Speed Optimization for short living applications

[Fruchuxs]

We figured out, that we can speed up the executionen of an short living application (like a console application) with Expression<TDelegate>.Compile(preferInterpretation: true), because of avoiding some JIT overhead.

Can you maybe add a rule to DryIoc, so we can switch between preferInterpretation: true and preferInterpretation: false for the compilation of the resolve trees?

[dadhi]

@Fruchuxs Hi, thanks for bringing it.

First, what platform did you target?

Next, there maybe a more work involved, considering that on some platforms DryIoc uses custom expression compiler.

But everything is possible. In FEC we probably immediately fallback to Expression.Compile when preferInterpretation: true is passed.

[Fruchuxs]

We also tested the FEC, but the speed up was in our cases negligible.
Thanks for your acceptance.

[dadhi]

@Fruchuxs, what platform did you test?

[Fruchuxs]

Oh sorry, forget to advise: .net core 2.1

[EamonNerbonne]

It's conceivable that on .net core 2.1 you might get a similar boost by enabling tiered compilation.

It's certainly worth trying out; if that really works well, it might render the interpretation of expression trees themselves moot (because the JIT would be interpreting IL, and the conversion to IL is fairly fast with FastExpressionCompiler).

[dadhi]

Hi @EamonNerbonne , are you suggesting to add <TieredCompilation>true</TieredCompilation> and check?

[EamonNerbonne]

Yeah, that would be a start ;-).

But I have no experience with that (new) feature, so can't tell you about any gotchas - I mean I hope there aren't any, but it's conceivable it doesn't work with dynamically generated IL or some such thing. Only one way to find out! ;-).

[Fruchuxs]

Will check but their measurements doesn't look promising.

[Fruchuxs]

COMPlus_TieredCompilation

true	false
1,2543458	1,1927924
1,1643181	1,7478098
1,1927924	1,3724727
1,1572949	1,413623
avg	avg
1,1921878	1,431674475

Probes in seconds.
In my case it seems mostly faster, thanks for the info.

[EamonNerbonne]

Just a little worried about the high variance in those numbers - it might just be chance.

[Fruchuxs]

Yes, this is a general problem on my working machine (it's a laptop), as suggested here. I assume it's a CPU boost / throttling Problem. But my measurements gives an idea that COMPlus_TieredCompilation=true is faster at cold start.

[dadhi]

@Fruchuxs ,

Cool actually. I was sceptical, but it is a visible difference. Moreover, it affects dynamic methods as well.

Thanks for trying!

[dadhi]

#45 should help.