Imagine there's a function which does a large amount of work on the CPU. We want to optimize it by splitting the work across two cores; so now the function creates a new thread, does half the work in each thread, and then has its original thread wait for the new thread to finish. (In a more modern system, the function might add a task to a global thread pool, but the basic concept is the same.) There is a relationship between the work done by these two threads, but the system doesn't know about it. That makes it much harder to solve systemic problems.