From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail104.syd.optusnet.com.au ([211.29.132.246]:33714 "EHLO mail104.syd.optusnet.com.au" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S2389044AbfHHAbe (ORCPT ); Wed, 7 Aug 2019 20:31:34 -0400 Date: Thu, 8 Aug 2019 10:30:25 +1000 From: Dave Chinner Subject: Re: [PATCH] [Regression, v5.0] mm: boosted kswapd reclaim b0rks system cache balance Message-ID: <20190808003025.GU7777@dread.disaster.area> References: <20190807091858.2857-1-david@fromorbit.com> <20190807093056.GS11812@dhcp22.suse.cz> <20190807150316.GL2708@suse.de> <20190807220817.GN7777@dread.disaster.area> <20190807235534.GK2739@techsingularity.net> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20190807235534.GK2739@techsingularity.net> Sender: linux-xfs-owner@vger.kernel.org List-ID: List-Id: xfs To: Mel Gorman Cc: Michal Hocko , linux-mm@kvack.org, linux-xfs@vger.kernel.org, Vlastimil Babka On Thu, Aug 08, 2019 at 12:55:34AM +0100, Mel Gorman wrote: > On Thu, Aug 08, 2019 at 08:08:17AM +1000, Dave Chinner wrote: > > On Wed, Aug 07, 2019 at 04:03:16PM +0100, Mel Gorman wrote: > > > On Wed, Aug 07, 2019 at 11:30:56AM +0200, Michal Hocko wrote: > > > The boosting was not intended to target THP specifically -- it was meant > > > to help recover early from any fragmentation-related event for any user > > > that might need it. Hence, it's not tied to THP but even with THP > > > disabled, the boosting will still take effect. > > > > > > One band-aid would be to disable watermark boosting entirely when THP is > > > disabled but that feels wrong. However, I would be interested in hearing > > > if sysctl vm.watermark_boost_factor=0 has the same effect as your patch. > > > > > > > > Ok, it still runs it out of page cache, but it doesn't drive page > > cache reclaim as hard once there's none left. The IO patterns are > > less peaky, context switch rates are increased from ~3k/s to 15k/s > > but remain pretty steady. > > > > Test ran 5s faster and file rate improved by ~2%. So it's better > > once the page cache is largerly fully reclaimed, but it doesn't > > prevent the page cache from being reclaimed instead of inodes.... > > > > Ok. Ideally you would also confirm the patch itself works as you want. > It *should* but an actual confirmation would be nice. Yup, I'll get to that later today. Cheers, Dave. -- Dave Chinner david@fromorbit.com