Why data sharing is expensive
As we have just seen, concurrent (simultaneous) access of the shared data is a real performance killer. Intuitively, it makes sense: in order to avoid a data race, only one thread can operate on the shared data at any given time. We can accomplish this with a mutex or use an atomic operation if one is available. Either way, when one thread is, say, incrementing the shared variable, all other threads have to wait. Our measurements in the last section confirm it.
However, before taking any action based on observations and experiments, it is critically important to understand precisely what we measured and what can be concluded with certainty.
It is easy to describe what was observed: incrementing a shared variable from multiple threads at the same time does not scale at all and, in fact, is slower than using just one thread. This is true for both atomic shared variables and non-atomic variables guarded by a mutex. We have not tried to measure unguarded...
