Serializing object classes is 10x slower than data classes

[chrisbanes]

Copying over my Android Microbenchmarks results (ran on Google Pixel 3):

Benchmark	Time (ns)	Allocations
object_reflection_serializer	56,077	27
object_reflection_json	69,069	40
object_intrinsic_serializer	56,211	30
object_intrinsic_json	118,768	73
dataclass_reflection_serializer	2,450	8
dataclass_reflection_json	6,462	28
dataclass_intrinsic_serializer	2,687	11
dataclass_intrinsic_json	9,927	42

Also interesting is that the intrinsic is usually (marginally) slower than using reflection.

To Reproduce
All benchmarks are here: https://github.com/chrisbanes/kotlin-serialization-object-perf

Expected behavior
I'd expect objects to be faster than data classes (or at least similar).

Environment

• Kotlin version: 1.7.0
• Library version: 1.3.3 (1.4-RC is very similar numbers though)
• Kotlin platforms: JVM/Android

[qwwdfsad]

The problem has multiple layers in it.

First of all, intrinsics are not yet implemented (#1348), which is our biggest issue right now. I suggest reading through the issue to understand why it's important.

The second problem is that we have seemingly the same code having a significant difference in its performance characteristics.

All lookups in all benchmarks use reflection under the hood and use basically the same code path (more or less).
Optimistically, we assume that data classes are serialized much more frequently, thus it is prioritized on the serializer() code path and will always be faster.

Though, it shouldn't be ten times faster. I've ran my own benchmarks under the profiler and pin-pointed the root cause:

kotlinx.serialization/core/jvmMain/src/kotlinx/serialization/internal/Platform.kt

Lines 104 to 115 in 16a85df

	private fun <T : Any> invokeSerializerOnCompanion(jClass: Class<*>, vararg args: KSerializer<Any?>): KSerializer<T>? {
	val companion = jClass.companionOrNull() ?: return null
	return try {
	val types = if (args.isEmpty()) emptyArray() else Array(args.size) { KSerializer::class.java }
	companion.javaClass.getDeclaredMethod("serializer", *types)
	.invoke(companion, *args) as? KSerializer<T>
	} catch (e: NoSuchMethodException) {
	null
	} catch (e: InvocationTargetException) {
	val cause = e.cause ?: throw e
	throw InvocationTargetException(cause, cause.message ?: e.message)
	}

The problem here is simple -- we either have to use getDeclaredMethod and pay for it throwing an exception if the companion doesn't have serializer method (which is the case for object!), thus x10 performance loss on "uncommon path".

Or, we can use getDeclaredMethods().find { ... } as we did before (#952). Then, lookup of object serializer will be as good as lookup of regular class serializer. The problem is, regular lookups will become at least 2.5 times slower.

Java API won't let us do any better here, so this is a tradeoff we are willing to accept. Hopefully, proper intrinsics (#1348) will make all these problems disappear, so closing as "won't fix"

[qwwdfsad]

Just in case, flamegraph of object serializer lookup:

[chrisbanes]

Optimistically, we assume that data classes are serialized much more frequently

Yup, makes total sense at the library level. Thanks for attaching the flame graph. Very interesting!

For now, we're going to workaround this by checking for data::class.objectInstance and avoiding any serializer work when there is an instance. That has a cost, but it is miniscule compared to the serializer (370ns vs 56,000ns with my testing)

[qwwdfsad]

Thanks for reporting!

This report got #2004 rolling as well as our work on speeding up some [obvious] kotlin-reflect parts

[sandwwraith]

NB: You can also use generated serializer() method whenever possible (e.g. ObjectArgs.serializer()) to eliminate any reflection at all

[qwwdfsad]

I've tweaked reflection here and there (https://youtrack.jetbrains.com/issue/KT-50705/Use-ClassValue-to-cache-KClass-objects-in-kotlin-reflect and subtasks) and now there is no Kotlin-specific reflection whatsoever. The changes are targeting 1.8.0 release.

Along with #2004 (@shanshin works on that right now), we expect the difference between reified and .serialized() version to be almost negligible