Re: [PATCH 2/7] writeback: switch to per-bdi threads for flushing data

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

On Tue, 2009-03-17 at 10:38 +1100, Dave Chinner wrote:
> On Mon, Mar 16, 2009 at 08:33:21AM +0100, Jens Axboe wrote:
> > On Mon, Mar 16 2009, Dave Chinner wrote:
> > > On Fri, Mar 13, 2009 at 11:54:46AM +0100, Jens Axboe wrote:
> > > > On Thu, Mar 12 2009, Andrew Morton wrote:
> > > > > On Thu, 12 Mar 2009 15:33:43 +0100 Jens Axboe <jens.axboe@oracle.com> wrote:
> > > > > Bear in mind that the XFS guys found that one thread per fs had
> > > > > insufficient CPU power to keep up with fast devices.
> > > > 
> > > > Yes, I definitely want to experiment with > 1 thread per device in the
> > > > near future.
> > > 
> > > The question here is how to do this efficiently. Even if XFS is
> > > operating on a single device, it is not optimal just to throw
> > > multiple threads at the bdi. Ideally we want a thread per region
> > > (allocation group) of the filesystem as each allocation group has
> > > it's own inode cache (radix tree) to traverse. These traversals can
> > > be done completely in parallel and won't contend either at the
> > > traversal level or in the IO hardware....
> > > 
> > > i.e. what I'd like to see is the ability so any new flushing
> > > mechanism to be able to offload responsibility of tracking,
> > > traversing and flushing of dirty inodes to the filesystem.
> > > Filesystems that don't do such things could use a generic
> > > bdi-based implementation.
> > > 
> > > FWIW, we also want to avoid the current pattern of flushing
> > > data, then the inode, then data, then the inode, ....
> > > By offloading into the filesystem, this writeback ordering can
> > > be done as efficiently as possible for each given filesystem.
> > > XFs already has all the hooks to be able to do this
> > > effectively....
> > > 
> > > I know that Christoph was doing some work towards this end;
> > > perhaps he can throw his 2c worth in here...
> > 
> > This is very useful feedback, thanks Dave. So on the filesystem vs bdi
> > side, XFS could register a bdi per allocation group.
> 
> How do multiple bdis on a single block device interact?

The main difference is that dirty page tracking for balance_dirty_pages
and friends is done per-bdi.  So, you'll end up with uneven memory
pressure on ags that don't have much dirty data, but hopefully that's a
good thing.

-chris