[Top] [All Lists]

Re: [PATCH 8/8] mm: vmscan: Do not writeback filesystem pages from kswap

To: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Subject: Re: [PATCH 8/8] mm: vmscan: Do not writeback filesystem pages from kswapd
From: Mel Gorman <mgorman@xxxxxxx>
Date: Fri, 22 Jul 2011 14:31:38 +0100
Cc: Linux-MM <linux-mm@xxxxxxxxx>, LKML <linux-kernel@xxxxxxxxxxxxxxx>, XFS <xfs@xxxxxxxxxxx>, Dave Chinner <david@xxxxxxxxxxxxx>, 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: <1311339432.27400.36.camel@twins>
References: <1311265730-5324-1-git-send-email-mgorman@xxxxxxx> <1311265730-5324-9-git-send-email-mgorman@xxxxxxx> <1311339432.27400.36.camel@twins>
User-agent: Mutt/1.5.21 (2010-09-15)
On Fri, Jul 22, 2011 at 02:57:12PM +0200, Peter Zijlstra wrote:
> On Thu, 2011-07-21 at 17:28 +0100, Mel Gorman wrote:
> > Assuming that flusher threads will always write back dirty pages promptly
> > then it is always faster for reclaimers to wait for flushers. This patch
> > prevents kswapd writing back any filesystem pages. 
> That is a somewhat sort changelog for such a big assumption ;-)

That is an understatement but the impact of the patch is discussed in
detail in the leader. On NUMA, this patch has a negative impact so
I put no effort into the changelog. The patch is part of the series
because it was specifically asked for.

> I think it can use a few extra words to explain the need to clean pages
> from @zone vs writeback picks whatever fits best on disk and how that
> works out wrt the assumption.

At the time of writing the changelog, I knew that flushers were
not finding pages from the correct zones quickly enough in the NUMA
usecase. The changelog documents the assumptions testing shows them to
be false.

> What requirements does this place on writeback and how does it meet
> them.

It places a requirement on writeback to prioritise pages from zones
under memory pressure. It doesn't meet them. I mention in the leader
that I think patch 8 should be dropped which is why the changelog

Mel Gorman

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