Review request: (turn off transparent huge pages) 8024838: Significant slowdown due to transparent huge pages

[Andrew Hughes]

----- Original Message -----
> On 10/11/13 10:32 PM, Andrew Hughes wrote:
> > ----- Original Message -----
> >> Please review this change to turn off transparent huge pages by default.
> >>
> >> http://cr.openjdk.java.net/~stefank/8024838/webrev.00/
> >>
> >> 8024838: Significant slowdown due to transparent huge pages
> >> Summary: Don't turn on transparent huge pages
> >> (-XX:+UseTransparentHugePages) unless explicitly specified on the
> >> command line. This has the effect that large pages are never turned on
> >> Linux unless the user has explicitly enabled any of the large pages
> >> flags: -XX:+UseLargePages, -XX:+UseTransparentHugePages,
> >> -XX:+UseHugeTLBFS, and -XX:+UseSHM.
> >> Reviewed-by: TBD1, TBD2
> >>
> >> thanks,
> >> StefanK
> >>
> > Can you explain where you see this slowdown?
> 
> One of the issues are described in the bug:
> https://bugs.openjdk.java.net/browse/JDK-8024838
> 
> There's another issues we're Nashorn runs 10% slower with CompressedOops
> turned on, most likely due to THP in the Compressed Class Space. That
> bug is marked Confidential, so I can't give you a link. =(
> 
> Another issue is the extra memory usage for small work loads. That bug
> is, apparently, also marked Confidential.
> 
> See also:
> https://blogs.oracle.com/linux/entry/performance_issues_with_transparent_huge
> 
> >    I'm a little worried that
> > this does not seem to have been questioned in this review.
> 
> Turning transparent huge pages on by default was done prematurely. If we
> do more performance evaluations in the future, this decision could be
> reevaluated.
> 
> If anyone still wants large pages they are free to use either of:
>   -XX:+UseTransparentHugePages
>   -XX:+UseHugeTLBFS
>   -XX:+UseSHM
> 
> thanks,
> StefanK
> >
> > Thanks,
> 
> 

Thanks for the explanation.  I'm just anticipating that we're inevitably
going to get asked about this when it works its way through into a HotSpot
in OpenJDK 7 some time from now.  The original bug just made it sound like
it was one developer's machine, which seemed a bit curious.

Thanks,
-- 
Andrew :)

Free Java Software Engineer
Red Hat, Inc. (http://www.redhat.com)

PGP Key: 248BDC07 (https://keys.indymedia.org/)
Fingerprint = EC5A 1F5E C0AD 1D15 8F1F  8F91 3B96 A578 248B DC07