Increased Runtime Duration for Java Functions in .NET Android Compared to Xamarin Android

[MrAlbin]

Android framework version

net9.0-android, net8.0-android

Affected platform version

VS 2022 17.12.3

Description

I recently converted a Xamarin Android app to .NET Android (net8.0-android). I noticed that while the startup performance has improved, other functions are taking longer to execute at runtime. This is particularly evident when executing Java functions. In a simple comparison between Xamarin and .NET, I measured the following times for executing Java.Lang.Math.Sqrt:

• Xamarin: 14 µs
• .NET: 25 µs

How can this significant performance difference be explained?

Attached is a speedscope file that I recorded for the .NET app.
dotnet-dsrouter.exe_20241208_140545.speedscope.json

Steps to Reproduce

N/A

Did you find any workaround?

N/A

Relevant log output

[jpobst]

Are you testing Debug or Release configurations? One change I'm aware of is that Debug uses an interpreter instead of a JIT for a better debugging experience.

You can disable this with <AndroidUseInterpreter>false</AndroidUseInterpreter>.

@jonathanpeppers Any other thoughts?

[jonathanpeppers]

@MrAlbin you don't mention what configuration you were using to test this. AOT, LLVM, Interpreter, etc.? Debug mode has an interpreter by default in .NET 6+, while Xamarin had no interpreter at all.

Can you share the BenchmarkDotNet project(s)? I know there are currently a few workarounds to even run BenchmarkDotNet on Android.

[MrAlbin]

I use Release configuration.

The sample app can be found here: https://github.com/MrAlbin/Benchmark

[jonathanpeppers]

I noticed you run BenchmarkDotNet in a custom way:

• https://github.com/MrAlbin/Benchmark/blob/master/Benchmark.Shared/BenchmarkCreator.cs

How do things compare if you run the standard BenchmarkDotNet pipeline like this:

• https://github.com/dotnet/maui/blob/1c4f5a31732aea6ee07344a652d9ca259e7ec50f/src/Core/tests/Benchmarks.Droid/MainInstrumentation.cs#L64-L70

BenchmarkDotNet does some statistics that is over my head, so I trust they do a better job at the math than I would by manually using Stopwatch.

[jonathanpeppers]

Oh wait, you aren't using BDN at all. I just saw classes named Benchmark and thought you were.

[MrAlbin]

I had initially tried BenchmarkDotNet but had issues getting it running with Xamarin Android. Thus I have done it without any benchmark library.