json/tests/benchmarks/json_view/others/README.md

json_document against other JSON libraries

compare.py compares json_document with yyjson, simdjson and Boost.JSON. The programs here add sonic-cpp, RapidJSON and Glaze in C++, and sonic-rs, simd-json and serde_json in Rust, for the post How 3.13.0 got faster. They are not built by CMake or run by CI.

What is measured

The programs compare use cases, not library functions. A library may implement a use case in every reasonable way (another API, SAX events, in-situ parsing, a struct binding); each such way is a variant, named <library>: <variant> (yyjson: insitu (copy), RapidJSON: SAX, nlohmann json_document: accept). The report shows each library's fastest correct variant. nlohmann json (basic_json) and nlohmann json_document (with json_view) count as two libraries.

use case	what it does
validate	is this text valid JSON? returns a bool
sum	visit every value, convert every number, touch every string and key; sum all numbers in document order
select	read a few fields of every record (the fields of bench_view.cpp, e.g. id, user.screen_name and retweet_count of every status)
edit	change some values of every record and write the document as compact JSON (see below)
minify	text in, compact text out
write	write a document parsed beforehand as compact JSON; only the writing is timed

The edits:

file	edit
twitter.json	for every element of statuses: retweet_count + 1, erase entities, add "seen": true
status	the same edit on the one status
citm_catalog.json	for every element of performances: start + 1, erase seatMapImage, add "seen": true
canada.json	features[0].properties.name = "Kanada"; append [0.5,0.25] to the first ring (geometry.coordinates[0]) of every feature
jeopardy.json	for every record: value = null, erase air_date, add "seen": true

The files are twitter.json, citm_catalog.json, canada.json (nativejson-benchmark), jeopardy.json, the first status of twitter.json (status; written by json::dump(), timed in batches of 200), and twitter_min and citm_catalog_min: twitter.json and citm_catalog.json with all whitespace outside strings removed, built in memory at startup by the same small function in both programs (strip_ws; the bytes inside strings are untouched). The minified files use the select and edit of their originals.

Input contract

The input is a const std::string holding the file, with at least 64 bytes of spare capacity (reserve), so libraries that need padding (simdjson) can use it without a copy. A variant that modifies the input (in-situ parsing: yyjson YYJSON_READ_INSITU, RapidJSON ParseInsitu, simd-json) copies it inside the timed call; its name says (copy). Each timed call starts from the text and frees everything it built, except variants named (reused) or (reused allocator), which keep a parser, document or allocator across calls (simdjson's reused parsers, json_document::read(), Boost.JSON's reset parser and basic_parser, yyjson's dynamic allocator, RapidJSON's user-buffer allocators, simd-json's buffers). These are reported but never chosen as a library's best; summarize.py reports each library's best reused variant separately. In Rust, the input is the file as bytes (&[u8]), so every library has to validate UTF-8 itself.

Correctness rules

Before timing, every variant runs on the file and must produce the right result:

• validate: true.
• sum: the same number of values, the same number of string and key bytes, and a sum of all numbers that is bit-identical to json_document's (document order), so every number is converted exactly; variants that visit objects in another order (nlohmann json: parse, simd-json and serde_json values) are compared within 1e-9.
• select: the same value as json_document. A variant that does not convert numbers ((raw numbers), (numbers: none)) has to convert the numbers it selects exactly itself.
• edit, minify, write: the output, read with json::parse, compares equal (==) to the expected document (for edit, the edit applied to a json value; floats compare exactly, so every float must be written exactly). In Rust, serde_json::Value (with float_roundtrip) is the reference.

A variant that gives a wrong result on a file, or crashes on it, is printed as excluded: <file> <use case> | <variant> | <reason> and not timed.

Validation levels

Except for write, every variant is also probed with four malformed inputs derived from each file: the file truncated by one byte (after trailing whitespace), the file followed by ]x (trailing garbage), a stray , before the first } (before the last ] in a file without objects), and a 0xFF byte inside a string value (the first in the second half of the file, or the last one; never a key, so that a variant that accepts the byte still finds the keys it looks up; a file without string values gets no 0xFF probe). The probes run in a child process, so a crash is caught. The result is the variant's validation level, printed in its cell (sonic-cpp: Document [no UTF-8] 380.1us):

level	meaning
full	rejects all four probes
no UTF-8	rejects the structural probes, accepts the 0xFF byte
partial	accepts the truncated input, the trailing garbage or the stray comma, or crashes on any probe
n/a	write: there is no input to validate (counts as full)

What a variant accepts is printed as level: <file> <use case> | <variant> | <level> | <what>. A variant's level in the summary is the worst it got for any file. UTF-8 validation is enabled where a library has a switch (RapidJSON kParseValidateEncodingFlag; the (no encoding check) variants leave it off). sonic-cpp 1.0.2 cannot validate UTF-8 at all; serde_json does not validate UTF-8 in the parts it skips (IgnoredAny, unknown struct fields) unless the input is checked with std::str::from_utf8 first (the from_utf8 + ... variants).

The sum and select reference values are printed as check ... reference: lines; summarize.py warns if the C++ and Rust logs disagree.

The four probes do not cover everything a parser can get wrong. --conformance <dir> runs every validate variant on every file of JSONTestSuite's test_parsing (json_test_data has it as nst_json_testsuite2/test_parsing) and prints conformance: <file> | <variant> | accepted|rejected|crashed; a y_ file must be accepted, an n_ file rejected, and an i_ file may go either way. bench_all runs each call in a child process (a call that hangs for 60 s counts as crashed); a stack overflow aborts bench_rs, so conformance-rs.sh runs it one file per process and, for a file that aborts, one variant per process (CONF_VARIANT). The trailing garbage probe was added after this check: Glaze reads anything after the value unless validate_trailing_whitespace is set.

Variants

Level on the Mac in brackets where it is not full (P = partial, U = no UTF-8); every write variant is n/a.

library	validate	sum	select	edit	minify	write
nlohmann json_document	accept, parse, read (reused)	parse, read (reused)	parse, structs (from_json), read (reused)	editable, editable (source numbers) (json_editable_document: set/erase, dump with number_format::source)	parse + dump, parse + dump (source numbers), read + dump (reused)	dump, dump (source numbers)
nlohmann json	accept, parse	parse, sax	parse, structs (from_json)	parse, ordered_json	parse + dump	dump
yyjson	read, insitu (copy), read (raw numbers), read (reused allocator)	read, insitu (copy), read (reused allocator)	read, insitu (copy), read (raw numbers), read (reused allocator)	mut copy, insitu (copy) + mut copy (yyjson_doc_mut_copy)	read + write, insitu (copy) + write, read (raw numbers) + write, read + write (reused allocator)	write
simdjson	DOM, DOM (reused)	DOM, DOM (reused), On-Demand, On-Demand (reused)	DOM, DOM (reused), On-Demand [P], On-Demand (reused) [P], On-Demand structs [P], On-Demand (key selectors) [P]	n/a (no mutable document)	DOM + to_string, DOM (reused) + to_string, minify [P]	to_string
RapidJSON	SAX (null handler), Document, insitu (copy), SAX (raw numbers), Document (reused allocator)	Document, SAX, insitu (copy), Document (reused allocator)	Document, insitu (copy), Document (raw numbers), Document (reused allocator)	Document, insitu (copy)	Document + Writer, SAX to Writer, SAX to Writer (raw numbers), Document + Writer (reused allocator)	Writer
RapidJSON, (no encoding check)	the same [U]	the same [U]	the same [U]	the same [U]	the same [U]
sonic-cpp	Document [U]	Document [U]	Document [U]	Document [U]	Document + Serialize [U]	Serialize
Glaze	validate_json, generic [P], generic (double) [P], generic (strict), generic (double, strict)	generic [P], generic (double) [P], generic (strict), generic (double, strict), lazy_json [P], lazy_json (string_view) (x)	structs [P], structs (validate_skipped), generic [P], generic (double) [P], generic (strict), generic (double, strict), lazy_json [P], lazy_json (string_view) (x)	generic [P], generic (double) (x), generic (strict), generic (double, strict) (x)	generic [P], generic (double) (x), generic (strict), generic (double, strict) (x), minify_json [P]	generic, generic (double) (x), generic (strict), generic (double, strict) (x)
Boost.JSON	SAX (basic_parser), parse, SAX (numbers: none), parse (numbers: none), SAX (reused), parser (reused)	parse, SAX, SAX (reused), parser (reused)	parse, parse (numbers: none) (x), parser (reused), parse_into structs	parse	parse + serialize, parser (reused) + serialize	serialize
serde_json	IgnoredAny [U], from_utf8 + IgnoredAny, Value	visitor, Value	structs [U], from_utf8 + structs, Value	Value	Value	Value
sonic-rs	IgnoredAny, Value	visitor, Value	structs, Value	Value	Value	Value
simd-json	tape (copy), borrowed (copy), tape (reused)	visitor (copy), borrowed (copy), borrowed (reused)	structs (copy), borrowed (copy)	owned (copy), borrowed (copy)	borrowed (copy)	borrowed

(x): excluded for a wrong result on some file, and so in the summary for every file: generic (double) loses the 64-bit ids of twitter.json when it writes them; lazy_json (string_view) reads the raw bytes of a string, so escaped strings have the wrong length (sum, and the " in 19,077 jeopardy.json categories in select); numbers: none reads every number as 0.

structs read into types that declare only the fields select uses and hold the same data: nlohmann/json's from_json() of the NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE macro (from a json value or from a json_document's view) reads into nsel, where the arrays and objects that are only counted (seatCategories, the values of events, coordinates) are a count_only holding their size. Glaze reads into the same types (reflected; count_only has a reader that skips and counts the elements; error_on_unknown_keys = false, and with validate_skipped to validate what it skips), and so does simdjson On-Demand (get<T>() with tag_invoke, fields in document order, count_elements()/count_fields()). serde derive reads into the same fields with a Count that counts with IgnoredAny (and a BTreeMap, as std::map). Boost.JSON's parse_into has no type that skips any value, so its counted members only count what it converts into them: seatCategories into a counting sequence of an empty described struct, coordinates into three levels of counting sequences down to double, and the events values into an empty described struct, so their members are not counted. key selectors are simdjson 5.0's object::for_each<"k1", ...>, which walks each object once and matches keys with a compile-time perfect hash. visitor is a serde DeserializeSeed that sums without building a value; SAX and sax are the libraries' event interfaces with a summing (or, for validate, an empty) handler. RapidJSON parses with kParseFullPrecisionFlag (correctly rounded floats; without it, sums and written floats are wrong) and Boost.JSON with number_precision::precise, as every other library rounds correctly by default.

Variants named (raw numbers) or (numbers: none) use a library's mode that validates numbers without converting them (yyjson YYJSON_READ_NUMBER_AS_RAW, RapidJSON kParseNumbersAsStringsFlag, Boost.JSON number_precision::none), as json_document converts a number only when it is read. In select they convert the selected numbers themselves (std::from_chars); Boost.JSON's none leaves no number text in its DOM, so it cannot. yyjson writes a raw number verbatim, so read (raw numbers) + write is exact. sonic-cpp 1.0.2 has no parse flags, and Glaze's generic DOM has no raw number mode (validate_json converts no number anyway). (reused allocator): yyjson's yyjson_alc_dyn_new(), RapidJSON's MemoryPoolAllocators over user buffers for the values and the parse stack; Boost.JSON's parser (reused) builds its value in a monotonic_resource over a buffer kept across calls. These buffers grow after a call that needed more, so the timed calls do not allocate. generic (double) is glz::generic, which stores every number as a double. The Glaze (strict) variants read with validate_trailing_whitespace (else Glaze accepts anything after the value) and write with escape_control_characters (else Glaze writes a string's control characters, e.g. U+0001, unescaped, which is not JSON; gsoc-2018.json has some).

library	version	notes
json_document, json	this checkout
yyjson	0.13.0
simdjson	5.0.8
sonic-cpp	1.0.2
RapidJSON	master 24b5e7a8b27f (2025-02-05)
Glaze	9.1.0	generic_u64 and generic (owning DOMs, member order kept)
Boost.JSON	1.92.0	monotonic resource
sonic-rs, simd-json, serde_json	see rust/Cargo.lock	serde_json with float_roundtrip

SHA-256 of the downloads:

55f8a675397c505f9bee6c160e1c26b0085868f185d04d7a7727531fdbff4bd4  glaze-v9.1.0.tar.gz
30edf9e47c417f653546c39295fdb66ba5f89fb37eb3780122c6ee66605e5837  sonic-cpp v1.0.2.tar.gz
2d2601a82d2d3b7e143a3c8d43ef616671391034bc46891a9816b79cf2d3e7a8  rapidjson 24b5e7a8b27f42fa16b96fc70aade9106cf7102f.tar.gz
34e0f62a2bc11ab20d601e8ca1cc2b2079503aa45119a19133d89d19b94a0fae  yyjson 0.13.0.tar.gz
356cc3a2d6a3d0d45bf5eba813e58ef4f2f6d9cc2a25d8aada2563bc13755789  simdjson-v5.0.8.tar.gz
c4a3b310ddd2472416e091067166b0713be97c63f38c212c484ada022fd296ce  boost_1_92_0.tar.gz

Timing

Every variant of a use case runs interleaved in every round, each timed call after an untimed one, and the best round is reported (the method of bench_view.cpp). The order of the variants changes in every round (a shuffle with a fixed seed; the first round keeps the listed order): on the Mac, a variant's time can depend on which variant ran before it, by up to 15 %, and a fixed order would give each variant the same neighbour in every round. jeopardy.json (50 MB) gets a fifth of the rounds (at least 3, at most the given number).

Building and running

All libraries except yyjson and simdjson are used header-only. With the downloads unpacked into DL (compare.py --download fetches yyjson and Boost into its build directory, and simdjson 4.6.11, so take simdjson 5.0.8 from its release archive):

cc  -O3 -DNDEBUG -c $DL/yyjson-0.13.0/src/yyjson.c -o yyjson.o
c++ -std=c++17 -O3 -DNDEBUG -c $DL/simdjson-5.0.8/singleheader/simdjson.cpp -o simdjson.o
c++ -std=c++23 -O3 -DNDEBUG -DRAPIDJSON_NEON \
    -I ../../../../include -I $DL/yyjson-0.13.0/src -I $DL/simdjson-5.0.8/singleheader -I $DL/boost_1_92_0 \
    -I $DL/glaze-9.1.0/include -I $DL/rapidjson/include -I $DL/sonic/include \
    bench_all.cpp yyjson.o simdjson.o -o bench_all
./bench_all <json_test_data> 30 > mac-cpp-1.log

cd rust && cargo build --release --locked && ./target/release/bench_rs <json_test_data> 30 > ../mac-rs-1.log

python3 summarize.py mac-cpp-*.log --rust mac-rs-*.log --json mac.json

With GCC 16 and Homebrew Clang on macOS, build the three steps with the same flags plus --sysroot=$(xcrun --show-sdk-path) (gcc-16, g++-16) or -isysroot $(xcrun --show-sdk-path) (/opt/homebrew/opt/llvm/bin/clang, clang++). The key selectors and simdjson's tag_invoke need C++20, so bench_all.cpp needs -std=c++20 or newer.

Both programs take <json_test_data> [rounds] [file] [use case]: ./bench_all <json_test_data> 30 canada edit times only the edit of canada.json; "" stands for all files; the file names are twitter, citm_catalog, canada, jeopardy, status, twitter_min and citm_catalog_min. The checks cover only the selected files and use cases. bench_all uses fork() for the probes and the crash check (macOS, Linux).

--files <dir> [rounds] [file] [use case] runs validate, sum, minify and write (select and edit are defined only for the files above) on every *.json file of a directory, named by their stem; files under 500 KB are timed in batches of about 500 KB of input. It is meant for files the libraries were not tuned on, such as the jsonexamples of simdjson-data without twitter.json, citm_catalog.json and canada.json. summarize.py takes such logs as well.

On x86-64, use -march=haswell -DRAPIDJSON_SSE42 instead of -DRAPIDJSON_NEON (sonic-cpp needs AVX2), pin the process to one core, and time one file per process. Glaze 9 needs std::expected, which Clang 18 and older cannot use from GCC 13's standard library (Clang 19 and newer can); -DBENCH_NO_GLAZE=1 builds without it. For Rust, build with RUSTFLAGS="-C target-cpu=haswell" to match -march=haswell.

Reading the results

A log line holds the file, the use case, and one cell per variant: | <library>: <variant> [<level>] <time>us (<ratio>x), the ratio relative to the first variant of the line. summarize.py prints, for every use case and file, (a) per library the best full variant and, if faster, the best variant at any level, with their ratios to the best full nlohmann json_document variant ((reused) and (reused allocator) variants are never a library's best; a reused: line below the library shows its best reused variant in the same way), (b) all variants with their levels and times, and (c) what the variants below full accept, and the excluded variants. --json writes {use case: {file: {library: {"best": {"variant", "us", "level"} or null, "best_any": {...}, "best_reused": {...} or null, "best_reused_any": {...} or null, "variants": {variant: {"us", "level"}}, "excluded": {...}}}}} for the chart scripts (best_reused: the best full reused variant; best_reused_any: at any level). Several logs are combined by taking each cell's best time; a variant excluded for one file is left out for every file of that use case. On a shared machine, differences of a few percent are noise: run twice and compare.