Re: [PATCH -next v3 0/9] rbtree: Cache leftmost node internally

[Davidlohr Bueso]

On Wed, 19 Jul 2017, Andrew Morton wrote:

> On Thu, 29 Jun 2017 10:15:44 -0700 Davidlohr Bueso <dave@xxxxxxxxxxxx> wrote:
>
> > Changes from v2 (https://lkml.org/lkml/2017/6/8/857):
> > - Fixed 0day reported crash for drm_mm selftest program. We were
> > not correctly using the cached version of rbtree with the allocated
> > nodes.
> > - Added cfq patch to use internal rbtree caching.
> > - Added Christian's and Jan's reviews.
> >
> > Changes from v1 (https://marc.info/?l=linux-kernel&m=149611025616685):
> > - No longer rfc.
> > - Removed bogus semimcolon in rb_first_cached()
> > - Updated missing interval tree user drivers/infiniband/hw/hfi1/
> > - Removed redundant @cached arg in when erasing a node.
> > - Added more patches that make use of rb_first_cached(), which I
> >   thought might be worth it: procfs and epoll.
> > - Cc more people for patch 5, which touches drivers such as infiniband
> > and gpu. The rest of the changes are pretty covered with the current
> > Cc'ed maintainers and mm folks.
> >
> > Hi,
> >
> > Here's a proposal for extending rbtrees to internally cache the leftmost
> > node such that we can have fast overlap check optimization for all interval
> > tree users[1]. The benefits of this series are that:
> >
> > (i)   Unify users that do internal leftmost node caching.
>
> That's nice.  Except the series adds more lines than it removes.
>
> > (ii)  Optimize all interval tree users.
>
> Was any attempt made to quantify the benefit?

Yes, but ultimately it will depend a lot on the workload and size of the tree.
For bare numbers, on a Xeon E5-2450 @ 2.10GHz the cost of a rb_first() was
~60 cycles for 100 nodes, and ~75 cycles with 1000 nodes. fwiw.

Thanks,
Davidlohr