In practice, how would one go about measuring the percentage of time spent synchronizing? I imagine it's hard to start a timer and begin synchronizing at the same time. And secondly, that coordinating timing information would itself add synchronization costs.

Are there good "rules of thumb" that parallel programmers use to determine how to partition tasks?

@sareyan, I think good profilers can often take care of this. They'd simply sample the program's execution at some [possibly regular] intervals, and they can keep track of the estimated percentage of time spent on synchronization.

It can also be possible to run a dummy version of the program where synchronization returns immediately, but this may need application-specific logic, since in general it won't be well defined.