Re: [PATCH 11/12] vmscan: Write out dirty pages in batch

[Dave Chinner <david@fromorbit.com>]

On Mon, Jun 14, 2010 at 06:39:57PM -0700, Andrew Morton wrote:
> On Tue, 15 Jun 2010 10:39:43 +1000 Dave Chinner <david@fromorbit.com> wrote:
> 
> > On Mon, Jun 14, 2010 at 04:21:43PM -0700, Andrew Morton wrote:
> > > On Tue, 15 Jun 2010 09:11:44 +1000
> > > > 10-15% reduction in seeks on simple kernel compile workloads. This
> > > > shows that if we optimise IO patterns at higher layers where the
> > > > sort window is much, much larger than in the IO scheduler, then
> > > > overall system performance improves....
> > > 
> > > Yup.
> > > 
> > > But then, this all really should be done at the block layer so other
> > > io-submitting-paths can benefit from it.
> > 
> > That was what we did in the past with really, really deep IO
> > scheduler queues. That leads to IO latency and OOM problems because
> > we could lock gigabytes of memory away under IO and take minutes to
> > clean it.
> > 
> > Besides, there really isn't the right context in the block layer to
> > be able to queue and prioritise large amounts of IO without
> > significant penalties to some higher layer operation.
> > 
> > > IOW, maybe "the sort queue is the submission queue" wasn't a good idea.
> > 
> > Perhaps, but IMO sorting should be done where the context allows it
> > to be done most efficiently. Sorting is most effective when ever a
> > significant queue of IO is formed, whether it be in the filesystem,
> > the VFS, the VM or the block layer because the IO stack is very much
> > a GIGO queue.
> > 
> > Simply put, there's nothing the lower layers can do to optimise bad
> > IO patterns from the higher layers because they have small sort
> > windows which are necessary to keep IO latency in check. Hence if
> > the higher layers feed the lower layers crap they simply don't have
> > the context or depth to perform the same level of optimistations we
> > can do easily higher up the stack.
> > 
> > IOWs, IMO anywhere there is a context with significant queue of IO,
> > that's where we should be doing a better job of sorting before that
> > IO is dispatched to the lower layers. This is still no guarantee of
> > better IO (e.g. if the filesystem fragments the file) but it does
> > give the lower layers a far better chance at optimal allocation and
> > scheduling of IO...
> 
> None of what you said had much to do with what I said.
> 
> What you've described are implementation problems in the current block
> layer because it conflates "sorting" with "queueing".  I'm saying "fix
> that".

You can't sort until you've queued.

> And...  sorting at the block layer will always be superior to sorting
> at the pagecache layer because the block layer sorts at the physical
> block level and can handle not-well-laid-out files and can sort and merge
> pages from different address_spaces.

Yes it, can do that. And it still does that even if the higher
layers sort their I/O dispatch better,

Filesystems try very hard to allocate adjacent logical offsets in a
file in adjacent physical blocks on disk - that's the whole point of
extent-indexed filesystems. Hence with modern filesystems there is
generally a direct correlation between the page {mapping,index}
tuple and the physical location of the mapped block.

i.e. there is generally zero physical correlation between pages in
different mappings, but there is a high physical correlation
between the index of pages on the same mapping. Hence by sorting
where we have a {mapping,index} context, we push out IO that is
much more likely to be in contiguous physical chunks that the
current random page shootdown.

We optimise applications to use these sorts of correlations all the
time to improve IO patterns. Why can't we make the same sort of
optmisations to the IO that the VM issues?

> Still, I suspect none of it will improve anything anyway.  Those pages
> are still dirty, possibly-locked and need to go to disk.  It doesn't
> matter from the MM POV whether they sit in some VM list or in the
> request queue.

Oh, but it does.....

> Possibly there may be some benefit to not putting so many of these
> unreclaimable pages into the queue all at the the same time.  But
> that's a shortcoming in the block code: we should be able to shove
> arbitrary numbers of dirty page (segments) into the queue and not gum
> the system up.  Don't try to work around that in the VM.

I think you know perfectly well why the system gums up when we
increase block layer queue depth: it's the fact that the _VM_ relies
on block layer queue congestion to limit the amount of dirty memory
in the system.

We've got a feedback loop between the block layer and the VM that
only works if block device queues are kept shallow. Keeping the
number of dirty pages under control is a VM responsibility, but it
is putting limitations on the block layer to ensure that the VM
works correctly.  If you want the block layer to have deep queues,
then someone needs to fix the VM not to require knowledge of the
internal operation of the block layer for correct operation.

Adding a few lines of code to sort a list in the VM is far, far
easier than redesigning the write throttling code....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com