* Linus Torvalds <torva...@linux-foundation.org> wrote:
> On Fri, Feb 21, 2014 at 12:53 PM, Davidlohr Bueso <davidl...@hp.com> wrote:
> >
> > I think you are right. I just reran some of the tests and things are
> > pretty much the same, so we could get rid of it.
>
> Ok, I'd prefer the simpler model of just a single per-thread hashed
> lookup, and then we could perhaps try something more complex if there
> are particular loads that really matter. I suspect there is more
> upside to playing with the hashing of the per-thread cache (making it
> three bits, whatever) than with some global thing.
>
> >> Also, the hash you use for the vmacache index is *particularly* odd.
> >>
> >> int idx = (addr >> 10) & 3;
> >>
> >> you're using the top two bits of the address *within* the page.
> >> There's a lot of places that round addresses down to pages, and in
> >> general it just looks really odd to use an offset within a page as an
> >> index, since in some patterns (linear accesses, whatever), the page
> >> faults will always be to the beginning of the page, so index 0 ends up
> >> being special.
> >
> > Ah, this comes from tediously looking at access patterns. I actually
> > printed pages of them. I agree that it is weird, and I'm by no means
> > against changing it. However, the results are just too good, specially
> > for ebizzy, so I decided to keep it, at least for now. I am open to
> > alternatives.
>
> Hmm. Numbers talk, bullshit walks. So if you have the numbers that say
> this is actually a good model..
>
> I guess that for any particular page, only the first access address
> matters. And if it really is a "somewhat linear", and the first access
> tends to hit in the first part of the page, and the cache index tends
> to cluster towards idx=0. And for linear accesses, I guess *any*
> clustering is actually a good thing, since spreading things out just
> defeats the fact that linear accesses also tend to hit in the same
> vma.
>
> And if you have truly fairly random accesses, then presumably their
> offsets within the page are fairly random too, and so hashing by
> offset within page might work well to spread out the vma cache
> lookups.
>
> So I guess I can rationalize it. [...]
Davidlohr: it would be nice to stick a comment about the (post facto)
rationale into the changelog or the code (or both).
Thanks,
Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
