fwJson Benchmark — FiWorks Labs

Results

Overall Ranking

1st

simdjson

C++ — SIMD-accelerated

2nd

yyjson

C — pure scalar, no SIMD

3rd

fwJson

C — in-place zero-copy

4th

RapidJSON

C++ — widely used

5th

cJSON

C — simple & portable

Analysis

Key Findings

fwJson beats RapidJSON on every file tested — by 15% on float-heavy data up to 97% on string-heavy data. The advantage grows with string content because in-place parsing eliminates all string copying.

simdjson and yyjson are 2–3× faster than fwJson — simdjson uses SIMD instructions to process 16–64 bytes at a time. yyjson achieves similar speed with pure scalar C through compact memory layout, bump allocators, and unrolled loops.

cJSON is 2.5–7× slower than fwJson — it uses malloc per node and copies all strings, which adds up on large documents.

Throughput

Visual Comparison

canada.json — 2.2 MB, float-heavy coordinates

simdjson

606 MB/s

yyjson

623 MB/s

fwJson

385 MB/s

RapidJSON

292 MB/s

cJSON

52 MB/s

citm_catalog.json — 1.7 MB, mixed types

simdjson

1,759 MB/s

yyjson

1,448 MB/s

fwJson

821 MB/s

RapidJSON

714 MB/s

cJSON

241 MB/s

twitter1-full.json — 1.0 MB, string-heavy (fwJson's strength)

simdjson

1,031 MB/s

yyjson

945 MB/s

fwJson

374 MB/s

RapidJSON

190 MB/s

cJSON

84 MB/s

epa1-full.json — 195 KB, medium-sized NGSI-LD-like data

simdjson

1,154 MB/s

yyjson

1,120 MB/s

fwJson

371 MB/s

RapidJSON

212 MB/s

cJSON

137 MB/s

Data

Full Results

File	Size	fwJson	RapidJSON	simdjson	yyjson	cJSON
canada.json	2,198 KB	5,572 us	7,341 (1.32x)	3,593 (0.64x)	3,446 (0.62x)	40,986 (7.36x)
citm_catalog.json	1,687 KB	2,008 us	2,307 (1.15x)	936 (0.47x)	1,137 (0.57x)	6,842 (3.41x)
twitter1-full.json	1,032 KB	2,691 us	5,300 (1.97x)	977 (0.36x)	1,066 (0.40x)	11,997 (4.46x)
salessandri-small-dict.json	805 KB	1,525 us	2,186 (1.43x)	735 (0.48x)	638 (0.42x)	4,060 (2.66x)
snomed1-full.json	490 KB	1,520 us	2,477 (1.63x)	573 (0.38x)	499 (0.33x)	5,638 (3.71x)
books1-full.json	428 KB	1,383 us	2,586 (1.87x)	579 (0.42x)	484 (0.35x)	4,791 (3.46x)
epa1-full.json	194 KB	512 us	897 (1.75x)	165 (0.32x)	170 (0.33x)	1,383 (2.70x)
ngsild-batch-upsert.json	4.7 KB	8.8 us	14.0 (1.59x)	3.1 (0.35x)	3.8 (0.43x)	21.9 (2.48x)

Times in microseconds (lower is better). Ratio vs fwJson in parentheses (<1.00 = faster, >1.00 = slower).
100 iterations per file. CPU time via CLOCK_THREAD_CPUTIME_ID. All parsers compiled with -O2.

Analysis

Why fwJson Beats RapidJSON

In-place parsing — fwJson modifies the input buffer directly, replacing quote characters with null terminators. String values point into the original buffer with zero copying.
Minimal allocation — uses a pre-allocated buffer pool (fwAlloc) instead of per-node malloc calls.
Simple C implementation — no C++ template overhead, virtual dispatch, or exception handling.
Biggest advantage on string-heavy data — twitter1-full.json shows a 97% speed advantage because in-place parsing eliminates the most work on string-heavy documents.

Study

What Makes simdjson & yyjson Faster

simdjson

Uses SIMD instructions (SSE4.2, AVX2, ARM NEON) to process 16–64 bytes at a time
Two-pass architecture: first pass indexes all structural characters via SIMD, second pass builds the tree
Branchless number parsing avoids branch misprediction penalties
Runtime CPU detection with automatic fallback to the best available instruction set

yyjson

No SIMD at all — achieves near-simdjson speed with pure scalar C
Linear bump allocator — single large allocation, pointer increment per node
Compact 32-byte node representation
Manually unrolled parsing loops
Precomputed lookup tables for character classification
Immutable parse result allows simpler allocation strategy

Portability: yyjson's techniques are 100% portable across all CPU architectures. simdjson requires SIMD support but includes a scalar fallback. For the fw-libs, yyjson-style optimizations are the recommended path forward — maximum performance with zero portability risk.

Method

Methodology

Timer: CLOCK_THREAD_CPUTIME_ID — CPU time consumed by the calling thread only
Iterations: 100 per file, reporting average
Optimization: all parsers compiled with -O2
Parse mode: full DOM parse for all libraries (simdjson uses DOM parser, not on-demand)
Fairness: fwJson gets a fresh input copy each iteration (in-place parsing modifies input); all parsers allocate and free their trees each iteration
Logging: all FLOG trace macros compiled out (#ifdef DEBUG guard)

Versions

fwJson 0.10.2
RapidJSON 1.0.2
simdjson (system package, libsimdjson-dev)
yyjson (latest from GitHub master)
cJSON (system package, libcjson-dev)

Reproduce

sudo apt install libsimdjson-dev libcjson-dev
cd fwJson/bench
make
./run_bench.sh 100