-
Bug
-
Resolution: Fixed
-
P3
-
11, 14
-
b25
In G1's prediction update for the cost of copying a byte, it currently uses the *used* amount of bytes for calculating this value.
I.e.
copy_cost_per_byte = used number of bytes / time of object copy phase
The used number of bytes is impacted by PLAB waste, region end waste and others, particularly because when we retire survivor regions we "use" them as a whole.
Simple applications show a difference between actually copied bytes and used in the range of 5-10%, which means that the copy cost is inflated by that value.
I.e.
copy_cost_per_byte = used number of bytes / time of object copy phase
The used number of bytes is impacted by PLAB waste, region end waste and others, particularly because when we retire survivor regions we "use" them as a whole.
Simple applications show a difference between actually copied bytes and used in the range of 5-10%, which means that the copy cost is inflated by that value.
- links to
-
Review openjdk/jdk11u-dev/927