You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@harmony.apache.org by Nataly Naumova <na...@gmail.com> on 2008/03/07 11:42:21 UTC
Re: [drlvm][startup][performance][classloader] mapping bootstrap jars directly into memory (HARMONY-5281)
Tim,
I've measured the startup performance for simple applications, speedup
was ~2-3% on simple graphics aplication with lots of AWT/SWING
classes. But I think the idea is quite pretty, because of we get rid
of a big amount of system calls, such as lseek, read etc. As we're
working just with pointers directly.
On Thu, Mar 6, 2008 at 7:45 PM, Tim Ellison <t....@gmail.com> wrote:
> Pavel Pervov (JIRA) wrote:
> > Please, note, that memory footprint of DRLVM will increase by size of
> > JRE, namely, ~20 megs.
>
> Eek, that's quite significant! What's the speed-up by doing this?
>
> Regards,
> Tim
>
--
Nataly Naumova,
Intel Middleware Products Division
Re: [drlvm][startup][performance][classloader] mapping bootstrap jars directly into memory (HARMONY-5281)
Posted by Pavel Pervov <pm...@gmail.com>.
Egor,
I'm not aware of any regular measurements of Harmony memory footprint.
Pavel.
On 10 Mar 2008 18:05:51 +0300, Egor Pasko <eg...@gmail.com> wrote:
> On the 0x401 day of Apache Harmony Nathan Beyer wrote:
> > On Fri, Mar 7, 2008 at 4:46 AM, Tim Ellison <t....@gmail.com> wrote:
> > > Nataly Naumova wrote:
> > > > I've measured the startup performance for simple applications, speedup
> > > > was ~2-3% on simple graphics aplication with lots of AWT/SWING
> > > > classes. But I think the idea is quite pretty, because of we get rid
> > > > of a big amount of system calls, such as lseek, read etc. As we're
> > > > working just with pointers directly.
> > >
> > > Understood, but to increase the runtime footprint by 20Mb is bad IMHO.
> >
> > What's the total footprint? What was it before? The RI has a ~25M
> > footprint for general uses the last time I checked.
>
> Do we measure DRLVM footprint regularly? Is there something in BTI to
> facilitate this? Say, we can have some reference startup footprints
> like Eclipse, JEdit (more?). Did anybody compare them to RI?
>
> To detect bottlenecks it is also important to know what part of the
> footprint is heap, what is JITted code, what is other stuff. Is
> there any handy function to measure this? (if at performance cost,
> it's OK)
>
> Offtopicly, I regularly worry about footprint impact of Jitrino
> optimizations, where inliner is most notable, you know. Looks as if
> everybody sees the picture, but not me :)
>
> --
> Egor Pasko
>
>
--
Pavel Pervov,
Intel Enterprise Solutions Software Division
Re: [drlvm][startup][performance][classloader] mapping bootstrap jars directly into memory (HARMONY-5281)
Posted by Egor Pasko <eg...@gmail.com>.
On the 0x401 day of Apache Harmony Nathan Beyer wrote:
> On Fri, Mar 7, 2008 at 4:46 AM, Tim Ellison <t....@gmail.com> wrote:
> > Nataly Naumova wrote:
> > > I've measured the startup performance for simple applications, speedup
> > > was ~2-3% on simple graphics aplication with lots of AWT/SWING
> > > classes. But I think the idea is quite pretty, because of we get rid
> > > of a big amount of system calls, such as lseek, read etc. As we're
> > > working just with pointers directly.
> >
> > Understood, but to increase the runtime footprint by 20Mb is bad IMHO.
>
> What's the total footprint? What was it before? The RI has a ~25M
> footprint for general uses the last time I checked.
Do we measure DRLVM footprint regularly? Is there something in BTI to
facilitate this? Say, we can have some reference startup footprints
like Eclipse, JEdit (more?). Did anybody compare them to RI?
To detect bottlenecks it is also important to know what part of the
footprint is heap, what is JITted code, what is other stuff. Is
there any handy function to measure this? (if at performance cost,
it's OK)
Offtopicly, I regularly worry about footprint impact of Jitrino
optimizations, where inliner is most notable, you know. Looks as if
everybody sees the picture, but not me :)
--
Egor Pasko
Re: [drlvm][startup][performance][classloader] mapping bootstrap jars directly into memory (HARMONY-5281)
Posted by Nathan Beyer <nb...@gmail.com>.
On Fri, Mar 7, 2008 at 4:46 AM, Tim Ellison <t....@gmail.com> wrote:
> Nataly Naumova wrote:
> > I've measured the startup performance for simple applications, speedup
> > was ~2-3% on simple graphics aplication with lots of AWT/SWING
> > classes. But I think the idea is quite pretty, because of we get rid
> > of a big amount of system calls, such as lseek, read etc. As we're
> > working just with pointers directly.
>
> Understood, but to increase the runtime footprint by 20Mb is bad IMHO.
What's the total footprint? What was it before? The RI has a ~25M
footprint for general uses the last time I checked.
-Nathan
>
> Tim
>
>
>
>
> > On Thu, Mar 6, 2008 at 7:45 PM, Tim Ellison <t....@gmail.com> wrote:
> >> Pavel Pervov (JIRA) wrote:
> >> > Please, note, that memory footprint of DRLVM will increase by size of
> >> > JRE, namely, ~20 megs.
> >>
> >> Eek, that's quite significant! What's the speed-up by doing this?
> >>
> >> Regards,
> >> Tim
> >>
> >
> >
> >
>
Re: [drlvm][startup][performance][classloader] mapping bootstrap
jars directly into memory (HARMONY-5281)
Posted by Tim Ellison <t....@gmail.com>.
Nataly Naumova wrote:
> I've measured the startup performance for simple applications, speedup
> was ~2-3% on simple graphics aplication with lots of AWT/SWING
> classes. But I think the idea is quite pretty, because of we get rid
> of a big amount of system calls, such as lseek, read etc. As we're
> working just with pointers directly.
Understood, but to increase the runtime footprint by 20Mb is bad IMHO.
Tim
> On Thu, Mar 6, 2008 at 7:45 PM, Tim Ellison <t....@gmail.com> wrote:
>> Pavel Pervov (JIRA) wrote:
>> > Please, note, that memory footprint of DRLVM will increase by size of
>> > JRE, namely, ~20 megs.
>>
>> Eek, that's quite significant! What's the speed-up by doing this?
>>
>> Regards,
>> Tim
>>
>
>
>
Re: [drlvm][startup][performance][classloader] mapping bootstrap jars directly into memory (HARMONY-5281)
Posted by Xiao-Feng Li <xi...@gmail.com>.
On Thu, Mar 13, 2008 at 7:01 PM, Alexei Fedotov
<al...@gmail.com> wrote:
> Thanks, Pavel!
> I believe that your idea of making a questionable option configurable
> is the best way to resolve these concerns.
Yes, agreed.
Thanks,
xiaofeng
>
>
> On Tue, Mar 11, 2008 at 4:13 PM, Pavel Pervov <pm...@gmail.com> wrote:
> > Xiao-Feng,
> >
> > Yes, in the patch I've also added command line option to change that
> > behaviour (actually, patch _adds_ jar mapping, not _changes_ file
> > reads to mapping).
> > I made it on by default in the patch but I can change it to off before
> > committing and keep this functionality as a potential performance
> > optimization on some scenarios.
> >
> > Pavel.
> >
> >
> >
> > On 3/7/08, Xiao-Feng Li <xi...@gmail.com> wrote:
> > > On Fri, Mar 7, 2008 at 6:42 PM, Nataly Naumova
> > > <na...@gmail.com> wrote:
> > > > Tim,
> > > >
> > > > I've measured the startup performance for simple applications, speedup
> > > > was ~2-3% on simple graphics aplication with lots of AWT/SWING
> > > > classes. But I think the idea is quite pretty, because of we get rid
> > > > of a big amount of system calls, such as lseek, read etc. As we're
> > > > working just with pointers directly.
> > >
> > > Can it be a runtime option chosen by the user?
> > >
> > > Thanks,
> > > xiaofeng
> > > >
> > > >
> > > > On Thu, Mar 6, 2008 at 7:45 PM, Tim Ellison <t....@gmail.com> wrote:
> > > > > Pavel Pervov (JIRA) wrote:
> > > > > > Please, note, that memory footprint of DRLVM will increase by size of
> > > > > > JRE, namely, ~20 megs.
> > > > >
> > > > > Eek, that's quite significant! What's the speed-up by doing this?
> > > > >
> > > > > Regards,
> > > > > Tim
> > > > >
> > > >
> > > >
> > > >
> > > > --
> > > > Nataly Naumova,
> > > > Intel Middleware Products Division
> > > >
> > >
> > >
> > >
> > > --
> > > http://xiao-feng.blogspot.com
> > >
> >
> >
> >
> >
> > --
> > Pavel Pervov,
> > Intel Enterprise Solutions Software Division
> >
>
>
>
> --
> With best regards,
> Alexei
>
--
http://xiao-feng.blogspot.com
Re: [drlvm][startup][performance][classloader] mapping bootstrap jars directly into memory (HARMONY-5281)
Posted by Alexei Fedotov <al...@gmail.com>.
Thanks, Pavel!
I believe that your idea of making a questionable option configurable
is the best way to resolve these concerns.
On Tue, Mar 11, 2008 at 4:13 PM, Pavel Pervov <pm...@gmail.com> wrote:
> Xiao-Feng,
>
> Yes, in the patch I've also added command line option to change that
> behaviour (actually, patch _adds_ jar mapping, not _changes_ file
> reads to mapping).
> I made it on by default in the patch but I can change it to off before
> committing and keep this functionality as a potential performance
> optimization on some scenarios.
>
> Pavel.
>
>
>
> On 3/7/08, Xiao-Feng Li <xi...@gmail.com> wrote:
> > On Fri, Mar 7, 2008 at 6:42 PM, Nataly Naumova
> > <na...@gmail.com> wrote:
> > > Tim,
> > >
> > > I've measured the startup performance for simple applications, speedup
> > > was ~2-3% on simple graphics aplication with lots of AWT/SWING
> > > classes. But I think the idea is quite pretty, because of we get rid
> > > of a big amount of system calls, such as lseek, read etc. As we're
> > > working just with pointers directly.
> >
> > Can it be a runtime option chosen by the user?
> >
> > Thanks,
> > xiaofeng
> > >
> > >
> > > On Thu, Mar 6, 2008 at 7:45 PM, Tim Ellison <t....@gmail.com> wrote:
> > > > Pavel Pervov (JIRA) wrote:
> > > > > Please, note, that memory footprint of DRLVM will increase by size of
> > > > > JRE, namely, ~20 megs.
> > > >
> > > > Eek, that's quite significant! What's the speed-up by doing this?
> > > >
> > > > Regards,
> > > > Tim
> > > >
> > >
> > >
> > >
> > > --
> > > Nataly Naumova,
> > > Intel Middleware Products Division
> > >
> >
> >
> >
> > --
> > http://xiao-feng.blogspot.com
> >
>
>
>
>
> --
> Pavel Pervov,
> Intel Enterprise Solutions Software Division
>
--
With best regards,
Alexei
Re: [drlvm][startup][performance][classloader] mapping bootstrap jars directly into memory (HARMONY-5281)
Posted by Pavel Pervov <pm...@gmail.com>.
Xiao-Feng,
Yes, in the patch I've also added command line option to change that
behaviour (actually, patch _adds_ jar mapping, not _changes_ file
reads to mapping).
I made it on by default in the patch but I can change it to off before
committing and keep this functionality as a potential performance
optimization on some scenarios.
Pavel.
On 3/7/08, Xiao-Feng Li <xi...@gmail.com> wrote:
> On Fri, Mar 7, 2008 at 6:42 PM, Nataly Naumova
> <na...@gmail.com> wrote:
> > Tim,
> >
> > I've measured the startup performance for simple applications, speedup
> > was ~2-3% on simple graphics aplication with lots of AWT/SWING
> > classes. But I think the idea is quite pretty, because of we get rid
> > of a big amount of system calls, such as lseek, read etc. As we're
> > working just with pointers directly.
>
> Can it be a runtime option chosen by the user?
>
> Thanks,
> xiaofeng
> >
> >
> > On Thu, Mar 6, 2008 at 7:45 PM, Tim Ellison <t....@gmail.com> wrote:
> > > Pavel Pervov (JIRA) wrote:
> > > > Please, note, that memory footprint of DRLVM will increase by size of
> > > > JRE, namely, ~20 megs.
> > >
> > > Eek, that's quite significant! What's the speed-up by doing this?
> > >
> > > Regards,
> > > Tim
> > >
> >
> >
> >
> > --
> > Nataly Naumova,
> > Intel Middleware Products Division
> >
>
>
>
> --
> http://xiao-feng.blogspot.com
>
--
Pavel Pervov,
Intel Enterprise Solutions Software Division
Re: [drlvm][startup][performance][classloader] mapping bootstrap jars directly into memory (HARMONY-5281)
Posted by Xiao-Feng Li <xi...@gmail.com>.
On Fri, Mar 7, 2008 at 6:42 PM, Nataly Naumova
<na...@gmail.com> wrote:
> Tim,
>
> I've measured the startup performance for simple applications, speedup
> was ~2-3% on simple graphics aplication with lots of AWT/SWING
> classes. But I think the idea is quite pretty, because of we get rid
> of a big amount of system calls, such as lseek, read etc. As we're
> working just with pointers directly.
Can it be a runtime option chosen by the user?
Thanks,
xiaofeng
>
>
> On Thu, Mar 6, 2008 at 7:45 PM, Tim Ellison <t....@gmail.com> wrote:
> > Pavel Pervov (JIRA) wrote:
> > > Please, note, that memory footprint of DRLVM will increase by size of
> > > JRE, namely, ~20 megs.
> >
> > Eek, that's quite significant! What's the speed-up by doing this?
> >
> > Regards,
> > Tim
> >
>
>
>
> --
> Nataly Naumova,
> Intel Middleware Products Division
>
--
http://xiao-feng.blogspot.com