I instrumented the test regime to flag test cases where the GC time was more than 20% of the ?user time. I ran the same test on Linux 5.7.1 and the Mac Catalina machine.

As per screenshot (Linux on left, Mac on right; user times in parentheses, gc times in seconds in red), the 5.8 GC is definitely worse. (I appreciate that occasionally the GC will kick in with "bad" times during a long series of very quick tests; I expect this explains the "bad" time in the Linux series.)

I'm using Timer.getCPUTimer to get these measurements, getting nongc.usr and gc.usr.

Michael