On Thu, May 10 2007, Alan Cox wrote:
> On Thu, 10 May 2007 16:17:57 +0200 (MEST)
> Jan Engelhardt <jengelh@xxxxxxxxxxxxxxx> wrote:
>
> >
> > On May 9 2007 08:49, Jens Axboe wrote:
> > >On Tue, May 08 2007, Andrew Morton wrote:
> > >> > +#define bio_list_for_each(bio, bl) \
> > >> > + for (bio = (bl)->head; bio && ({ prefetch(bio->bi_next); 1; }); \
> > >> > + bio = bio->bi_next)
> > >> > +
> > >
> > >Besides, manual prefetching is very rarely a win. I dabbled with some
> > >benchmarks a few weeks back (with the doubly linked lists), and in most
> > >cases it was actually a loss. So I'd vote for just removing the
> > >prefetch() above.
> >
> > So is the prefetching in the basic ADTs (e.g. linux/list.h) a loss too?
>
> Depends on the box it seems. On the newest systems the processor
> prefetching seems to be very much smarter. On a "classic" AMD Athlon the
> prefetching made the scheduler about 1.5% faster...
It very much depends on the box, indeed. The ones I tested on were
_slower_ with the prefetching, perhaps the dumber CPU's will benefit. In
the long run, I don't think the manual prefetching is a good idea.
--
Jens Axboe
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
