lkml.org 
[lkml]   [2012]   [May]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] mm: consider all swapped back pages in used-once logic
On Mon, May 21, 2012 at 10:51:49AM +0800, Zheng Liu wrote:
> On Thu, May 17, 2012 at 09:54:25PM +0200, Johannes Weiner wrote:
> > On Thu, May 17, 2012 at 11:13:53AM +0200, Michal Hocko wrote:
> > > [64574746 vmscan: detect mapped file pages used only once] made mapped pages
> > > have another round in inactive list because they might be just short
> > > lived and so we could consider them again next time. This heuristic
> > > helps to reduce pressure on the active list with a streaming IO
> > > worklods.
> > > This patch fixes a regression introduced by this commit for heavy shmem
> > > based workloads because unlike Anon pages, which are excluded from this
> > > heuristic because they are usually long lived, shmem pages are handled
> > > as a regular page cache.
> > > This doesn't work quite well, unfortunately, if the workload is mostly
> > > backed by shmem (in memory database sitting on 80% of memory) with a
> > > streaming IO in the background (backup - up to 20% of memory). Anon
> > > inactive list is full of (dirty) shmem pages when watermarks are
> > > hit. Shmem pages are kept in the inactive list (they are referenced)
> > > in the first round and it is hard to reclaim anything else so we reach
> > > lower scanning priorities very quickly which leads to an excessive swap
> > > out.
> > >
> > > Let's fix this by excluding all swap backed pages (they tend to be long
> > > lived wrt. the regular page cache anyway) from used-once heuristic and
> > > rather activate them if they are referenced.
> >
> > Yes, the algorithm only makes sense for file cache, which is easy to
> > reclaim. Thanks for the fix!
>
> Hi Johannes,
>
> Out of curiosity, I notice that, in this patch (64574746), the commit log
> said that this patch aims to reduce the impact of pages used only once.
> Could you please tell why you think these pages will flood the active
> list? How do you find this problem?

Applications that use mmap for large, linear used-once IO. Reclaim
used to just activate every mapped file page it encountered for the
first time (activate referenced ones, but they all start referenced) .
This resulted in horrible reclaim latency as most pages in memory
where active.

> Actually, we met a huge regression in our product system. This
> application uses mmap/munmap and read/write simultaneously. Meanwhile
> it wants to keep mapped file pages in memory as much as possible. But
> this patch causes that mapped file pages are reclaimed frequently. So I
> want to know whether or not this patch consider this situation. Thank
> you.

Is it because the read()/write() IO is high throughput and pushes
pages through the LRU lists faster than the mmap pages are referenced?

Are the mmap pages executable or shared between tasks? If so, does
the kernel you are using include '34dbc67 vmscan: promote shared file
mapped pages' and 'c909e99 vmscan: activate executable pages after
first usage'?

All of this is very lame. I see no way to automatically detect when
you really want to keep mapped pages over unmapped ones. And making
this assumption hurt some loads, while not making it now hurts others.


\
 
 \ /
  Last update: 2012-05-21 10:21    [from the cache]
©2003-2011 Jasper Spaans