Round-robin part thread activation doesn't work due to an unfair semaphore

runnablePartThreadLimit is guaranteed by a semaphore with the same capacity. Part threads release and re-acquire the semaphore after each move. Due to the semaphore being initialized with non-fair setting, it is possible that active threads re-acquire the semaphore even if there are other threads already waiting for the permit. This leads to temporary part thread starvation.