It is application dependent. Short sleeps are usually intended as waits,
so the application assumes it will wake up no earlier than after the
requested time. This is why the OS specific sleep functions always round
up, never down. And this is why I think the epicsTimer should behave the
same way.

In case an application wants a statistically optimized behavior, it can
easily get the quantum and do the subtraction itself.

I know this is symmetric (the application could add half a quantum when
it wants the round-up behavior), but I would strongly suggest that the
semantics of the epicsTimer follow the semantics of all reasonable OS's
sleep implementations, and always round up to the next quantum.