why mono runtime's performance is lowwer than java(or dotnetcore) runtime?can you(or we) optimize mono runtime

[introspection3]

why mono runtime's performance is slower than java(or dotnetcore) runtime?can you(or we) optimize mono runtime

[lewurm]

It depends on the workload. Please post a specific (minimal) program where you experience performance issues, so we can look into it.

[introspection3]

please see this
https://github.com/kostya/benchmarks

[introspection3]

please reopen this issue

[introspection3]

@lewurm please reopen this issue
please see this
https://github.com/kostya/benchmarks

[Therzok]

I did a quick look at the bf2 benchmark, and I saw the following:

Kotlin does a dry warming run, so it doesn't include JIT time: https://github.com/kostya/benchmarks/blob/master/brainfuck2/bf2.kt#L89

The C# version does not:
https://github.com/kostya/benchmarks/blob/master/brainfuck2/bf.cs#L85

And the build and run file:
https://github.com/kostya/benchmarks/blob/master/brainfuck2/build.sh#L5-L6

On Mono, it uses mcs -debug -optimize+ (thus debug code), instead of csc /optimize+, while on .NET Core it uses a release build.

Edit: mcs -optimize+ is mostly a NOP.

[Therzok]

I'd say the benchmarks are rather incorrectly setup and don't reflect real stats between different runtimes.

It also uses mono -O=all, instead of the default set of optimizations, which can incur JIT overhead.

[Therzok]

On my machine using mcs:

➜  brainfuck2 git:(master) time mono ./bf.exe bench.b
ZYXWVUTSRQPONMLKJIHGFEDCBA
mono ./bf.exe bench.b  4.36s user 0.01s system 99% cpu 4.384 total

With csc /optimize+:

➜  brainfuck2 git:(master) time mono ./bf.exe bench.b
ZYXWVUTSRQPONMLKJIHGFEDCBA
mono ./bf.exe bench.b  4.34s user 0.02s system 99% cpu 4.371 total

Then I looked at the IL generated by that code, and noticed that the hot functions in the code are not inlined at all:

diff --git a/brainfuck2/bf.cs b/brainfuck2/bf.cs
index a93e6a3..a411c46 100644
--- a/brainfuck2/bf.cs
+++ b/brainfuck2/bf.cs
@@ -2,6 +2,7 @@ using System;
 using System.IO;
 using System.Linq;
 using System.Collections.Generic;
+using System.Runtime.CompilerServices;
 using System.Text;
 
 namespace Test
@@ -29,7 +30,9 @@ namespace Test
         }
 
         public int Get() { return tape[pos]; }
+		[MethodImpl(MethodImplOptions.AggressiveInlining)]
         public void Inc(int x) { tape[pos] += x; }
+		[MethodImpl(MethodImplOptions.AggressiveInlining)]
         public void Move(int x) { pos += x; while (pos >= tape.Length) Array.Resize(ref tape, tape.Length*2); }
     }

So I applied this patch and the runtime was cut down to:

➜  brainfuck2 git:(master) ✗ time mono ./bf.exe bench.b
ZYXWVUTSRQPONMLKJIHGFEDCBA
mono ./bf.exe bench.b  2.93s user 0.02s system 99% cpu 2.958 total

Even without a warm run.

[Therzok]

Using the same code:

➜  brainfuck2 git:(master) ✗ time dotnet bin/Release/netcoreapp2.0/brainfuck2.dll bench.b
ZYXWVUTSRQPONMLKJIHGFEDCBA
dotnet bin/Release/netcoreapp2.0/brainfuck2.dll bench.b  2.54s user 0.02s system 99% cpu 2.567 total

I get this with .netcore on the same machine for comparison.

[Therzok]

And my assumption was right, doing -O=all does increase runtime of the app:

➜  brainfuck2 git:(master) ✗ time mono -O=all ./bf.exe bench.b
ZYXWVUTSRQPONMLKJIHGFEDCBA
mono -O=all ./bf.exe bench.b  3.01s user 0.02s system 99% cpu 3.042 total