lkml.org 
[lkml]   [2016]   [Apr]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[RFC PATCH 00/27] Move LRU page reclaim from zones to nodes v3
Date
This is the latest version of a series that moves LRUs from the zones
to the node that is based upon 4.6-rc1. Conceptually, this is simple but
there are a lot of details. Some of the broad motivations for this are;

1. The residency of a page partially depends on what zone the page was
allocated from. This is partially combatted by the fair zone allocation
policy but that is a partial solution that introduces overhead in the
page allocator paths.

2. Currently, reclaim on node 0 behaves slightly different to node 1. For
example, direct reclaim scans in zonelist order and reclaims even if
the zone is over the high watermark regardless of the age of pages
in that LRU. Kswapd on the other hand starts reclaim on the highest
unbalanced zone. A difference in distribution of file/anon pages due
to when they were allocated results can result in a difference in
again. While the fair zone allocation policy mitigates some of the
problems here, the page reclaim results on a multi-zone node will
always be different to a single-zone node.
it was scheduled on as a result.

3. kswapd and the page allocator scan zones in the opposite order to
avoid interfering with each other but it's sensitive to timing. This
mitigates the page allocator using pages that were allocated very recently
in the ideal case but it's sensitive to timing. When kswapd is allocating
from lower zones then it's great but during the rebalancing of the highest
zone, the page allocator and kswapd interfere with each other. It's worse
if the highest zone is small and difficult to balance.

4. slab shrinkers are node-based which makes it harder to identify the exact
relationship between slab reclaim and LRU reclaim.

The reason we have zone-based reclaim is that we used to have
large highmem zones in common configurations and it was necessary
to quickly find ZONE_NORMAL pages for reclaim. Today, this is much
less of a concern as machines with lots of memory will (or should) use
64-bit kernels. Combinations of 32-bit hardware and 64-bit hardware are
rare. Machines that do use highmem should have relatively low highmem:lowmem
ratios than we worried about in the past.

Conceptually, moving to node LRUs should be easier to understand. The
page allocator plays fewer tricks to game reclaim and reclaim behaves
similarly on all nodes.

It was tested on a UMA (16 cores single socket) and a NUMA machine (48
cores, 2 sockets). In most cases, only the UMA results are presented as
the NUMA machine takes an excessive amount of time to complete tests.

There may be an obvious difference in the number of
allocations from each zone as the fair zone allocation policy is removed
towards the end of the series. In cases where the working set exceeds memory,
the differences will be small but on small workloads it'll be very obvious.
For example, these are the allocation stats on a workload that is doing small
amounts of dd.

4.6.0-rc1 4.6.0-rc1
vanilla nodelru-v3
DMA allocs 0 0
DMA32 allocs 1961196 0
Normal allocs 3355799 5247180
Movable allocs 0 0

The key reason why this is not a problem is that kswapd will sleep if any
applicable zone for a classzone is free. If it tried to balance all zones
then there would be excessive reclaim.

bonnie
------

This was configured to do an IO test with a working set 2*RAM using the
ext4 filesystem. For both machines, there was no significant performance
difference between them but this is the result for the UMA machine

bonnie
bonnie
4.6.0-rc1 4.6.0-rc1
vanilla nodelru-v3r10
Hmean SeqOut Char 53306.32 ( 0.00%) 79027.86 ( 48.25%)
Hmean SeqOut Block 87796.15 ( 0.00%) 87881.69 ( 0.10%)
Hmean SeqOut Rewrite 35996.31 ( 0.00%) 36355.59 ( 1.00%)
Hmean SeqIn Char 38789.17 ( 0.00%) 76356.20 ( 96.85%)
Hmean SeqIn Block 105315.39 ( 0.00%) 105514.07 ( 0.19%)
Hmean Random seeks 329.80 ( 0.00%) 334.36 ( 1.38%)
Hmean SeqCreate ops 4.62 ( 0.00%) 4.62 ( 0.00%)
Hmean SeqCreate read 4.62 ( 0.00%) 4.62 ( 0.00%)
Hmean SeqCreate del 599.29 ( 0.00%) 1580.23 (163.68%)
Hmean RandCreate ops 5.00 ( 0.00%) 5.00 ( 0.00%)
Hmean RandCreate read 5.00 ( 0.00%) 4.62 ( -7.69%)
Hmean RandCreate del 629.51 ( 0.00%) 1634.55 (159.66%)

4.6.0-rc1 4.6.0-rc1
vanillanodelru-v3r10
User 2049.02 1078.82
System 294.25 181.00
Elapsed 6960.58 6021.58

Note that the massive gains shown here are possible an anomaly. It has been noted
that in some cases, bonnie gets an artifical boost due to dumb reclaim luck. There
is no guarantee this result would be reproducible on the same machine let alone
any other machine. That said, the VM stats are interesting;

However, the overall VM stats are interesting


4.5.0-rc3 4.5.0-rc3
mmotm-20160209 nodelru-v2
Swap Ins 14 0
Swap Outs 873 0
DMA allocs 0 0
DMA32 allocs 38259888 36320496
Normal allocs 64762073 66488556
Movable allocs 0 0
Allocation stalls 3584 0
Direct pages scanned 736769 0
Kswapd pages scanned 77818637 78836064
Kswapd pages reclaimed 77782378 78812260
Direct pages reclaimed 736548 0
Kswapd efficiency 99% 99%
Kswapd velocity 11179.907 13092.256
Direct efficiency 99% 100%
Direct velocity 105.849 0.000

The series does not swap the workload and it never stalls on direct reclaim. There
is a slight increase in kswapd scans but it's offset by the elimination of direct
scans and the overall scanning velocity is not noticably higher. While it's not
reported here, the overall IO stats and CPU usage over time are very similar. kswapd
CPU usage is slightly elevated but (0.5% usage to roughly 1.2% usage over time) but
that is acceptable given the lack of direct reclaim.

tiobench
--------

tiobench is a flawed benchmark but it's very important in this case. tiobench
benefited from a bug prior to the fair zone allocation policy that allowed
old pages to be artificially preserved. The visible impact was that performance
exceeded the physical capabilities of the disk. With this patch applied the results are

tiobench Throughput
tiobench Throughput
4.6.0-rc1 4.6.0-rc1
vanilla nodelru-v3
Hmean PotentialReadSpeed 85.84 ( 0.00%) 86.20 ( 0.42%)
Hmean SeqRead-MB/sec-1 84.48 ( 0.00%) 84.60 ( 0.14%)
Hmean SeqRead-MB/sec-2 75.69 ( 0.00%) 75.44 ( -0.34%)
Hmean SeqRead-MB/sec-4 77.35 ( 0.00%) 77.62 ( 0.35%)
Hmean SeqRead-MB/sec-8 68.29 ( 0.00%) 68.58 ( 0.43%)
Hmean SeqRead-MB/sec-16 62.82 ( 0.00%) 62.72 ( -0.15%)
Hmean RandRead-MB/sec-1 0.93 ( 0.00%) 0.88 ( -4.69%)
Hmean RandRead-MB/sec-2 1.11 ( 0.00%) 1.08 ( -3.20%)
Hmean RandRead-MB/sec-4 1.52 ( 0.00%) 1.48 ( -2.86%)
Hmean RandRead-MB/sec-8 1.70 ( 0.00%) 1.70 ( -0.26%)
Hmean RandRead-MB/sec-16 1.96 ( 0.00%) 1.91 ( -2.49%)
Hmean SeqWrite-MB/sec-1 83.01 ( 0.00%) 83.07 ( 0.07%)
Hmean SeqWrite-MB/sec-2 77.80 ( 0.00%) 78.20 ( 0.52%)
Hmean SeqWrite-MB/sec-4 81.68 ( 0.00%) 81.72 ( 0.05%)
Hmean SeqWrite-MB/sec-8 78.17 ( 0.00%) 78.41 ( 0.31%)
Hmean SeqWrite-MB/sec-16 80.08 ( 0.00%) 80.08 ( 0.01%)
Hmean RandWrite-MB/sec-1 1.17 ( 0.00%) 1.17 ( -0.03%)
Hmean RandWrite-MB/sec-2 1.02 ( 0.00%) 1.06 ( 4.21%)
Hmean RandWrite-MB/sec-4 1.02 ( 0.00%) 1.04 ( 2.32%)
Hmean RandWrite-MB/sec-8 0.95 ( 0.00%) 0.97 ( 1.75%)
Hmean RandWrite-MB/sec-16 0.95 ( 0.00%) 0.96 ( 0.97%)

Note that the performance is almost identical allowing us to conclude that
the correct reclaim behaviour granted by the fair zone allocation policy
is preserved.

stutter
-------

stutter simulates a simple workload. One part uses a lot of anonymous
memory, a second measures mmap latency and a third copies a large file.
The primary metric is checking for mmap latency.

stutter
4.6.0-rc1 4.6.0-rc1
vanilla nodelru-v3
Min mmap 13.4442 ( 0.00%) 13.6705 ( -1.68%)
1st-qrtle mmap 38.0442 ( 0.00%) 37.7842 ( 0.68%)
2nd-qrtle mmap 78.5109 ( 0.00%) 40.3648 ( 48.59%)
3rd-qrtle mmap 86.7806 ( 0.00%) 46.2499 ( 46.70%)
Max-90% mmap 89.7028 ( 0.00%) 86.5790 ( 3.48%)
Max-93% mmap 90.6776 ( 0.00%) 89.5367 ( 1.26%)
Max-95% mmap 91.1678 ( 0.00%) 90.3138 ( 0.94%)
Max-99% mmap 92.0036 ( 0.00%) 93.2003 ( -1.30%)
Max mmap 167.0073 ( 0.00%) 94.5935 ( 43.36%)
Mean mmap 68.7672 ( 0.00%) 48.9853 ( 28.77%)
Best99%Mean mmap 68.5246 ( 0.00%) 48.5354 ( 29.17%)
Best95%Mean mmap 67.5540 ( 0.00%) 46.7102 ( 30.86%)
Best90%Mean mmap 66.2798 ( 0.00%) 44.3547 ( 33.08%)
Best50%Mean mmap 50.7730 ( 0.00%) 37.1298 ( 26.87%)
Best10%Mean mmap 35.8311 ( 0.00%) 33.6910 ( 5.97%)
Best5%Mean mmap 34.0159 ( 0.00%) 31.4259 ( 7.61%)
Best1%Mean mmap 22.1306 ( 0.00%) 24.8851 (-12.45%)

4.6.0-rc1 4.6.0-rc1
vanillanodelru-v3r10
User 1.51 0.97
System 138.03 122.58
Elapsed 2420.90 2394.80

The VM stats in this case were not that intresting and are very roughly comparable.

Page allocator intensive workloads showed few differences as the cost
of the fair zone allocation policy does not dominate from a userspace
perspective but a microbench of just the allocator shows a difference

4.6.0-rc1 4.6.0-rc1
vanilla nodelru-v3
Min total-odr0-1 725.00 ( 0.00%) 697.00 ( 3.86%)
Min total-odr0-2 559.00 ( 0.00%) 527.00 ( 5.72%)
Min total-odr0-4 459.00 ( 0.00%) 436.00 ( 5.01%)
Min total-odr0-8 403.00 ( 0.00%) 391.00 ( 2.98%)
Min total-odr0-16 329.00 ( 0.00%) 366.00 (-11.25%)
Min total-odr0-32 365.00 ( 0.00%) 355.00 ( 2.74%)
Min total-odr0-64 297.00 ( 0.00%) 348.00 (-17.17%)
Min total-odr0-128 752.00 ( 0.00%) 344.00 ( 54.26%)
Min total-odr0-256 385.00 ( 0.00%) 379.00 ( 1.56%)
Min total-odr0-512 899.00 ( 0.00%) 414.00 ( 53.95%)
Min total-odr0-1024 763.00 ( 0.00%) 530.00 ( 30.54%)
Min total-odr0-2048 982.00 ( 0.00%) 469.00 ( 52.24%)
Min total-odr0-4096 928.00 ( 0.00%) 526.00 ( 43.32%)
Min total-odr0-8192 1007.00 ( 0.00%) 768.00 ( 23.73%)
Min total-odr0-16384 375.00 ( 0.00%) 366.00 ( 2.40%)

This series is not without its hazards. There are at least three areas
that I'm concerned with even though I could not reproduce any problems in
that area.

1. Reclaim/compaction is going to be affected because the amount of reclaim is
no longer targetted at a specific zone. Compaction works on a per-zone basis
so there is no guarantee that reclaiming a few THP's worth page pages will
have a positive impact on compaction success rates.

2. The Slab/LRU reclaim ratio is affected because the frequency the shrinkers
are called is now different. This may or may not be a problem but if it
is, it'll be because shrinkers are not called enough and some balancing
is required.

3. The anon/file reclaim ratio may be affected. Pages about to be dirtied are
distributed between zones and the fair zone allocation policy used to do
something very similar for anon. The distribution is now different but not
necessarily in any way that matters but it's still worth bearing in mind.

Documentation/cgroup-v1/memcg_test.txt | 4 +-
Documentation/cgroup-v1/memory.txt | 4 +-
arch/s390/appldata/appldata_mem.c | 2 +-
arch/tile/mm/pgtable.c | 18 +-
drivers/base/node.c | 73 +--
drivers/staging/android/lowmemorykiller.c | 12 +-
fs/fs-writeback.c | 4 +-
fs/fuse/file.c | 8 +-
fs/nfs/internal.h | 2 +-
fs/nfs/write.c | 2 +-
fs/proc/meminfo.c | 14 +-
include/linux/backing-dev.h | 2 +-
include/linux/memcontrol.h | 30 +-
include/linux/mm_inline.h | 4 +-
include/linux/mm_types.h | 2 +-
include/linux/mmzone.h | 156 +++---
include/linux/swap.h | 13 +-
include/linux/topology.h | 2 +-
include/linux/vm_event_item.h | 14 +-
include/linux/vmstat.h | 112 +++-
include/linux/writeback.h | 2 +-
include/trace/events/vmscan.h | 40 +-
include/trace/events/writeback.h | 10 +-
kernel/power/snapshot.c | 10 +-
kernel/sysctl.c | 4 +-
mm/backing-dev.c | 14 +-
mm/compaction.c | 24 +-
mm/filemap.c | 14 +-
mm/huge_memory.c | 14 +-
mm/internal.h | 11 +-
mm/memcontrol.c | 235 ++++-----
mm/memory-failure.c | 4 +-
mm/memory_hotplug.c | 7 +-
mm/mempolicy.c | 2 +-
mm/migrate.c | 35 +-
mm/mlock.c | 12 +-
mm/page-writeback.c | 119 ++---
mm/page_alloc.c | 293 +++++------
mm/page_idle.c | 4 +-
mm/rmap.c | 15 +-
mm/shmem.c | 12 +-
mm/swap.c | 66 +--
mm/swap_state.c | 4 +-
mm/util.c | 4 +-
mm/vmscan.c | 838 ++++++++++++++----------------
mm/vmstat.c | 367 ++++++++++---
mm/workingset.c | 53 +-
47 files changed, 1473 insertions(+), 1218 deletions(-)

--
2.6.4

\
 
 \ /
  Last update: 2016-04-06 13:41    [W:0.108 / U:22.312 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site