Anything in hs25-b47 or b47 that would cause a significant performance regression?
Alan.Bateman at oracle.com
Sun Sep 15 08:05:26 PDT 2013
On 12/09/2013 12:10, Alan Bateman wrote:
> This is a long shot but does anyone know of any performance
> regressions that might have crept in recently, specifically in
> jdk8-b105 or jdk8-b106 (hs25-b47 or hs25-b48).
> jdk8/tl was sync'ed up to jdk8-b106 (it had been at b104) last Friday
> and since then several of us have noticed significant slowdowns
> (20-30%) in the time is takes to run the jdk tests. The jdk tests run
> in agent VM mode that should mostly eliminate startup for most of the
> tests. As jdk8/tl is always ahead of jdk8/jdk8 with library changes
> then it means that the only interesting changes that the sync up
> brought in is the HotSpot changes.
> I don't think anyone of has had the time to attempt to bisect this yet
> (or know if this is all or only some OS/arch) and this mail is just to
> see whether it rings any bells with anyone.
Just to follow up on this, it seems to be due to large pages, most
likely the changes in JDK-8007074. Running with -XX:-UseLargePages
restores the performance, at least on Linux x64 (Ubuntu 12.04LTS).
More information about the hotspot-dev