Request for Review (M) - 6858051: Create GC worker threads dynamically

[Jon Masamitsu]

Changes:

In order to reduce the dependency of WorkGang on AdaptiveSizePolicy,
removed AdaptiveSizePolicy::initial_number_of_workers() and duplicated 
the code
in GCTaskManager::initialize() and AbstractWorkGang::initialize_workers().

Moved AdaptiveSizePolicy::add_workers() to WorkerManager::add_workers() in
file shared/workerManager.hpp

Removed guard DisableStartThread from add_workers() and always start the
workers.

Removed the bool return value from WorkerManager::add_workers (and let 
that rippled down
to the callers).

Changed WorkerManager::add_workers() to return the number of created workers
(removed inout on parameters) and let the callers of 
WorkerManager::add_workers()
adjust their _active_workers to be consistent with the number of created 
workers.

Changed number of workers created at initialization from 2 to 1.

Please let me know if I forgot anything.

Delta: http://cr.openjdk.java.net/~jmasa/6858051/webrev_delta_01_02/
Full:http://cr.openjdk.java.net/~jmasa/6858051/webrev.02/

Thanks.

Jon



On 04/26/2016 10:06 PM, Jon Masamitsu wrote:
> 6858051: Create GC worker threads dynamically
> https://bugs.openjdk.java.net/browse/JDK-6858051
>
> This change creates a minimal number of GC workers at JVM initialization
> and then adds additional workers as they are needed.  This has been 
> made part of the
> UseDynamicNumberOfGCThreads feature.  When a parallel task is about to
> be executed, the number of workers needed to execute the task is 
> calculated (call
> that number N).  If that number of workers already exist, then no 
> additional workers
> are created.  If fewer than N exists, then additional workers are 
> created.  Workers
> are not destroyed if not needed for the task.
>
> The UseParallelGC way of executing parallel tasks is different from 
> the other
> collectors so it has a somewhat separate implementation.
>
> http://cr.openjdk.java.net/~jmasa/6858051/webrev.00/
>
> Testing: gc_test_suite with and without UseDynamicNumberOfGCThreads.
> Performance testing with the prototype did not show any regression with
> UseDynamicNumberOfGCThreads turned on.     RBT testing for hotspot_gc
> is in progress.
>
> Thanks.
>
> Jon
>