Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
G1 has a smoother memory profile than parallelGC, which is desirable for workers that users run locally like javac. Unfortunately, completely getting rid of the parallelGC option regresses Java build performance. This is because Turbine does not run as a worker—see https://github.com/bazelbuild/bazel/issues/8006—and G1 seems to do worse than parallelGC for short-lived programs. So, I simply moved the parallelGC option to the Turbine-specific jvm opts. I ran bazel-bench to test the performance this change building //src:bazel_nojdk: ``` metric mean median stddev pval wall: 475.842s ( -0.05%) 474.997s ( -0.09%) 1.203s 0.00000 cpu: 73.063s ( -2.28%) 72.980s ( -2.21%) 1.499s 0.40000 system: 19.427s ( +0.66%) 19.450s ( +1.09%) 0.111s 0.40000 memory: 89.667MB ( -0.74%) 90.000MB ( +0.00%) 0.471MB 0.00000 ``` Closes #12598. PiperOrigin-RevId: 3456162
- Loading branch information