Whenever multiple threads are using shared data, there will be contention. Contention hurts performance and sometimes the overhead caused by contention can make a parallel algorithm work slower than a single-threaded alternative.
Using a lock that causes a wait and a context switch is an obvious performance penalty, but what is not equally obvious is that both locks and atomics disable optimizations in the code generated by the compiler, and they do so at runtime when the CPU executes the code. This is necessary in order to guarantee sequential consistency. But remember, the solution to such problems is never to ignore synchronization and therefore introduce data races. Data races mean undefined behavior, and having a fast ...
