A First Look at Starlink Performance初探 Starlink 的效能

1. Where this paper sits1. 這篇論文的位置

This first peer-reviewed Starlink consumer-terminal study opens the measurement era in Chapter 3. Earlier Handley, Hypatia and motif studies establish the modeling era. Earlier LEO latency estimates came from simulators using FCC filings. Published measurements of user experience were still needed.

The paper is deliberately modest in scope. This occurs because it reports what one dish saw rather than proposing a mechanism. The paper establishes a baseline that later work confirms or refines. Mohan et al. confirm 40–50 ms idle medians across many countries.

Later 15-second scheduler studies explain loss and delay structure that Michel et al. describe statistically. It belongs in the same reading as Kassem et al., which appeared at IMC 2022 alongside it with a browser-extension method and the congestion-control comparison. The two are frequently conflated, so one purpose of this digest is to separate what each actually showed.

這是第一篇經同儕審查的 Starlink 消費級終端量測，開啟了第 3 章拿來與 Handley、Hypatia 及 motif 論文的「建模時代」對照的「量測時代」。在它之前，LEO 星座的每一個延遲數字都來自餵入 FCC 申請文件的模擬器，而真實使用者究竟體驗到什麼，量測答案仍待發表。這篇論文刻意保持謙遜的範圍，它報告一支終端看到的行為，機制則留待後續研究，而其持久價值在於訂下了之後研究或予以確認、或加以細化的基準，例如 Mohan et al. 在許多國家確認了 40–50 ms 的閒置中位數，而 15 秒排程器的論文則細化了 Michel et al. 只能以統計描述的遺失與延遲結構。它應與同在 IMC 2022 發表、採用瀏覽器擴充套件方法並包含壅塞控制比較的 Kassem et al. 一起閱讀，兩者經常被混為一談，所以本導讀的目的之一就是釐清各自到底證明了什麼。

2. The problem, measured2. 問題，用量測說話

The authors compare LEO behavior with terrestrial broadband and GEO satellite Internet. Three access paths from the same building provide the comparison. The idle latency result is the clearest.作者為自己設下的問題是：LEO 服務的行為更像地面寬頻，還是更像它要取代的 GEO 衛星網路？對照設計以同一棟建築的三條存取路徑直接回答。閒置延遲的結果最清楚。

Latency under load is where the GEO comparison stops being flattering. The authors attribute part of the upload inflation to quiche lacking pacing. Sender-stack behavior therefore contributes to the measured result.負載下的延遲是 GEO 對照限制更清楚的條件。作者把部分上傳膨脹歸因於當時 quiche 缺少 pacing，這提醒我們傳送端協定堆疊也是量測的一部分。

3. The design3. 設計

The design is a controlled measurement plan. Its strength is the three-way control. The tools are ordinary and that is the point. Traceroute, Tracebox and Wehe round out the path and middlebox analysis. The figure below shows the three paths and the headline latency ladder they produced.本研究的設計重點是量測計畫，而其強處在於三方對照。工具都很平常，而這正是重點。traceroute、Tracebox 與 Wehe 則補足路徑與中介設備分析。下圖顯示三條路徑與它們產出的延遲階梯。

4. Evaluation setup4. 實驗設定

Traceroutes to San Francisco, Singapore and European anchors shared European exits. The authors inferred inactive laser links for this terminal at that time. ISL rate, latency and topology therefore lie outside the study. The open data set is published. This is rare for this literature and lets a reader re-check every distribution.星間鏈路位於設定範圍之外，因為到舊金山與新加坡的 traceroute 與到歐洲錨點的 traceroute 走同樣的歐洲出口，作者將此解讀為當時這支終端的雷射鏈路處於停用狀態，所以論文中把 ISL 速率、延遲與拓樸列於範圍之外。資料集已公開，這在這類文獻中罕見，讓讀者能重新檢驗每一個分布。

5. What was measured5. 量到了什麼

6. Limits and open questions6. 限制與待解的問題

• One dish, one site, one season. Each distribution describes one Belgian terminal from late 2021 to mid 2022. The authors state this scope in their limitations ^{[48, §4]}. Each distribution applies to this terminal and observation period. Mohan et al. later found country-level median differences of tens of milliseconds associated with PoP location.一支終端、一個地點、一個季節。論文裡的每個分布都描述 2021 年底到 2022 年中比利時的單一終端，作者在限制一節^{[48, §4]}也如此說明。這裡的數字適用於該終端與觀測時段，國家或殼層整體數據需要更廣的取樣，而 Mohan et al. 之後顯示，本地與遠端 PoP 的國家之間中位數可差數十毫秒。
• CUBIC-only transport measurements. All QUIC tests used CUBIC inside quiche, and the paper evaluates CUBIC exclusively. Kassem et al. report BBR at about half the UDP-measured Starlink capacity, compared with over 90% on campus Wi-Fi. This IMC 2022 result is their Fig. 8. Cite Kassem et al. and see digest imc22-kassem.壅塞控制範圍限於 CUBIC。所有 QUIC 測試都在 quiche 內使用 CUBIC，論文中BBR 位於測試範圍之外。IMC 2022 那個「BBR 在 Starlink 上只達到 UDP 量到容量的約一半、在校園 Wi-Fi 上卻超過 90%」的結果，是 Kassem et al. 的 Fig. 8，見導讀 imc22-kassem，引用時應歸於他們。
• The probing resolution leaves scheduler behavior open. Five-minute pings and per-packet QUIC traces show message-test loss bursts lasting 104–127 ms at the 95th and 99th percentiles. Finer probes are needed to identify scheduler cadence. Tanveer et al. reveal a 15-second cycle at seconds 12, 27, 42 and 57 using 20 ms probes a year later. The drop-front packet-count queue that Garcia et al. characterized in 2026. Read the loss statistics as a description, with mechanism attribution requiring separate evidence.排程器機制需要更細的探測。每五分鐘一次的 ping 與 QUIC 流內的逐封包追蹤，足以顯示訊息測試中第 95 與第 99 百分位 104–127 ms 的遺失叢發，揭露下列現象則需要更細的探測： Tanveer et al. 一年後以 20 ms 探測發現的第 12、27、42、57 秒的 15 秒節律，以及 Garcia et al. 在 2026 年刻畫的 drop-front 封包計數佇列。請把遺失統計當成行為描述，機制解釋則參考後續研究。
• Exit paths provide indirect evidence about ISL use. The claim that lasers were inactive rests on traceroute exits. It is specific to a European terminal in 2022. Thus, it has a scope distinct from the ISL-dependent regions that Izhikevich et al. and Mohan et al. later documented.ISL 的使用是推論，直接量測仍待補充。雷射停用的說法建立在 traceroute 的出口上，且只適用於 2022 年的歐洲終端，其適用範圍限於該終端與時段。其他情境請參考 Izhikevich et al. 與 Mohan et al. 之後記錄的依賴 ISL 的地區。
• The numbers are dated. SpaceX reports US median latency of 33 ms in March 2024 and 25.7 ms in July 2025 after buffer and scheduler changes. These company measurements run from router to PoP. Their evidence scope differs from peer-reviewed studies. Thus, the 46–52 ms academic figure and the operator figure have separate measurement definitions and retain their own scope. company數字已有時效。SpaceX 報告美國中位延遲在緩衝與排程調整後於 2024 年 3 月降到 33 ms、2025 年 7 月降到 25.7 ms，但這些是營運商自己量的路由器到 PoP，屬於公司聲明，同儕審查適用於上述學術量測，所以 46–52 ms 的學術數字與營運商數字各有量測端點與時段，應各自引用。company
• Speedtest and QUIC disagree by design. The multi-connection speed-test median is 178 Mbit/s, while controlled QUIC downloads reach 100–150 Mbit/s. The paper records the gap and leaves its cause open. Later studies combining crowd-sourced speed tests and single-flow experiments encounter the same distinction.Speedtest 與 QUIC 在設計上就會差異。178 Mbit/s 中位數來自多連線測速，而受控的 QUIC 下載停在 100–150 Mbit/s，論文指出，原因仍待進一步分析，而這個差距在之後每一篇混用群眾測速與單流實驗的研究中都會再出現。

7. How it connects7. 與其他內容的關聯

• Section 3.3: the Belgian traceroutes exiting through Dutch and German PoPs are the textbook bent pipe. The absence of ISL evidence in 2022 anchors the chapter's statement that single-hop relay is the dominant mode in gateway-dense regions.第 3.3 節：比利時 traceroute 經荷蘭與德國 PoP 出口正是教科書式的彎管，而 2022 年缺乏 ISL 證據支撐了該章「在閘道密集地區單跳中繼是主要模式」的敘述。
• Section 3.4 starts its performance table with these measurements: 46–52 ms idle RTT, 178/17 Mbit/s throughput and 1.56%/0.40% loss.第 3.4 節：46–52 ms 的閒置中位數、178/17 Mbit/s 的上下行速率差距與 1.56%/0.40% 的遺失率，是該章實測效能表格的第一列。
• Section 3.5: the loss bursts and the gap between minimum and median RTT are what the 15-second schedule later explains.第 3.5 節：遺失叢發以及最小與中位 RTT 之間的差距，正是之後 15 秒排程所解釋的現象。
• Section 6.1: the loss and RTT-inflation numbers are the sender's view that the congestion-control chapter starts from.第 6.1 節：遺失與 RTT 膨脹的數字就是壅塞控制一章起步時傳送端所見。
• Sibling digest imc22-kassem covers the CUBIC-versus-BBR result and rain effects. www24-mohan extends idle-latency and bufferbloat observations to 34 countries. conext23-tanveer explains the scheduler mechanism.相關導讀：imc22-kassem 收錄同一會議的 CUBIC 對 BBR 與降雨結果，www24-mohan 把閒置延遲與緩衝膨脹的圖像擴展到 34 國，而 conext23-tanveer 補上了這篇論文需要更細探測才能解析的排程器。

Key numbers關鍵數字