latency-histogram-bin · research · also latency bucket

Latency histogram bin

A latency histogram bin counts observations assigned to a defined time interval, whose representative value need not equal any individual observation.

Histogram precision depends on bin boundaries, not the number of digits printed for a representative. In fio JSON+, duration keys represent intervals; preserve units, latency statistic and histogram layout before combining counts. [1]

A quantile selected from counts belongs to the selected interval. It is not thereby an exact latency or a confidence interval for a population quantile. A threshold inside that interval may leave the number of violations unresolved. Saturating overflow bins need separate treatment because their representative does not cap the observations. [2]

Hesela supplies a constructed count example, tested arithmetic and source hashes; no hardware benchmark was run. Counts are combined without reconstructing individual samples, and integer rank comparisons avoid floating-point threshold rounding. [3]

Relations

related

Quantile, Completion latency (fio), Tail latency

Citations

1. fio manual: JSON+ bin midpoints, log_hist_msec and log_hist_coarseness; accessed 21 October 2026. (accessed 2026-10-21)
2. fio 3.41 stat.c: plat_val_to_idx, plat_idx_to_val and count-based percentile selection. (accessed 2026-10-21)
3. Hesela: histogram precision, strict-threshold bounds and reproducible arithmetic (accessed 2026-10-21)

JSON: /data/terms/latency-histogram-bin.json