Democratizing LEO Satellite Network Measurement (LEO HitchHiking)讓 LEO 衛星網路量測平民化（LEO HitchHiking）
HitchHiking isolates a remote Starlink segment using paired probes. Examine validation against a local dish and the inferred ISL behavior in gateway-sparse regions.HitchHiking 以成對探測分離遠端 Starlink 區段。導讀檢視本地終端驗證，以及閘道稀疏地區的 ISL 行為推論。
Izhikevich and colleagues find exposed Starlink customer services through Censys, then use traceroute to identify hops around the satellite segment. Subtracting paired ping RTTs isolates that segment from terrestrial transit. Validation against a San Diego dish supports accuracy and spike detection, with broader coverage than contemporary Atlas probes [35, §4–6, Tab. 1]. Regional results reveal large latency differences and intermittent ISL detours, supported by geometric bounds and Starlink engineer confirmation [35, §7]. Coverage remains biased toward service-hosting customers, and satellite and gateway identities remain hidden.
Izhikevich 與同事透過 Censys 找到公開的 Starlink 客戶服務，再以 traceroute 辨識衛星區段兩側的跳點。成對 ping 的 RTT 相減，可將衛星區段從地面傳輸中分離。聖地牙哥終端驗證支持其準確度與尖峰偵測，涵蓋也廣於當時的 Atlas 探針[35, §4–6, Tab. 1]。區域結果呈現明顯延遲差異與間歇 ISL 繞行，依據為幾何下界及 Starlink 工程師確認[35, §7]。樣本偏向架設服務的客戶，衛星與閘道身分則持續隱藏於隧道中。
1. Where this paper sits1. 這篇論文的位置
Earlier measurements relied on dishes and a small set of Atlas probes, limiting geographic coverage. HitchHiking expands the unit of observation to reachable customers. Chapter 7 compares its endpoint coverage with Atlas and LENS, while Chapter 3 uses its estimate of laser-routing frequency in gateway-sparse regions. The work appeared in Proceedings of the ACM on Measurement and Analysis of Computing Systems, volume 8 issue 1, at SIGMETRICS 2024. The tables retain source wording differences for accuracy and coverage, with this digest using the stated multiplicative comparisons.
較早量測依賴終端與少量 Atlas 探針，地理涵蓋有限，HitchHiking 則將觀測對象擴展至可達客戶。第 7 章比較它與 Atlas、LENS 的端點涵蓋，第 3 章採用它對閘道稀疏地區雷射路由頻率的估計。論文刊於 Proceedings of the ACM on Measurement and Analysis of Computing Systems 第 8 卷第 1 期，並於 SIGMETRICS 2024 報告。表格保留準確度與涵蓋率的來源措辭差異，本導讀採其中標明的倍數比較。
2. The problem, measured2. 問題，用量測說話
The study addresses sparse observation coverage and the high resources needed for full-scale emulation [35, §6.1.3, §6.2]. It finds customer services through Censys, filters addresses by PTR names and removes known performance-enhancing proxies [35, §5]. The endpoint pool expands during the study, and reverse-DNS PoP codes such as “atlagax” locate the service exits [35, Tab. 1, §6.2]. The census table records filtering stages, regional shares and PoP counts so that coverage bias remains explicit.
研究針對觀測涵蓋稀疏，以及全尺度模擬平台資源需求高的問題[35, §6.1.3, §6.2]。方法透過 Censys 找出客戶服務，依 PTR 名稱過濾位址，再移除已知效能強化代理[35, §5]。研究期間端點集合持續擴大，反向 DNS 中如「atlagax」的 PoP 代碼可定位服務出口[35, Tab. 1, §6.2]。清單表列出過濾階段、區域比例及 PoP 數，使涵蓋偏差清楚可見。
| Item項目 | Value數值 | Conditions and interpretation條件與解讀 |
|---|---|---|
| Atlas coverageAtlas 涵蓋 | 58 probes, 5 countries, 60 s | Nearest probe over 300 miles from the test PoP. Full StarryNet emulation would require over 2 TB RAM.最近探針距測試 PoP 超過 300 英里，全尺度 StarryNet 需超過 2 TB 記憶體。 |
| Censys snapshotCensys 快照 | 10 May 2023: 4,521 / 2,051 | Exposed services and addresses in AS14593.AS14593 的公開服務數與位址數。 |
| Filtering stages過濾階段 | 1,790 → 1,629 | PTR customer filtering, then proxy removal. Peplink accounts for about 9% of services.先以 PTR 過濾客戶，再移除代理，Peplink 約占服務的 9%。 |
| Study pool研究端點集合 | ≈2,400 / 3,500 | Customer addresses and candidates, covering 13 countries and 20 PoP cities in May–June 2023.客戶位址與候選數，2023 年 5–6 月涵蓋 13 國及 20 個 PoP 城市。 |
| Regional service shares區域服務比例 | 59%, 7%, 6%, 3% | US, Australia, Germany and England respectively.依序為美國、澳洲、德國、英格蘭。 |
| US PoP services美國 PoP 服務數 | 243 / 210 / 186 / 182 / 173 / 144 | Seattle, Atlanta, Dallas, Chicago, Los Angeles and New York respectively.依序為西雅圖、亞特蘭大、達拉斯、芝加哥、洛杉磯、紐約。 |
| Other PoP services其他 PoP 服務數 | 148 / 124 | Sydney and Frankfurt respectively.依序為雪梨與法蘭克福。 |
| Measurement cadence量測頻率 | 2 probes/s, 5 min, 15 s | Paired streams and smoothing window.成對串流及平滑窗口。 |
| Fixed-satellite spikes固定衛星尖峰 | 2 of 5, zero drops | Observed spikes with the same serving satellite and packet delivery preserved.服務衛星相同，且封包持續送達時觀察到的尖峰。 |
| Nearest gateway use最近閘道使用 | ≥20% | Observed while the second-nearest station gives the best Hypatia comparison.觀察值，Hypatia 最佳比較設定則使用第二近站。 |
| Geometry overhead幾何開銷 | 12 / 13 ms | Selection and indirect ISL worst cases. The latter uses 1,243 miles spacing in a 22-satellite orbit.選擇與間接 ISL 的最差情況，後者採 22 顆衛星軌道中 1,243 英里間距。 |
| Validation azimuth驗證方位角 | <158° or >338°, offset −22° | Appendix E permits front-facing satellites within this constraint.附錄 E 在此限制內使用朝向前方的衛星。 |
| Spike shape from later work後續研究尖峰形狀 | ≈74 ms in first 140 ms | Casparsen et al. Tanveer et al. identify :12/:27/:42/:57 timing.Casparsen 等人的結果，Tanveer 等人則辨識 :12/:27/:42/:57 節律。 |
| Accuracy wording準確度措辭 | “up to 80% more accurate” / “1.8 times” | Required source quotations for the comparison.比較所需的來源原文引述。 |
| Coverage wording涵蓋措辭 | “an order of magnitude” / “28 times” / “more than 10 times” | Required source quotations. The digest uses 28×.必要來源原文引述，導讀採 28 倍。 |
| Earlier study coverage較早研究涵蓋 | 98 Atlas probes / 13 LENS dishes | Mohan et al. and LENS comparison.Mohan 等人與 LENS 的比較。 |
| OneWeb sampleOneWeb 樣本 | ≈20 endpoints | Appendix F, validation pending.附錄 F，仍待驗證。 |
| Ground-segment census地面段清單 | 21 → 49 PoPs, 5.98 million routers | 2023 versus December 2025. Lagos was the sole African PoP in October 2023.2023 年對照 2025 年 12 月，後者涵蓋 598 萬台路由器。2023 年 10 月拉哥斯是非洲唯一 PoP。 |
| Later company latency較晚公司延遲數據 | 2024–25 | SpaceX statements measure to the PoP.SpaceX 聲明量到 PoP。 |
3. The design3. 設計
Outside-in traceroute identifies the last responding hop before Starlink's tunnel and the first responding device after it. TTL-limited pings measure the earlier hop, ordinary pings measure the later one, and smoothed RTT differences isolate the satellite segment [35, §4, Fig. 3]. A San Diego dish and its obstruction map validate accuracy and sustained-spike detection [35, §6.1]. Some spikes occur while the serving satellite stays fixed, with zero packet drops, suggesting an additional scheduled mechanism. Hypatia's best tested ground-station configuration gives larger errors and misses the sustained spikes [35, §6.1.2]. Later studies characterize the scheduler behind these observations.
由外向內的 traceroute 找出 Starlink 隧道前最後一個回應跳點，以及隧道後第一個回應裝置。限制 TTL 的 ping 量測前者，一般 ping 量測後者，平滑後的 RTT 差分離出衛星區段[35, §4, Fig. 3]。聖地牙哥終端與遮蔽圖驗證準確度及持續尖峰偵測[35, §6.1]。部分尖峰發生時服務衛星保持固定，封包遺失也為零，呈現額外排程機制的線索。Hypatia 在測試中最佳的地面站設定仍有較大誤差，且漏掉持續尖峰[35, §6.1.2]，後續研究再刻畫其背後的排程器。
Original figure note原始圖說出處
Drawn for this primer from Izhikevich et al., Proc. ACM MACS 2024, §4 (Fig. 3), §6.1, §7.2 (Figs. 10–11).4. Evaluation setup4. 實驗設定
A single host measures the production network, with a San Diego dish as ground truth. Hypatia supplies a simulation baseline using its best tested ground-station choice [35, §6.1.2]. The method classifies paths using a geometric lower bound for two-satellite ISL RTT [35, §7.2]. A separate overhead calculation estimates satellite selection and an indirect ISL hop from assumed orbital spacing [35, §7.2]. Terminal azimuth limits in Appendix E belong to validation, while the remote method relies on reachable hops rather than equipment identity.
研究以單一主機量測營運網路，並以聖地牙哥終端提供真值，Hypatia 則採測試中最佳的地面站選擇作模擬基準[35, §6.1.2]。方法以兩衛星 ISL RTT 的幾何下界分類路徑[35, §7.2]，另依假設軌道間距計算衛星選擇與間接 ISL 跳段的開銷[35, §7.2]。附錄 E 的終端方位角限制屬於驗證，遠端方法則依賴可達跳點與差分。
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| About 2,400 customer endpoints in 13 countries at 20 PoP cities, 28× the 58 RIPE Atlas Starlink probes [35, Tab. 1, §6.2]13 國 20 個 PoP 城市約 2,400 個客戶端點，是 RIPE Atlas 58 個 Starlink 探針的 28 倍[35, Tab. 1, §6.2] | Exposed services found by Censys. 59% on US PoPs. Only 7 addresses overlapped with Atlas. Coverage is about 0.1% of roughly 2 M customers and biased toward people hosting services.由 Censys 找到的暴露服務。59% 在美國 PoP。只有 7 個位址與 Atlas 重疊。涵蓋約 200 萬客戶的 0.1%，且偏向架設服務的人。 |
| 96% of RTTs lie within about 10 ms, and 50% within 3 ms. All sustained spikes are captured, with one false positive over ≈10,000 s [35, §6.1]96% 的 RTT 在真值終端約 10 ms（一個標準差）內、50% 在 3 ms 內，100% 的持續尖峰被抓到，約 10,000 s 內一次誤報[35, §6.1] | Single validation dish in San Diego. Sustained spikes defined as more than 2σ for at least 15 s.聖地牙哥的單一驗證終端。持續尖峰定義為超過 2σ 且至少 15 s。 |
| Accuracy improves 1.8× over Hypatia, whose mean error is 7.6 ms and sustained-spike predictions remain absent [35, §6.1.2]準確度比 Hypatia 高 1.8 倍，後者平均誤差為 7.6 ms，並漏掉全部持續尖峰[35, §6.1.2] | Hypatia with second-nearest-ground-station configuration. The introduction's "up to 80% more accurate" is the same comparison worded differently.Hypatia 採第二近地面站設定。引言中「準確度高達 80%」是同一比較的另一種措辭。 |
| Nigeria: over 70% of RTTs fall below the two-satellite ISL minimum. Engineer-confirmed ISL periods produce 2–5× spikes about one-third of the time. Average 149 ms (SD 109) against Mexico 28 ms (SD 4), Brazil 104 (SD 27), Georgia US 58 (SD 21) [35, §7.2, Figs. 10–11]奈及利亞：超過 70% 的 RTT 低於兩顆衛星 ISL 最小值，但約三分之一時間有 2–5 倍尖峰，經 Starlink 工程師確認為 ISL 路由。平均 149 ms（SD 109），對照墨西哥 28 ms（SD 4）、巴西 104（SD 27）、美國喬治亞州 58（SD 21）[35, §7.2, Figs. 10–11] | Lagos was the only African PoP in October 2023 and Nigeria had two gateways about 80 miles apart. The Nigerian median fell about 25% from roughly 400 to 300 ms between May and June 2023 [35, Fig. 12].2023 年 10 月拉哥斯是唯一的非洲 PoP，奈及利亞有兩個相距約 80 英里的閘道。奈及利亞的中位數在 2023 年 5–6 月間從約 400 降到 300 ms，下降約 25%[35, Fig. 12]。 |
| Offshore Brazil customer averaged 96 ms where a direct ISL path would give 14 ms, with sustained spikes 16% of the time. Seychelles yacht: minimum RTT 181 ms versus a 40 ms geometric direct-ISL estimate. Rerouting covers 42% of time, adding 150 ms above minimum [35, §7.2–7.3]巴西外海客戶平均 96 ms，而直接 ISL 路徑應為 14 ms，16% 的時間有持續尖峰。塞席爾的一艘遊艇最小 RTT 為 181 ms，直接 ISL 路徑應為 40 ms，42% 的時間被重新路由至高於最小值 150 ms[35, §7.2–7.3] | Direct-ISL values are geometric estimates. Zigzag routing could reach 80 ms. These observations describe individual customers.「直接 ISL」值為幾何估計，之字形路由可合理達 80 ms，觀察範圍涵蓋單一客戶。 |
| Minimum RTT grows more than 3× with distance to the assigned PoP. A US Virgin Islands customer about 1,600 miles from Atlanta had nearly twice the minimum RTT of nearby customers. At least 70% of customers saw at least one sustained spike per day [35, §7.1, §7.4, Fig. 7]最小 RTT 隨到指派 PoP 的距離增加超過 3 倍。距亞特蘭大約 1,600 英里的美屬維京群島客戶最小 RTT 接近鄰近客戶的兩倍。至少 70% 的客戶每天至少看到一次持續尖峰[35, §7.1, §7.4, Fig. 7] | Every traceroute passes through the assigned PoP, making its distance the main geographic reference for the delay floor.所有 traceroute 皆經指派 PoP，延遲下限以該 PoP 距離為主要地理依據。 |
6. Limits and open questions6. 限制與待研究的問題
- The sample covers service-hosting customers, and automatic proxy removal targets Peplink [35, §8]. Other middleboxes may remain and add latency. Regional averages therefore describe this selected sample.樣本涵蓋架設服務的客戶，自動代理移除則針對 Peplink[35, §8]。其他中介設備可能持續存在並增加延遲，因此地區平均描述的是這組選定樣本。
- Validation covers a single San Diego dish, in a gateway-dense region with a nearby PoP. Applying its accuracy bound to ISL-dependent regions remains an assumption.驗證涵蓋聖地牙哥的一支終端，當地閘道密集且 PoP 鄰近。將其準確度界限延伸至依賴 ISL 的區域，仍屬模型假設。
- The MPLS tunnel hides satellite and gateway identities. Geometric minimum-delay comparisons help classify paths, but direct path observation requires additional visibility. Nigerian ISL attribution combines this bound with informal Starlink engineer confirmation.MPLS 隧道隱藏衛星及閘道身分。幾何最小延遲比較有助於分類路徑，直接觀察則需要更多可見資訊。奈及利亞的 ISL 歸因結合該下界與 Starlink 工程師的口頭確認。
- Scheduler interpretation comes from later studies. This paper observes repeated sustained spikes, including events with a fixed serving satellite. Cite Tanveer et al. for global timing and Casparsen et al. for the short spike shape.排程器解讀來自後續研究。本文觀察到反覆持續尖峰，其中部分事件的服務衛星保持固定。全球節律應引 Tanveer 等人，短期尖峰形狀則引 Casparsen 等人。
- Source wording varies for accuracy and coverage gains. The table preserves each wording, while the digest uses multiplicative ratios consistently. The OneWeb appendix endpoints still await validation.準確度及涵蓋增益的來源措辭各有版本，表格保留各種寫法，導讀則一致採用倍數比例。OneWeb 附錄端點仍待驗證。
- The results describe a dated ground-segment map. Later PoP expansion changes the scope of Nigerian, Brazilian and island comparisons. SpaceX's later latency claims measure to the PoP, whereas this study probes exposed services. company結果描述指定時期的地面段地圖，後續 PoP 擴建改變了奈及利亞、巴西與島嶼比較的適用範圍。SpaceX 較晚的延遲聲明量到 PoP，本研究則探測公開服務。company
7. How it connects7. 與其他內容的關聯
- Section 3.3 uses Nigerian observations to estimate bent-pipe and intermittent laser-routing use, with spikes illustrating detour cost.第 3.3 節採奈及利亞觀察估計彎管與間歇雷射路由的使用情況，尖峰則呈現繞行成本。
- Section 3.4: the 28 ms Mexico and 149 ms Nigeria averages bracket the chapter's table of measured RTTs by region.第 3.4 節：墨西哥 28 ms 與奈及利亞 149 ms 的平均值，框出該章各地區實測 RTT 表格的兩端。
- Section 3.5 relates fixed-satellite spikes to the later explanation of scheduled reconfiguration.第 3.5 節：固定衛星時的持續尖峰，提供排程影響的早期證據。
- Section 7.2: HitchHiking expands endpoint coverage. Its 1.8× accuracy gain illustrates the importance of measured spikes when evaluating simulators.第 7.2 節：HitchHiking 是觀測點階梯的最頂層，其對 Hypatia 1.8 倍的準確度優勢是該章模擬器漏掉真實尖峰的例子。
- Sibling digests: www24-mohan uses 98 Atlas probes. imc20-hypatia supplies the simulation baseline. pimrc23-pan lists the 21 contemporary PoPs. conext23-tanveer explains the spike cadence.相關導讀：www24-mohan 使用本文所倍增的 98 個 Atlas 探針，imc20-hypatia 是作為基準的模擬器，pimrc23-pan 列出同年的 21 個 PoP，而 conext23-tanveer 解釋尖峰節律。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| ≈ 2,400 · 13 · 20 | Customer endpoints · countries · PoP cities, May–June 2023 [35, Tab. 1, §6.2] | 客戶端點 · 國家 · PoP 城市，2023 年 5–6 月[35, Tab. 1, §6.2] | |
| 96% / 50% | Share of RTTs within ≈ 10 ms / 3 ms of a ground-truth dish [35, §6.1] | RTT 落在真值終端 ≈ 10 ms／3 ms 內的比例[35, §6.1] | |
| 1.8× · 28× | Accuracy gain over Hypatia · vantage points relative to 58 Atlas probes [35, §6.1.2, §6.2] | 對 Hypatia 的準確度增益 · 相對 58 個 Atlas 探針的觀測點倍數[35, §6.1.2, §6.2] | |
| 28 / 149 ms | Average RTT in Mexico (SD 4) and Nigeria (SD 109) [35, Fig. 11] | 墨西哥（SD 4）與奈及利亞（SD 109）的平均 RTT[35, Fig. 11] | |
| > 70% · ≈ ⅓ | Nigerian RTTs below the two-satellite ISL minimum · share of time with 2–5× ISL spikes [35, §7.2, Fig. 10] | 奈及利亞低於兩顆衛星 ISL 最小值的 RTT 比例 · 有 2–5 倍 ISL 尖峰的時間比例[35, §7.2, Fig. 10] | |
| 181 vs 40 ms | Seychelles yacht minimum RTT against the direct-ISL estimate [35, §7.3] | 塞席爾遊艇的最小 RTT 對直接 ISL 估計值[35, §7.3] | |
| 7.6 ms · 12 / 13 ms | Hypatia's average error · worst-case satellite-selection and indirect-ISL overheads [35, §6.1.2, §7.2] | Hypatia 的平均誤差 · 衛星選擇與間接 ISL 的最差開銷[35, §6.1.2, §7.2] |