Appeared in retriage for 23-b21, seen on linun-aarch64, linux-x64 and osx-aarch64.
Comments
JDK-8338977 aims to restore the out-of-the-box performance.
26-08-2024
At startup, the benchmark calls System.gc twice. In the baseline, this introduces two extra young garbage collections (GCs), increasing the GC-time/mutator-time ratio. To respect the GCTimeRatio flag, the young generation and heap are expanded aggressively. This expansion results in fewer GC pauses (700 vs. 1700), explaining the performance difference. In other words, this is an implicit space-time trade-off: the baseline has larger heap capacity and fewer GCs, while the new code uses smaller heap capacity and more GCs.
One can control the young generation and heap sizes to make the space-time trade-off explicitly if fine-tuning is desirable. Therefore, I think this ticket can be closed as not-an-issue.