xfs
[Top] [All Lists]

Re: [PATCH 2/5] mm: vmscan: Do not writeback filesystem pages in kswapd

To: Dave Chinner <david@xxxxxxxxxxxxx>
Subject: Re: [PATCH 2/5] mm: vmscan: Do not writeback filesystem pages in kswapd except in high priority
From: Mel Gorman <mgorman@xxxxxxx>
Date: Thu, 14 Jul 2011 14:17:45 +0100
Cc: Linux-MM <linux-mm@xxxxxxxxx>, LKML <linux-kernel@xxxxxxxxxxxxxxx>, XFS <xfs@xxxxxxxxxxx>, Christoph Hellwig <hch@xxxxxxxxxxxxx>, Johannes Weiner <jweiner@xxxxxxxxxx>, Wu Fengguang <fengguang.wu@xxxxxxxxx>, Jan Kara <jack@xxxxxxx>, Rik van Riel <riel@xxxxxxxxxx>, Minchan Kim <minchan.kim@xxxxxxxxx>
In-reply-to: <20110714115220.GB21663@dastard>
References: <1310567487-15367-1-git-send-email-mgorman@xxxxxxx> <1310567487-15367-3-git-send-email-mgorman@xxxxxxx> <20110713233743.GV23038@dastard> <20110714062947.GO7529@xxxxxxx> <20110714115220.GB21663@dastard>
User-agent: Mutt/1.5.21 (2010-09-15)
On Thu, Jul 14, 2011 at 09:52:21PM +1000, Dave Chinner wrote:
> On Thu, Jul 14, 2011 at 07:29:47AM +0100, Mel Gorman wrote:
> > On Thu, Jul 14, 2011 at 09:37:43AM +1000, Dave Chinner wrote:
> > > On Wed, Jul 13, 2011 at 03:31:24PM +0100, Mel Gorman wrote:
> > > > It is preferable that no dirty pages are dispatched for cleaning from
> > > > the page reclaim path. At normal priorities, this patch prevents kswapd
> > > > writing pages.
> > > > 
> > > > However, page reclaim does have a requirement that pages be freed
> > > > in a particular zone. If it is failing to make sufficient progress
> > > > (reclaiming < SWAP_CLUSTER_MAX at any priority priority), the priority
> > > > is raised to scan more pages. A priority of DEF_PRIORITY - 3 is
> > > > considered to tbe the point where kswapd is getting into trouble
> > > > reclaiming pages. If this priority is reached, kswapd will dispatch
> > > > pages for writing.
> > > > 
> > > > Signed-off-by: Mel Gorman <mgorman@xxxxxxx>
> > > 
> > > Seems reasonable, but btrfs still will ignore this writeback from
> > > kswapd, and it doesn't fall over.
> > 
> > At least there are no reports of it falling over :)
> 
> However you want to spin it.
> 

I regret that it is coming across as spin. My primary concern is
that if we get OOM-related bugs due to this series later that it'll
be difficult to pinpoint whether the whole series is at fault or whether
preventing kswapd writing any pages was at fault.

> > > Given that data point, I'd like to
> > > see the results when you stop kswapd from doing writeback altogether
> > > as well.
> > > 
> > 
> > The results for this test will be identical because the ftrace results
> > show that kswapd is already writing 0 filesystem pages.
> 
> You mean these numbers:
> 
> Kswapd reclaim write file async I/O           4483       4286 0          1    
>       0          0
> 
> Which shows that kswapd, under this workload has been improved to
> the point that it doesn't need to do IO. Yes, you've addressed the
> one problematic workload, but the numbers do not provide the answers
> to the fundamental question that have been raised during
> discussions. i.e. do we even need IO at all from reclaim?
> 

I don't know and at best will only be able to test with a single
disk which is why I wanted to separate this series from a complete
preventing of kswapd writing pages. I may be able to get access to
a machine with more disks but it'll take time.

> > Where it makes a difference is when the system is under enough
> > pressure that it is failing to reclaim any memory and is in danger
> > of prematurely triggering the OOM killer. Andrea outlined some of
> > the concerns before at http://lkml.org/lkml/2010/6/15/246
> 
> So put the system under more pressure such that with this patch
> series memory reclaim still writes from kswapd. Can you even get it
> to that stage, and if you can, does the system OOM more or less if
> you don't do file IO from reclaim?
> 

I can setup such a tests, it'll be at least next week before I
configure such a test and get it queued. It'll probably take a few
days to run then because more iterations will be required to pinpoint
where the OOM threshold is.  I know from the past that pushing a
system near OOM causes a non-deterministic number of triggers that
depend heavily on what was killed so the only real choice is to start
light and increase the load until boom which is time consuming.

Even then, the test will be inconclusive because it'll be just one
or two machines that I'll have to test on. There will be important
corner cases that I won't be able to test for.  For example;

  o small lowest zone that is critical for operation of some reason and
    the pages must be cleaned from there even though there is a large
    amount of memory overall

  o small highest zone causing high kswapd usage as it fails to balance
    continually due to pages being dirtied constantly and the window
    between when flushers clean the page and kswapd reclaim the page
    being too big. I might be able to simulate this one but bugs of
    this nature tend to be workload specific and affect some machines
    worse than others

  o Machines with many nodes and dirty pages spread semi-randomly
    on all nodes. If the flusher thread is not cleaning pages from
    a particular node that is under memory pressure due to affinity,
    processes will stall for long periods of time until the relevant
    inodes expire and gets cleaned. This will be particularly
    problematic if zone_reclaim is enabled

Questions about scenarios like this are going to cause problems in
review because it's reasonable to ask if any of them can occur and
we can't give an iron-clad answer.

> > > Can you try removing it altogether and seeing what that does to your
> > > test results? i.e
> > > 
> > >                   if (page_is_file_cache(page)) {
> > >                           inc_zone_page_state(page, NR_VMSCAN_WRITE_SKIP);
> > >                           goto keep_locked;
> > >                   }
> > 
> > It won't do anything, it'll still be writing 0 filesystem-backed pages.
> > 
> > Because of the possibility for the OOM killer triggering prematurely due
> > to the inability of kswapd to write pages, I'd prefer to separate such a
> > change by at least one release so that if there is an increase in OOM
> > reports, it'll be obvious what was the culprit.
> 
> I'm not asking for release quality patches or even when such fixes
> would roll out.
> 

Very well. I was hoping to start with just this series and handle the
complete disabling of writing later but it can wait a few weeks too. It
was always a stretch that the next merge window was going to be hit.

> What you've shown here is that memory reclaim can be more efficient
> without issuing IO itself under medium memory pressure. Now the
> question is whether it can do so under heavy, sustained, near OOM
> memory pressure?
> 
> IOWs, what I want to see is whether the fundamental principle of
> IO-less reclaim can be validated as workable or struck down.  This
> patchset demonstrates that IO-less reclaim is superior for a
> workload that produces medium levels of sustained IO-based memory
> pressure, which leads to the conclusion that the approach has merit
> and needs further investigation.
> 
> It's that next step that I'm asking you to test now. What form
> potential changes take or when they are released is irrelevant to me
> at this point, because we still haven't determined if the
> fundamental concept is completely sound or not. If the concept is
> sound I'm quite happy to wait until the implementation is fully
> baked before it gets rolled out....
> 

I'll setup a suitable test next week then.

-- 
Mel Gorman
SUSE Labs

<Prev in Thread] Current Thread [Next in Thread>