Re: [PATCH] sort: use posix_fadvise to announce access patterns on files opened for reading

[Pádraig Brady]

On 03/03/10 11:43, Jim Meyering wrote:
> Pádraig Brady wrote:
> > On 02/03/10 22:51, Pádraig Brady wrote:
> > > I'll apply the patch with SEQUENTIAL enabled,
> > > and summarise in a comment the results we've uncovered.
> >
> > I'll push the attached later today, unless there are objections.
>
> Nice work.
> Thanks to both of you!
>
> You must have a few small stand-alone tests that run timings.
> If you post one or two, I'll run them on a variety of
> hardware and newer kernels.

shuf -i 1-100000000 -n 10000000 > file

prep_cache() {
  echo 1 > /proc/sys/vm/drop_caches
  time sort < /dev/null
}

#This should show benefit on faster flash devices
prep_cache; time LANG=C sort -T/dev/shm      < file > /dev/null

#This should show some benefit on all devices
prep_cache; time LANG=C sort -T/dev/shm -S1M < file > /dev/null

cheers,
Pádraig.

p.s. It would be nice to nuke a particular file in the cache
for testing at least. I wonder would a flag on dd be useful.
I previously noticed a /proc/filecache interface for doing this
but it never materialized.