harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Mikhail Fursov" <mike.fur...@gmail.com>
Subject Re: [perf] Comparative benchmarking
Date Thu, 08 Nov 2007 14:55:02 GMT
Aleksey, if JNI stub has no significant impact here we can check
another solution, that was already proposed by Maxim: process
'Math.sqrt' instruction as API magic in JIT. It will eliminate all
negative side-effects related to CC. What do you think about it?

Another possible solution here is results caching in case of
quasi-constant parameters. Do you know if RI has such an optimization?


On Nov 8, 2007 4:47 PM, Aleksey Shipilev <aleksey.shipilev@gmail.com> wrote:
> Hi, Yuri,
>
> I had applied LightJNI to Harmony-r589548 and here come the results
> for sqrt() tests again. I have reproduced the boost on Dacapo:jython
> and Dacapo:lusearch, so I presume the patch was applied correctly. I
> have also specify the interested method explicitly:
>   -XX:jit.arg.codegen.light_jni.light_jni_methods=::Java_java_lang_Math_sqrt
>
> In short, LightJNI brings another 5% of boost, but Harmony is still
> 300% slower than Sun 1.6 on sqrt() microtest.
>
> That's probably due to presence of call. There are really have two
> directions where to go:
>  a. inline hot native methods into JITted code (sic!) :)
>  b. implement hot native methods as magics
>
> Details:
>
> == Harmony-r589548 (clean)
> ========================================================
> Dacapo:lusearch: 2900 msecs
>
> Warmup finished: 4.527292719905836E10
> Math Result = 2.1081849486439312E10 in 12344ms
> StrictMath Result = 2.1081849486439312E10 in 12109ms
>
>
> == Harmony-r589548 (+ LightJNI)
> =====================================================
> Dacapo:lusearch: 2900 msecs
>
> Warmup finished: 4.527292719905836E10
> Math Result = 2.1081849486439312E10 in 12078ms
> StrictMath Result = 2.1081849486439312E10 in 11921ms
>
>
> == Harmony-r589548 (+ intrinsic in native code)
> ===========================================
> Dacapo:lusearch: 2200 msecs
>
> Warmup finished: 4.527292719913238E10
> Math Result = 2.1081849486508232E10 in 4140ms
> StrictMath Result = 2.1081849486439312E10 in 12172ms
>
>
> == Harmony-r589548 (+ intrinsic in native code + LightJNI)
> ===================================
> Dacapo:lusearch: 2100 msecs
>
> Warmup finished: 4.527292719913238E10
> Math Result = 2.1081849486508232E10 in 3904 ms
> StrictMath Result = 2.1081849486439312E10 in 12047ms
>
>
> Thanks,
> Aleksey.
> ESSD, Intel.
>
>
> On Nov 7, 2007 7:34 PM, Yuri Kashnikoff <yuri.kashnikoff@gmail.com> wrote:
> > > Yuri Kashnikoff did this patch and I hope we will commit it before M4.
> > > Yuri, can you share your patch with us?
> > Yes, Mikhail. LightJNI available by
> > https://issues.apache.org/jira/browse/HARMONY-5084.
> > All information about LightJNI provided(just use the link).
>



-- 
Mikhail Fursov

Mime
View raw message