A Multifaceted Look at Starlink Performance多面向檢視 Starlink 的效能
Starlink latency depends strongly on PoP placement and load. Examine the global measurements, application effects and controlled evidence for scheduled reconfiguration.Starlink 延遲深受 PoP 位置與負載影響，導讀檢視全球量測、應用效應及受控終端對排程式重新配置的證據。
Mohan and colleagues combine M-Lab tests, RIPE Atlas probes, application sessions and controlled dishes to map Starlink performance across regions. They measure geographic latency differences, load-induced bufferbloat, asymmetric goodput and higher delay on islands that rely on distant gateways [49, §3–6, Figs. 5–16]. Aligned changes at Edinburgh and Munich persist even with a single satellite visible, supporting a global scheduling explanation. The paper appeared at the Web Conference 2024. ISL use is inferred from latency geography, while direct path identification requires further visibility.
Mohan 與同事結合 M-Lab 測試、RIPE Atlas 探針、應用測試及受控終端，建立跨區域的 Starlink 效能圖像。量測呈現地理延遲差異、負載造成的緩衝膨脹、上下行 goodput 差距，以及依賴遠方閘道的島嶼較高延遲[49, §3–6, Figs. 5–16]。愛丁堡與慕尼黑的同步變動，在僅一顆衛星可見時仍持續，支持全球排程解釋。論文發表於 Web Conference 2024，ISL 使用則由延遲地理推論，直接辨識路徑仍需更多可見資訊。
1. Where this paper sits1. 這篇論文的位置
This study tests whether earlier Belgian and British latency and bufferbloat findings extend across regions. It combines crowd-sourced tests, distributed probes and controlled dishes, linking earlier single-site work to later measurement methods. HitchHiking expands reachable endpoint coverage, while Tanveer et al. explains the global timing observed here. The primer uses this Web Conference 2024 paper for measured performance and the bent-pipe latency budget. PoP placement is central to its explanation of latency outside North America and Europe.
本研究檢驗較早比利時與英國的延遲及緩衝膨脹發現，如何延伸至其他區域。群眾測試、分散探針與受控終端，將單點研究銜接至後續量測方法。HitchHiking 擴大可達端點涵蓋，Tanveer 等人則解釋此處觀察到的全球節律。本教材以這篇 Web Conference 2024 論文建立實測效能及彎管延遲預算，PoP 位置是北美與歐洲以外延遲解釋的核心。
2. The problem, measured2. 問題，用量測說話
M-Lab identifies Starlink through AS14593 and uses major national mobile operators as terrestrial baselines. Atlas probes provide controlled cloud targets [49, §3]. Regional RTT distributions show higher delays and longer tails in some European and South American locations [49, §4, Figs. 5–7]. Goodput and TCP-inferred loss vary inversely [49, Figs. 9, 21]. The Manila PoP opening provides a before-and-after comparison: local traffic avoids the former Japan detour and latency falls sharply [49, Fig. 19]. The tables preserve the regional percentiles and measurement conditions.
M-Lab 透過 AS14593 辨識 Starlink，並以各國主要行動業者作地面基準，Atlas 探針則提供受控雲端目標[49, §3]。區域 RTT 分布顯示，部分歐洲與南美地點具有較高延遲及較長尾端[49, §4, Figs. 5–7]，goodput 與 TCP 推得的遺失率呈反向變化[49, Figs. 9, 21]。馬尼拉 PoP 啟用提供了前後比較：本地流量略過原先日本繞行後，延遲大幅下降[49, Fig. 19]。表格保留各區百分位與量測條件。
Download load raises RTT more than upload load, locating the deep queue on the downlink side [49, §4, Fig. 8]. Zoom measurements connect packet loss with additional forward-error-correction traffic and directional delay differences [49, §5, Fig. 10]. Luna gaming results place Starlink closer to cellular than wired service under the tested conditions [49, §5, Tab. 1]. These application examples illustrate mechanisms, while broader population claims require more sessions.
下載負載比上傳更明顯地拉高 RTT，將深佇列定位於下行側[49, §4, Fig. 8]。Zoom 量測將封包遺失連結至額外的前向錯誤更正流量，以及各方向的延遲差異[49, §5, Fig. 10]。Luna 遊戲在測試條件下的 Starlink 結果，較接近行動服務[49, §5, Tab. 1]。這些應用案例用於說明機制，延伸至整體使用者則需要更多測試場次。
| Item項目 | Value數值 | Conditions and interpretation條件與解讀 |
|---|---|---|
| M-Lab archiveM-Lab 檔案庫 | 19.2 M tests / 34 countries | Starts June 2021. Each country uses its top three mobile operators as terrestrial baselines.自 2021 年 6 月起，各國以當地前三大行動業者為地面基準。 |
| Atlas seriesAtlas 序列 | 98 probes / 21 countries / 1.8 M measurements | December 2022–September 2023, to 145 cloud endpoints from seven providers.2022 年 12 月至 2023 年 9 月，涵蓋七家供應商的 145 個雲端端點。 |
| Regional p75 RTT區域 p75 RTT | Seattle <50 ms, Rome/Paris +≈20 ms | Rome and Paris compared with Dublin, London and Berlin. European tails exceed 100 ms.羅馬與巴黎相對都柏林、倫敦、柏林，歐洲尾端超過 100 ms。 |
| South American RTT南美 RTT | p75 >100 ms, tails ≈200 ms | Figure-based regional distributions.由圖讀取的區域分布。 |
| Download versus upload load下載對照上傳負載 | p60 ≈200 ms / ≤100 ms | Download and upload RTT respectively.依序為下載與上傳 RTT。 |
| Zoom uplink delayZoom 上行延遲 | 52 ± 14 / 27 ± 7 ms | Starlink and fiber, one-way.Starlink 與光纖的單向延遲。 |
| Zoom downlink delayZoom 下行延遲 | 35 ± 11 / 32 ± 7 ms | Starlink and fiber, one-way.Starlink 與光纖的單向延遲。 |
| Zoom FECZoom 錯誤更正 | 146 ± 99 / 2 ± 2 kbps | Starlink and fiber, one call.Starlink 與光纖，單次通話。 |
| Luna idle RTTLuna 閒置 RTT | 9 / 46 / 40 ms | Terrestrial, cellular and Starlink respectively.依序為地面、行動、Starlink。 |
| Luna game delayLuna 遊戲延遲 | 133.53 / 165.82 / 167.13 ms | Same order, one 150-minute session.相同順序，一次 150 分鐘測試。 |
| Luna 1080p timeLuna 1080p 時間比例 | 100% / 94.11% / 99.45% | Same order and session.相同順序與場次。 |
| Terminal–gateway segment終端至閘道區段 | Mostly <1,200 km | Crowd-sourced gateway location model.使用群眾蒐集閘道位置的模型。 |
| Cellular comparison行動比較 | ≈1.5× | Bent-pipe delay relative to the cellular last mile.彎管延遲相對行動最後一哩。 |
| Gateway–PoP source ranges閘道至 PoP 來源範圍 | ≈5 ms global, 7–8 ms US/Canada, ≥6 ms North America | Main text versus Appendix D, summarized as 6–8 ms for North America.正文與附錄 D，北美概括為 6–8 ms。 |
| Controlled sky views受控天空視野 | 53° / 70° and 97.6° | German dish versus shielded Scottish dish, with one-second gRPC polling.德國終端與加遮罩的蘇格蘭終端，採一秒 gRPC 輪詢。 |
| Island inference島嶼推論 | ≈9,000 km, ≈60 ms advantage | Réunion–Frankfurt, relative to fiber. Island class RTT is ≈100 ms.留尼旺至法蘭克福相對光纖，島嶼類型 RTT 約 100 ms。 |
| Later spike measurement後續尖峰量測 | 500 Hz, ≈74 ms in first 140 ms | Casparsen et al., above the period mean.Casparsen 等人，相對週期平均值。 |
| Later company measurements較晚公司量測 | 33 ms / 25.7 ms | US peak-hour medians in March 2024 and July 2025, measured to PoP after fq_codel and buffer changes.2024 年 3 月及 2025 年 7 月美國尖峰中位數，fq_codel 與緩衝調整後量到 PoP。 |
| Company telemetry interval公司遙測間隔 | 15 s | SpaceX router telemetry statement.SpaceX 路由器遙測聲明。 |
3. The design3. 設計
The measurement layers provide complementary evidence. M-Lab gives broad coverage with coarse public-IP location, Atlas supplies controlled targets and long time series, and local dishes add detailed instrumentation. Crowd-sourced gateway and PoP locations divide RTT into terminal-to-gateway, gateway-to-PoP and PoP-to-target segments [49, §6.1, Figs. 12–13, App. D]. The resulting bent-pipe budget supplies the baseline for later idle-RTT comparisons. Each method's location, access-technology and time-resolution limits remain part of the interpretation.
各量測層提供互補證據。M-Lab 涵蓋廣，位置依公開 IP 粗略推定，Atlas 提供受控目標與長期序列，本地終端則加入詳細儀器。群眾蒐集的閘道與 PoP 位置，將 RTT 分為終端至閘道、閘道至 PoP，以及 PoP 至目標三段[49, §6.1, Figs. 12–13, App. D]。得到的彎管預算提供後續閒置 RTT 比較的基準，各方法的位置、存取技術及時間解析度限制仍需一併解讀。
Original figure note原始圖說出處
Drawn for this primer from Mohan et al., WWW 2024, §3, §4 (Fig. 8), §6.1 (Figs. 12–14), Appendix D.4. Evaluation setup4. 實驗設定
All results come from production-network measurements. The archive, Atlas series and single application sessions cover separate periods, while the two local dishes use controlled sky views [49, §3, §6.2]. The Réunion–Frankfurt latency advantage over fiber supports an ISL inference, with satellite identities hidden [49, §6.1, Fig. 14]. Islands lacking nearby gateways require distant ground exits. Good weather, fixed terminals and one-second polling define the controlled experiment's operating and timing scope.
結果全部來自營運網路量測，檔案庫、Atlas 序列與單次應用測試涵蓋各自期間，兩支本地終端則採受控視野[49, §3, §6.2]。留尼旺至法蘭克福相對光纖的延遲優勢，支持 ISL 推論，衛星身分仍隱藏於路徑中[49, §6.1, Fig. 14]。附近閘道缺席的島嶼需要遠方地面出口，良好天氣、固定終端及一秒輪詢，則界定受控實驗的操作與時間範圍。
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Median minimum RTT about 40–50 ms in most countries. European tails above 100 ms. South American 75th percentiles above 100 ms [49, §4, Figs. 5–7]多數國家最小 RTT 中位數約 40–50 ms。歐洲尾端超過 100 ms。南美第 75 百分位超過 100 ms[49, §4, Figs. 5–7] | Atlas cloud targets and M-Lab provide these values, read from figures because per-country tables remain absent.Atlas 雲端量測加上 M-Lab 提供這些值，逐國數據由圖讀取。 |
| Bent-pipe latency 36–48 ms, median near 40 ms. Gateway-to-PoP about 5 ms (6–8 ms in North America). About 100 ms on gateway-less islands [49, §6.1, Figs. 12–15, App. D]彎管延遲為 36–48 ms、中位近 40 ms，閘道至 PoP 約 5 ms（北美 6–8 ms），閘道缺席島嶼約 100 ms。[49, §6.1, Figs. 12–15, App. D] | Decomposition relies on crowd-sourced gateway and PoP locations. The text gives 7–8 ms for the US and Canada. Appendix D gives a lower bound of 6 ms.分解依賴群眾蒐集的閘道與 PoP 位置，正文的美加範圍與附錄 D 的北美下界各有表述，來源表保留兩者。 |
| RTT inflates 2–4× to 400–500 ms during downloads. Upload-load 60th-percentile RTT at or below 100 ms [49, §4, Fig. 8]下載時 RTT 升為 2–4 倍至 400–500 ms，上傳負載下第 60 百分位 RTT 至多為 100 ms[49, §4, Fig. 8] | M-Lab ndt7 uses BBR. Cross-controller comparisons require a separate study.M-Lab ndt7 使用 BBR，其他控制器的比較需另外評估。 |
| 75th-percentile goodput 50–100 Mbps down, 4–12 Mbps up. 75th-percentile TCP loss 4–8% [49, §4, Figs. 9, 21]第 75 百分位 goodput 下載 50–100 Mbps、上傳 4–12 Mbps。第 75 百分位 TCP 遺失 4–8%[49, §4, Figs. 9, 21] | Public-IP geolocation identifies crowd-sourced tests. Access technology behind the dish remains unspecified, and loss is inferred from TCP.群眾測試依公開 IP 定位，有線及 Wi-Fi 類型待區分，遺失則由 TCP 推得。 |
| Zoom uplink one-way delay 52 ± 14 ms against 27 ± 7 ms on fiber, FEC 146 ± 99 against 2 ± 2 kbps. Luna game delay 167.13 ms against 133.53 ms wired [49, §5, Fig. 10, Tab. 1]Zoom 上行單向延遲 52 ± 14 ms 對光纖 27 ± 7 ms，FEC 146 ± 99 對 2 ± 2 kbps。Luna 遊戲延遲 167.13 ms 對有線 133.53 ms[49, §5, Fig. 10, Tab. 1] | One call and one 150-minute game session illustrate mechanisms. Population statistics require broader sampling.一通通話與一場 150 分鐘遊戲，用於說明機制，整體統計仍需更廣樣本。 |
| Latency and throughput shift at 15-second boundaries aligned between Edinburgh and Munich, even with a single satellite in view [49, §6.2, Fig. 16]延遲與吞吐量在愛丁堡與慕尼黑對齊的 15 秒邊界切換，即使只有一顆衛星可見[49, §6.2, Fig. 16] | Two dishes use one-second gRPC polling with clock drift. Restricted views admit 13% of candidate satellites. Connection endings align with boundaries, while starts vary.兩支終端使用一秒 gRPC 輪詢，站點間存在時鐘漂移。受限視野允許 13% 的候選衛星，連線結束對齊邊界，開始時間則各有差異。 |
| Manila PoP in May 2023 cut Philippine in-country latency about 90% after a 50–70 ms detour via Japan [49, Fig. 19]2023 年 5 月馬尼拉 PoP 啟用後，菲律賓境內延遲下降約 90%，此前經日本繞行代價 50–70 ms[49, Fig. 19] | Before/after comparison on the same probes. The strongest single piece of evidence that PoP placement sets latency.同一組探針的前後比較。是 PoP 位置決定延遲最有力的單一證據。 |
6. Limits and open questions6. 限制與待研究的問題
- Broad tests and controlled sites answer separate questions. Public-IP locations and unspecified access technology limit interpretation of the crowd sample. Controlled mechanisms come from fixed dishes in good weather, so their operating scope remains local.廣泛測試與受控站點回答各自範圍的問題。公開 IP 定位及待確認的存取技術，限制群眾樣本的解讀。受控機制則來自好天氣下的固定終端，因此操作範圍維持於當地條件。
- Geography supports indirect ISL evidence. The island comparisons reveal delay consistent with laser routing, while identities and global traffic shares require additional observations. HitchHiking later adds engineer-confirmed Nigerian routing evidence.地理提供間接 ISL 證據，島嶼比較呈現與雷射路由相符的延遲，身分及全球流量比例仍需更多觀察。HitchHiking 後續加入工程師確認的奈及利亞路由證據。
- One-second polling and clock drift bound the scheduling evidence. Cross-site alignment locates global boundaries, while later high-frequency probes resolve the short spike shape. The authors describe their timing explanation as a theory supported by observations [49, §6.2].一秒輪詢與時鐘漂移界定了排程證據的解析度。跨站點對齊可定位全球邊界，後續高頻探測才解析短期尖峰形狀。作者將此時間解釋稱為觀察支持的理論[49, §6.2]。
- BBR defines the congestion-control scope. The bufferbloat results describe ndt7 under that controller. Kassem et al. provides the IMC 2022 BBR–CUBIC comparison.BBR 界定了壅塞控制範圍，緩衝膨脹結果描述 ndt7 在此控制器下的行為。IMC 2022 的 BBR 與 CUBIC 比較應參閱 Kassem 等人。
- Gateway-to-PoP source statements give a regional range and a separate lower bound. Cite their locations with the range, preserving geographic scope.閘道至 PoP 的來源陳述包含區域範圍及另一項下界。引用時應附位置並提供範圍，同時保留地理條件。
- Later SpaceX reports use a PoP endpoint after router queue-management changes. Their dates, values and telemetry interval appear in the evidence table. Comparing them with academic cloud-target RTT requires matched endpoints and conditions. company較晚的 SpaceX 報告採用路由器佇列管理調整後的 PoP 端點，日期、數值與遙測間隔集中列於證據表。與學術雲端目標 RTT 比較時，需對齊端點及條件。company
7. How it connects7. 與其他內容的關聯
- Section 3.3 uses the segmented bent-pipe budget and the island detour comparison.第 3.3 節：彎管分解、5 ms 的閘道到 PoP 一段與 100 ms 的 ISL 島嶼就是該章的延遲預算。
- Section 3.4 collects regional RTT, goodput, loss and bufferbloat measurements in its performance table.第 3.4 節的效能表集中列出區域 RTT、goodput、遺失與緩衝膨脹量測。
- Section 3.5 uses Edinburgh–Munich alignment and fixed-satellite observations to support a global schedule.第 3.5 節：愛丁堡與慕尼黑的對齊，支持全球排程邊界的解釋。
- Section 6.1 and Section 6.5: the bufferbloat and the Zoom FEC and delay figures are the application-layer symptoms that the transport chapter explains.第 6.1 節與第 6.5 節：緩衝膨脹以及 Zoom 的 FEC 與延遲數字，是傳輸層一章所解釋的應用層症狀。
- Section 7.2 compares Atlas probe coverage with HitchHiking's 2,400-endpoint sample.第 7.2 節比較 Atlas 探針涵蓋與 HitchHiking 較大的端點樣本。
- Sibling digests: imc22-michel supplies the single-dish baseline. sigmetrics24-hitchhiking expands ISL-region coverage. conext23-tanveer explains scheduling. imc26-fries is the same group's later video-conferencing study.相關導讀：imc22-michel 是本文所推廣的單終端基準，sigmetrics24-hitchhiking 觸及本文只能窺見的依賴 ISL 的地區，conext23-tanveer 給出排程器機制，而 imc26-fries 是同一團隊之後的視訊會議研究。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 19.2 M · 98 · 2 | M-Lab tests (34 countries) · Atlas probes (21 countries) · controlled dishes [49, §3] | M-Lab 測試（34 國）· Atlas 探針（21 國）· 受控終端[49, §3] | |
| 40–50 ms | Median minimum RTT in most countries, Dec 2022–Sep 2023 [49, §4, Figs. 5–7] | 多數國家的最小 RTT 中位數，2022 年 12 月–2023 年 9 月[49, §4, Figs. 5–7] | |
| 36–48 ms · ≈ 5 ms | Bent-pipe latency across countries · gateway-to-PoP leg, 6–8 ms in North America [49, §6.1, App. D] | 各國彎管延遲 · 閘道到 PoP 一段，北美 6–8 ms[49, §6.1, App. D] | |
| ≈ 100 ms | Bent-pipe class latency on gateway-less islands (US Virgin Islands, Réunion, Falklands) [49, §6.1] | 閘道缺席島嶼的彎管級延遲，涵蓋美屬維京群島、留尼旺及福克蘭[49, §6.1] | |
| 2–4× → 400–500 ms | RTT inflation during downloads [49, §4, Fig. 8] | 下載期間的 RTT 膨脹[49, §4, Fig. 8] | |
| 50–100 / 4–12 Mbps | 75th-percentile goodput down/up from M-Lab [49, §4, Fig. 9] | M-Lab 第 75 百分位 goodput 下載／上傳[49, §4, Fig. 9] | |
| 52 ± 14 vs 27 ± 7 ms | Zoom uplink one-way delay, Starlink vs fiber. FEC 146 ± 99 vs 2 ± 2 kbps [49, §5, Fig. 10] | Zoom 上行單向延遲，Starlink 對光纖。FEC 146 ± 99 對 2 ± 2 kbps[49, §5, Fig. 10] | |
| ≈ 60 ms · ≈ 90% | Réunion beats fiber to Frankfurt by this margin [49, Fig. 14] · Philippine in-country latency cut by the Manila PoP [49, Fig. 19] | 留尼旺到法蘭克福比光纖快此幅度[49, Fig. 14]· 馬尼拉 PoP 使菲律賓境內延遲下降的比例[49, Fig. 19] |