lkml.org 
[lkml]   [2016]   [Aug]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH 07/34] mm, vmscan: make kswapd reclaim in terms of nodes
    > Patch "mm: vmscan: Begin reclaiming pages on a per-node basis" started
    > thinking of reclaim in terms of nodes but kswapd is still zone-centric. This
    > patch gets rid of many of the node-based versus zone-based decisions.
    >
    > o A node is considered balanced when any eligible lower zone is balanced.
    > This eliminates one class of age-inversion problem because we avoid
    > reclaiming a newer page just because it's in the wrong zone
    > o pgdat_balanced disappears because we now only care about one zone being
    > balanced.
    > o Some anomalies related to writeback and congestion tracking being based on
    > zones disappear.
    > o kswapd no longer has to take care to reclaim zones in the reverse order
    > that the page allocator uses.
    > o Most importantly of all, reclaim from node 0 with multiple zones will
    > have similar aging and reclaiming characteristics as every
    > other node.
    >
    > Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
    > Acked-by: Johannes Weiner <hannes@cmpxchg.org>
    > Acked-by: Vlastimil Babka <vbabka@suse.cz>

    This patch seems to hurt FA_DUMP functionality. This behaviour is not
    seen on v4.7 but only after this patch.

    So when a kernel on a multinode machine with memblock_reserve() such
    that most of the nodes have zero available memory, kswapd seems to be
    consuming 100% of the time.

    This is independent of CONFIG_DEFERRED_STRUCT_PAGE, i.e this problem is
    seen even with parallel page struct initialization disabled.


    top - 13:48:52 up 1:07, 3 users, load average: 15.25, 15.32, 21.18
    Tasks: 11080 total, 16 running, 11064 sleeping, 0 stopped, 0 zombie
    %Cpu(s): 0.0 us, 2.7 sy, 0.0 ni, 97.3 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
    KiB Mem: 15929941+total, 8637824 used, 15843563+free, 2304 buffers
    KiB Swap: 91898816 total, 0 used, 91898816 free. 1381312 cached Mem

    PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
    10824 root 20 0 0 0 0 R 100.000 0.000 65:30.76 kswapd2
    10837 root 20 0 0 0 0 R 100.000 0.000 65:31.17 kswapd15
    10823 root 20 0 0 0 0 R 97.059 0.000 65:30.85 kswapd1
    10825 root 20 0 0 0 0 R 97.059 0.000 65:31.10 kswapd3
    10826 root 20 0 0 0 0 R 97.059 0.000 65:31.18 kswapd4
    10827 root 20 0 0 0 0 R 97.059 0.000 65:31.08 kswapd5
    10828 root 20 0 0 0 0 R 97.059 0.000 65:30.91 kswapd6
    10829 root 20 0 0 0 0 R 97.059 0.000 65:31.17 kswapd7
    10830 root 20 0 0 0 0 R 97.059 0.000 65:31.17 kswapd8
    10831 root 20 0 0 0 0 R 97.059 0.000 65:31.18 kswapd9
    10832 root 20 0 0 0 0 R 97.059 0.000 65:31.12 kswapd10
    10833 root 20 0 0 0 0 R 97.059 0.000 65:31.19 kswapd11
    10834 root 20 0 0 0 0 R 97.059 0.000 65:31.13 kswapd12
    10835 root 20 0 0 0 0 R 97.059 0.000 65:31.09 kswapd13
    10836 root 20 0 0 0 0 R 97.059 0.000 65:31.18 kswapd14
    277155 srikar 20 0 16960 13760 3264 R 52.941 0.001 0:00.37 top

    top - 13:48:55 up 1:07, 3 users, load average: 15.23, 15.32, 21.15
    Tasks: 11080 total, 16 running, 11064 sleeping, 0 stopped, 0 zombie
    %Cpu(s): 0.0 us, 1.0 sy, 0.0 ni, 99.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
    KiB Mem: 15929941+total, 8637824 used, 15843563+free, 2304 buffers
    KiB Swap: 91898816 total, 0 used, 91898816 free. 1381312 cached Mem

    PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
    10836 root 20 0 0 0 0 R 100.000 0.000 65:33.39 kswapd14
    10823 root 20 0 0 0 0 R 100.000 0.000 65:33.05 kswapd1
    10824 root 20 0 0 0 0 R 100.000 0.000 65:32.96 kswapd2
    10825 root 20 0 0 0 0 R 100.000 0.000 65:33.30 kswapd3
    10826 root 20 0 0 0 0 R 100.000 0.000 65:33.38 kswapd4
    10827 root 20 0 0 0 0 R 100.000 0.000 65:33.28 kswapd5
    10828 root 20 0 0 0 0 R 100.000 0.000 65:33.11 kswapd6
    10829 root 20 0 0 0 0 R 100.000 0.000 65:33.37 kswapd7
    10830 root 20 0 0 0 0 R 100.000 0.000 65:33.37 kswapd8
    10831 root 20 0 0 0 0 R 100.000 0.000 65:33.38 kswapd9
    10832 root 20 0 0 0 0 R 100.000 0.000 65:33.32 kswapd10
    10833 root 20 0 0 0 0 R 100.000 0.000 65:33.39 kswapd11
    10834 root 20 0 0 0 0 R 100.000 0.000 65:33.33 kswapd12
    10835 root 20 0 0 0 0 R 100.000 0.000 65:33.29 kswapd13
    10837 root 20 0 0 0 0 R 100.000 0.000 65:33.37 kswapd15
    277155 srikar 20 0 17536 14912 3264 R 9.091 0.001 0:00.57 top
    1092 root rt 0 0 0 0 S 0.455 0.000 0:00.08 watchdog/178

    Please see that there is no used swap space. However 15 kswapd threads
    corresponding to 15 out of the 16 nodes are running full throttle. The
    only node 0 has memory, other nodes memory is fully reserved.

    git bisect output
    I tried git bisect between v4.7 and v4.8-rc3 filtered to mm/vmscan.c

    # bad: [d7f05528eedb047efe2288cff777676b028747b6] mm, vmscan: account for skipped pages as a partial scan
    # good: [b1123ea6d3b3da25af5c8a9d843bd07ab63213f4] mm: balloon: use general non-lru movable page feature
    git bisect start 'HEAD' 'b1123ea6' '--' 'mm/vmscan.c'
    # bad: [c4a25635b60d08853a3e4eaae3ab34419a36cfa2] mm: move vmscan writes and file write accounting to the node
    git bisect bad c4a25635b60d08853a3e4eaae3ab34419a36cfa2
    # bad: [38087d9b0360987a6db46c2c2c4ece37cd048abe] mm, vmscan: simplify the logic deciding whether kswapd sleeps
    git bisect bad 38087d9b0360987a6db46c2c2c4ece37cd048abe
    # good: [b2e18757f2c9d1cdd746a882e9878852fdec9501] mm, vmscan: begin reclaiming pages on a per-node basis
    git bisect good b2e18757f2c9d1cdd746a882e9878852fdec9501
    # bad: [1d82de618ddde0f1164e640f79af152f01994c18] mm, vmscan: make kswapd reclaim in terms of nodes
    git bisect bad 1d82de618ddde0f1164e640f79af152f01994c18
    # good: [f7b60926ebc05944f73d93ffaf6690503b796a88] mm, vmscan: have kswapd only scan based on the highest requested zone
    git bisect good f7b60926ebc05944f73d93ffaf6690503b796a88
    # first bad commit: [1d82de618ddde0f1164e640f79af152f01994c18] mm, vmscan: make kswapd reclaim in terms of nodes

    Here is perf top output on the kernel where kswapd is hogging cpu.

    - 93.50% 0.01% [kernel] [k] kswapd
    - kswapd
    - 114.31% shrink_node
    - 111.51% shrink_node_memcg
    - pgdat_reclaimable
    - 95.51% pgdat_reclaimable_pages
    - 86.34% pgdat_reclaimable_pages
    - 6.69% _find_next_bit.part.0
    - 2.47% find_next_bit
    - 14.46% pgdat_reclaimable
    1.13% _find_next_bit.part.0
    + 0.30% find_next_bit
    - 2.38% shrink_slab
    - super_cache_count
    - 0
    - __list_lru_count_one.isra.1
    _raw_spin_lock
    - 28.04% pgdat_reclaimable
    - 23.97% pgdat_reclaimable_pages
    - 21.66% pgdat_reclaimable_pages
    - 1.69% _find_next_bit.part.0
    0.63% find_next_bit
    - 3.70% pgdat_reclaimable
    0.29% _find_next_bit.part.0
    - 16.33% zone_balanced
    - zone_watermark_ok_safe
    - 14.86% zone_watermark_ok_safe
    1.15% _find_next_bit.part.0
    0.31% find_next_bit
    - 2.72% prepare_kswapd_sleep
    - zone_balanced
    - zone_watermark_ok_safe
    zone_watermark_ok_safe
    - 80.72% 10.51% [kernel] [k] pgdat_reclaimable
    - 140.49% pgdat_reclaimable
    - 138.40% pgdat_reclaimable_pages
    - 125.10% pgdat_reclaimable_pages
    - 9.71% _find_next_bit.part.0
    - 3.59% find_next_bit
    1.64% _find_next_bit.part.0
    + 0.44% find_next_bit
    - 21.03% ret_from_kernel_thread
    kthread
    - kswapd
    - 16.75% shrink_node
    shrink_node_memcg
    - 4.28% pgdat_reclaimable
    pgdat_reclaimable
    - 69.17% 62.48% [kernel] [k] pgdat_reclaimable_pages
    - 145.91% ret_from_kernel_thread
    kthread
    - 15.61% pgdat_reclaimable_pages
    - 11.33% _find_next_bit.part.0
    - 4.19% find_next_bit
    - 66.18% 0.01% [kernel] [k] shrink_node
    - shrink_node
    - 157.54% shrink_node_memcg
    - pgdat_reclaimable
    - 134.94% pgdat_reclaimable_pages
    - 121.99% pgdat_reclaimable_pages
    - 9.46% _find_next_bit.part.0
    - 3.49% find_next_bit
    - 20.44% pgdat_reclaimable
    1.59% _find_next_bit.part.0
    + 0.42% find_next_bit
    - 3.37% shrink_slab
    - super_cache_count
    - 0
    - __list_lru_count_one.isra.1
    _raw_spin_lock
    - 64.56% 0.03% [kernel] [k] shrink_node_memcg
    - shrink_node_memcg
    - pgdat_reclaimable
    - 138.31% pgdat_reclaimable_pages
    - 125.04% pgdat_reclaimable_pages
    - 9.69% _find_next_bit.part.0
    - 3.58% find_next_bit
    - 20.95% pgdat_reclaimable
    1.63% _find_next_bit.part.0
    + 0.43% find_next_bit
    53.73% 0.00% [kernel] [k] kthread
    53.73% 0.00% [kernel] [k] ret_from_kernel_thread
    - 11.04% 10.04% [kernel] [k] zone_watermark_ok_safe
    - 146.80% ret_from_kernel_thread
    kthread
    - kswapd
    - 125.81% zone_balanced
    zone_watermark_ok_safe
    - 20.97% prepare_kswapd_sleep
    zone_balanced
    zone_watermark_ok_safe
    - 14.55% zone_watermark_ok_safe
    11.38% _find_next_bit.part.0
    3.06% find_next_bit
    - 11.03% 0.00% [kernel] [k] zone_balanced
    - zone_balanced
    - zone_watermark_ok_safe
    145.84% zone_watermark_ok_safe
    11.31% _find_next_bit.part.0
    3.04% find_next_bit


    --
    Thanks and Regards
    Srikar Dronamraju

    \
     
     \ /
      Last update: 2016-09-17 09:58    [W:4.324 / U:0.168 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site