的 XOR 分離目前軌跡，每十分鐘重置使相鄰軌跡分開。 |
| 5 | 動態時間規整比對 SGP4 預測，在 500 次人工核對中超過 99% 一致。 |
| 6 | iRTT 每 20 ms 探測 PoP 同址伺服器，iPerf3 維持 50% 上行負載，NTP 對齊時鐘以聯合分析延遲與衛星身分。 |
3. The design3. 設計
Common wall-clock boundaries support global scheduling. Selection preferences vary with elevation, azimuth and obstruction, under the northern sites’ GEO-arc geometry.
共同時鐘邊界支持全球排程，選擇偏好則隨仰角、方位與遮蔽而變，適用於北方站點的 GEO 弧幾何。
| # | Reported values and conditions |
|---|---|
| 1 | Joint latency and satellite identification reveal a schedule and a selection policy. |
| 2 | All four sites change latency regimes every 15 s at :12/:27/:42/:57. |
| 3 | Adjacent windows differ under Mann–Whitney U at p < 0.05. |
| 4 | Parallel bands several milliseconds apart suggest round-robin satellite frame allocation [13, §3–4, Fig. 2]. |
| 5 | Shared wall-clock boundaries across continents support a global controller rather than orbital-motion timing. |
| 6 | Above the 25° mask, each slot has 35–44 candidates, and selected satellites have a 22.9° median elevation advantage. |
| 7 | The 45–90° band supplies 80% of selections from 30% of candidates [13, §5.1, Fig. 4]. |
| 8 | Northern candidates supply 82% of selections from 58% of availability. |
| 9 | An obstructed Ithaca terminal selects 9.7% from the north-west, against 55.4% at clear sites [13, §5.1, Fig. 5]. |
| 10 | The authors connect higher northern selections to GEO-arc exclusion southward at low elevation above about 40°N, plus shorter paths and lower power. |
| # | 數值、條件與證據 |
|---|---|
| 1 | 延遲與衛星身分的聯合觀察揭露排程及選擇策略。 |
| 2 | 四站都每 15 s 於 :12/:27/:42/:57 切換延遲，鄰接窗口的 Mann–Whitney U 達 p < 0.05，平行帶相隔數毫秒，支持輪詢訊框分配解讀 [13, §3–4, Fig. 2]。 |
| 3 | 跨洲共享時鐘邊界支持全球控制器的解釋。 |
| 4 | 25° 遮罩以上，每時槽有 35–44 候選，選中衛星仰角中位優勢為 22.9°。 |
| 5 | 45–90° 帶從 30% 候選提供 80% 選擇 [13, §5.1, Fig. 4]，北方則從 58% 可用衛星提供 82%。 |
| 6 | 受遮蔽 Ithaca 終端從西北選擇 9.7%，視野通暢站點為 55.4% [13, §5.1, Fig. 5]。 |
| 7 | 作者將高仰角、北方偏好連至約 40°N 以上南方低仰角的 GEO 弧排除區，以及較短距離與較低功率。 |
Original figure note原始圖說出處
Drawn for this primer from Tanveer et al., CoNEXT 2023 Companion, §3–5, Figs. 2, 4, 5, spike shape from Casparsen et al., arXiv 2601.08439, Fig. 3 (preprint).Sunlight and fleet age provide further selection signals. Partial predictability from terminal-visible features also indicates additional hidden scheduler inputs.
日照與艦隊年齡提供更多選擇訊號，終端特徵的部分可預測性也顯示其他隱藏排程輸入。
| # | Reported values and conditions |
|---|---|
| 1 | Selection probability also rises with launch date by roughly one month of fleet age. |
| 2 | When sunlit and eclipsed candidates coexist, sunlit satellites win 72.3%. |
| 3 | Eclipsed candidates win when they occupy at least 35% of the set and sit about 25° higher [13, §5.2–5.3, Fig. 6]. |
| 4 | A random forest uses local time and sky-cluster counts, with an 80/20 split and five-fold cross-validation. |
| 5 | It ranks the chosen satellite in its top five 65% of the time, against a 22% top-k baseline [13, §6, Fig. 7]. |
| 6 | This supports a learnable policy with additional hidden inputs such as satellite load and terminal density. |
| # | 數值、條件與證據 |
|---|---|
| 1 | 選擇機率也隨發射日期增加，約相當於一個月的艦隊年齡。 |
| 2 | 日照與地影候選同時存在時，日照衛星勝出 72.3%，地影候選則在占至少 35%、仰角高約 25° 時勝出 [13, §5.2–5.3, Fig. 6]。 |
| 3 | 隨機森林採當地時間與天空群集數，以 80/20 切分、五折交叉驗證訓練，前五名命中率為 65%，top-k 基準為 22% [13, §6, Fig. 7]。 |
| 4 | 這支持可學習策略，也顯示衛星負載與終端密度等隱藏輸入的作用。 |
4. Evaluation setup4. 實驗設定
The study measures the production user-access path directly. Four bridge-mode terminals sit behind Raspberry Pis at northern mid-to-high latitudes, where GEO-arc exclusion applies. Their sites are Western Europe and the north-eastern, midwestern and north-western United States. PoP-colocated servers minimize terrestrial delay [13, §3].
Public SGP4 elements have kilometer-scale position error, and fifteen-second map sampling limits track resolution. Collection duration and latency-sample count require clarification. Text and caption also assign Fig. 2 to separate terminals, midwestern US and Europe, leaving a source inconsistency.
研究直接量測正式網路的使用者存取路徑。四支橋接終端接在 Raspberry Pi 後方，位於適用 GEO 弧排除的中高北緯，依序是西歐及美國東北、中西部、西北部。PoP 同址伺服器縮小地面延遲 [13, §3]。公開 SGP4 元素有公里級位置誤差，十五秒遮蔽圖取樣限制軌跡解析度。蒐集時長與延遲樣本數仍待說明，正文與圖說也各把 Fig. 2 指為美國中西部、歐洲終端，形成來源差異。
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Latency regimes shift every 15 s, at seconds 12, 27, 42 and 57, across all vantage points. Consecutive windows differ (Mann–Whitney U, p < .05) [13, §3–4, Fig. 2]延遲狀態在所有觀測點每 15 s 於第 12、27、42、57 秒切換，相鄰窗口各異（Mann–Whitney U，p < .05）[13, §3–4, Fig. 2] | iRTT probes every 20 ms to PoP-colocated servers establish cadence. Later studies resolve spike shape and confirm global alignment at higher sampling rates.iRTT 每 20 ms 探測 PoP 同址伺服器以建立節律，後續研究再以較高取樣解析尖峰並確認全球對齊。 |
| Within-window latency forms parallel bands a few ms apart, attributed to round-robin frame allocation [13, §4]窗口內延遲形成相距數 ms 的平行帶，歸因於輪詢式訊框分配[13, §4] | The frame-allocation interpretation fits Humphreys et al.’s 1.333 ms frames and Garcia et al.’s later 1:3 access model. Direct validation remains open here.訊框分配解讀符合 Humphreys 等人的 1.333 ms 訊框與 Garcia 等人的後續 1:3 存取模型，此處直接驗證仍待完成。 |
| Selected satellites have median elevation 22.9° higher than unselected, 45–90° holds 30% of candidates but 80% of selections [13, §5.1, Fig. 4]選中衛星的仰角中位數比其他候選高 22.9°，45–90° 占候選 30%、選擇 80%[13, §5.1, Fig. 4] | All four terminals lie above roughly 40°N where the GEO-arc exclusion forces the preference, terminals nearer the equator may behave differently.四支終端都在約北緯 40° 以上，GEO 弧排除區迫使此偏好，較接近赤道的終端可能有各異行為。 |
| Northern satellites account for 58% of availability and 82% of selections. An obstructed Ithaca terminal selected its blocked northwest region 9.7% of the time. The corresponding share elsewhere was 55.4% [13, §5.1, Fig. 5]58% 的可用衛星在北方但 82% 的選擇是北方，受遮蔽的 Ithaca 終端從其被擋住的西北方只獲得 9.7% 的選擇，其他終端為 55.4%[13, §5.1, Fig. 5] | The obstruction result shows the scheduler respects the terminal's reported sky, which is itself a scheduler input.遮蔽結果顯示排程器尊重終端回報的天空，而這本身就是排程器的輸入。 |
| Sunlit satellites chosen 72.3% of the time when both kinds are available, selection probability rises with launch date [13, §5.2–5.3, Fig. 6]兩類都可用時受日照衛星有 72.3% 的時間被選中，選擇機率隨發射日期上升[13, §5.2–5.3, Fig. 6] | Sample counts are required to assess the strength of these secondary preferences.評估次要偏好的強度需要樣本數。 |
| Random forest predicts the chosen satellite in its top five 65% of the time against a 22% baseline [13, §6, Fig. 7]隨機森林有 65% 的時間在前五名內預測出所選衛星，基準為 22%[13, §6, Fig. 7] | The 80/20 split and five-fold cross-validation use terminal-visible features. Satellite load and terminal density are hidden inputs.80/20 切分與五折交叉驗證採終端可見特徵，衛星負載、終端密度則屬隱藏輸入。 |
6. Limits and open questions6. 限制與待解的問題
- The four-terminal evidence concerns northern mid-latitudes, with duration and sample count awaiting clarification. GEO-arc geometry predicts other elevation preferences near the equator and in the southern hemisphere, which require separate tests. Fries et al. later found 69.3% boundary-aligned handovers at McMurdo against at least 89% in Germany, the Netherlands and Norway, establishing geographic differences.四支終端的證據限於北半球中緯度，時長與樣本數仍待說明。GEO 弧幾何預測赤道附近與南半球的仰角偏好會有差異，需另行測試。Fries 等人後來量到 McMurdo 邊界對齊換手為 69.3%，德、荷、挪至少 89%，建立了地理差異。
- The paper establishes cadence through 20 ms probing. Casparsen et al.’s 2026 measurements use a 500 Hz sampling rate to resolve transitions. The initial rise reaches about 74 ms above the mean within 140 ms. Another rise occurs in the final 75 ms. The primer attributes that shape to their preprint. preprint本論文以 20 ms 探測建立節律。Casparsen 等人在 2026 年用 500 Hz 解析轉換：開頭 140 ms 內約高出均值 74 ms，最後 75 ms 另有上升。本教材將這項形狀歸於該預印本。preprint
- Scheduler internals remain inferred. Round-robin allocation, GEO-arc rationale and global control fit external observations. SpaceX’s March 2024 engineering statement names per-beam “fronthaul scheduling” as a delay source, while further algorithm detail awaits publication. company排程器內部仍屬推論，輪詢分配、GEO 弧理由與全球控制模型皆用來解釋外部觀察。SpaceX 2024 年 3 月工程聲明把逐波束「前傳排程」列為延遲來源，更多演算法細節仍待公開。company
- Identification requires ten-minute resets to keep XOR tracks separate. This bounds continuous natural-handover tracking, while dynamic time warping absorbs kilometer-scale SGP4 error rather than measuring it independently.辨識每十分鐘重置終端以分離 XOR 軌跡，因此連續追蹤自然換手的時長受限。動態時間規整吸收公里級 SGP4 誤差，獨立誤差量測仍待完成。
- The scheduling interval measures fifteen-second reassignment. A handover’s own delay cost requires a separate measurement. Fries and Garcia later distinguish reconfiguration spikes, dish handovers and drop-front queue behavior in 2026.排程間隔衡量十五秒重新指派，換手本身的延遲成本需要另行量測。Fries 與 Garcia 在 2026 年進一步區分重配尖峰、終端換手與前端丟棄佇列。
- This CoNEXT 2023 Companion paper describes one scheduler version. SpaceX reports later scheduling and buffering changes reducing US peak-hour median latency from 48.5 to 33 ms at the PoP. Continued elevation and azimuth preferences after those changes require testing. companyCoNEXT 2023 Companion 短文描述當時版本的排程器。SpaceX 報告後續排程、緩衝調整使美國尖峰 PoP 中位延遲從 48.5 降至 33 ms，調整後的仰角與方位偏好仍需測試。company
7. How it connects7. 與其他內容的關聯
- Section 3.5: this paper is the primary source for the :12/:27/:42/:57 schedule and the 45–90° elevation preference drawn in the chapter's minute timeline.第 3.5 節：本文是該章一分鐘時間線中所畫 :12/:27/:42/:57 排程與 45–90° 仰角偏好的主要來源。
- Section 3.4: the regime shifts explain why idle-RTT distributions in every measurement study are banded rather than smooth.第 3.4 節：狀態切換解釋閒置 RTT 分布的帶狀結構。
- Section 3.3: the PoP-colocated servers isolate the access link, which is the bent pipe's first leg.第 3.3 節：PoP 同址的伺服器分離出存取鏈路，即彎管的第一段。
- Section 6.1: the 15-second step is the non-congestive disturbance that SaTCP, the HotNets 2024 patches and LeoCC are built to recognise.第 6.1 節：15 秒的階躍是 SaTCP、HotNets 2024 的修補與 LeoCC 被設計來辨識的非壅塞擾動。
- Section 7.2 presents obstruction-map identification as a measurement method. The observed selection policy differs from the nearest-satellite and longest-visibility handover rules assumed in simulators.第 7.2 節：遮蔽圖辨識本身就是一種量測技術，而實證的選擇策略與模擬器假設的最近衛星及最長可見換手規則相矛盾。
- Related pages: www24-mohan confirms global boundary alignment, and sigmetrics24-hitchhiking detects the same spikes through external probes. hotnets24-lai-cc measures their capacity effect. imc26-fries aligns handovers with them.相關導讀：www24-mohan 確認邊界全球對齊，sigmetrics24-hitchhiking 透過外部探測看到相同尖峰，hotnets24-lai-cc 量測其容量效應，而 imc26-fries 把換手對齊到這些邊界。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| :12 :27 :42 :57 | Seconds of each minute at which latency regimes shift, all four vantage points [13, §3–4, Fig. 2] | 每分鐘延遲狀態切換的秒數，四個觀測點皆然[13, §3–4, Fig. 2] | |
| 15 s · 20 ms | Reassignment period · probe interval of iRTT to PoP-colocated servers [13, §3] | 重新指派週期 · iRTT 到 PoP 同址伺服器的探測間隔[13, §3] | |
| +22.9° | Median elevation advantage of selected over unselected satellites [13, §5.1, Fig. 4] | 選中衛星相對其他候選的仰角中位優勢[13, §5.1, Fig. 4] | |
| 30% → 80% | Share of candidates in 45–90° → share of selections from that band [13, §5.1, Fig. 4] | 45–90° 內候選比例 → 來自該帶的選擇比例[13, §5.1, Fig. 4] | |
| 58% → 82% | Northern share of candidates → northern share of selections [13, §5.1, Fig. 5] | 北方候選比例 → 北方選擇比例[13, §5.1, Fig. 5] | |
| 72.3% | Share of picks that were sunlit when both sunlit and eclipsed candidates existed [13, §5.3] | 受日照與地影候選同時存在時，選中受日照者的比例[13, §5.3] | |
| 65% vs 22% | Random-forest top-5 accuracy against top-k baseline [13, §6, Fig. 7] | 隨機森林前五名準確率對 top-k 基準[13, §6, Fig. 7] | |
| 35–44 · 25° | Satellites in view per slot · minimum elevation for terminals and gateways [13, §2, §5.1] | 每時槽可見衛星數 · 終端與閘道的最低仰角[13, §2, §5.1] | |
| 123 × 123 px · 45 px | Obstruction-map size · radius spanning 25°–90° elevation [13, §3] | 遮蔽圖尺寸 · 涵蓋 25°–90° 仰角的半徑[13, §3] |