From: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
To: Rik van Riel <riel@redhat.com>
Subject: Re: [PATCH]vmscan: handle underflow for get_scan_ratio
Cc: kosaki.motohiro@jp.fujitsu.com, Andrew Morton <akpm@linux-foundation.org>,
       Shaohua Li <shaohua.li@intel.com>,
       "Wu, Fengguang" <fengguang.wu@intel.com>,
       "linux-mm@kvack.org" <linux-mm@kvack.org>,
       "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
In-Reply-To: <20100413144519.D107.A69D9226@jp.fujitsu.com>
References: <4BC3DA2B.3070605@redhat.com> <20100413144519.D107.A69D9226@jp.fujitsu.com>
Message-Id: <20100413175414.D110.A69D9226@jp.fujitsu.com>
MIME-Version: 1.0
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 8bit
Date: Tue, 13 Apr 2010 17:55:52 +0900 (JST)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2109
Lines: 52

> > > I'm surprised this ack a bit. Rik, do you have any improvement plan about
> > > streaming io detection logic?
> > > I think the patch have a slightly marginal benefit, it help to<1% scan
> > > ratio case. but it have big regression, it cause streaming io (e.g. backup
> > > operation) makes tons swap.
> > 
> > How?  From the description I believe it took 16GB in
> > a zone before we start scanning anon pages when
> > reclaiming at DEF_PRIORITY?
> > 
> > Would that casue a problem?
> 
> Please remember, 2.6.27 has following +1 scanning modifier.
> 
>   zone->nr_scan_active += (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
>                                                                          ^^^^
> 
> and, early (ano not yet merged) SplitLRU VM has similar +1. likes
> 
>          scan = zone_nr_lru_pages(zone, sc, l);
>          scan >>= priority;
>          scan = (scan * percent[file]) / 100 + 1;
>                                              ^^^
> 
> We didn't think only one page scanning is not big matter. but it was not
> correct. we got streaming io bug report. the above +1 makes annoying swap
> io. because some server need big backup operation rather much much than
> physical memory size.
> 
> example, If vm are dropping 1TB use once pages, 0.1% anon scanning makes
> 1GB scan. and almost server only have some gigabyte swap although it
> has >1TB memory.
> 
> If my memory is not correct, please correct me.
> 
> My point is, greater or smaller than 16GB isn't essential. all patches 
> should have big worth than the downside. The description said "the impact 
> sounds not a big deal", nobody disagree it. but it's worth is more little.
> I don't imagine this patch improve anything.

And now I've merged this patch into my local vmscan patch queue.
After solving streaming io issue, I'll put it to mainline.

Thanks.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/