On Tue, 2024-08-06 at 21:22 +0200, Mateusz Guzik wrote:
> On Tue, Aug 6, 2024 at 9:11 PM Andi Kleen <ak@xxxxxxxxxxxxxxx> wrote:
> >
> > Mateusz Guzik <mjguzik@xxxxxxxxx> writes:
> > >
> > > I would bench with that myself, but I temporarily don't have
> > > handy
> > > access to bigger hw. Even so, the below is completely optional
> > > and
> > > perhaps more of a suggestion for the future :)
> > >
> > > I hacked up the test case based on tests/open1.c.
> >
> > Don't you need two test cases? One where the file exists and one
> > where it doesn't. Because the "doesn't exist" will likely be slower
> > than before because it will do the lookups twice,
> > and it will likely even slow single threaded.
> >
> > I assume the penalty will also depend on the number of entries
> > in the path.
> >
> > That all seem to be an important considerations in judging the
> > benefits
> > of the patch.
> >
>
> This is why I suggested separately running "unlink1" which is
> guaranteed to create a file every time -- all iterations will fail
> the
> proposed fast path.
>
> Unless you meant a mixed variant where only some of the threads
> create
> files. Perhaps worthwhile to add, not hard to do (one can switch the
> mode based on passed worker number).
>
Well...
# ./unlink1_processes -t 70 -s 100
average:
v6.10: 114455
v6.10 + patch: 149513
I suspect what's happening here is that this patch relieves contention
for the inode_lock and that allows the unlinks to proceed faster.
Running it with a single process though:
average:
v6.10: 200106
v6.10 + patch: 199188
So, ~.4% degradation there? That doesn't seem too bad given the gain in
the threaded test.
--
Jeff Layton <jlayton@xxxxxxxxxx>
