linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [PATCH 00/27] Move LRU page reclaim from zones to nodes v7
@ 2016-06-21 11:43 Mel Gorman
  2016-06-21 11:43 ` [PATCH 01/27] mm, vmstat: Add infrastructure for per-node vmstats Mel Gorman
                   ` (2 more replies)
  0 siblings, 3 replies; 11+ messages in thread
From: Mel Gorman @ 2016-06-21 11:43 UTC (permalink / raw)
  To: Andrew Morton, Linux-MM
  Cc: Rik van Riel, Vlastimil Babka, Johannes Weiner, LKML, Mel Gorman

The bulk of the updates are in response to review from Vlastimil Babka
and received a lot more testing than v6.

Changelog since v6
o Correct reclaim_idx when direct reclaiming for memcg
o Also account LRU pages per zone for compaction/reclaim
o Add page_pgdat helper with more efficient lookup
o Init pgdat LRU lock only once
o Slight optimisation to wake_all_kswapds
o Always wake kcompactd when kswapd is going to sleep
o Rebase to mmotm as of June 15th, 2016

Changelog since v5
o Rebase and adjust to changes

Changelog since v4
o Rebase on top of v3 of page allocator optimisation series

Changelog since v3
o Rebase on top of the page allocator optimisation series
o Remove RFC tag

This is the latest version of a series that moves LRUs from the zones to
the node that is based upon 4.6-rc3 plus the page allocator optimisation
series. 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. 

The series has been tested on a 16 core UMA machine and a 2-socket 48 core
NUMA machine. The UMA results are presented in most cases as the NUMA machine
behaved similarly.

pagealloc
---------

This is a microbenchmark that shows the benefit of removing the fair zone
allocation policy. It was tested uip to order-4 but only orders 0 and 1 are
shown as the other orders were comparable.

                                           4.7.0-rc3                  4.7.0-rc3
                                      mmotm-20160615              nodelru-v7r17
Min      total-odr0-1               485.00 (  0.00%)           462.00 (  4.74%)
Min      total-odr0-2               354.00 (  0.00%)           341.00 (  3.67%)
Min      total-odr0-4               285.00 (  0.00%)           277.00 (  2.81%)
Min      total-odr0-8               249.00 (  0.00%)           240.00 (  3.61%)
Min      total-odr0-16              230.00 (  0.00%)           224.00 (  2.61%)
Min      total-odr0-32              222.00 (  0.00%)           215.00 (  3.15%)
Min      total-odr0-64              216.00 (  0.00%)           210.00 (  2.78%)
Min      total-odr0-128             214.00 (  0.00%)           208.00 (  2.80%)
Min      total-odr0-256             248.00 (  0.00%)           233.00 (  6.05%)
Min      total-odr0-512             277.00 (  0.00%)           270.00 (  2.53%)
Min      total-odr0-1024            294.00 (  0.00%)           284.00 (  3.40%)
Min      total-odr0-2048            308.00 (  0.00%)           298.00 (  3.25%)
Min      total-odr0-4096            318.00 (  0.00%)           307.00 (  3.46%)
Min      total-odr0-8192            322.00 (  0.00%)           308.00 (  4.35%)
Min      total-odr0-16384           324.00 (  0.00%)           309.00 (  4.63%)
Min      total-odr1-1               729.00 (  0.00%)           686.00 (  5.90%)
Min      total-odr1-2               533.00 (  0.00%)           520.00 (  2.44%)
Min      total-odr1-4               434.00 (  0.00%)           415.00 (  4.38%)
Min      total-odr1-8               390.00 (  0.00%)           364.00 (  6.67%)
Min      total-odr1-16              359.00 (  0.00%)           335.00 (  6.69%)
Min      total-odr1-32              356.00 (  0.00%)           327.00 (  8.15%)
Min      total-odr1-64              356.00 (  0.00%)           321.00 (  9.83%)
Min      total-odr1-128             356.00 (  0.00%)           333.00 (  6.46%)
Min      total-odr1-256             354.00 (  0.00%)           337.00 (  4.80%)
Min      total-odr1-512             366.00 (  0.00%)           340.00 (  7.10%)
Min      total-odr1-1024            373.00 (  0.00%)           354.00 (  5.09%)
Min      total-odr1-2048            375.00 (  0.00%)           354.00 (  5.60%)
Min      total-odr1-4096            374.00 (  0.00%)           354.00 (  5.35%)
Min      total-odr1-8192            370.00 (  0.00%)           355.00 (  4.05%)

This shows a steady improvement throughout. The primary benefit is from
reduced system CPU usage which is obvious from the overall times;

           4.7.0-rc3   4.7.0-rc3
        mmotm-20160615 nodelru-v7
User          174.06      174.58
System       2656.78     2485.21
Elapsed      2885.07     2713.67

The vmstats also showed that the fair zone allocation policy was definitely
removed as can be seen here;

                             4.7.0-rc3   4.7.0-rc3
                          mmotm-20160615nodelru-v7r17
DMA32 allocs               28794408561           0
Normal allocs              48431969998 77226313470
Movable allocs                       0           0

tiobench on ext4
----------------

tiobench is a benchmark that artifically benefits if old pages remain resident
while new pages get reclaimed. The fair zone allocation policy mitigates this
problem so pages age fairly. While the benchmark has problems, it is important
that tiobench performance remains constant as it implies that page aging
problems that the fair zone allocation policy fixes are not re-introduced.

                                         4.7.0-rc3             4.7.0-rc3
                                    mmotm-20160615         nodelru-v7r17
Min      PotentialReadSpeed        90.24 (  0.00%)       90.14 ( -0.11%)
Min      SeqRead-MB/sec-1          80.63 (  0.00%)       83.09 (  3.05%)
Min      SeqRead-MB/sec-2          71.91 (  0.00%)       72.44 (  0.74%)
Min      SeqRead-MB/sec-4          75.20 (  0.00%)       74.32 ( -1.17%)
Min      SeqRead-MB/sec-8          65.30 (  0.00%)       65.21 ( -0.14%)
Min      SeqRead-MB/sec-16         62.62 (  0.00%)       62.12 ( -0.80%)
Min      RandRead-MB/sec-1          0.90 (  0.00%)        0.94 (  4.44%)
Min      RandRead-MB/sec-2          0.96 (  0.00%)        0.97 (  1.04%)
Min      RandRead-MB/sec-4          1.43 (  0.00%)        1.41 ( -1.40%)
Min      RandRead-MB/sec-8          1.67 (  0.00%)        1.72 (  2.99%)
Min      RandRead-MB/sec-16         1.77 (  0.00%)        1.86 (  5.08%)
Min      SeqWrite-MB/sec-1         78.12 (  0.00%)       79.78 (  2.12%)
Min      SeqWrite-MB/sec-2         72.74 (  0.00%)       73.23 (  0.67%)
Min      SeqWrite-MB/sec-4         79.40 (  0.00%)       78.32 ( -1.36%)
Min      SeqWrite-MB/sec-8         73.18 (  0.00%)       71.40 ( -2.43%)
Min      SeqWrite-MB/sec-16        75.82 (  0.00%)       75.24 ( -0.76%)
Min      RandWrite-MB/sec-1         1.18 (  0.00%)        1.15 ( -2.54%)
Min      RandWrite-MB/sec-2         1.05 (  0.00%)        0.99 ( -5.71%)
Min      RandWrite-MB/sec-4         1.00 (  0.00%)        0.96 ( -4.00%)
Min      RandWrite-MB/sec-8         0.91 (  0.00%)        0.92 (  1.10%)
Min      RandWrite-MB/sec-16        0.92 (  0.00%)        0.92 (  0.00%)

This shows that the series has little or not impact on tiobench which is
desirable. It indicates that the fair zone allocation policy was removed
in a manner that didn't reintroduce one class of page aging bug. There
were only minor differences in overall reclaim activity

                             4.7.0-rc3   4.7.0-rc3
                          mmotm-20160615nodelru-v7r17
Minor Faults                    640992      640721
Major Faults                       728         623
Swap Ins                             0           0
Swap Outs                            0           0
DMA allocs                           0           0
DMA32 allocs                  46174282    44219717
Normal allocs                 77949344    79858024
Movable allocs                       0           0
Allocation stalls                   38          76
Direct pages scanned             17463       34865
Kswapd pages scanned          93331163    93302388
Kswapd pages reclaimed        93328173    93299677
Direct pages reclaimed           17463       34865
Kswapd efficiency                  99%         99%
Kswapd velocity              13729.855   13755.612
Direct efficiency                 100%        100%
Direct velocity                  2.569       5.140
Percentage direct scans             0%          0%
Page writes by reclaim               0           0
Page writes file                     0           0
Page writes anon                     0           0
Page reclaim immediate              54          36

kswapd activity was roughly comparable. There was slight differences
in direct reclaim activity but negligible in the context of the overall
workload (velocity of 5 pages per second with the patches applied, 2 pages
per second in the baseline kernel).

pgbench read-only large configuration on ext4
---------------------------------------------

pgbench is a database benchmark that can be sensitive to page reclaim
decisions. This also checks if removing the fair zone allocation policy
is safe

pgbench Transactions
                        4.7.0-rc3             4.7.0-rc3
                   mmotm-20160615         nodelru-v7r17
Hmean    1       191.00 (  0.00%)      193.67 (  1.40%)
Hmean    5       338.59 (  0.00%)      336.99 ( -0.47%)
Hmean    12      374.03 (  0.00%)      386.15 (  3.24%)
Hmean    21      372.24 (  0.00%)      372.02 ( -0.06%)
Hmean    30      383.98 (  0.00%)      370.69 ( -3.46%)
Hmean    32      431.01 (  0.00%)      438.47 (  1.73%)

Negligible differences again. As with tiobench, overall reclaim activity
was comparable.

bonnie++ on ext4
----------------

No interesting performance difference, negligible differences on reclaim
stats.

paralleldd on ext4
------------------

This workload uses varying numbers of dd instances to read large amounts of
data from disk.

paralleldd
                               4.7.0-rc3             4.7.0-rc3
                          mmotm-20160615         nodelru-v7r17
Amean    Elapsd-1       181.57 (  0.00%)      179.63 (  1.07%)
Amean    Elapsd-3       188.29 (  0.00%)      183.68 (  2.45%)
Amean    Elapsd-5       188.02 (  0.00%)      181.73 (  3.35%)
Amean    Elapsd-7       186.07 (  0.00%)      184.11 (  1.05%)
Amean    Elapsd-12      188.16 (  0.00%)      183.51 (  2.47%)
Amean    Elapsd-16      189.03 (  0.00%)      181.27 (  4.10%)

           4.7.0-rc3   4.7.0-rc3
        mmotm-20160615nodelru-v7r17
User         1439.23     1433.37
System       8332.31     8216.01
Elapsed      3619.80     3532.69

There is a slight gain in performance, some of which is from the reduced system
CPU usage. There areminor differences in reclaim activity but nothing significant

                             4.7.0-rc3   4.7.0-rc3
                          mmotm-20160615nodelru-v7r17
Minor Faults                    362486      358215
Major Faults                      1143        1113
Swap Ins                            26           0
Swap Outs                         2920         482
DMA allocs                           0           0
DMA32 allocs                  31568814    28598887
Normal allocs                 46539922    49514444
Movable allocs                       0           0
Allocation stalls                    0           0
Direct pages scanned                 0           0
Kswapd pages scanned          40886878    40849710
Kswapd pages reclaimed        40869923    40835207
Direct pages reclaimed               0           0
Kswapd efficiency                  99%         99%
Kswapd velocity              11295.342   11563.344
Direct efficiency                 100%        100%
Direct velocity                  0.000       0.000
Slabs scanned                   131673      126099
Direct inode steals                 57          60
Kswapd inode steals                762          18

It basically shows that kswapd was active at roughly the same rate in
both kernels. There was also comparable slab scanning activity and direct
reclaim was avoided in both cases. There appears to be a large difference
in numbers of inodes reclaimed but the workload has few active inodes and
is likely a timing artifact. It's interesting to note that the node-lru
did not swap in any pages but given the low swap activity, it's unlikely
to be significant.

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
                             4.7.0-rc3             4.7.0-rc3
                        mmotm-20160615         nodelru-v7r17
Min         mmap     16.8422 (  0.00%)     15.9821 (  5.11%)
1st-qrtle   mmap     57.8709 (  0.00%)     58.0794 ( -0.36%)
2nd-qrtle   mmap     58.4335 (  0.00%)     59.4286 ( -1.70%)
3rd-qrtle   mmap     58.6957 (  0.00%)     59.6862 ( -1.69%)
Max-90%     mmap     58.9388 (  0.00%)     59.8759 ( -1.59%)
Max-93%     mmap     59.0505 (  0.00%)     59.9333 ( -1.50%)
Max-95%     mmap     59.1877 (  0.00%)     59.9844 ( -1.35%)
Max-99%     mmap     60.3237 (  0.00%)     60.2337 (  0.15%)
Max         mmap    285.6454 (  0.00%)    135.6006 ( 52.53%)
Mean        mmap     57.8366 (  0.00%)     58.4884 ( -1.13%)

This shows that there is a slight impact on mmap latency but that
the worst-case outlier is much improved. As the problem with this
benchmark used to be that the kernel stalled for minutes, this
difference is negligible.

Some of the vmstats are interesting

                             4.7.0-rc3   4.7.0-rc3
                          mmotm-20160615nodelru-v7r17
Swap Ins                            58          42
Swap Outs                            0           0
Allocation stalls                   16           0
Direct pages scanned              1374           0
Kswapd pages scanned          42454910    41782544
Kswapd pages reclaimed        41571035    41781833
Direct pages reclaimed            1167           0
Kswapd efficiency                  97%         99%
Kswapd velocity              14774.479   14223.796
Direct efficiency                  84%        100%
Direct velocity                  0.478       0.000
Percentage direct scans             0%          0%
Page writes by reclaim          696918           0
Page writes file                696918           0
Page writes anon                     0           0
Page reclaim immediate            2940         137
Sector Reads                  81644424    81699544
Sector Writes                 99193620    98862160
Page rescued immediate               0           0
Slabs scanned                  1279838       22640

kswapd and direct reclaim activity are similar but the node LRU series
did not attempt to trigger any page writes from reclaim context.

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                |  32 +-
 include/linux/mm.h                        |   5 +
 include/linux/mm_inline.h                 |  21 +-
 include/linux/mm_types.h                  |   2 +-
 include/linux/mmzone.h                    | 158 +++---
 include/linux/swap.h                      |  23 +-
 include/linux/topology.h                  |   2 +-
 include/linux/vm_event_item.h             |  14 +-
 include/linux/vmstat.h                    | 111 +++-
 include/linux/writeback.h                 |   2 +-
 include/trace/events/vmscan.h             |  63 ++-
 include/trace/events/writeback.h          |  10 +-
 kernel/power/snapshot.c                   |  10 +-
 kernel/sysctl.c                           |   4 +-
 mm/backing-dev.c                          |  15 +-
 mm/compaction.c                           |  28 +-
 mm/filemap.c                              |  14 +-
 mm/huge_memory.c                          |  33 +-
 mm/internal.h                             |  11 +-
 mm/memcontrol.c                           | 246 ++++----
 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                       | 124 ++--
 mm/page_alloc.c                           | 268 ++++-----
 mm/page_idle.c                            |   4 +-
 mm/rmap.c                                 |  14 +-
 mm/shmem.c                                |  12 +-
 mm/swap.c                                 |  66 +--
 mm/swap_state.c                           |   4 +-
 mm/util.c                                 |   4 +-
 mm/vmscan.c                               | 901 +++++++++++++++---------------
 mm/vmstat.c                               | 376 ++++++++++---
 mm/workingset.c                           |  54 +-
 48 files changed, 1573 insertions(+), 1263 deletions(-)

-- 
2.6.4

^ permalink raw reply	[flat|nested] 11+ messages in thread
* [PATCH 00/27] Move LRU page reclaim from zones to nodes v7
@ 2016-06-21 14:15 Mel Gorman
  2016-06-21 14:15 ` [PATCH 01/27] mm, vmstat: Add infrastructure for per-node vmstats Mel Gorman
  0 siblings, 1 reply; 11+ messages in thread
From: Mel Gorman @ 2016-06-21 14:15 UTC (permalink / raw)
  To: Andrew Morton, Linux-MM
  Cc: Rik van Riel, Vlastimil Babka, Johannes Weiner, LKML, Mel Gorman

(sorry for resend, the previous attempt didn't go through fully for
some reason)

The bulk of the updates are in response to review from Vlastimil Babka
and received a lot more testing than v6.

Changelog since v6
o Correct reclaim_idx when direct reclaiming for memcg
o Also account LRU pages per zone for compaction/reclaim
o Add page_pgdat helper with more efficient lookup
o Init pgdat LRU lock only once
o Slight optimisation to wake_all_kswapds
o Always wake kcompactd when kswapd is going to sleep
o Rebase to mmotm as of June 15th, 2016

Changelog since v5
o Rebase and adjust to changes

Changelog since v4
o Rebase on top of v3 of page allocator optimisation series

Changelog since v3
o Rebase on top of the page allocator optimisation series
o Remove RFC tag

This is the latest version of a series that moves LRUs from the zones to
the node that is based upon 4.6-rc3 plus the page allocator optimisation
series. 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. 

The series has been tested on a 16 core UMA machine and a 2-socket 48 core
NUMA machine. The UMA results are presented in most cases as the NUMA machine
behaved similarly.

pagealloc
---------

This is a microbenchmark that shows the benefit of removing the fair zone
allocation policy. It was tested uip to order-4 but only orders 0 and 1 are
shown as the other orders were comparable.

                                           4.7.0-rc3                  4.7.0-rc3
                                      mmotm-20160615              nodelru-v7r17
Min      total-odr0-1               485.00 (  0.00%)           462.00 (  4.74%)
Min      total-odr0-2               354.00 (  0.00%)           341.00 (  3.67%)
Min      total-odr0-4               285.00 (  0.00%)           277.00 (  2.81%)
Min      total-odr0-8               249.00 (  0.00%)           240.00 (  3.61%)
Min      total-odr0-16              230.00 (  0.00%)           224.00 (  2.61%)
Min      total-odr0-32              222.00 (  0.00%)           215.00 (  3.15%)
Min      total-odr0-64              216.00 (  0.00%)           210.00 (  2.78%)
Min      total-odr0-128             214.00 (  0.00%)           208.00 (  2.80%)
Min      total-odr0-256             248.00 (  0.00%)           233.00 (  6.05%)
Min      total-odr0-512             277.00 (  0.00%)           270.00 (  2.53%)
Min      total-odr0-1024            294.00 (  0.00%)           284.00 (  3.40%)
Min      total-odr0-2048            308.00 (  0.00%)           298.00 (  3.25%)
Min      total-odr0-4096            318.00 (  0.00%)           307.00 (  3.46%)
Min      total-odr0-8192            322.00 (  0.00%)           308.00 (  4.35%)
Min      total-odr0-16384           324.00 (  0.00%)           309.00 (  4.63%)
Min      total-odr1-1               729.00 (  0.00%)           686.00 (  5.90%)
Min      total-odr1-2               533.00 (  0.00%)           520.00 (  2.44%)
Min      total-odr1-4               434.00 (  0.00%)           415.00 (  4.38%)
Min      total-odr1-8               390.00 (  0.00%)           364.00 (  6.67%)
Min      total-odr1-16              359.00 (  0.00%)           335.00 (  6.69%)
Min      total-odr1-32              356.00 (  0.00%)           327.00 (  8.15%)
Min      total-odr1-64              356.00 (  0.00%)           321.00 (  9.83%)
Min      total-odr1-128             356.00 (  0.00%)           333.00 (  6.46%)
Min      total-odr1-256             354.00 (  0.00%)           337.00 (  4.80%)
Min      total-odr1-512             366.00 (  0.00%)           340.00 (  7.10%)
Min      total-odr1-1024            373.00 (  0.00%)           354.00 (  5.09%)
Min      total-odr1-2048            375.00 (  0.00%)           354.00 (  5.60%)
Min      total-odr1-4096            374.00 (  0.00%)           354.00 (  5.35%)
Min      total-odr1-8192            370.00 (  0.00%)           355.00 (  4.05%)

This shows a steady improvement throughout. The primary benefit is from
reduced system CPU usage which is obvious from the overall times;

           4.7.0-rc3   4.7.0-rc3
        mmotm-20160615 nodelru-v7
User          174.06      174.58
System       2656.78     2485.21
Elapsed      2885.07     2713.67

The vmstats also showed that the fair zone allocation policy was definitely
removed as can be seen here;

                             4.7.0-rc3   4.7.0-rc3
                          mmotm-20160615nodelru-v7r17
DMA32 allocs               28794408561           0
Normal allocs              48431969998 77226313470
Movable allocs                       0           0

tiobench on ext4
----------------

tiobench is a benchmark that artifically benefits if old pages remain resident
while new pages get reclaimed. The fair zone allocation policy mitigates this
problem so pages age fairly. While the benchmark has problems, it is important
that tiobench performance remains constant as it implies that page aging
problems that the fair zone allocation policy fixes are not re-introduced.

                                         4.7.0-rc3             4.7.0-rc3
                                    mmotm-20160615         nodelru-v7r17
Min      PotentialReadSpeed        90.24 (  0.00%)       90.14 ( -0.11%)
Min      SeqRead-MB/sec-1          80.63 (  0.00%)       83.09 (  3.05%)
Min      SeqRead-MB/sec-2          71.91 (  0.00%)       72.44 (  0.74%)
Min      SeqRead-MB/sec-4          75.20 (  0.00%)       74.32 ( -1.17%)
Min      SeqRead-MB/sec-8          65.30 (  0.00%)       65.21 ( -0.14%)
Min      SeqRead-MB/sec-16         62.62 (  0.00%)       62.12 ( -0.80%)
Min      RandRead-MB/sec-1          0.90 (  0.00%)        0.94 (  4.44%)
Min      RandRead-MB/sec-2          0.96 (  0.00%)        0.97 (  1.04%)
Min      RandRead-MB/sec-4          1.43 (  0.00%)        1.41 ( -1.40%)
Min      RandRead-MB/sec-8          1.67 (  0.00%)        1.72 (  2.99%)
Min      RandRead-MB/sec-16         1.77 (  0.00%)        1.86 (  5.08%)
Min      SeqWrite-MB/sec-1         78.12 (  0.00%)       79.78 (  2.12%)
Min      SeqWrite-MB/sec-2         72.74 (  0.00%)       73.23 (  0.67%)
Min      SeqWrite-MB/sec-4         79.40 (  0.00%)       78.32 ( -1.36%)
Min      SeqWrite-MB/sec-8         73.18 (  0.00%)       71.40 ( -2.43%)
Min      SeqWrite-MB/sec-16        75.82 (  0.00%)       75.24 ( -0.76%)
Min      RandWrite-MB/sec-1         1.18 (  0.00%)        1.15 ( -2.54%)
Min      RandWrite-MB/sec-2         1.05 (  0.00%)        0.99 ( -5.71%)
Min      RandWrite-MB/sec-4         1.00 (  0.00%)        0.96 ( -4.00%)
Min      RandWrite-MB/sec-8         0.91 (  0.00%)        0.92 (  1.10%)
Min      RandWrite-MB/sec-16        0.92 (  0.00%)        0.92 (  0.00%)

This shows that the series has little or not impact on tiobench which is
desirable. It indicates that the fair zone allocation policy was removed
in a manner that didn't reintroduce one class of page aging bug. There
were only minor differences in overall reclaim activity

                             4.7.0-rc3   4.7.0-rc3
                          mmotm-20160615nodelru-v7r17
Minor Faults                    640992      640721
Major Faults                       728         623
Swap Ins                             0           0
Swap Outs                            0           0
DMA allocs                           0           0
DMA32 allocs                  46174282    44219717
Normal allocs                 77949344    79858024
Movable allocs                       0           0
Allocation stalls                   38          76
Direct pages scanned             17463       34865
Kswapd pages scanned          93331163    93302388
Kswapd pages reclaimed        93328173    93299677
Direct pages reclaimed           17463       34865
Kswapd efficiency                  99%         99%
Kswapd velocity              13729.855   13755.612
Direct efficiency                 100%        100%
Direct velocity                  2.569       5.140
Percentage direct scans             0%          0%
Page writes by reclaim               0           0
Page writes file                     0           0
Page writes anon                     0           0
Page reclaim immediate              54          36

kswapd activity was roughly comparable. There was slight differences
in direct reclaim activity but negligible in the context of the overall
workload (velocity of 5 pages per second with the patches applied, 2 pages
per second in the baseline kernel).

pgbench read-only large configuration on ext4
---------------------------------------------

pgbench is a database benchmark that can be sensitive to page reclaim
decisions. This also checks if removing the fair zone allocation policy
is safe

pgbench Transactions
                        4.7.0-rc3             4.7.0-rc3
                   mmotm-20160615         nodelru-v7r17
Hmean    1       191.00 (  0.00%)      193.67 (  1.40%)
Hmean    5       338.59 (  0.00%)      336.99 ( -0.47%)
Hmean    12      374.03 (  0.00%)      386.15 (  3.24%)
Hmean    21      372.24 (  0.00%)      372.02 ( -0.06%)
Hmean    30      383.98 (  0.00%)      370.69 ( -3.46%)
Hmean    32      431.01 (  0.00%)      438.47 (  1.73%)

Negligible differences again. As with tiobench, overall reclaim activity
was comparable.

bonnie++ on ext4
----------------

No interesting performance difference, negligible differences on reclaim
stats.

paralleldd on ext4
------------------

This workload uses varying numbers of dd instances to read large amounts of
data from disk.

paralleldd
                               4.7.0-rc3             4.7.0-rc3
                          mmotm-20160615         nodelru-v7r17
Amean    Elapsd-1       181.57 (  0.00%)      179.63 (  1.07%)
Amean    Elapsd-3       188.29 (  0.00%)      183.68 (  2.45%)
Amean    Elapsd-5       188.02 (  0.00%)      181.73 (  3.35%)
Amean    Elapsd-7       186.07 (  0.00%)      184.11 (  1.05%)
Amean    Elapsd-12      188.16 (  0.00%)      183.51 (  2.47%)
Amean    Elapsd-16      189.03 (  0.00%)      181.27 (  4.10%)

           4.7.0-rc3   4.7.0-rc3
        mmotm-20160615nodelru-v7r17
User         1439.23     1433.37
System       8332.31     8216.01
Elapsed      3619.80     3532.69

There is a slight gain in performance, some of which is from the reduced system
CPU usage. There areminor differences in reclaim activity but nothing significant

                             4.7.0-rc3   4.7.0-rc3
                          mmotm-20160615nodelru-v7r17
Minor Faults                    362486      358215
Major Faults                      1143        1113
Swap Ins                            26           0
Swap Outs                         2920         482
DMA allocs                           0           0
DMA32 allocs                  31568814    28598887
Normal allocs                 46539922    49514444
Movable allocs                       0           0
Allocation stalls                    0           0
Direct pages scanned                 0           0
Kswapd pages scanned          40886878    40849710
Kswapd pages reclaimed        40869923    40835207
Direct pages reclaimed               0           0
Kswapd efficiency                  99%         99%
Kswapd velocity              11295.342   11563.344
Direct efficiency                 100%        100%
Direct velocity                  0.000       0.000
Slabs scanned                   131673      126099
Direct inode steals                 57          60
Kswapd inode steals                762          18

It basically shows that kswapd was active at roughly the same rate in
both kernels. There was also comparable slab scanning activity and direct
reclaim was avoided in both cases. There appears to be a large difference
in numbers of inodes reclaimed but the workload has few active inodes and
is likely a timing artifact. It's interesting to note that the node-lru
did not swap in any pages but given the low swap activity, it's unlikely
to be significant.

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
                             4.7.0-rc3             4.7.0-rc3
                        mmotm-20160615         nodelru-v7r17
Min         mmap     16.8422 (  0.00%)     15.9821 (  5.11%)
1st-qrtle   mmap     57.8709 (  0.00%)     58.0794 ( -0.36%)
2nd-qrtle   mmap     58.4335 (  0.00%)     59.4286 ( -1.70%)
3rd-qrtle   mmap     58.6957 (  0.00%)     59.6862 ( -1.69%)
Max-90%     mmap     58.9388 (  0.00%)     59.8759 ( -1.59%)
Max-93%     mmap     59.0505 (  0.00%)     59.9333 ( -1.50%)
Max-95%     mmap     59.1877 (  0.00%)     59.9844 ( -1.35%)
Max-99%     mmap     60.3237 (  0.00%)     60.2337 (  0.15%)
Max         mmap    285.6454 (  0.00%)    135.6006 ( 52.53%)
Mean        mmap     57.8366 (  0.00%)     58.4884 ( -1.13%)

This shows that there is a slight impact on mmap latency but that
the worst-case outlier is much improved. As the problem with this
benchmark used to be that the kernel stalled for minutes, this
difference is negligible.

Some of the vmstats are interesting

                             4.7.0-rc3   4.7.0-rc3
                          mmotm-20160615nodelru-v7r17
Swap Ins                            58          42
Swap Outs                            0           0
Allocation stalls                   16           0
Direct pages scanned              1374           0
Kswapd pages scanned          42454910    41782544
Kswapd pages reclaimed        41571035    41781833
Direct pages reclaimed            1167           0
Kswapd efficiency                  97%         99%
Kswapd velocity              14774.479   14223.796
Direct efficiency                  84%        100%
Direct velocity                  0.478       0.000
Percentage direct scans             0%          0%
Page writes by reclaim          696918           0
Page writes file                696918           0
Page writes anon                     0           0
Page reclaim immediate            2940         137
Sector Reads                  81644424    81699544
Sector Writes                 99193620    98862160
Page rescued immediate               0           0
Slabs scanned                  1279838       22640

kswapd and direct reclaim activity are similar but the node LRU series
did not attempt to trigger any page writes from reclaim context.

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                |  32 +-
 include/linux/mm.h                        |   5 +
 include/linux/mm_inline.h                 |  21 +-
 include/linux/mm_types.h                  |   2 +-
 include/linux/mmzone.h                    | 158 +++---
 include/linux/swap.h                      |  23 +-
 include/linux/topology.h                  |   2 +-
 include/linux/vm_event_item.h             |  14 +-
 include/linux/vmstat.h                    | 111 +++-
 include/linux/writeback.h                 |   2 +-
 include/trace/events/vmscan.h             |  63 ++-
 include/trace/events/writeback.h          |  10 +-
 kernel/power/snapshot.c                   |  10 +-
 kernel/sysctl.c                           |   4 +-
 mm/backing-dev.c                          |  15 +-
 mm/compaction.c                           |  28 +-
 mm/filemap.c                              |  14 +-
 mm/huge_memory.c                          |  33 +-
 mm/internal.h                             |  11 +-
 mm/memcontrol.c                           | 246 ++++----
 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                       | 124 ++--
 mm/page_alloc.c                           | 268 ++++-----
 mm/page_idle.c                            |   4 +-
 mm/rmap.c                                 |  14 +-
 mm/shmem.c                                |  12 +-
 mm/swap.c                                 |  66 +--
 mm/swap_state.c                           |   4 +-
 mm/util.c                                 |   4 +-
 mm/vmscan.c                               | 901 +++++++++++++++---------------
 mm/vmstat.c                               | 376 ++++++++++---
 mm/workingset.c                           |  54 +-
 48 files changed, 1573 insertions(+), 1263 deletions(-)

-- 
2.6.4

^ permalink raw reply	[flat|nested] 11+ messages in thread
* [PATCH 00/27] Move LRU page reclaim from zones to nodes v6
@ 2016-06-09 18:04 Mel Gorman
  2016-06-09 18:04 ` [PATCH 01/27] mm, vmstat: Add infrastructure for per-node vmstats Mel Gorman
  0 siblings, 1 reply; 11+ messages in thread
From: Mel Gorman @ 2016-06-09 18:04 UTC (permalink / raw)
  To: Andrew Morton, Linux-MM
  Cc: Rik van Riel, Vlastimil Babka, Johannes Weiner, LKML, Mel Gorman

This is only lightly tested as I've had stability problems during boot
that have nothing to do with the series. It's based on mmots as of June
6th. Very little has changed with the big exception of "mm, vmscan:
Move LRU lists to node" because it had to adapt to per-zone changes in
should_reclaim_retry and compaction_zonelist_suitable.

Changelog since v5
o Rebase and adjust to changes

Changelog since v4
o Rebase on top of v3 of page allocator optimisation series

Changelog since v3
o Rebase on top of the page allocator optimisation series
o Remove RFC tag

This is the latest version of a series that moves LRUs from the zones to
the node that is based upon 4.6-rc3 plus the page allocator optimisation
series. 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. 

The series got basic testing this time on a UMA machine. The page allocator
microbenchmark highlights the gain from removing the fair zone allocation
policy

                                           4.7.0-rc2                  4.7.0-rc2
                                      mmotm-20160606               nodelru-v6r2
Min      total-odr0-1               500.00 (  0.00%)           475.00 (  5.00%)
Min      total-odr0-2               358.00 (  0.00%)           343.00 (  4.19%)
Min      total-odr0-4               292.00 (  0.00%)           279.00 (  4.45%)
Min      total-odr0-8               253.00 (  0.00%)           242.00 (  4.35%)
Min      total-odr0-16              275.00 (  0.00%)           226.00 ( 17.82%)
Min      total-odr0-32              225.00 (  0.00%)           215.00 (  4.44%)
Min      total-odr0-64              219.00 (  0.00%)           210.00 (  4.11%)
Min      total-odr0-128             216.00 (  0.00%)           207.00 (  4.17%)
Min      total-odr0-256             243.00 (  0.00%)           246.00 ( -1.23%)
Min      total-odr0-512             276.00 (  0.00%)           265.00 (  3.99%)
Min      total-odr0-1024            290.00 (  0.00%)           287.00 (  1.03%)
Min      total-odr0-2048            303.00 (  0.00%)           296.00 (  2.31%)
Min      total-odr0-4096            312.00 (  0.00%)           310.00 (  0.64%)
Min      total-odr0-8192            320.00 (  0.00%)           308.00 (  3.75%)
Min      total-odr0-16384           320.00 (  0.00%)           308.00 (  3.75%)
Min      total-odr1-1               737.00 (  0.00%)           707.00 (  4.07%)
Min      total-odr1-2               547.00 (  0.00%)           521.00 (  4.75%)
Min      total-odr1-4               620.00 (  0.00%)           418.00 ( 32.58%)
Min      total-odr1-8               386.00 (  0.00%)           367.00 (  4.92%)
Min      total-odr1-16              361.00 (  0.00%)           340.00 (  5.82%)
Min      total-odr1-32              352.00 (  0.00%)           328.00 (  6.82%)
Min      total-odr1-64              345.00 (  0.00%)           324.00 (  6.09%)
Min      total-odr1-128             347.00 (  0.00%)           328.00 (  5.48%)
Min      total-odr1-256             347.00 (  0.00%)           329.00 (  5.19%)
Min      total-odr1-512             354.00 (  0.00%)           332.00 (  6.21%)
Min      total-odr1-1024            355.00 (  0.00%)           337.00 (  5.07%)
Min      total-odr1-2048            358.00 (  0.00%)           345.00 (  3.63%)
Min      total-odr1-4096            360.00 (  0.00%)           346.00 (  3.89%)
Min      total-odr1-8192            360.00 (  0.00%)           347.00 (  3.61%)

A basic IO benchmark based on varying numbers of dd running in parallel
showed nothing interesting other than differences in what zones were
scanned due to the fair zone allocation policy being removed.

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                 |   2 +-
 include/linux/mm_types.h                  |   2 +-
 include/linux/mmzone.h                    | 157 +++---
 include/linux/swap.h                      |  15 +-
 include/linux/topology.h                  |   2 +-
 include/linux/vm_event_item.h             |  14 +-
 include/linux/vmstat.h                    | 111 +++-
 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                          |  15 +-
 mm/compaction.c                           |  39 +-
 mm/filemap.c                              |  14 +-
 mm/huge_memory.c                          |  33 +-
 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                       | 124 +++--
 mm/page_alloc.c                           | 271 +++++-----
 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                               | 829 +++++++++++++++---------------
 mm/vmstat.c                               | 374 +++++++++++---
 mm/workingset.c                           |  52 +-
 47 files changed, 1489 insertions(+), 1217 deletions(-)

-- 
2.6.4

^ permalink raw reply	[flat|nested] 11+ messages in thread
* [PATCH 00/27] Move LRU page reclaim from zones to nodes v5
@ 2016-04-15  9:13 Mel Gorman
  2016-04-15  9:13 ` [PATCH 01/27] mm, vmstat: Add infrastructure for per-node vmstats Mel Gorman
  0 siblings, 1 reply; 11+ messages in thread
From: Mel Gorman @ 2016-04-15  9:13 UTC (permalink / raw)
  To: Andrew Morton, Linux-MM
  Cc: Rik van Riel, Vlastimil Babka, Johannes Weiner,
	Jesper Dangaard Brouer, LKML, Mel Gorman

Changelog since v4
o Rebase on top of v3 of page allocator optimisation series

Changelog since v3
o Rebase on top of the page allocator optimisation series
o Remove RFC tag

This is the latest version of a series that moves LRUs from the zones to
the node that is based upon 4.6-rc3 plus the page allocator optimisation
series. 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
                                           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                    | 111 +++-
 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                           | 289 +++++-----
 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                               | 847 ++++++++++++++----------------
 mm/vmstat.c                               | 369 ++++++++++---
 mm/workingset.c                           |  53 +-
 47 files changed, 1476 insertions(+), 1221 deletions(-)

-- 
2.6.4

^ permalink raw reply	[flat|nested] 11+ messages in thread

end of thread, other threads:[~2016-06-21 14:16 UTC | newest]

Thread overview: 11+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-06-21 11:43 [PATCH 00/27] Move LRU page reclaim from zones to nodes v7 Mel Gorman
2016-06-21 11:43 ` [PATCH 01/27] mm, vmstat: Add infrastructure for per-node vmstats Mel Gorman
2016-06-21 11:43 ` [PATCH 02/27] mm, vmscan: Move lru_lock to the node Mel Gorman
2016-06-21 11:43 ` [PATCH 03/27] mm, vmscan: Move LRU lists to node Mel Gorman
  -- strict thread matches above, loose matches on Subject: below --
2016-06-21 14:15 [PATCH 00/27] Move LRU page reclaim from zones to nodes v7 Mel Gorman
2016-06-21 14:15 ` [PATCH 01/27] mm, vmstat: Add infrastructure for per-node vmstats Mel Gorman
2016-06-09 18:04 [PATCH 00/27] Move LRU page reclaim from zones to nodes v6 Mel Gorman
2016-06-09 18:04 ` [PATCH 01/27] mm, vmstat: Add infrastructure for per-node vmstats Mel Gorman
2016-06-10 13:31   ` Vlastimil Babka
2016-06-10 13:47     ` Mel Gorman
2016-06-13 17:26   ` Christoph Lameter
2016-06-14 14:25     ` Mel Gorman
2016-04-15  9:13 [PATCH 00/27] Move LRU page reclaim from zones to nodes v5 Mel Gorman
2016-04-15  9:13 ` [PATCH 01/27] mm, vmstat: Add infrastructure for per-node vmstats Mel Gorman

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).