Hi Staffan,
I think this looks good.
Two minor comments about the @run commands.
Both of these will use the SerialOld collector and since the test really
only test the old generation I think one is enough:
* @run main/othervm -XX:+UseSerialGC -Xmn8m ResetPeakMemoryUsage
* @run main/othervm -XX:+UseParNewGC -Xmn8m ResetPeakMemoryUsage
I would suggest to keep the SerialGC version.
Also, I think the G1 @run command should work with the same settings as
the other GCs. To be on the safe side it might be worth setting the
region size to make sure it matches the young gen size. So:
* @run main/othervm -XX:+UseG1GC -Xmn8m -XX:G1HeapRegionSize=1m
ResetPeakMemoryUsage
Thanks,
Bengt
On 5/6/13 5:17 PM, Staffan Larsen wrote:
This test had a number of problems reported against it:
http://bugs.sun.com/view_bug.do?bug_id=7148492
http://bugs.sun.com/view_bug.do?bug_id=6980985
http://bugs.sun.com/view_bug.do?bug_id=7181907
This overhaul of the test tries to address all of these and improve the test
robustness. The changes I have done are:
- Make sure the allocated large object is not easy to optimize away by storing
it in a public field.
- Removed the dynamic calculation of the large object's size since this
sometimes caused problems.
- Added GC configuration parameters to make sure the large object is allocated
in the old gen.
- Added GC configuration parameters to run the test with all GC algorithms.
- Changed the printMemoryUsage() method to take MemoryUsages as parameters so
it prints the same values as were used for comparisons.
- Changed some incorrect '>' to '<' in printouts.
- Removed the assumption that peak usage after the GC should be the same as before the
GC. This is not true since "stuff" may have been moved from younger generations
into the old gen thereby increasing the heap.
webrev: http://cr.openjdk.java.net/~sla/7148492/webrev.00/
I've been running the test through JPRT a couple of times to try to ensure that
it always passes.
Thanks,
/Staffan
