On Wed, 7 Oct 2009, Linus Torvalds wrote:
>
> Every extra line you fetch pushes out another line. No amount of
> prefetching will hide that.
Side note: what it _will_ do is hide the cost, and spread it out to other
places where you don't see it any more, or you see it but can't do
anything about it.
And that's not a good thing. You end up with "muddy" profiles that don't
have nice clear problem spots any more, so you may think that your profile
looks better ("Look ma, no nasty hotspots!"), but actual performance
hasn't improved one whit, and may well have decreased.
This, btw, is exactly the kind of thing we saw with some of the
non-temporal work, when we used nontemporal stores to copy pages on COW
faults, or when doing pre-zeroing of pages. You get rid of some of the
hot-spots in the kernel, and you then replace them with user space taking
the cache misses in random spots instead. The kernel profile looks better,
and system time may go down, but actual performace never went down - you
just moved your cache miss cost from one place to another.
There's no question that prefetching cannot help, but it helps only if
it's about fetching data that you would need anyway early. In contrast, if
the option is "don't touch the other cacheline at all", prefetching is
_always_ a loss. No ifs, buts and maybes about it.
Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
