From: Mel Gorman <mgorman@techsingularity.net> To: Linus Torvalds <torvalds@linux-foundation.org> Cc: Dave Chinner <david@fromorbit.com>, Michal Hocko <mhocko@suse.cz>, Minchan Kim <minchan@kernel.org>, Vladimir Davydov <vdavydov@virtuozzo.com>, Johannes Weiner <hannes@cmpxchg.org>, Vlastimil Babka <vbabka@suse.cz>, Andrew Morton <akpm@linux-foundation.org>, Bob Peterson <rpeterso@redhat.com>, "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>, "Huang, Ying" <ying.huang@intel.com>, Christoph Hellwig <hch@lst.de>, Wu Fengguang <fengguang.wu@intel.com>, LKP <lkp@01.org>, Tejun Heo <tj@kernel.org>, LKML <linux-kernel@vger.kernel.org> Subject: Re: [LKP] [lkp] [xfs] 68a9f5e700: aim7.jobs-per-min -13.6% regression Date: Fri, 19 Aug 2016 11:49:46 +0100 [thread overview] Message-ID: <20160819104946.GL8119@techsingularity.net> (raw) In-Reply-To: <CA+55aFzqmNRWyczFHDCgjv2w6ZyD_AUJL1O1ZRmDs+DHT2=2cQ@mail.gmail.com> On Thu, Aug 18, 2016 at 03:25:40PM -0700, Linus Torvalds wrote: > >> In fact, looking at the __page_cache_alloc(), we already have that > >> "spread pages out" logic. I'm assuming Dave doesn't actually have that > >> bit set (I don't think it's the default), but I'm also envisioning > >> that maybe we could extend on that notion, and try to spread out > >> allocations in general, but keep page allocations from one particular > >> mapping within one node. > > > > CONFIG_CPUSETS=y > > > > But I don't have any cpusets configured (unless systemd is doing > > something wacky under the covers) so the page spread bit should not > > be set. > > Yeah, but even when it's not set we just do a generic alloc_pages(), > which is just going to fill up all nodes. Not perhaps quite as "spread > out", but there's obviously no attempt to try to be node-aware either. > There is a slight difference. Reads should fill the nodes in turn but dirty pages (__GFP_WRITE) get distributed to balance the number of dirty pages on each node to avoid hitting dirty balance limits prematurely. Yesterday I tried a patch that avoids distributing to remote nodes close to the high watermark to avoid waking remote kswapd instances. It added a lot of overhead to the fast path (3%) which hurts every writer but did not reduce contention enough on the special case of writing a single large file. As an aside, the dirty distribution check itself is very expensive so I prototyped something that does the expensive calculations on a vmstat update. Not sure if it'll work but it's a side issue. > So _if_ we come up with some reasonable way to say "let's keep the > pages of this mapping together", we could try to do it in that > numa-aware __page_cache_alloc(). > > It *could* be as simple/stupid as just saying "let's allocate the page > cache for new pages from the current node" - and if the process that > dirties pages just stays around on one single node, that might already > be sufficient. > > So just for testing purposes, you could try changing that > > return alloc_pages(gfp, 0); > > in __page_cache_alloc() into something like > > return alloc_pages_node(cpu_to_node(raw_smp_processor_id())), gfp, 0); > > or something. > The test would be interesting but I believe that keeping heavy writers on one node will force them to stall early on dirty balancing even if there is plenty of free memory on other nodes. -- Mel Gorman SUSE Labs
WARNING: multiple messages have this Message-ID (diff)
From: Mel Gorman <mgorman@techsingularity.net> To: lkp@lists.01.org Subject: Re: [xfs] 68a9f5e700: aim7.jobs-per-min -13.6% regression Date: Fri, 19 Aug 2016 11:49:46 +0100 [thread overview] Message-ID: <20160819104946.GL8119@techsingularity.net> (raw) In-Reply-To: <CA+55aFzqmNRWyczFHDCgjv2w6ZyD_AUJL1O1ZRmDs+DHT2=2cQ@mail.gmail.com> [-- Attachment #1: Type: text/plain, Size: 2515 bytes --] On Thu, Aug 18, 2016 at 03:25:40PM -0700, Linus Torvalds wrote: > >> In fact, looking at the __page_cache_alloc(), we already have that > >> "spread pages out" logic. I'm assuming Dave doesn't actually have that > >> bit set (I don't think it's the default), but I'm also envisioning > >> that maybe we could extend on that notion, and try to spread out > >> allocations in general, but keep page allocations from one particular > >> mapping within one node. > > > > CONFIG_CPUSETS=y > > > > But I don't have any cpusets configured (unless systemd is doing > > something wacky under the covers) so the page spread bit should not > > be set. > > Yeah, but even when it's not set we just do a generic alloc_pages(), > which is just going to fill up all nodes. Not perhaps quite as "spread > out", but there's obviously no attempt to try to be node-aware either. > There is a slight difference. Reads should fill the nodes in turn but dirty pages (__GFP_WRITE) get distributed to balance the number of dirty pages on each node to avoid hitting dirty balance limits prematurely. Yesterday I tried a patch that avoids distributing to remote nodes close to the high watermark to avoid waking remote kswapd instances. It added a lot of overhead to the fast path (3%) which hurts every writer but did not reduce contention enough on the special case of writing a single large file. As an aside, the dirty distribution check itself is very expensive so I prototyped something that does the expensive calculations on a vmstat update. Not sure if it'll work but it's a side issue. > So _if_ we come up with some reasonable way to say "let's keep the > pages of this mapping together", we could try to do it in that > numa-aware __page_cache_alloc(). > > It *could* be as simple/stupid as just saying "let's allocate the page > cache for new pages from the current node" - and if the process that > dirties pages just stays around on one single node, that might already > be sufficient. > > So just for testing purposes, you could try changing that > > return alloc_pages(gfp, 0); > > in __page_cache_alloc() into something like > > return alloc_pages_node(cpu_to_node(raw_smp_processor_id())), gfp, 0); > > or something. > The test would be interesting but I believe that keeping heavy writers on one node will force them to stall early on dirty balancing even if there is plenty of free memory on other nodes. -- Mel Gorman SUSE Labs
next prev parent reply other threads:[~2016-08-19 10:50 UTC|newest] Thread overview: 219+ messages / expand[flat|nested] mbox.gz Atom feed top 2016-08-09 14:33 [lkp] [xfs] 68a9f5e700: aim7.jobs-per-min -13.6% regression kernel test robot 2016-08-09 14:33 ` kernel test robot 2016-08-10 18:24 ` [lkp] " Linus Torvalds 2016-08-10 18:24 ` Linus Torvalds 2016-08-10 23:08 ` [lkp] " Dave Chinner 2016-08-10 23:08 ` Dave Chinner 2016-08-10 23:51 ` [lkp] " Linus Torvalds 2016-08-10 23:51 ` Linus Torvalds 2016-08-10 23:58 ` [LKP] [lkp] " Huang, Ying 2016-08-10 23:58 ` Huang, Ying 2016-08-11 0:11 ` [LKP] [lkp] " Huang, Ying 2016-08-11 0:11 ` Huang, Ying 2016-08-11 0:23 ` [LKP] [lkp] " Linus Torvalds 2016-08-11 0:23 ` Linus Torvalds 2016-08-11 0:33 ` [LKP] [lkp] " Huang, Ying 2016-08-11 0:33 ` Huang, Ying 2016-08-11 1:00 ` [LKP] [lkp] " Linus Torvalds 2016-08-11 1:00 ` Linus Torvalds 2016-08-11 4:46 ` [LKP] [lkp] " Dave Chinner 2016-08-11 4:46 ` Dave Chinner 2016-08-15 17:22 ` [LKP] [lkp] " Huang, Ying 2016-08-15 17:22 ` Huang, Ying 2016-08-16 0:08 ` [LKP] [lkp] " Dave Chinner 2016-08-16 0:08 ` Dave Chinner 2016-08-11 15:57 ` [LKP] [lkp] " Christoph Hellwig 2016-08-11 15:57 ` Christoph Hellwig 2016-08-11 16:55 ` [LKP] [lkp] " Linus Torvalds 2016-08-11 16:55 ` Linus Torvalds 2016-08-11 17:51 ` [LKP] [lkp] " Huang, Ying 2016-08-11 17:51 ` Huang, Ying 2016-08-11 19:51 ` [LKP] [lkp] " Linus Torvalds 2016-08-11 19:51 ` Linus Torvalds 2016-08-11 20:00 ` [LKP] [lkp] " Christoph Hellwig 2016-08-11 20:00 ` Christoph Hellwig 2016-08-11 20:35 ` [LKP] [lkp] " Linus Torvalds 2016-08-11 20:35 ` Linus Torvalds 2016-08-11 22:16 ` [LKP] [lkp] " Al Viro 2016-08-11 22:16 ` Al Viro 2016-08-11 22:30 ` [LKP] [lkp] " Linus Torvalds 2016-08-11 22:30 ` Linus Torvalds 2016-08-11 21:16 ` [LKP] [lkp] " Huang, Ying 2016-08-11 21:16 ` Huang, Ying 2016-08-11 21:40 ` [LKP] [lkp] " Linus Torvalds 2016-08-11 21:40 ` Linus Torvalds 2016-08-11 22:08 ` [LKP] [lkp] " Christoph Hellwig 2016-08-11 22:08 ` Christoph Hellwig 2016-08-12 0:54 ` [LKP] [lkp] " Dave Chinner 2016-08-12 0:54 ` Dave Chinner 2016-08-12 2:23 ` [LKP] [lkp] " Dave Chinner 2016-08-12 2:23 ` Dave Chinner 2016-08-12 2:32 ` [LKP] [lkp] " Linus Torvalds 2016-08-12 2:32 ` Linus Torvalds 2016-08-12 2:52 ` [LKP] [lkp] " Christoph Hellwig 2016-08-12 2:52 ` Christoph Hellwig 2016-08-12 3:20 ` [LKP] [lkp] " Linus Torvalds 2016-08-12 3:20 ` Linus Torvalds 2016-08-12 4:16 ` [LKP] [lkp] " Dave Chinner 2016-08-12 4:16 ` Dave Chinner 2016-08-12 5:02 ` [LKP] [lkp] " Linus Torvalds 2016-08-12 5:02 ` Linus Torvalds 2016-08-12 6:04 ` [LKP] [lkp] " Dave Chinner 2016-08-12 6:04 ` Dave Chinner 2016-08-12 6:29 ` [LKP] [lkp] " Ye Xiaolong 2016-08-12 6:29 ` Ye Xiaolong 2016-08-12 8:51 ` [LKP] [lkp] " Ye Xiaolong 2016-08-12 8:51 ` Ye Xiaolong 2016-08-12 10:02 ` [LKP] [lkp] " Dave Chinner 2016-08-12 10:02 ` Dave Chinner 2016-08-12 10:43 ` Fengguang Wu 2016-08-12 10:43 ` Fengguang Wu 2016-08-13 0:30 ` [LKP] [lkp] " Christoph Hellwig 2016-08-13 0:30 ` Christoph Hellwig 2016-08-13 21:48 ` [LKP] [lkp] " Christoph Hellwig 2016-08-13 21:48 ` Christoph Hellwig 2016-08-13 22:07 ` [LKP] [lkp] " Fengguang Wu 2016-08-13 22:07 ` Fengguang Wu 2016-08-13 22:15 ` [LKP] [lkp] " Christoph Hellwig 2016-08-13 22:15 ` Christoph Hellwig 2016-08-13 22:51 ` [LKP] [lkp] " Fengguang Wu 2016-08-13 22:51 ` Fengguang Wu 2016-08-14 14:50 ` [LKP] [lkp] " Fengguang Wu 2016-08-14 14:50 ` Fengguang Wu 2016-08-14 16:17 ` [LKP] [lkp] " Christoph Hellwig 2016-08-14 16:17 ` Christoph Hellwig 2016-08-14 23:46 ` [LKP] [lkp] " Dave Chinner 2016-08-14 23:46 ` Dave Chinner 2016-08-14 23:57 ` [LKP] [lkp] " Fengguang Wu 2016-08-14 23:57 ` Fengguang Wu 2016-08-15 14:14 ` [LKP] [lkp] " Fengguang Wu 2016-08-15 14:14 ` Fengguang Wu 2016-08-15 21:22 ` [LKP] [lkp] " Dave Chinner 2016-08-15 21:22 ` Dave Chinner 2016-08-16 12:20 ` [LKP] [lkp] " Fengguang Wu 2016-08-16 12:20 ` Fengguang Wu 2016-08-15 20:30 ` [LKP] [lkp] " Huang, Ying 2016-08-15 20:30 ` Huang, Ying 2016-08-22 22:09 ` [LKP] [lkp] " Huang, Ying 2016-08-22 22:09 ` Huang, Ying 2016-09-26 6:25 ` [LKP] [lkp] " Huang, Ying 2016-09-26 6:25 ` Huang, Ying 2016-09-26 14:55 ` [LKP] [lkp] " Christoph Hellwig 2016-09-26 14:55 ` Christoph Hellwig 2016-09-27 0:52 ` [LKP] [lkp] " Huang, Ying 2016-09-27 0:52 ` Huang, Ying 2016-08-16 13:25 ` [LKP] [lkp] " Fengguang Wu 2016-08-16 13:25 ` Fengguang Wu 2016-08-13 23:32 ` [LKP] [lkp] " Dave Chinner 2016-08-13 23:32 ` Dave Chinner 2016-08-12 2:27 ` [LKP] [lkp] " Linus Torvalds 2016-08-12 2:27 ` Linus Torvalds 2016-08-12 3:56 ` [LKP] [lkp] " Dave Chinner 2016-08-12 3:56 ` Dave Chinner 2016-08-12 18:03 ` [LKP] [lkp] " Linus Torvalds 2016-08-12 18:03 ` Linus Torvalds 2016-08-13 23:58 ` [LKP] [lkp] " Fengguang Wu 2016-08-13 23:58 ` Fengguang Wu 2016-08-15 0:48 ` [LKP] [lkp] " Dave Chinner 2016-08-15 0:48 ` Dave Chinner 2016-08-15 1:37 ` [LKP] [lkp] " Linus Torvalds 2016-08-15 1:37 ` Linus Torvalds 2016-08-15 2:28 ` [LKP] [lkp] " Dave Chinner 2016-08-15 2:28 ` Dave Chinner 2016-08-15 2:53 ` [LKP] [lkp] " Linus Torvalds 2016-08-15 2:53 ` Linus Torvalds 2016-08-15 5:00 ` [LKP] [lkp] " Dave Chinner 2016-08-15 5:00 ` Dave Chinner [not found] ` <CA+55aFwva2Xffai+Eqv1Jn_NGryk3YJ2i5JoHOQnbQv6qVPAsw@mail.gmail.com> [not found] ` <CA+55aFy14nUnJQ_GdF=j8Fa9xiH70c6fY2G3q5HQ01+8z1z3qQ@mail.gmail.com> 2016-08-15 5:12 ` Linus Torvalds 2016-08-15 22:22 ` [LKP] [lkp] " Dave Chinner 2016-08-15 22:22 ` Dave Chinner 2016-08-15 22:42 ` [LKP] [lkp] " Dave Chinner 2016-08-15 22:42 ` Dave Chinner 2016-08-15 23:20 ` [LKP] [lkp] " Linus Torvalds 2016-08-15 23:20 ` Linus Torvalds 2016-08-15 23:48 ` [LKP] [lkp] " Linus Torvalds 2016-08-15 23:48 ` Linus Torvalds 2016-08-16 0:44 ` [LKP] [lkp] " Dave Chinner 2016-08-16 0:44 ` Dave Chinner 2016-08-16 15:05 ` [LKP] [lkp] " Mel Gorman 2016-08-16 15:05 ` Mel Gorman 2016-08-16 17:47 ` [LKP] [lkp] " Linus Torvalds 2016-08-16 17:47 ` Linus Torvalds 2016-08-17 15:48 ` [LKP] [lkp] " Michal Hocko 2016-08-17 15:48 ` Michal Hocko 2016-08-17 16:42 ` [LKP] [lkp] " Michal Hocko 2016-08-17 16:42 ` Michal Hocko 2016-08-17 15:49 ` [LKP] [lkp] " Mel Gorman 2016-08-17 15:49 ` Mel Gorman 2016-08-18 0:45 ` [LKP] [lkp] " Mel Gorman 2016-08-18 0:45 ` Mel Gorman 2016-08-18 7:11 ` [LKP] [lkp] " Dave Chinner 2016-08-18 7:11 ` Dave Chinner 2016-08-18 13:24 ` [LKP] [lkp] " Mel Gorman 2016-08-18 13:24 ` Mel Gorman 2016-08-18 17:55 ` [LKP] [lkp] " Linus Torvalds 2016-08-18 17:55 ` Linus Torvalds 2016-08-18 21:19 ` [LKP] [lkp] " Dave Chinner 2016-08-18 21:19 ` Dave Chinner 2016-08-18 22:25 ` [LKP] [lkp] " Linus Torvalds 2016-08-18 22:25 ` Linus Torvalds 2016-08-19 9:00 ` [LKP] [lkp] " Michal Hocko 2016-08-19 9:00 ` Michal Hocko 2016-08-19 10:49 ` Mel Gorman [this message] 2016-08-19 10:49 ` Mel Gorman 2016-08-19 23:48 ` [LKP] [lkp] " Dave Chinner 2016-08-19 23:48 ` Dave Chinner 2016-08-20 1:08 ` [LKP] [lkp] " Linus Torvalds 2016-08-20 1:08 ` Linus Torvalds 2016-08-20 12:16 ` [LKP] [lkp] " Mel Gorman 2016-08-20 12:16 ` Mel Gorman 2016-08-19 15:08 ` [LKP] [lkp] " Mel Gorman 2016-08-19 15:08 ` Mel Gorman 2016-09-01 23:32 ` [LKP] [lkp] " Dave Chinner 2016-09-01 23:32 ` Dave Chinner 2016-09-06 15:37 ` [LKP] [lkp] " Mel Gorman 2016-09-06 15:37 ` Mel Gorman 2016-09-06 15:52 ` [LKP] [lkp] " Huang, Ying 2016-09-06 15:52 ` Huang, Ying 2016-08-24 15:40 ` [LKP] [lkp] " Huang, Ying 2016-08-24 15:40 ` Huang, Ying 2016-08-25 9:37 ` [LKP] [lkp] " Mel Gorman 2016-08-25 9:37 ` Mel Gorman 2016-08-18 2:44 ` [LKP] [lkp] " Dave Chinner 2016-08-18 2:44 ` Dave Chinner 2016-08-16 0:15 ` [LKP] [lkp] " Linus Torvalds 2016-08-16 0:15 ` Linus Torvalds 2016-08-16 0:38 ` [LKP] [lkp] " Dave Chinner 2016-08-16 0:38 ` Dave Chinner 2016-08-16 0:50 ` [LKP] [lkp] " Linus Torvalds 2016-08-16 0:50 ` Linus Torvalds 2016-08-16 0:19 ` [LKP] [lkp] " Dave Chinner 2016-08-16 0:19 ` Dave Chinner 2016-08-16 1:51 ` [LKP] [lkp] " Linus Torvalds 2016-08-16 1:51 ` Linus Torvalds 2016-08-16 22:02 ` [LKP] [lkp] " Dave Chinner 2016-08-16 22:02 ` Dave Chinner 2016-08-16 23:23 ` [LKP] [lkp] " Linus Torvalds 2016-08-16 23:23 ` Linus Torvalds 2016-08-15 23:01 ` [LKP] [lkp] " Linus Torvalds 2016-08-15 23:01 ` Linus Torvalds 2016-08-16 0:17 ` [LKP] [lkp] " Dave Chinner 2016-08-16 0:17 ` Dave Chinner 2016-08-16 0:45 ` [LKP] [lkp] " Linus Torvalds 2016-08-16 0:45 ` Linus Torvalds 2016-08-15 5:03 ` [LKP] [lkp] " Ingo Molnar 2016-08-15 5:03 ` Ingo Molnar 2016-08-17 16:24 ` [LKP] [lkp] " Peter Zijlstra 2016-08-17 16:24 ` Peter Zijlstra 2016-08-15 12:58 ` [LKP] [lkp] " Fengguang Wu 2016-08-15 12:58 ` Fengguang Wu 2016-08-11 1:16 ` [LKP] [lkp] " Dave Chinner 2016-08-11 1:16 ` Dave Chinner 2016-08-11 1:32 ` [LKP] [lkp] " Dave Chinner 2016-08-11 1:32 ` Dave Chinner 2016-08-11 2:36 ` [LKP] [lkp] " Ye Xiaolong 2016-08-11 2:36 ` Ye Xiaolong 2016-08-11 3:05 ` [LKP] [lkp] " Dave Chinner 2016-08-11 3:05 ` Dave Chinner 2016-08-12 1:26 ` [LKP] [lkp] " Dave Chinner 2016-08-12 1:26 ` Dave Chinner
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=20160819104946.GL8119@techsingularity.net \ --to=mgorman@techsingularity.net \ --cc=akpm@linux-foundation.org \ --cc=david@fromorbit.com \ --cc=fengguang.wu@intel.com \ --cc=hannes@cmpxchg.org \ --cc=hch@lst.de \ --cc=kirill.shutemov@linux.intel.com \ --cc=linux-kernel@vger.kernel.org \ --cc=lkp@01.org \ --cc=mhocko@suse.cz \ --cc=minchan@kernel.org \ --cc=rpeterso@redhat.com \ --cc=tj@kernel.org \ --cc=torvalds@linux-foundation.org \ --cc=vbabka@suse.cz \ --cc=vdavydov@virtuozzo.com \ --cc=ying.huang@intel.com \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes, see mirroring instructions on how to clone and mirror all data and code used by this external index.