Request for review (S): 7005808 G1: re-enable ReduceInitialCardMarks for G1

Bengt Rutisson bengt.rutisson at
Tue Oct 4 12:47:08 UTC 2011


On 2011-09-29 10:05, Ramki Ramakrishna wrote:
> Looks good to me too. Apropos of a comment in one of the CR's in the 
> trail,
> would it also be worthwhile to run a test with OpenDS?

I finally got OpenDS up and running on a Linux x64 machine. I ran a two 
hour run, which corresponded to 116 iterations. Then I repeated that 
procedure, so in total I have run 232 iterations of OpenDS without 
seeing any issues. I have not looked closely at the results (don't 
really know what to look for), but the harness has not complained about 
any crashes or hangs and I have not found any hs_err files.

Is this enough OpenDS testing?

If yes, then I will push this change as soon as hotspot-gc opens for 
hs23 stuff for G1.


> -- ramki
> On 9/28/2011 6:42 AM, Jon Masamitsu wrote:
>> Looks good.
>> On 09/28/11 05:00, Bengt Rutisson wrote:
>>> Hi all,
>>> Could I have a couple of reviews for this small change?
>>> The code change is very small. I am just enabling the 
>>> ReduceInitialCardMarks by default for G1. The ReduceInitialCardMarks 
>>> was disabled since it caused some stability issues. The theory is 
>>> that those issues arose from the fact that there were bugs that 
>>> would allocate new objects outside the young gen. These bugs have 
>>> now been fixed and we should be able to use ReduceInitialCardMarks 
>>> again.
>>> Most of the work on this fix has been going in to testing it. I have 
>>> run the following tests:
>>> 2 x RefWorkload
>>> 1 x Full GC test suite
>>> 7 x GC test suite Dacapo2006
>>> 16 x GC test suite GCBasher
>>> 4 x JPRT - All platforms
>>> 7 x JPRT - Windows only
>>> I choose the test distribution based on Tony's feedback that the 
>>> issues related to ReduceInitialCardMarks used to be intermittent in 
>>> JPRT, and mostly occurred on Windows, and that GCBasher and 
>>> Dacapo2006 used to have issues with it as well.
>>> I have not seen any failures in the tests above. I also did not see 
>>> any regressions in RefWorkLoad. Unfortunately I did not see any 
>>> improvement either, but I think this change is still worth doing to 
>>> make G1 behave the same way as the other collectors wrt 
>>> ReduceInitialCardMarks.
>>> CR:
>>> 7005808 G1: re-enable ReduceInitialCardMarks for G1
>>> This is for hs23, so I will not push it until next week.
>>> Bengt

More information about the hotspot-gc-dev mailing list