Google GitGoogle Git
Sign in
apache/fory/HEAD/./benchmarks/kotlin
tree: 0d55765a7bacaffcbaf361e864c2e2b8cb51af4b
  1. src/
  2. .gitignore
  3. build.gradle.kts
  4. gradle.properties
  5. README.md
  6. run_json_benchmark.py
  7. settings.gradle.kts
benchmarks/kotlin/README.md

Kotlin JSON Benchmarks

This standalone Gradle/JMH project compares Fory JSON Kotlin, kotlinx.serialization, Moshi, and Jackson Kotlin on one immutable Eishay MediaContent model. It is a repository benchmark project, not a published Fory artifact.

The model uses only val properties and has no public zero-argument constructor. All four libraries consume the same model and fixture; no library-specific data transfer object or measured model conversion is used. The fixture SHA-256 is 8faba2f57ab397f319aced5cf1e8411a76785557d4c7d1703ec9d540354310a1.

Compared operations

The suite contains exactly 16 methods: four libraries multiplied by String serialization, UTF-8 byte serialization, String deserialization, and UTF-8 byte deserialization.

Each library uses a retained declared-type or generated serializer API:

  • Fory retains jsonTypeRef<MediaContent>(), disables asynchronous compilation for deterministic setup, enables null emission, and warms all four generated paths before measurement.
  • kotlinx.serialization retains MediaContent.serializer() and uses its String and stream APIs.
  • Moshi retains its KSP-generated adapter and uses its String and Okio buffer APIs.
  • Jackson Kotlin retains one ObjectReader and one ObjectWriter and uses its direct String and byte APIs.

The final byte materialization required by a library remains inside its measured byte-serialization method. Deserialization likewise includes any fresh in-memory stream or buffer required by that library. No byte method routes through a prebuilt String.

Correctness gates

Before timing, setup verifies that every library:

  • decodes the exact fixture from String and UTF-8 bytes to the independent expected object;
  • emits structurally equivalent JSON from String and byte APIs; and
  • round-trips its own String and byte output.

The Gradle build also fails unless the Moshi KSP adapter is present for every object model. Fory uses its normal HotSpot metadata path in this benchmark; Android retention rules do not participate in the measured runtime.

Build and run

Install the current fory-json-kotlin artifact in Maven local first. Use Gradle 9.3.0 and a JDK 17 or later toolchain.

Build the correctness and JMH artifacts:

gradle --no-daemon test verifyGeneratedJsonArtifacts jmhJar

Run the standard JMH benchmark:

python run_json_benchmark.py --output-dir reports/json

The runner invokes JMH once for all 16 methods. By default, each method uses three 2-second warmup iterations and five 2-second measurement iterations. Standard JMH JSON is written to benchmark_results.json; the full process log is retained in benchmark.log. For a quick harness check without collecting publishable numbers, pass --warmup-iterations 1 --iterations 1 --duration 100ms.

Use --prepare-only for CI correctness and Moshi adapter-generation checks without performance timing.

See the published Kotlin JSON benchmark report. The published page explicitly remains pending until a complete measured run is available; the tooling does not synthesize results or publish reports automatically.