lkml.org 
[lkml]   [2010]   [Sep]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 10/10] vmscan: Kick flusher threads to clean pages when reclaim is encountering dirty pages
On Mon, Sep 13, 2010 at 10:10:46PM +0800, Mel Gorman wrote:
> On Mon, Sep 13, 2010 at 09:48:45PM +0800, Wu Fengguang wrote:
> > > + /*
> > > + * If reclaim is encountering dirty pages, it may be because
> > > + * dirty pages are reaching the end of the LRU even though the
> > > + * dirty_ratio may be satisified. In this case, wake flusher
> > > + * threads to pro-actively clean up to a maximum of
> > > + * 4 * SWAP_CLUSTER_MAX amount of data (usually 1/2MB) unless
> > > + * !may_writepage indicates that this is a direct reclaimer in
> > > + * laptop mode avoiding disk spin-ups
> > > + */
> > > + if (file && nr_dirty_seen && sc->may_writepage)
> > > + wakeup_flusher_threads(nr_writeback_pages(nr_dirty));
> >
> > wakeup_flusher_threads() works, but seems not the pertinent one.
> >
> > - locally, it needs some luck to clean the pages that direct reclaim is waiting on
>
> There is a certain amount of luck involved but it's depending on there being a
> correlation between old inodes and old pages on the LRU list. As long as that
> correlation is accurate, some relevant pages will get cleaned. Testing on
> previously released versions of this patch did show that the percentage of
> dirty pages encountered during reclaim were reduced as a result of this patch.

Yup.

> > - globally, it cleans up some dirty pages, however some heavy dirtier
> > may quickly create new ones..
> >
> > So how about taking the approaches in these patches?
> >
> > - "[PATCH 4/4] vmscan: transfer async file writeback to the flusher"
> > - "[PATCH 15/17] mm: lower soft dirty limits on memory pressure"
> >
>
> There is a lot going on in those patches. It's going to take me a while to
> figure them out and formulate an opinion.

OK. I also need some time off for doing other works :)

> > In particular the first patch should work very nicely with memcg, as
> > all pages of an inode typically belong to the same memcg. So doing
> > write-around helps clean lots of dirty pages in the target LRU list in
> > one shot.
> >
>
> It might but as there is also a correlation between old dirty inodes and
> the location of dirty pages, it is tricky to predict if it is better and
> if so, by how much.

It at least guarantees to clean the one page pageout() is running into :)
Others will depend on the locality/sequentiality of the workload. But
as the write-around pages are in the same LRU lists, the vmscan code
will hit them sooner or later.

Thanks,
Fengguang



\
 
 \ /
  Last update: 2010-09-13 16:43    [W:0.069 / U:4.052 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site