It might be related to some sort of OS memory leak triggered by some random event, rebooting the server often resolves and stablizes it (based on experience rather than insider knowledge). I know of a server where an instance restart will trigger ever increasing PREEMPTIVE_OS_AUTHENTICATIONOPS wait type times, a server reboot stablizes it. When it reaches between 12-15% the impact becomes noticeable.