This one is silly .. but an interesting lesson in cpu design changes.
There is no concurrency whatsoever here. The only thing
that may make it failing is the duration of the sleep itself,
which seems to be very small. On my computer for example I apparently cannot measure time differences of time less than 15ms.
In fact, if I change the sleep time to be less than 15ms,
the test starts to randomly fail, and consistenly fails if I use
less than 5ms. Mine is a windows xp box.