faf-json-bench is a Rust JSON serialization and deserialization benchmarking tool.
Only Linux is supported.
I was looking for benchmarks and couldn't find any that included all the libraries I wanted tested.
You must comment out incompatible RUSTFLAGS in ./cargo/config. There will usually only be two ("-Clinker=/usr/bin/clang-16","-Clink-arg=-fuse-ld=/usr/bin/ld.lld"). These are so I can use clang / llvm when running the test but aren't necessary to see results on the benchmark.
After that, if using clang, run it with ./clang_build_and_run.sh, otherwise run it with cargo build --release && sudo nice -20 ./target/release/faf-json-bench
sonic-rs to_vec 785,789,775 bytes/sec
sonic-rs to_writer 929,728,115 bytes/sec
serde_json to_vec 668,032,647 bytes/sec
serde_json to_writer 770,139,985 bytes/sec
serde_json_core to_vec 598,008,355 bytes/sec
serde_json_core to_slice 652,524,808 bytes/sec
nanoserde serialize_json 302,670,533 bytes/sec
nanoserde ser_json 515,064,108 bytes/sec
simd_json to_vec 1,165,577,270 bytes/sec
simd_json to_writer 1,262,988,965 bytes/sec
simd_json_derive json_vec 1,056,956,455 bytes/sec
simd_json_derive json_write 1,334,914,161 bytes/sec
sonic-rs to_vec 691,168,747 bytes/sec
sonic-rs to_writer 915,829,798 bytes/sec
serde_json to_vec 608,365,940 bytes/sec
serde_json to_writer 759,952,561 bytes/sec
serde_json_core to_vec 633,651,573 bytes/sec
serde_json_core to_slice 652,826,200 bytes/sec
nanoserde serialize_json 280,116,027 bytes/sec
nanoserde ser_json 495,267,370 bytes/sec
simd_json to_vec 927,485,208 bytes/sec
simd_json to_writer 1,234,747,046 bytes/sec
simd_json_derive json_vec 863,682,707 bytes/sec
simd_json_derive json_write 1,330,641,520 bytes/sec
sonic-rs to_vec 686,959,477 bytes/sec
sonic-rs to_writer 912,113,921 bytes/sec
serde_json to_vec 587,635,143 bytes/sec
serde_json to_writer 748,133,447 bytes/sec
serde_json_core to_vec 599,925,430 bytes/sec
serde_json_core to_slice 644,265,518 bytes/sec
nanoserde serialize_json 300,077,492 bytes/sec
nanoserde ser_json 515,323,623 bytes/sec
simd_json to_vec 938,033,529 bytes/sec
simd_json to_writer 1,263,035,453 bytes/sec
simd_json_derive json_vec 869,011,511 bytes/sec
simd_json_derive json_write 1,335,306,561 bytes/sec
sonic-rs to_vec 788,752,848 bytes/sec
sonic-rs to_writer 921,088,350 bytes/sec
serde_json to_vec 681,721,777 bytes/sec
serde_json to_writer 760,149,347 bytes/sec
serde_json_core to_vec 568,485,797 bytes/sec
serde_json_core to_slice 644,376,763 bytes/sec
nanoserde serialize_json 346,096,400 bytes/sec
nanoserde ser_json 515,741,321 bytes/sec
simd_json to_vec 1,222,398,979 bytes/sec
simd_json to_writer 1,235,489,129 bytes/sec
simd_json_derive json_vec 1,052,723,187 bytes/sec
simd_json_derive json_write 1,335,253,937 bytes/sec
- LLVM / Clang (v16)
- MiMalloc Allocator
- Elevated priority
- Minimized system-wide noise. I always take precautions to ensure my benchmarking isn't adversely affected by other running programs by not having anything open/running, including most background services
- Compiled with flags specified in
./cargo/config, which include"-Ctarget-cpu=native"- - To ensure results aren't optimized away, I always check the output after serializing / deserializing for correctness
- To ensure, loops aren't unpredictably vectorized, I check the time on each iteration instead of doing a fixed number of iterations. My time check uses VDSO on Linux which is very fast. I thought Rust's standard library also used VDSO on Linux but when I switched to my own implementation times improved by over 10%
- Currently only testing with 'small input'
System info for context
Unless there is some unexpected popular demand, I only want to support what I actually use and would prefer not to add the complexity to support other platforms. Also, I often use techniques for performance reasons that are Linux specific and getting similar results on other platforms may be complex.
- Choice of allocator matters the most when using methods that allocate a buffer (.. yep)
- For
serde_json_core to_vec, checking the len withassert!(bytes_len == 26);after the serialization actually makes the operation faster, probably because it is a compiler hint - Clang is faster than gcc for some tests but not others. For example, the
serde_json to_writeris slightly faster with gcc but bothserde-json-coretests are much faster with Clang
Contributions are welcome, but please discuss before submitting a pull request. If a discussion leads to a pull request, please reference the #issue in the pull request. Unsolicited pull requests will not be reviewed nor merged.
All code is licensed under AGPL 3.0 unless an individual source file specifies otherwise.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent