Free ratio based heap shrinking in the parallel collector

Jon Masamitsu jon.masamitsu at
Fri May 14 05:44:49 UTC 2010

On 5/13/10 4:34 PM, Hiroshi Yamauchi wrote:
> I assume by 'close' you mean between UseSerialGC and UseParallelGC
> with -UseAdaptiveSizePolicy.
> I need more elaboration. How would you do the testing described above
> (eg how/when to measure them and what to run)?

I would start by running some benchmarks with your your changes and with 
and looking at the number of GC that are done and what the heap looks 
like at
the end of the execution.  When PrintGCDetails is set, the heap is printed
on exit.   GCOld would be  a good benchmark to start with because it 
comes to
steady state quickly and runs a specified amount of work.  I would 
expect that
the number of GC's would be within 10% of each other.   I would also hope
the generations to be within 20% of each other.  If not we should look at
the details to see why the sizes are not closer.   Let's see how that 
goes before
deciding what else we should do.

GChisto is a good tool to compare the GC counts.  There are also some
scripts available (PrintGCStats) to do the same if you prefer.

More information about the hotspot-gc-dev mailing list