Speed difference between EA versions and official distributions.

[Mikael Gerdin]

On Wednesday 25 June 2014 15.16.09 Ahmet A. Akin wrote:
> Hello David.
> 
> On Wed, Jun 25, 2014 at 2:40 PM, David Holmes <david.holmes at oracle.com>
> 
> wrote:
> > Hi Ahmet,
> > 
> > On 25/06/2014 9:15 PM, Ahmet A. Akin wrote:
> >> Hello,
> >> 
> >> I hope this is the correct place to ask.
> > 
> > No, the discuss list is for "General discussion about the OpenJDK
> > Community":
> > 
> > http://mail.openjdk.java.net/mailman/listinfo
> > 
> > I've cc'd the hotspot-dev list as a better starting point, and the discuss
> > list should be dropped from replies.
> > 
> >  Recently I tested our Java speech
> >  
> >> recognition engine speed with several Java versions.
> >> 
> >> Two JDK 8 EA builds ( 1.8.0-ea-b109 and 1.8.0_20-ea-b19 )
> >> One Stable Oracle JDK 8 release: ( 1.8.0_05-b13 )
> >> One Stable Open JDK 7 release: ( IcedTea 7u55-2.4.7   )
> >> 
> >> And I observed that both Java-8 EA versions runs %6-8 faster than the
> >> stable Oracle JDK 8 and Open JDK 7 releases. What can be the reason? Do
> >> EA
> >> builds and Oracle official releases use different compilers? Or is there
> >> something else I should be checking? What do you suggest?
> > 
> > There can be a lot of factors involved but EA releases are not
> > deliberately different other than they are, by their nature, not as
> > complete as the GA release will be. The compiler might be different if a
> > compiler update is planned for a release and you happen to have taken an
> > EA
> > build from before the change.
> > 
> > Sometimes problems are discovered with code that requires a fix that
> > lowers performance prior to GA. Developers then look for ways to get the
> > performance back while maintaining correctness. So, as a purely
> > hypothetical example,  8-b109 may have had a fast but incorrect piece of
> > code, 8-b132 (the GA build) may have had a correct but slower version of
> > the code, and then for 8u20 a new fix was found that improved the
> > performance.
> > 
> > An EA build may have a partially implemented feature, and once the feature
> > is complete, the performance has changed.
> > 
> > To know exactly what the differences are would require a lot of detailed
> > benchmarking of each build to create a performance profile for the release
> > and then try to see what changes may have contributed to the performance
> > deltas.
> 
> First, sorry for sending this to the wrong list and thanks for the logical
> explanation. Unfortunately code and algorithms are quite large and it will
> be hard to demonstrate the slow down with a sample code. There are cache
> sensitive numerical calculations, millions of hash, graph search
> operations. But anyway I will give some time and try to write a small test
> code for this.

A quick experiment could be to run 1.8.0_05-b13 with 
-XX:+EliminateAutoBox
That optimization was disabled because of a bug late in 8 and the bug was 
fixed in 1.8.0_20 and the flag was re-enabled.

/Mikael

> 
> Regards,
> 
> Ahmet
> 
> > Cheers,
> > David
> > 
> >  Environment: Ubuntu 12.04 64-bit, Intel i7-3770 3.4Ghz x 8
> >  
> >> Best,
> >> 
> >> Ahmet