Re:Pessimism about parallelism
by mattnewport on Wednesday December 26, 2018 @10:39AM
(#18762631)
> Most desktop software does not use more than 1 core because most desktop software does not need anything near 1 full core. How much processing can it take to show you a button?
It seems like this ought to be true but in practice it's not, though the problem appears to be more due to developers not paying enough attention to performance than to fundamental limitations of the hardware. I can think of very few applications that I use on a day to day basis that do not have multiple areas where I wish they performed better.
That does not change the fact that some software does need more, and people are wise to dimension their computers taking the largest usual load into account, not the average one.
That said, the point about memory locality is good. It is just not as important as the author makes - data tends to have locality, even physics pushes it there. But the point about non-parallelizable algorithms isn't good; who cares about the property of your algorithm? If the problems allow for parallelism (with local data), dump the algorithm and get one that scales.
Also interestingly, the authors does bother to link the limits to the von Neumann architecture, but didn't think about using the extra cores to JIT-compile the binaries into some other architecture...