ssrjson-benchmark 0.0.1a0

Benchmark for ssrJSON

ssrJSON-benchmark

The ssrJSON benchmark repository.

Benchmark Results

The benchmark results can be found in results. Contributing your benchmark result is welcomed.

Quick jump for

• x86-64-v2, SSE4.2
• x86-64-v3, AVX2
• x86-64-v4, AVX512

Usage

To generate a benchmark report, you need to install ssrJSON either by fetched PyPi or built from source, and toolkit(ssrjson_benchmark) from this repo by:

python -m build
pip install dist/*.whl

Then run the benchmark script:

python benchmark.py

Benchmark options

• -m output in Markdown instead of PDF.
• -f <json_path> used exists benchmark json result.
• --process-bytes <bytes_num> Total process bytes per test, default 1e8.

Notes

• This repository conducts benchmarking using json, orjson, and ssrJSON. The dumps benchmark produces str objects, comparing three operations: json.dumps, orjson.dumps followed by decode, and ssrjson.dumps. The dumps_to_bytes benchmark produces bytes objects, comparing three functions: json.dumps followed by encode, orjson.dumps, and ssrjson.dumps_to_bytes.

• The ssrJSON built with the BUILD_BENCHMARK option includes several additional C functions specifically designed for executing benchmarks. These functions utilize high-precision timing APIs, and within the loop, only the time spent on the actual PyObject_Call invocations is measured.

• When orjson handles non-ASCII strings, if the cache of the PyUnicodeObject’s UTF-8 representation does not exist, it invokes the PyUnicode_AsUTF8AndSize function to obtain the UTF-8 encoding. This function then caches the UTF-8 representation within the PyUnicodeObject. If the same PyUnicodeObject undergoes repeated encode-decode operations, subsequent calls after the initial one will execute more quickly due to this caching. However, in real-world production scenarios, it is uncommon to perform JSON encode-decode repeatedly on the exact same string object; even identical strings are unlikely to be the same object instance. To achieve benchmark results that better reflect practical use cases, we employ ssrjson.run_unicode_accumulate_benchmark and benchmark_invalidate_dump_cache functions, which ensure that new PyUnicodeObjects are different for each input every time.

• The performance of JSON encoding is primarily constrained by the speed of writing to the buffer, whereas decoding performance is mainly limited by the frequent invocation of CPython interfaces for object creation. During decoding, both ssrJSON and orjson employ short key caching to reduce the number of object creations, and this caching mechanism is global in both cases. As a result, decoding benchmark tests may not accurately reflect the conditions encountered in real-world production environments.

• The files simple_object.json and simple_object_zh.json do not represent real-world data; they are solely used to compare the performance of the fast path. Therefore, the benchmark results should not be interpreted as indicative of actual performance.