Active Probing Methodology: Why We Use P50/P95 Instead of Averages

Fairness Starts with a Unified Test Object

If you want to compare CDN providers fairly, the first rule is simple: every provider must serve the same test object. We host the same standard test files on every CDN endpoint (500B / 50KB / 1MB) so the measurement scope stays consistent.

What We Measure

• Latency: DNS lookup + TCP handshake + TLS + time to first byte (TTFB)

• Throughput: real transfer bandwidth while downloading larger files

• Uptime: the success rate of probing requests that return HTTP 200

Why We Avoid Simple Averages

Latency distributions have a long tail. A single burst of network jitter can add hundreds of milliseconds and drag the arithmetic mean upward in a misleading way.

$$P95 = \text{the value that 95% of requests are faster than}$$

That is why we focus mainly on:

• P50 (median): the experience of a typical user

• P95: a worse-case-but-still-common experience, which exposes stability

Using percentiles keeps a small number of abnormal samples from distorting the whole conclusion.

Geographic Distribution of Probes

A single probe location only reflects one region. We launch measurements from multiple geographies around the world, including mainland China egress nodes, because the real user experience for mainland visitors can only be observed from domestic probe vantage points.

Refresh Cadence and Data Freshness

Probing runs hourly, then aggregates into 24-hour / 7-day / 30-day trends. Every value that appears on the page includes sample size and last-updated time. A ranking without sample support is not trustworthy.