Re: [RFC] page-writeback: move indoes from one superblock together

[Wu Fengguang <fengguang.wu@intel.com>]

On Thu, Sep 24, 2009 at 09:29:50PM +0800, Jens Axboe wrote:
> On Thu, Sep 24 2009, Wu Fengguang wrote:
> > On Thu, Sep 24, 2009 at 08:35:19PM +0800, Jens Axboe wrote:
> > > On Thu, Sep 24 2009, Wu Fengguang wrote:
> > > > On Thu, Sep 24, 2009 at 02:54:20PM +0800, Li, Shaohua wrote:
> > > > > __mark_inode_dirty adds inode to wb dirty list in random order. If a disk has
> > > > > several partitions, writeback might keep spindle moving between partitions.
> > > > > To reduce the move, better write big chunk of one partition and then move to
> > > > > another. Inodes from one fs usually are in one partion, so idealy move indoes
> > > > > from one fs together should reduce spindle move. This patch tries to address
> > > > > this. Before per-bdi writeback is added, the behavior is write indoes
> > > > > from one fs first and then another, so the patch restores previous behavior.
> > > > > The loop in the patch is a bit ugly, should we add a dirty list for each
> > > > > superblock in bdi_writeback?
> > > > > 
> > > > > Test in a two partition disk with attached fio script shows about 3% ~ 6%
> > > > > improvement.
> > > > 
> > > > A side note: given the noticeable performance gain, I wonder if it
> > > > deserves to generalize the idea to do whole disk location ordered
> > > > writeback. That should benefit many small file workloads more than
> > > > 10%. Because this patch only sorted 2 partitions and inodes in 5s
> > > > time window, while the below patch will roughly divide the disk into
> > > > 5 areas and sort inodes in a larger 25s time window.
> > > > 
> > > >         http://lkml.org/lkml/2007/8/27/45
> > > > 
> > > > Judging from this old patch, the complexity cost would be about 250
> > > > lines of code (need a rbtree).
> > > 
> > > First of all, nice patch, I'll add it to the current tree. I too was
> > 
> > You mean Shaohua's patch? It should be a good addition for 2.6.32.
> 
> Yes indeed, the parent patch.
> 
> > In long term move_expired_inodes() needs some rework.  Because it
> > could be time consuming to move around all the inodes in a large
> > system, and thus hold inode_lock() for too long time (and this patch
> > scales up the locked time).
> 
> It does. As mentioned in my reply, for 100 inodes or less, it will still
> be faster than eg using an rbtree. But the more "reliable" runtime of an
> rbtree based solution is appealing, though. It's not hugely critical,
> though.

Agreed. Desktops are not big worries; servers rarely do many
partitions per disk.

> > So would need to split the list moves into smaller pieces in future,
> > or to change data structure.
> 
> Yes, those are the two options.
> 
> > > pondering using an rbtree for sb+dirty_time insertion and extraction.

Note that dirty_time may not be unique, so need some workaround.  And
the resulted rbtree implementation may not be more efficient than
several list traversals even for a very large list (as long as
superblocks numbers are low).

The good side is, once sb+dirty_time rbtree is implemented, it should
be trivial to switch the key to sb+inode_number (also may not be unique),
and to do location ordered writeback ;)

Thanks,
Fengguang

> > FYI Michael Rubin did some work on a rbtree implementation, just
> > in case you are interested:
> > 
> >         http://lkml.org/lkml/2008/1/15/25
> 
> Thanks, I'll take a look.
> 
> -- 
> Jens Axboe