Patch series "Reduce amount of time kswapd sleeps prematurely", v2.
The series is unusual in that the first patch fixes one problem and
introduces other issues that are noted in the changelog. Patch 2 makes
a minor modification that is worth considering on its own but leaves the
kernel in a state where it behaves badly. It's not until patch 3 that
there is an improvement against baseline.
This was mostly motivated by examining Chris Mason's "simoop" benchmark
which puts the VM under similar pressure to HADOOP. It has been
reported that the benchmark has regressed severely during the last
number of releases. While I cannot reproduce all the same problems
Chris experienced due to hardware limitations, there was a number of
problems on a 2-socket machine with a single disk.
simoop latencies
4.11.0-rc1 4.11.0-rc1
vanilla keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 125765.57 ( 49.08%)
These are latencies. Read/write are threads reading fixed-size random
blocks from a simulated database. The allocation latency is mmaping and
faulting regions of memory. The p50, 95 and p99 reports the worst
latencies for 50% of the samples, 95% and 99% respectively.
For example, the report indicates that while the test was running 99% of
writes completed 99.75% faster. It's worth noting that on a UMA machine
that no difference in performance with simoop was observed so milage
will vary.
It's noted that there is a slight impact to read latencies but it's
mostly due to IO scheduler decisions and offset by the large reduction
in other latencies.
This patch (of 3):
The check in prepare_kswapd_sleep needs to match the one in
balance_pgdat since the latter will return as soon as any one of the
zones in the classzone is above the watermark. This is specially
important for higher order allocations since balance_pgdat will
typically reset the order to zero relying on compaction to create the
higher order pages. Without this patch, prepare_kswapd_sleep fails to
wake up kcompactd since the zone balance check fails.
It was first reported against 4.9.7 that kswapd is failing to wake up
kcompactd due to a mismatch in the zone balance check between
balance_pgdat() and prepare_kswapd_sleep().
balance_pgdat() returns as soon as a single zone satisfies the
allocation but prepare_kswapd_sleep() requires all zones to do +the
same. This causes prepare_kswapd_sleep() to never succeed except in the
order == 0 case and consequently, wakeup_kcompactd() is never called.
For the machine that originally motivated this patch, the state of
compaction from /proc/vmstat looked this way after a day and a half +of
uptime:
compact_migrate_scanned 240496
compact_free_scanned 76238632
compact_isolated 123472
compact_stall 1791
compact_fail 29
compact_success 1762
compact_daemon_wake 0
After applying the patch and about 10 hours of uptime the state looks
like this:
compact_migrate_scanned 59927299
compact_free_scanned 2021075136
compact_isolated 640926
compact_stall 4
compact_fail 2
compact_success 2
compact_daemon_wake 5160
Further notes from Mel that motivated him to pick this patch up and
resend it;
It was observed for the simoop workload (pressures the VM similar to
HADOOP) that kswapd was failing to keep ahead of direct reclaim. The
investigation noted that there was a need to rationalise kswapd
decisions to reclaim with kswapd decisions to sleep. With this patch on
a 2-socket box, there was a 49% reduction in direct reclaim scanning.
However, the impact otherwise is extremely negative. Kswapd reclaim
efficiency dropped from 98% to 76%. simoop has three latency-related
metrics for read, write and allocation (an anonymous mmap and fault).
4.11.0-rc1 4.11.0-rc1
vanilla fixcheck-v2
Amean p50-Read 21670074.18 ( 0.00%) 20464344.18 ( 5.56%)
Amean p95-Read 25456267.64 ( 0.00%) 25721423.64 ( -1.04%)
Amean p99-Read 29369064.73 ( 0.00%) 30174230.76 ( -2.74%)
Amean p50-Write 1390.30 ( 0.00%) 1395.28 ( -0.36%)
Amean p95-Write 412901.57 ( 0.00%) 37737.74 ( 90.86%)
Amean p99-Write 6668722.09 ( 0.00%) 666489.04 ( 90.01%)
Amean p50-Allocation 78714.31 ( 0.00%) 86286.22 ( -9.62%)
Amean p95-Allocation 175533.51 ( 0.00%) 351812.27 (-100.42%)
Amean p99-Allocation 247003.02 ( 0.00%) 6291171.56 (-2447.00%)
Of greater concern is that the patch causes swapping and page writes
from kswapd context rose from 0 pages to 4189753 pages during the hour
the workload ran for. By and large, the patch has very bad behaviour
but easily missed as the impact on a UMA machine is negligible.
This patch is included with the data in case a bisection leads to this
area. This patch is also a pre-requisite for the rest of the series.
Link: http://lkml.kernel.org/r/20170309075657.25121-2-mgorman@techsingularity.net
Signed-off-by: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
333b0a459c