But Problem #1 is that you are measuring this the wrong way. You have too many dependencies on library calls that are irrelevant to what you are trying to measure. Take out ALL the semaphore code, re-test, and see what times you get. Or, run a 10,000 x loop around the semaphore. Then scale down to the nearest second so that microseconds mistimings aren't coming into play.

Problem #2 is that semaphores require bus-access and are relatively independent of Processor speed.