benchmark

ojson benchmark

Measures ojson across four access patterns and three document shapes. Run from the repo root:

make bench

make bench passes -o:speed -microarch:native.

Cases

case	what it measures
parse	parse the whole document, reusing one Reader
extract	parse, then read a few known fields by path
access	parse once, then read fields from every array element
oneshot	full setup, parse, read all fields and teardown, every call

parse, extract and oneshot are the patterns ojson is built for. access is the opposite: it holds one parsed document and reads from it repeatedly, which is what a DOM parser is built for. It is here to keep that cost visible.

Test data

twitter.json, canada.json and citm_catalog.json are the canonical nativejson-benchmark files, also used by simdjson and RapidJSON. They cover string-heavy, number-heavy and large-nested shapes respectively.

Results

Measured on an i9-11900K (Arch Linux, Odin dev-2026-04, -o:speed -microarch:x86-64-v3). Each figure is the median of 5 trials.

case	document	size	per-op	throughput
parse	twitter.json	616.7 KB	266.96 µs	2.4 GB/s
parse	canada.json	2.1 MB	1.32 ms	1.7 GB/s
parse	citm_catalog.json	1.6 MB	704.62 µs	2.5 GB/s
extract	twitter.json	616.7 KB	276.21 µs	2.3 GB/s
extract	canada.json	2.1 MB	1.40 ms	1.6 GB/s
extract	citm_catalog.json	1.6 MB	712.67 µs	2.4 GB/s
access	twitter.json	616.7 KB	3.19 µs	-
access	citm_catalog.json	1.6 MB	5.65 µs	-
oneshot	small struct	103 B	228 ns	-

Throughput is omitted for access, where no parsing happens inside the timed loop, and for oneshot, where the 103-byte payload makes it meaningless. The harness prints a trailing checksum so the optimiser cannot discard the work.