tom.rodriguez at oracle.com
Thu May 5 17:04:20 PDT 2011
It's mainly targeted at apps that are heavy users of the concurrency APIs on machines with multiple sockets. On a single socket machine the overhead isn't very great and if you don't have much concurrency then the overhead of the conditional check will swamp any benefit you might get. The description in the bug report actually came from http://blogs.sun.com/dave/entry/false_sharing_induced_by_card. I believe there were specific workloads where there were benefits but I don't think it's any standard benchmark. Doug Lea was the primary requestor of this feature because of scalability problem he had seen.
On May 5, 2011, at 3:54 PM, David Dabbs wrote:
> I just read the RFE details
> for the newly-added hotspot flag, -XX:+UseCondCardMarks.
> Is this relevant only for large MP machines such as the
> "4-socket system where each socket contains a Niagara T2+ UltraSPARC
> processor having 64 logical processors"
> referenced in the writeup? The switch is marked product() and not
> experimental, so I'll give it a
> try on some workloads. Just wondering if there's some benchmark or test case
> that is known to perform
> better (on MP hardware) with the flag enabled.
> Many thanks,
More information about the hotspot-compiler-dev