aboutsummaryrefslogtreecommitdiff
path: root/include
diff options
context:
space:
mode:
authorJohannes Weiner2017-02-24 14:56:11 -0800
committerLinus Torvalds2017-02-24 17:46:54 -0800
commit1276ad68e2491d1ceeb65f55d790f9277593c459 (patch)
treed39ac1de4a2d61258e82dccc07bc8aba2b99771e /include
parent64527f5d540ad496718c7bca5e9387cf6cf94e8c (diff)
mm: vmscan: scan dirty pages even in laptop mode
Patch series "mm: vmscan: fix kswapd writeback regression". We noticed a regression on multiple hadoop workloads when moving from 3.10 to 4.0 and 4.6, which involves kswapd getting tangled up in page writeout, causing direct reclaim herds that also don't make progress. I tracked it down to the thrash avoidance efforts after 3.10 that make the kernel better at keeping use-once cache and use-many cache sorted on the inactive and active list, with more aggressive protection of the active list as long as there is inactive cache. Unfortunately, our workload's use-once cache is mostly from streaming writes. Waiting for writes to avoid potential reloads in the future is not a good tradeoff. These patches do the following: 1. Wake the flushers when kswapd sees a lump of dirty pages. It's possible to be below the dirty background limit and still have cache velocity push them through the LRU. So start a-flushin'. 2. Let kswapd only write pages that have been rotated twice. This makes sure we really tried to get all the clean pages on the inactive list before resorting to horrible LRU-order writeback. 3. Move rotating dirty pages off the inactive list. Instead of churning or waiting on page writeback, we'll go after clean active cache. This might lead to thrashing, but in this state memory demand outstrips IO speed anyway, and reads are faster than writes. Mel backported the series to 4.10-rc5 with one minor conflict and ran a couple of tests on it. Mix of read/write random workload didn't show anything interesting. Write-only database didn't show much difference in performance but there were slight reductions in IO -- probably in the noise. simoop did show big differences although not as big as Mel expected. This is Chris Mason's workload that similate the VM activity of hadoop. Mel won't go through the full details but over the samples measured during an hour it reported 4.10.0-rc5 4.10.0-rc5 vanilla johannes-v1r1 Amean p50-Read 21346531.56 ( 0.00%) 21697513.24 ( -1.64%) Amean p95-Read 24700518.40 ( 0.00%) 25743268.98 ( -4.22%) Amean p99-Read 27959842.13 ( 0.00%) 28963271.11 ( -3.59%) Amean p50-Write 1138.04 ( 0.00%) 989.82 ( 13.02%) Amean p95-Write 1106643.48 ( 0.00%) 12104.00 ( 98.91%) Amean p99-Write 1569213.22 ( 0.00%) 36343.38 ( 97.68%) Amean p50-Allocation 85159.82 ( 0.00%) 79120.70 ( 7.09%) Amean p95-Allocation 204222.58 ( 0.00%) 129018.43 ( 36.82%) Amean p99-Allocation 278070.04 ( 0.00%) 183354.43 ( 34.06%) Amean final-p50-Read 21266432.00 ( 0.00%) 21921792.00 ( -3.08%) Amean final-p95-Read 24870912.00 ( 0.00%) 26116096.00 ( -5.01%) Amean final-p99-Read 28147712.00 ( 0.00%) 29523968.00 ( -4.89%) Amean final-p50-Write 1130.00 ( 0.00%) 977.00 ( 13.54%) Amean final-p95-Write 1033216.00 ( 0.00%) 2980.00 ( 99.71%) Amean final-p99-Write 1517568.00 ( 0.00%) 32672.00 ( 97.85%) Amean final-p50-Allocation 86656.00 ( 0.00%) 78464.00 ( 9.45%) Amean final-p95-Allocation 211712.00 ( 0.00%) 116608.00 ( 44.92%) Amean final-p99-Allocation 287232.00 ( 0.00%) 168704.00 ( 41.27%) The latencies are actually completely horrific in comparison to 4.4 (and 4.10-rc5 is worse than 4.9 according to historical data for reasons Mel hasn't analysed yet). Still, 95% of write latency (p95-write) is halved by the series and allocation latency is way down. Direct reclaim activity is one fifth of what it was according to vmstats. Kswapd activity is higher but this is not necessarily surprising. Kswapd efficiency is unchanged at 99% (99% of pages scanned were reclaimed) but direct reclaim efficiency went from 77% to 99% In the vanilla kernel, 627MB of data was written back from reclaim context. With the series, no data was written back. With or without the patch, pages are being immediately reclaimed after writeback completes. However, with the patch, only 1/8th of the pages are reclaimed like this. This patch (of 5): We have an elaborate dirty/writeback throttling mechanism inside the reclaim scanner, but for that to work the pages have to go through shrink_page_list() and get counted for what they are. Otherwise, we mess up the LRU order and don't match reclaim speed to writeback. Especially during deactivation, there is never a reason to skip dirty pages; nothing is even trying to write them out from there. Don't mess up the LRU order for nothing, shuffle these pages along. Link: http://lkml.kernel.org/r/20170123181641.23938-2-hannes@cmpxchg.org Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Minchan Kim <minchan@kernel.org> Acked-by: Michal Hocko <mhocko@suse.com> Acked-by: Mel Gorman <mgorman@suse.de> Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'include')
-rw-r--r--include/linux/mmzone.h2
1 files changed, 0 insertions, 2 deletions
diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 82fc632fd11d..8e02b3750fe0 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -236,8 +236,6 @@ struct lruvec {
#define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
#define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
-/* Isolate clean file */
-#define ISOLATE_CLEAN ((__force isolate_mode_t)0x1)
/* Isolate unmapped file */
#define ISOLATE_UNMAPPED ((__force isolate_mode_t)0x2)
/* Isolate for asynchronous migration */