This is one of the reasons using things like boost's boost::shared_ptr (which uses atomics for ref counting) can have a huge time overhead in multithreaded code

This is one of the reasons using things like boost's boost::shared_ptr (which uses atomics for ref counting) can have a huge time overhead in multithreaded code. Often people think atomics are much faster than mutexes for simply checking or incrementing single count/bool values or pointer swapping (and they can be), but they can also be just as slow, depending on the use-case and situation. In certain situations, they can even be slower! as using them in spinlocks doesn't allow the thread to yield, wasting CPU time doing nothing.