On Sat, Oct 22, 2011 at 02:59:26PM +0800, Wu Fengguang wrote:
> > In general it looks better we first root case the "decreasing wrote
> > pages by writeback_single_inode() over time" problem before looking
> > into further steps..
> 
> The problem shows up in many of the regression cases. For example,
> the thresh=1G/ext4-100dd case with patches 1-4 applied will not only
> see nr_writeback occasionally dropped low (bdi_dirty_state-8:0.png),
> but also see repeated pattern of nr_wrote by writeback_single_inode()
> slowly decreasing to 0 during some period of dozens of seconds
> (writeback_single_inode.png).

Another example is the thresh=1G/xfs-100dd case with patches 1-4 and
Jan's requeue_io_wait-on-I_SYNC patch. The writeback_single_inode.png
graph shows steadily decreasing wrote pages in some period of 400
seconds. It's accompanied by decrease of nr_writeback, IO size and
impacts write bandwidth a lot.

Thanks,
Fengguang