In JDK-8159422 we tried to improve the lock contention on the global mark stack by increasing the size of the chunks operated on significantly and tightening the scope of the locks themselves.

Due to time and code infrastructure (reusable, hazard pointers, tagged pointers, ...) constraints we did not use these lock-free data structures yet.

Testing on large machines with a significant amount of arrays of j.l.O, we found that the steps taken are not sufficient: marking cycle length is significantly longer (50s instead of 20s), and even sometimes does not complete in time, going into full gc after 300s.

Remedy that.