Latency Percentile Calculator

How this reads your output

Values are sorted and percentiles taken by nearest rank, which is the method monitoring systems use and what an SLA is normally written against. Alongside the table you get a distribution plot, because a percentile table hides shape: two datasets with identical p50 and p99 can be a smooth spread or two separate clusters, and those are entirely different problems with entirely different causes.

What the results mean

p95 and p99

The value 95% or 99% of samples fall below. These describe the experience of your slowest requests, which averages hide completely.

Tail ratio

p99 divided by p50. Above ten, you have a long tail rather than a slow system, and the cause is usually queueing or a retry path.

Sample count

Below a hundred samples, p99 is decided by one or two values and moves wildly between runs.

Common problems and fixes

The average looks fine but users complain

Look at p95 and p99. An average is dominated by the common case and says nothing about the tail, so a service can average 80 ms while one request in twenty takes two seconds. Percentiles are the standard answer precisely because they describe the experience at the edge rather than in the middle.

p99 jumps around between measurement periods

Too few samples. At a hundred samples, p99 is literally the single worst value, so one slow request moves it entirely. Either gather more samples or report p95, which is far more stable at small sample sizes and still describes the tail.

Frequently asked questions

More in Interpreter

• True Jitter Calculator (RFC 3550)Measure real jitter — the variation between consecutive packets, not deviation from the average.
• Ping Output AnalyserPaste any ping run and get latency, jitter and loss with a plain-English verdict.
• VoIP & Gaming Quality GraderGrade a connection for calls, gaming and streaming against real thresholds.