RFR: 8297186: G1 triggers unnecessary full GCs when heap utilization is low

[Thomas Schatzl]

On Fri, 18 Nov 2022 00:15:05 GMT, Man Cao <manc at openjdk.org> wrote:

>> Hi all,
>> 
>>   can I have reviews for this change that better enforces that after gc there is at least one eden region to allocate into?
>> 
>> That avoids the problem described in the PR.
>> 
>> Thanks,
>>   Thomas
>
> src/hotspot/share/gc/g1/g1Policy.cpp line 283:
> 
>> 281:     uint desired_eden_length = MAX3(desired_eden_length_by_pause,
>> 282:                                     desired_eden_length_by_mmu,
>> 283:                                     MinDesiredEdenLength);
> 
> Does this mean `G1Policy::calculate_young_desired_length()` could return 0 instead of a minimum of 1 now? Comment on line 228 may need updating.
> 
> I noticed `young_list_desired_length()` is also used in `G1Policy::update_ihop_prediction()`. Could this change mess up with `_ihop_control->update_allocation_info()`?

No. Line 285 clamps the value to min/max of the young gen sizer; minimum size returned by it is 1 region, see `G1YoungSizer` code. So the original code to ensure a minimum region size of one has already been superfluous in the original implementation.
Since it serves no purpose, I removed it. I added an assertion at the end.

-------------

PR: https://git.openjdk.org/jdk/pull/11209