r205231

[Matthew Ahrens]

On Wed, Sep 18, 2013 at 3:52 PM, Pawel Jakub Dawidek <pjd at freebsd.org>wrote:

> On Wed, Sep 18, 2013 at 02:08:50PM +0300, Andriy Gapon wrote:
> >
> > Guys,
> >
> > I would like to ask your opinion of r205231, specifically the part that
> splits
> > the state lists.
> > The change is quite large.  I can admit that I do not fully understand
> it.  And
> > it introduces many differences to the upstream code which makes merges
> quite a
> > bit harder.
>
> AFAIR Kip observed this contention for workload with multiple sequential
> streams, like video streaming or something.
>
> If it really reduces contention and we can prove it then I think it
> would be nice to upstream the change. My guess would be that it does
> help, the hard part might be to find a case where it is clearly visible.
>
> I'm also not fun of hardcoding numer of locks to 16. It should scale
> with the number of cores at least.
>

Agreed, I'd be happy to review this for inclusion in illumos.  I responded
to Andriy's post on zfs at lists.illumos.org:

I have seen contention on arcs_mtx before.  To validate that the changes
help, it would be good to see lockstat (or equivalent) data before + after,
and a description of the workload.

--matt