Re: [PATCH] btrfs: lower metadata writeback threshold on low dirty threshold

[Chris Mason <chris.mason@oracle.com>]

On Thu, May 03, 2012 at 11:25:28AM +0200, Jan Kara wrote:
> On Thu 03-05-12 11:43:11, Wu Fengguang wrote:
> > This helps write performance when setting the dirty threshold to tiny numbers.
> > 
> >      3.4.0-rc2         3.4.0-rc2-btrfs4+
> >   ------------  ------------------------
> >          96.92        -0.4%        96.54  bay/thresh=1000M/btrfs-100dd-1-3.4.0-rc2
> >          98.47        +0.0%        98.50  bay/thresh=1000M/btrfs-10dd-1-3.4.0-rc2
> >          99.38        -0.3%        99.06  bay/thresh=1000M/btrfs-1dd-1-3.4.0-rc2
> >          98.04        -0.0%        98.02  bay/thresh=100M/btrfs-100dd-1-3.4.0-rc2
> >          98.68        +0.3%        98.98  bay/thresh=100M/btrfs-10dd-1-3.4.0-rc2
> >          99.34        -0.0%        99.31  bay/thresh=100M/btrfs-1dd-1-3.4.0-rc2
> >   ==>    88.98        +9.6%        97.53  bay/thresh=10M/btrfs-10dd-1-3.4.0-rc2
> >   ==>    86.99       +13.1%        98.39  bay/thresh=10M/btrfs-1dd-1-3.4.0-rc2
> >   ==>     2.75     +2442.4%        69.88  bay/thresh=1M/btrfs-10dd-1-3.4.0-rc2
> >   ==>     3.31     +2634.1%        90.54  bay/thresh=1M/btrfs-1dd-1-3.4.0-rc2
> > 
> > Signed-off-by: Fengguang Wu <fengguang.wu@intel.com>
> > ---
> >  fs/btrfs/disk-io.c |    3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > --- linux-next.orig/fs/btrfs/disk-io.c	2012-05-02 14:04:00.989262395 +0800
> > +++ linux-next/fs/btrfs/disk-io.c	2012-05-02 14:04:01.773262414 +0800
> > @@ -930,7 +930,8 @@ static int btree_writepages(struct addre
> >  
> >  		/* this is a bit racy, but that's ok */
> >  		num_dirty = root->fs_info->dirty_metadata_bytes;
> > -		if (num_dirty < thresh)
> > +		if (num_dirty < min(thresh,
> > +				    global_dirty_limit << (PAGE_CACHE_SHIFT-2)))
> >  			return 0;
> >  	}
> >  	return btree_write_cache_pages(mapping, wbc);
>   Frankly, that whole condition on WB_SYNC_NONE in btree_writepages() looks
> like a hack. I think we also had problems with this condition when we tried
> to change b_more_io list handling. I found rather terse commit message
> explaining the code:
> Btrfs: Limit btree writeback to prevent seeks

It is definitely a hack ;)  The basic point is that once we write a
metadata block, we have to cow it for any future changes.  So writing
the metadata has a pretty big impact on performance, and I'd rather
write everything else that is dirty first.  When that code was added I
was finding the metadata going to disk very soon under memory pressure.

I'm open to any ideas on this one.

-chris