Re: [PATCH 0/6] writeback time order/delay fixes take 3

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

On Thu, 23 Aug 2007 12:47:23 +1000
David Chinner <dgc@sgi.com> wrote:

> On Wed, Aug 22, 2007 at 08:42:01AM -0400, Chris Mason wrote:
> > I think we should assume a full scan of s_dirty is impossible in the
> > presence of concurrent writers.  We want to be able to pick a start
> > time (right now) and find all the inodes older than that start time.
> > New things will come in while we're scanning.  But perhaps that's
> > what you're saying...
> > 
> > At any rate, we've got two types of lists now.  One keeps track of
> > age and the other two keep track of what is currently being
> > written.  I would try two things:
> > 
> > 1) s_dirty stays a list for FIFO.  s_io becomes a radix tree that
> > indexes by inode number (or some arbitrary field the FS can set in
> > the inode).  Radix tree tags are used to indicate which things in
> > s_io are already in progress or are pending (hand waving because
> > I'm not sure exactly).
> > 
> > inodes are pulled off s_dirty and the corresponding slot in s_io is
> > tagged to indicate IO has started.  Any nearby inodes in s_io are
> > also sent down.
> 
> the problem with this approach is that it only looks at inode
> locality. Data locality is ignored completely here and the data for
> all the inodes that are close together could be splattered all over
> the drive. In that case, clustering by inode location is exactly the
> wrong thing to do.

Usually it won't be less wrong than clustering by time.

> 
> For example, XFs changes allocation strategy at 1TB for 32bit inode
> filesystems which makes the data get placed way away from the inodes.
> i.e. inodes in AGs below 1TB, all data in AGs > 1TB. clustering
> by inode number for data writeback is mostly useless in the >1TB
> case.

I agree we'll want a way to let the FS provide the clustering key.  But
for the first cut on the patch, I would suggest keeping it simple.

> 
> The inode32 for <1Tb and inode64 allocators both try to keep data
> close to the inode (i.e. in the same AG) so clustering by inode number
> might work better here.
> 
> Also, it might be worthwhile allowing the filesystem to supply a
> hint or mask for "closeness" for inode clustering. This would help
> the gernic code only try to cluster inode writes to inodes that
> fall into the same cluster as the first inode....

Yes, also a good idea after things are working.

> 
> > > Notes:
> > > (1) I'm not sure inode number is correlated to disk location in
> > >     filesystems other than ext2/3/4. Or parent dir?
> > 
> > In general, it is a better assumption than sorting by time.  It may
> > make sense to one day let the FS provide a clustering hint
> > (corresponding to the first block in the file?), but for starters it
> > makes sense to just go with the inode number.
> 
> Perhaps multiple hints are needed - one for data locality and one
> for inode cluster locality.

So, my feature creep idea would have been more data clustering.  I'm
mainly trying to solve this graph:

http://oss.oracle.com/~mason/compilebench/makej/compare-create-dirs-0.png

Where background writing of the block device inode is making ext3 do
seeky writes while directory trees.  My simple idea was to kick
off a 'I've just written block X' call back to the FS, where it may
decide to send down dirty chunks of the block device inode that also
happen to be dirty.

But, maintaining the kupdate max dirty time and congestion limits in
the face of all this clustering gets tricky.  So, I wasn't going to
suggest it until the basic machinery was working.

Fengguang, this isn't a small project ;)  But, lots of people will be
interested in the results.

-chris