harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Evgueni Brevnov" <evgueni.brev...@gmail.com>
Subject Re: [drlvm][performance] VM properties implementation is inefficient
Date Fri, 11 May 2007 12:59:21 GMT
Ilya,

You results is really interesting and surprising. Taking into account
that we spent more than 90% of the time in JITed code (at least should
do in server mode) 2.41% boost is huge. I wounder why property module
is used so intensively. Did you try to find a hotspot?

Another point is that VM string pool is heavily used in runtime during
constant pool resolution, class loading and so on.. From that point it
seems quit undesirable to pollute the pool with unrelated stuff like
properties since it can slow down application startup time.

Thanks
Evgueni

On 5/11/07, Ilya Berezhniuk <ilya.berezhniuk@gmail.com> wrote:
> Thanks a lot, Mikhail!
>
> Now the response from GC gurus is needed...
>
>
> 2007/5/11, Mikhail Fursov <mike.fursov@gmail.com>:
> >
> > On 5/10/07, Ilya Berezhniuk <ilya.berezhniuk@gmail.com> wrote:
> > >
> > >
> > > With this modifications, I've reached performance gain (measured with
> > > SPECjbb2005) 2.16% for default mode and 2.41% for server mode.
> > > I've submitted patch to H-3836 [1].
> >
> >
> > It's very good results for such a simple patch!
> >
> > This patch contains a lot of modifications in 'vmcore' module, and several
> > > small modifications in components like EM, GC, Jitrino, but I'm not
> > > familiar
> > > with these components.
> > > Could VM gurus and component owners check this patch for completeness
> > and
> > > consistency?
> >
> >
> > I checked JIT and EM parts. Patch is OK.  Neither JIT nor EM do no modify
> > property strings.
> >
> >
> >
> > --
> > Mikhail Fursov
> >
>
>
>
> --
>
> Ilya
>

Mime
View raw message