It is currently statically driven by:

  experimental(uint, ShenandoahParallelSafepointThreads, 4, \
          "Number of parallel threads used for safepoint prolog/epilog") \

This is troublesome either way:
 *) For small machines or containers where ergonomics figures out ParallelGCThreads=1, this trashes safepoint cleanup with excessive threads;
 *) For large machines where ergonomics figures out ParallelGCThreads >> 4, this holds back parallelism needed at cleanup.

Should be just driven by ParallelGCThreads. ZGC drives this with ParallelGCThreads. G1 implicitly does the same.