-
Enhancement
-
Resolution: Won't Fix
-
P4
-
20
Currently preventive gc prediction for regions needed for the GC is fairly inaccurate or too pessimistic.
Reasons include:
* the used waste percentage for estimating survivors from eden is way too pessimistic and too large. While the use of TargetPLABWastePct is reasonable, in practice particularly on larger heaps the waste is far smaller.
That applies to survivors from eden/survivor and old regions.
* survivor regions survivor estimation just takes the amount of memory survived into last gc into account. This does not take into account that the next gc may have a much less survivors. I.e. that value needs to be averaged out somehow.
These problems can cause lots of unnecessary premature (preventive) collections.
Reasons include:
* the used waste percentage for estimating survivors from eden is way too pessimistic and too large. While the use of TargetPLABWastePct is reasonable, in practice particularly on larger heaps the waste is far smaller.
That applies to survivors from eden/survivor and old regions.
* survivor regions survivor estimation just takes the amount of memory survived into last gc into account. This does not take into account that the next gc may have a much less survivors. I.e. that value needs to be averaged out somehow.
These problems can cause lots of unnecessary premature (preventive) collections.
- is blocked by
-
JDK-8291037 Move PLAB resizing mechanism to G1EvacStats
- Resolved
- relates to
-
JDK-8297639 Remove preventive GCs in G1
- Resolved