Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753447AbcDKIRV (ORCPT ); Mon, 11 Apr 2016 04:17:21 -0400 Received: from mx2.suse.de ([195.135.220.15]:39032 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753069AbcDKIRR (ORCPT ); Mon, 11 Apr 2016 04:17:17 -0400 Subject: Re: [PATCH v2 0/4] reduce latency of direct async compaction To: Joonsoo Kim References: <1459414236-9219-1-git-send-email-vbabka@suse.cz> <20160411070547.GA26116@js1304-P5Q-DELUXE> Cc: Andrew Morton , linux-mm@kvack.org, linux-kernel@vger.kernel.org, Mel Gorman , Rik van Riel , David Rientjes , Minchan Kim , Michal Hocko From: Vlastimil Babka Message-ID: <570B5D89.1010808@suse.cz> Date: Mon, 11 Apr 2016 10:17:13 +0200 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.7.0 MIME-Version: 1.0 In-Reply-To: <20160411070547.GA26116@js1304-P5Q-DELUXE> Content-Type: text/plain; charset=windows-1252; format=flowed Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 9548 Lines: 187 On 04/11/2016 09:05 AM, Joonsoo Kim wrote: > On Thu, Mar 31, 2016 at 10:50:32AM +0200, Vlastimil Babka wrote: >> The goal here is to reduce latency (and increase success) of direct async >> compaction by making it focus more on the goal of creating a high-order page, >> at some expense of thoroughness. >> >> This is based on an older attempt [1] which I didn't finish as it seemed that >> it increased longer-term fragmentation. Now it seems it doesn't, and we have >> kcompactd for that goal. The main patch (3) makes migration scanner skip whole >> order-aligned blocks as soon as isolation fails in them, as it takes just one >> unmigrated page to prevent a high-order buddy page from fully merging. >> >> Patch 4 then attempts to reduce the excessive freepage scanning (such as >> reported in [2]) by allocating migration targets directly from freelists. Here >> we just need to be sure that the free pages are not from the same block as the >> migrated pages. This is also limited to direct async compaction and is not >> meant to replace the more thorough free scanner for other scenarios. > > I don't like that another algorithm is introduced for async > compaction. As you know, we already suffer from corner case that async > compaction have (such as compaction deferring doesn't work if we only > do async compaction). It makes further analysis/improvement harder. Generally, > more difference on async compaction would cause more problem later. My idea is that async compaction could become "good enough" for majority of cases, and strive for minimum latency. If it has to be different for that goal, so be it. But of course it should not cause problems for the sync fallback/kcompactd work. > In suggested approach, possible risky places I think is finish condition > and deferring logic. Scanner meet position would be greatly affected > by system load. If there are no processes and async compaction > isn't aborted, freepage scanner will be at the end of the zone and > we can scan migratable page until we reach there. But, in the other case > that the system has some load, async compaction would be aborted easily and > freepage scanner will be at the some of point of the zone and > async compaction's scanning power can be limited a lot. Hmm, I thought that I've changed the migration scanner for the new mode to stop looking at free scanner position. Looks like I forgot/it got lost, but I definitely wanted to try that. > And, with different algorithm, it doesn't make sense to share same deferring > logic. Async compaction can succeed even if sync compaction continually fails. That makes sense. > I hope that we don't make async/sync compaction more diverse. I'd be > more happy if we can apply such a change to both async/sync direct > compaction. OK, perhaps for sync direct compaction it could be tried too. But I think not kcompactd, which has broader goals than making a single page of given order (well, not in the initial implementation, but I'm working on it :) But it just occured to me that even kcompactd could incorporate something like patch 3 to fight fragmentation. If we can't isolate a page, then migrating its buddy will only create order-0 freepage. That cannot help against fragmentation, only possibly make it worse if we have to split a larger page for migration target. The question is, to which order to extend this logic? >> >> [1] https://lkml.org/lkml/2014/7/16/988 >> [2] http://www.spinics.net/lists/linux-mm/msg97475.html >> >> Testing was done using stress-highalloc from mmtests, configured for order-4 >> GFP_KERNEL allocations: >> >> 4.6-rc1 4.6-rc1 4.6-rc1 >> patch2 patch3 patch4 >> Success 1 Min 24.00 ( 0.00%) 27.00 (-12.50%) 43.00 (-79.17%) >> Success 1 Mean 30.20 ( 0.00%) 31.60 ( -4.64%) 51.60 (-70.86%) >> Success 1 Max 37.00 ( 0.00%) 35.00 ( 5.41%) 73.00 (-97.30%) >> Success 2 Min 42.00 ( 0.00%) 32.00 ( 23.81%) 73.00 (-73.81%) >> Success 2 Mean 44.00 ( 0.00%) 44.80 ( -1.82%) 78.00 (-77.27%) >> Success 2 Max 48.00 ( 0.00%) 52.00 ( -8.33%) 81.00 (-68.75%) >> Success 3 Min 91.00 ( 0.00%) 92.00 ( -1.10%) 88.00 ( 3.30%) >> Success 3 Mean 92.20 ( 0.00%) 92.80 ( -0.65%) 91.00 ( 1.30%) >> Success 3 Max 94.00 ( 0.00%) 93.00 ( 1.06%) 94.00 ( 0.00%) >> >> While the eager skipping of unsuitable blocks from patch 3 didn't affect >> success rates, direct freepage allocation did improve them. > > Direct freepage allocation changes compaction algorithm a lot. It > removes limitation that we cannot get freepages from behind the > migration scanner so we can get freepage easily. It would be achieved > by other compaction algorithm changes (such as your pivot change or my > compaction algorithm change or this patchset). Pivot change or your algorithm would be definitely good for kcompactd. > For the long term, this > limitation should be removed for sync compaction (at least direct sync > compaction), too. What's the reason that you don't apply this algorithm > to other cases? Is there any change in fragmentation? I wanted to be on the safe side. As Mel pointed out, parallel compactions could be using same blocks for opposite purposes, so leave a fallback mode that's not prone to that. But I'm considering that pageblock skip bits could be repurposed as a "pageblock lock" for compaction. Michal's oom rework experiments show that the original purpose of the skip bits is causing problems when compaction is asked to "try really everything you can and either succeed, or report a real failure" and I suspect they aren't much better than a random pageblock skipping in reducing compaction latencies. And yeah, potential long-term fragmentation was another concern, but hopefully will be diminished by a more proactive kcompactd. So, it seems both you and Mel have doubts about Patch 4, but patches 1-3 could be acceptable for starters? > Thanks. > >> >> 4.6-rc1 4.6-rc1 4.6-rc1 >> patch2 patch3 patch4 >> User 2587.42 2566.53 2413.57 >> System 482.89 471.20 461.71 >> Elapsed 1395.68 1382.00 1392.87 >> >> Times are not so useful metric for this benchmark as main portion is the >> interfering kernel builds, but results do hint at reduced system times. >> >> 4.6-rc1 4.6-rc1 4.6-rc1 >> patch2 patch3 patch4 >> Direct pages scanned 163614 159608 123385 >> Kswapd pages scanned 2070139 2078790 2081385 >> Kswapd pages reclaimed 2061707 2069757 2073723 >> Direct pages reclaimed 163354 159505 122304 >> >> Reduced direct reclaim was unintended, but could be explained by more >> successful first attempt at (async) direct compaction, which is attempted >> before the first reclaim attempt in __alloc_pages_slowpath(). >> >> Compaction stalls 33052 39853 55091 >> Compaction success 12121 19773 37875 >> Compaction failures 20931 20079 17216 >> >> Compaction is indeed more successful, and thus less likely to get deferred, >> so there are also more direct compaction stalls. >> >> Page migrate success 3781876 3326819 2790838 >> Page migrate failure 45817 41774 38113 >> Compaction pages isolated 7868232 6941457 5025092 >> Compaction migrate scanned 168160492 127269354 87087993 >> Compaction migrate prescanned 0 0 0 >> Compaction free scanned 2522142582 2326342620 743205879 >> Compaction free direct alloc 0 0 920792 >> Compaction free dir. all. miss 0 0 5865 >> Compaction cost 5252 4476 3602 >> >> Patch 2 reduces migration scanned pages by 25% thanks to the eager skipping. >> Patch 3 reduces free scanned pages by 70%. The portion of direct allocation >> misses to all direct allocations is less than 1% which should be acceptable. >> Interestingly, patch 3 also reduces migration scanned pages by another 30% on >> top of patch 2. The reason is not clear, but we can rejoice nevertheless. > > s/Patch 2/Patch 3 > s/Patch 3/Patch 4 Thanks. >> Vlastimil Babka (4): >> mm, compaction: wrap calculating first and last pfn of pageblock >> mm, compaction: reduce spurious pcplist drains >> mm, compaction: skip blocks where isolation fails in async direct >> compaction >> mm, compaction: direct freepage allocation for async direct compaction >> >> include/linux/vm_event_item.h | 1 + >> mm/compaction.c | 189 ++++++++++++++++++++++++++++++++++-------- >> mm/internal.h | 5 ++ >> mm/page_alloc.c | 27 ++++++ >> mm/vmstat.c | 2 + >> 5 files changed, 191 insertions(+), 33 deletions(-) >> >> -- >> 2.7.3 >> >> -- >> To unsubscribe, send a message with 'unsubscribe linux-mm' in >> the body to majordomo@kvack.org. For more info on Linux MM, >> see: http://www.linux-mm.org/ . >> Don't email: email@kvack.org