lkml.org 
[lkml]   [2010]   [Aug]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH 0/2] Prioritise inodes and zones for writeback required by page reclaim
On Wed,  4 Aug 2010 15:38:29 +0100
Mel Gorman <mel@csn.ul.ie> wrote:

> Commenting on the series "Reduce writeback from page reclaim context V6"
> Andrew Morton noted;
>
> direct-reclaim wants to write a dirty page because that page is in the
> zone which the caller wants to allocate from! Telling the flusher threads
> to perform generic writeback will sometimes cause them to just gum the
> disk up with pages from different zones, making it even harder/slower to
> allocate a page from the zones we're interested in, no?
>
> On the machines used to test the series, there were relatively few zones
> and only one BDI so the scenario describes is a possibility. This series is
> a very early prototype series aimed at mitigating the problem.
>
> Patch 1 adds wakeup_flusher_threads_pages() which takes a list of pages
> from page reclaim. Each inode belonging to a page on the list is marked
> I_DIRTY_RECLAIM. When the flusher thread wakes, inodes with this tag are
> unconditionally moved to the wb->b_io list for writing.
>
> Patch 2 notes that writing back inodes does not necessarily write back
> pages belonging to the zone page reclaim is concerned with. In response, it
> adds a zone and counter to wb_writeback_work. As pages from the target zone
> are written, the zone-specific counter is updated. When the flusher thread
> then checks the zone counters if a specific zone is being targeted. While
> more pages may be written than necessary, the assumption is that the pages
> need cleaning eventually, the inode must be relatively old to have pages at
> the end of the LRU, the IO will be relatively efficient due to less random
> seeks and that pages from the target zone will still be cleaned.
>
> Testing did not show any significant differences in terms of reducing dirty
> file pages being written back but the lack of multiple BDIs and NUMA nodes in
> the test rig is a problem. Maybe someone else has access to a more suitable
> test rig.
>
> Any comment as to the suitability for such a direction?

um. Might work. Isn't pretty though.

But until we can demonstrate the problem or someone reports it, we
probably have more important issues to be looking at ;) I think that a
better approach is to try to trigger this problem as we develop and
test reclaim. And if we _can't_ demonstrate it, work out why the heck
not - either the code's smarter than we thought it was or the test is
no good.



\
 
 \ /
  Last update: 2010-08-05 01:01    [W:0.038 / U:0.056 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site