Pure Interpreter Speed

[Christian Humer]

Hi Stefan,

I quickly measured the interpreter performance for the first 20 runs for
deltablue on Truffle Javascript with the current tip.

I get on server (lower is better):
255,108,76,67,61,69,57,60,57,55,54,57,54,56,54,55,57,56,55,55
--- DeltaBlue finished (1418ms / 20 = 70.900ms/run)

And on client - graal tiered mode (lower is better):
1576,238,403,69,87,84,67,74,72,62,52,47,51,47,49,50,51,47,47,46 [ 1576 =>
46 ]
--- DeltaBlue finished (3219ms / 20 = 160.950ms/run)

The first runs on client-graal tiered are slower. As far as I know this is
expected as the graal compiler needs to warm up.
After warmup client-graal tiered is faster.




- Christian Humer

On Fri, Oct 24, 2014 at 11:01 AM, Stefan Marr <java at stefan-marr.de> wrote:

> Hi Bernhard:
>
> > On 24 Oct 2014, at 10:40, Bernhard Urban <Bernhard.Urban at jku.at> wrote:
> >
> > can you please try mx --vm server-nograal [...]? as the name tells, this
> is a server vm without graal modification, but built from the graal
> sourcetree.
>
> Sorry, no good news:
>
>  ../graal/mxtool/mx --vm server-nograal vm
> -Xbootclasspath/a:build/classes:../graal/truffle.jar som.vm.Universe -cp
> Smalltalk:Examples/Benchmarks/DeltaBlue
> Examples/Benchmarks/BenchmarkHarness.som DeltaBlue 100 0 1000
> Starting DeltaBlue benchmark ...
> DeltaBlue: iterations=1 runtime: 753905us
> DeltaBlue: iterations=1 runtime: 431390us
> DeltaBlue: iterations=1 runtime: 489044us
> DeltaBlue: iterations=1 runtime: 453485us
> DeltaBlue: iterations=1 runtime: 425499us
> …
>
> And just to make sure nothing else changed, I directly ran the normal one
> again:
>
> $ ../graal/mxtool/mx --vm server vm
> -G:TruffleCompileOnly=__something_that_really_doesnt_exist__so_that_nothing_is_compiled__
> -G:-TraceTruffleInlining -G:-TraceTruffleCompilation
> -Xbootclasspath/a:build/classes:../graal/truffle.jar som.vm.Universe -cp
> Smalltalk:Examples/Benchmarks/DeltaBlue
> Examples/Benchmarks/BenchmarkHarness.som DeltaBlue 100 0 1000
> Starting DeltaBlue benchmark ...
> DeltaBlue: iterations=1 runtime: 1336776us
> DeltaBlue: iterations=1 runtime: 1030942us
> DeltaBlue: iterations=1 runtime: 958083us
> DeltaBlue: iterations=1 runtime: 948710us
> DeltaBlue: iterations=1 runtime: 948175us
> DeltaBlue: iterations=1 runtime: 962985us
>
>
> > It would be a shame if the GraalTruffleRuntime would impose such an
> overhead. Overhead is expected, but I would say not in this order of
> magnitude.
>
> Well, it is a factor of 2 for the DeltaBlue on TruffleSOM.
> But, the other benchmarks I just retried look very similar.
>
> Best regards
> Stefan
>
> --
> Stefan Marr
> INRIA Lille - Nord Europe
> http://stefan-marr.de/research/
>
>
>
>