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: [drlvm][jit][win64] why jit perfomance so low?
Date Wed, 23 May 2007 09:57:46 GMT
Please check and commit the patch.
https://issues.apache.org/jira/browse/HARMONY-3651
Total compilation time in -Xem:opt mode for HWA with the patch applied was
reduced from 100 seconds to 35 seconds.But it's still 2 times more I have on
my 32-bit PC (17 seconds).

ps.
Use -XX:jit.arg.time=on to check Jitrino.OPT compilation timers if needed.




On 5/23/07, Xiao-Feng Li <xiaofeng.li@gmail.com> wrote:
>
> On 22 May 2007 12:53:37 -0400, Egor Pasko <egor.pasko@gmail.com> wrote:
> > On the 0x2DE day of Apache Harmony Mikhail Fursov wrote:
> > > All,
> > > I think this is not a good idea to make JIT always work in release
> mode.
> > > I agree that STL performance we have in JIT in debug build is also
> almost
> > > unusable on Windows64bit.
> > > I propose learning MSVC2005 STL options and check is we can disable
> some
> > > checks and speedup the JIT.
> > > Here is an example of such an option:
> > >
> http://groups.google.co.nz/group/microsoft.public.vc.stl/browse_thread/thread/24426056e5b256ea/0efe8dcc5f3ed97b?lnk=gst&q=debug+performance&rnum=2&hl=en#0efe8dcc5f3ed97b
> >
> > +1 to disable thorough STL checking in w/64/debug if possible
>
> +1
>
> > > Right now I can't test it because I run tests on my Windows64bit PC,
> but I
> > > will check this one and other options tomorrow.
> > >
> > >
> > > On 5/22/07, Xiao-Feng Li <xiaofeng.li@gmail.com> wrote:
> > > >
> > > > Vladimir, if the testing is really slow, I think it's ok to try the
> > > > release build on Win64. But I personally prefer to use debug build
> for
> > > > other cases if their times are ok, because debug build can help to
> > > > identify the reasons of any potential breaks.
> > > >
> > > > (Or we can use release build for only the JIT module, while keeping
> > > > other modules in debug build.)
> > > >
> > > > Thanks,
> > > > xiaofeng
> > > >
> > > > On 5/22/07, Vladimir Ivanov <ivavladimir@gmail.com> wrote:
> > > > > Does it means that we can use only 'release' build for testing
> while
> > > > > 'debug' build required more than 10 hours to run DRLVM tests?
> > > > >
> > > > >  thanks, Vladimir
> > > > >
> > > > > On 5/18/07, Mikhail Fursov <mike.fursov@gmail.com> wrote:
> > > > > > I think the numbers you mention are from debug build.
> > > > > >
> > > > > > The reason of significant compilation time increase when using
> Windows
> > > > > > x86_64 is MSVS2005 STL implementation. It does a lot of checks
> in
> > > > debug mode
> > > > > > that MSVS2003 version did not.
> > > > > >
> > > > > >
> > > > > > On 5/18/07, Vladimir Ivanov <ivavladimir@gmail.com> wrote:
> > > > > > >
> > > > > > > Hello everyone,
> > > > > > > I have a small question about JIT performance on Windows
> x86_64: why
> > > > > > > the DRLVM kernel test java.lang.ClassGenericsTest, for
> example,
> > > > > > > required in my env:
> > > > > > > ~20 seconds to finish in the default mode;
> > > > > > > ~400 seconds to finish in the 'opt' mode;
> > > > > > > ~2000 seconds to finish in the 'server' mode.
> > > > > > >
> > > > > > > I know that this platform was only enabled but may be someone
> from
> > > > JIT
> > > > > > > expert spends some time to improve performance a little
bit?
> > > > > > >
> > > > > > > Thanks, Vladimir
> > > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > > --
> > > > > > Mikhail Fursov
> > > > > >
> > > > >
> > > >
> > > >
> > > > --
> > > > http://xiao-feng.blogspot.com
> > > >
> > >
> > >
> > >
> > > --
> > > Mikhail Fursov
> >
> > --
> > Egor Pasko
> >
> >
>
>
> --
> http://xiao-feng.blogspot.com
>



-- 
Mikhail Fursov

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message