测量方法
测量范围
- 数据时效
- 2 组:<=1h
- 数据覆盖
- 120 个有效窗口 / 194 个已发布窗口
测量详情
- 方向
- 所有已发布线路;每行标明起点 → 终点
- IP 版本
- IPv4 与 IPv6 分别展示
- 测量层级
- 自有端点测量(Layer 1)
- 测量模式
- 常规测量
- 时间范围与时区
- 数据截至 2026-09-27 11:48 UTC (2026-09-27 19:48 北京时间);快照生成于 2026-09-27 11:48 UTC (2026-09-27 19:48 北京时间)
- 统计时段(cn_local_bucket)
- 各窗口注明时段,统一按北京时间划分
- 最近有效样本
- 借用节点仅显示时间范围,不公开精确的末次采样时间
- 本地测量状态
- no 组:in this snapshot carries instrument_contaminated, background_traffic or a partial session
- 观测或推断
- 本页区分观测值与推断值
时区与统计时段
时间同时注明 UTC 和北京时间(Asia/Shanghai,UTC+08:00)。 基线时段与长期趋势 按北京时间划分,不随显示时区改变。 下载数据包含 UTC 偏移量。
测量指标的定义、参数和版本说明。
术语定义
| 术语 | 定义 | 参数 |
|---|---|---|
| typical / 平常值 | The middle of this exact pair, direction, address family, profile and CST time bucket's own valid history. | the most recent 28 valid days of the series, with at least 7 distinct dates in the bucket itself; the valid-day count is series-level and the distinct-date count is the gate each bucket passes on 自身, because a weekend bucket only occurs on about 8 of any 28 valid days and a per-bucket day count would keep it building for fourteen weeks; no fallback 至 a mixed all-hours baseline |
| 通常范围 / 通常范围 | The historical prediction range for the same CST time。 Loss uses a count-scale estimator and RTT a quantile band over the same bucket; neither is ever computed on this side of the snapshot. | loss: count-scale interval; loss is published as a numerator and a denominator and never as an averaged percentage; rtt: 同时段分位数区间 |
| 高于通常范围 / 高于平常 | The current value is clearly outside that same-bucket range and enough packets were sent for the comparison 至 mean anything. One window never triggers it. | minimum 200 packets in a window before any loss figure is published |
| 正在建立基线 / 正在建立基线 | This time bucket does not yet have enough valid history. No status word is shown, and no other bucket's history is borrowed 至 make one. | the most recent 28 valid days of the series, with at least 7 distinct dates in the bucket itself; the valid-day count is series-level and the distinct-date count is the gate each bucket passes on 自身, because a weekend bucket only occurs on about 8 of any 28 valid days and a per-bucket day count would keep it building for fourteen weeks; no fallback 至 a mixed all-hours baseline |
| 无明显变化 / 未见明确变化 | The data does not support a stable change. That is not the same as proving nothing changed. | 未发布 |
| 路径变化 / 可见路径改变 | The normalised visible path differs between two snapshots. It does not mean a physical line switched. | inference, with confidence and dataset version on every row |
| cn_local_bucket | Which of the six China-side local time buckets a window falls in. The bucketing clock is always Asia/Shanghai and never the visitor's timezone, because the day/night pattern is the thing being studied. | Asia/Shanghai (no DST); the China-side node's local time, whatever the direction |
| 未应答探测 / 未收到回复 | The fraction of single-ended probes that drew no reply. It is never written as loss. | 未发布 |
统计时段
时区: Asia/Shanghai (no DST); the China-side node's local time, whatever the direction. Whichever direction a measurement runs in, the bucket is cut on the China-side node's local time — the daily rhythm on this link is driven by users in China, and averaging it away would make the site announce "loss higher than usual" every single evening, which is a clock, not a finding.
weekday_overnight— 工作日凌晨 00:00–08:00(北京时间)weekday_day— 工作日白天 08:00–19:00(北京时间)weekday_peak— 工作日晚高峰 19:00–24:00(北京时间)weekend_overnight— 周末凌晨 00:00–08:00(北京时间)weekend_day— 周末白天 08:00–19:00(北京时间)weekend_peak— 周末晚高峰 19:00–24:00(北京时间)
状态只与同一时段的历史数据比较。历史数据不足时,页面显示 正在建立基线 ,不作异常判断,也不使用全天或相邻时段的基线代替。
历史基线
丢包的通常范围按包数估计,并根据实际时间相关性校准区间宽度。RTT 使用同时段的分位数区间。受污染窗口、不完整会话和专项测量不进入基线。
单向指标的发布范围
线路页面展示的单向指标是 owd_variation,表示 单向排队增量:当前窗口的最小单向时延减去同方向近期历史数据拟合的时延下限。数值表示相对该方向下限增加的排队时间,不表示绝对单向时延。
两端使用广域网 NTP 校时,时钟偏差会受路径不对称影响。同方向的读数减去其历史下限,可消去恒定时钟偏差,用于比较两个方向的排队增量。RTT 是双向之和,无法单独说明哪个方向发生变化。
如何估计时延下限
- 时延下限采用 滚动 线性拟合,依据同方向近期窗口估计下限及其漂移斜率。斜率用于检查时钟质量,避免把时钟漂移误报为排队。
- 拟合历史范围至少为 24 小时。过短的历史范围可能完全处于晚高峰拥塞期间,导致下限偏高、排队增量偏低。每个窗口提供实际使用的历史跨度。
- 时钟跳变或可见路径变化后重新建立下限,避免把新路径的固有时延变化误判为排队。
- 无法建立下限时,窗口提供 缺测标记与原因 — in
owd_flags,下载数据使用相同原因标记,不估算替代值。
为何不提供通常范围
此指标直接与本方向的时延下限比较,不另设通常范围。 owd_variation 自带参考值:0 ms 表示达到该方向已观测到的最低时延,80 ms 表示高出该下限 80 ms。路径变化后下限需要至少一天数据重建,覆盖可能稀疏且不连续,因此目前不计算此指标的历史通常范围。
为何不公开绝对单向时延
绝对单向时延估计值保留在归档和研究下载中,不在公开图表展示。仅当两端时钟持续正常且变化显著大于时钟不确定性时,它可作为辅助证据。广域网 NTP 的路径不对称误差无法由本仪器严格限定,因此不得据此比较数毫秒差异、推断城市位置或定位某一跳的时延。
短突发测量的发布范围
每条线路每小时发送两次 ICMP 短突发,与常规测量使用同一目标。每次约 30 个包,速率约 50 包/秒,持续约 0.6 秒。20 ms 的包间隔可区分分散未应答和连续未应答;每秒一包的常规测量无法分辨这两种结构。
线路页面提供 连续未应答长度分布 ,以及乱序次数、最大乱序距离和反映突发扩散程度的无量纲比值。结果按北京时间时段汇总。包数不足 200 的时段只保留计数,不给出汇总指标。页面不展示会掩盖连续结构的平均未应答率。
这些结果来自单端 ICMP 观测,不称为丢包。未收到应答可能涉及 ICMP 限速、去程、回程或目标主机的应答处理,无法归因于某一方向。
逐包时间未获公开授权
归档保留逐包发送和接收时间、逐包 RTT,以及发送和接收包间隔的最小值、中位数和最大值。 这些细粒度记录已测量并归档,但不在公开页面和文件中发布。
节点所有者授权的是城市级汇总,以及省级网络向外的逐跳信息。突发包间隔可能反映家庭线路的并发活动,超出当前授权范围。包间隔统计和精确突发时间也不公开,仅保留北京时间统计时段。
线路页面列明保留而未公开的字段,区分授权限制与未测量。今后获得相应授权后,可通过新的分析版本发布归档中的值。
检测器与误报率
项目自有硬件和 ASN 上的两个端点运行相同测量与检测器,不涉及借用线路或 NAT。该对照通道的触发率作为实际误报率展示,仅用于仪器检查,不进入研究数据集。
No detector versions are published in this snapshot.
五种数据状态
如何理解数据状态
- 有效数据
ok— 该时段收发记录充分。丢包率为 0% 表示未观测到丢包。 - 探针离线
probe_offline— 该时段探针未运行。此状态与其他原因导致的数据缺失分开标记。 - 缺少数据
no_data— 该时段没有测量数据,图中保留空白。 - 测量受本地因素影响
contaminated— 测量可能受到本地因素影响。数据保留,但不用于默认结论和基线。 - 疑似受到干扰
interference_suspected— 该时段目标对其他地点的探针有响应,但未响应此中国侧节点。原因尚未确定。
文字和符号区分数据状态,纹理标记受影响的时段。 零丢包 表示有有效数据且未观测到丢包。图中的空白表示没有数据。
五种状态分别展示。探针离线、数据缺失和实测零丢包具有不同含义。 疑似受到干扰 仅在同一窗口存在相应观测事件时使用。此状态描述观测现象,不确定具体机制。
测量限制
- public.yaml is the only publication authority; a node with no entry appears in no file
- borrowed-node data is delayed at least one hour, and freshness is published as a coarse bucket rather than a timestamp
- coverage is published / 天, never hour × date
- forward traces are truncated by publish_path_from; with rtt_threshold, a hop is judged on its minimum observed RTT and a hop that straddles the threshold across traces is withheld entirely
- CPE, CGNAT, home egress and access hops never appear, under any field
- a truncated hop leaves no placeholder: published hops are renumbered and carry no original TTL
- the absolute one-way delay estimate is 未发布: design §7 admits it only as supporting evidence under continuous NTP health, forbids reading a few milliseconds, a city or a single hop's queueing off it, and does not claim its error is a strict bound — printed in milliseconds on a chart it would be read as the one-way delay, which this instrument cannot state
- the instant a direction's rolling floor was at its quietest is 未发布: it is a second-precision statement about when one 借用线路 was least busy
- Every window is cut by the sending side's send instant (sender_send_time); the receiver never defines a window.
- Loss is published as sent and 已接收 counts. Percentages are derived for display and never averaged.
- The null control channel: FMT↔LAX runs the same baseline and the same detectors on own hardware, 自有 ASN, no line owner and no NAT; its trigger rate is the empirical false-alarm rate quoted on each observation
- ok — a valid window; 零丢包 is this state with counts showing nothing was lost (from: )
- probe_offline — the probe was not running; this is never folded into no_data (from: abuse_paused, maintenance, probe_offline)
- no_data — nobody measured this window; the reason is named and the line is not drawn across it (from: artifact_incomplete, consent_review_overdue, controller_unreachable, controller_unreachable_from_cn, deferred_busy_line, disk_pressure, egress_asn_mismatch, egress_hijacked, grant_plan_mismatch, not_cross_border, owner_paused_optional, partial, probe_capability_unavailable, reflector_down, reflector_unreachable_all, skipped_no_consent, skipped_quota, target_unreachable_persistent, v6_access_shape_unknown, v6_prefix_changed, wan_ip_changed, wan_reconnect)
- contaminated — the data is kept and kept out of default conclusions and baselines (from: background_traffic, instrument_contaminated, link_wireless, nat_rebind, resolution_untrusted)
- interference_suspected — another class of observation, shown only with a matching §11 observation in the same window (from: blocked_suspected, reflector_unreachable_from_cn)
- 本站不提供综合质量分。RTT、丢包、连通性、事务耗时和路径变化分别展示。
- RTT 不拆分为两个绝对单向时延。线路页面仅展示各方向相对自身下限的排队增量。逐跳 RTT 也不等于该跳链路的时延。
- 不能根据 RTT 推断城市位置。
- 可见路径是对路由跟踪有响应的路径,不等于物理线路。可见路径变化不能证明物理线路切换。
- 单端探测报告未应答,不报告丢包。这是两种不同的测量。
版本记录
No methodology changes are recorded in this snapshot.
历史图表可使用原分析版本重现,不混合不同版本的结果。