Skip to contentSkip to conversation

c0f4dfd4f9: -53% perf-stat.cpu-migrations

rcu Fengguang Wu 2 participants 2 messages Jan 24, 2014 – Jan 27, 2014

lore
Share & tools
lore
#0 lore raw

Hi Paul,

Just FYI, we noticed -53% perf-stat.cpu-migrations in dd write tests on btrfs, which looks good. First good commit is

commit c0f4dfd4f90f1667d234d21f15153ea09a2eaa66

Author:     Paul E. McKenney <[email protected]>
AuthorDate: Fri Dec 28 11:30:36 2012 -0800
Commit:     Paul E. McKenney <[email protected]>
CommitDate: Tue Mar 26 08:04:51 2013 -0700

rcu: Make RCU_FAST_NO_HZ take advantage of numbered callbacks

Because RCU callbacks are now associated with the number of the grace period that they must wait for, CPUs can now take advance callbacks corresponding to grace periods that ended while a given CPU was in dyntick-idle mode. This eliminates the need to try forcing the RCU state machine while entering idle, thus reducing the CPU intensiveness of RCU_FAST_NO_HZ, which should increase its energy efficiency.

 Documentation/kernel-parameters.txt |  28 ++-
 include/linux/rcupdate.h            |   1 +
 init/Kconfig                        |  17 +-
 kernel/rcutree.c                    |  28 +--
 kernel/rcutree.h                    |  12 +-
 kernel/rcutree_plugin.h             | 374 ++++++++++--------------------------
 kernel/rcutree_trace.c              |   2 -
 7 files changed, 149 insertions(+), 313 deletions(-)

b11cc5760a9c48c c0f4dfd4f90f1667d234d21f1
--------------- -------------------------

     86878 ~138%     -90.3%       8397 ~152%   cpuidle.POLL.time
       154 ~16%     -87.3%         19 ~55%   cpuidle.POLL.usage
  12177976 ~ 4%     -85.6%    1748244 ~20%   cpuidle.C1-NHM.time
    381439 ~ 3%     -68.4%     120538 ~ 2%   softirqs.RCU
      0.53 ~87%    +161.8%       1.40 ~16%   perf-profile.cpu-cycles.copy_user_generic_string.__btrfs_buffered_write.btrfs_file_aio_write.do_sync_write.vfs_write
   5227241 ~ 4%     -58.3%    2180928 ~ 7%   cpuidle.C1E-NHM.time
      0.67 ~88%     +88.6%       1.26 ~21%   perf-profile.cpu-cycles.calc_csum_metadata_size.btrfs_delalloc_release_metadata.btrfs_clear_bit_hook.clear_state_bit.clear_extent_bit
    231531 ~ 2%     -48.3%     119653 ~ 2%   interrupts.LOC
     91019 ~ 2%     -40.2%      54404 ~ 2%   cpuidle.C3-NHM.usage
 1.991e+08 ~ 3%     -36.7%   1.26e+08 ~ 7%   cpuidle.C3-NHM.time
      7.07 ~ 4%     -32.7%       4.76 ~ 8%   turbostat.%c3
     23380 ~33%     +41.2%      33024 ~ 6%   proc-vmstat.kswapd_low_wmark_hit_quickly
     62805 ~ 3%     -28.4%      44960 ~ 2%   softirqs.SCHED
     64678 ~ 1%     -30.1%      45195 ~ 1%   softirqs.TIMER
     55051 ~ 3%     -22.2%      42823 ~ 2%   interrupts.0:IO-APIC-edge.timer
       920 ~ 4%     +19.2%       1097 ~ 5%   slabinfo.kmalloc-512.active_objs
       920 ~ 4%     +19.2%       1097 ~ 5%   slabinfo.kmalloc-512.num_objs
    361987 ~ 2%      +9.9%     397730 ~ 0%   cpuidle.C6-NHM.usage
      5.30 ~ 1%      -9.7%       4.78 ~ 1%   turbostat.%c1
    178105 ~ 3%     -53.5%      82837 ~ 1%   perf-stat.cpu-migrations
      5763 ~ 8%     -44.7%       3186 ~22%   vmstat.system.cs
   3566268 ~ 8%     -44.8%    1968744 ~21%   perf-stat.context-switches
       658 ~ 2%     -30.4%        458 ~ 0%   vmstat.system.in
  53376814 ~12%     -24.2%   40482438 ~22%   perf-stat.node-load-misses
 2.996e+10 ~ 3%     -10.8%  2.672e+10 ~ 3%   perf-stat.L1-icache-load-misses
 1.998e+09 ~ 4%     -11.6%  1.766e+09 ~ 2%   perf-stat.branch-misses
 1.005e+12 ~ 5%     -11.9%  8.852e+11 ~ 6%   perf-stat.stalled-cycles-frontend
 6.344e+08 ~ 2%      -6.8%  5.915e+08 ~ 2%   perf-stat.LLC-store-misses
 2.892e+10 ~ 2%      +5.4%  3.047e+10 ~ 3%   perf-stat.bus-cycles

perf-stat.cpu-migrations

90000 ++*----*----------*-*---*------*-----------------------------------+

         *   **   *.*.**.*     *  *.*.*   **.*.*.*.*   .*.                  |
   80000 ++                                         *.*   *                 |
         |                                                                  |
   70000 ++                                                                 |
         |                                                                  |
   60000 ++                                                                 |
         |                                                                  |
   50000 ++                                                                 |
         |                                                                  |
   40000 ++                                                                 |
         | O O                                                              |

30000 O+ O O O OO O O O O OO O O O OO O O O OO O O O O OO O O O OO O O
| O |
20000 ++-----------------------------------------------------------------+

#1 lore raw
Quoted 4 lines · Fengguang Wu

Hi Paul,

Just FYI, we noticed -53% perf-stat.cpu-migrations in dd write tests on btrfs, which looks good. First good commit is

Nice! ;-)

Thanx, Paul

Quoted 78 lines

commit c0f4dfd4f90f1667d234d21f15153ea09a2eaa66

Author:     Paul E. McKenney <[email protected]>
AuthorDate: Fri Dec 28 11:30:36 2012 -0800
Commit:     Paul E. McKenney <[email protected]>
CommitDate: Tue Mar 26 08:04:51 2013 -0700

rcu: Make RCU_FAST_NO_HZ take advantage of numbered callbacks

Because RCU callbacks are now associated with the number of the grace period that they must wait for, CPUs can now take advance callbacks corresponding to grace periods that ended while a given CPU was in dyntick-idle mode. This eliminates the need to try forcing the RCU state machine while entering idle, thus reducing the CPU intensiveness of RCU_FAST_NO_HZ, which should increase its energy efficiency.

 Documentation/kernel-parameters.txt |  28 ++-
 include/linux/rcupdate.h            |   1 +
 init/Kconfig                        |  17 +-
 kernel/rcutree.c                    |  28 +--
 kernel/rcutree.h                    |  12 +-
 kernel/rcutree_plugin.h             | 374 ++++++++++--------------------------
 kernel/rcutree_trace.c              |   2 -
 7 files changed, 149 insertions(+), 313 deletions(-)

b11cc5760a9c48c c0f4dfd4f90f1667d234d21f1
--------------- -------------------------

     86878 ~138%     -90.3%       8397 ~152%   cpuidle.POLL.time
       154 ~16%     -87.3%         19 ~55%   cpuidle.POLL.usage
  12177976 ~ 4%     -85.6%    1748244 ~20%   cpuidle.C1-NHM.time
    381439 ~ 3%     -68.4%     120538 ~ 2%   softirqs.RCU
      0.53 ~87%    +161.8%       1.40 ~16%   perf-profile.cpu-cycles.copy_user_generic_string.__btrfs_buffered_write.btrfs_file_aio_write.do_sync_write.vfs_write
   5227241 ~ 4%     -58.3%    2180928 ~ 7%   cpuidle.C1E-NHM.time
      0.67 ~88%     +88.6%       1.26 ~21%   perf-profile.cpu-cycles.calc_csum_metadata_size.btrfs_delalloc_release_metadata.btrfs_clear_bit_hook.clear_state_bit.clear_extent_bit
    231531 ~ 2%     -48.3%     119653 ~ 2%   interrupts.LOC
     91019 ~ 2%     -40.2%      54404 ~ 2%   cpuidle.C3-NHM.usage
 1.991e+08 ~ 3%     -36.7%   1.26e+08 ~ 7%   cpuidle.C3-NHM.time
      7.07 ~ 4%     -32.7%       4.76 ~ 8%   turbostat.%c3
     23380 ~33%     +41.2%      33024 ~ 6%   proc-vmstat.kswapd_low_wmark_hit_quickly
     62805 ~ 3%     -28.4%      44960 ~ 2%   softirqs.SCHED
     64678 ~ 1%     -30.1%      45195 ~ 1%   softirqs.TIMER
     55051 ~ 3%     -22.2%      42823 ~ 2%   interrupts.0:IO-APIC-edge.timer
       920 ~ 4%     +19.2%       1097 ~ 5%   slabinfo.kmalloc-512.active_objs
       920 ~ 4%     +19.2%       1097 ~ 5%   slabinfo.kmalloc-512.num_objs
    361987 ~ 2%      +9.9%     397730 ~ 0%   cpuidle.C6-NHM.usage
      5.30 ~ 1%      -9.7%       4.78 ~ 1%   turbostat.%c1
    178105 ~ 3%     -53.5%      82837 ~ 1%   perf-stat.cpu-migrations
      5763 ~ 8%     -44.7%       3186 ~22%   vmstat.system.cs
   3566268 ~ 8%     -44.8%    1968744 ~21%   perf-stat.context-switches
       658 ~ 2%     -30.4%        458 ~ 0%   vmstat.system.in
  53376814 ~12%     -24.2%   40482438 ~22%   perf-stat.node-load-misses
 2.996e+10 ~ 3%     -10.8%  2.672e+10 ~ 3%   perf-stat.L1-icache-load-misses
 1.998e+09 ~ 4%     -11.6%  1.766e+09 ~ 2%   perf-stat.branch-misses
 1.005e+12 ~ 5%     -11.9%  8.852e+11 ~ 6%   perf-stat.stalled-cycles-frontend
 6.344e+08 ~ 2%      -6.8%  5.915e+08 ~ 2%   perf-stat.LLC-store-misses
 2.892e+10 ~ 2%      +5.4%  3.047e+10 ~ 3%   perf-stat.bus-cycles

perf-stat.cpu-migrations

90000 ++*----*----------*-*---*------*-----------------------------------+

         *   **   *.*.**.*     *  *.*.*   **.*.*.*.*   .*.                  |
   80000 ++                                         *.*   *                 |
         |                                                                  |
   70000 ++                                                                 |
         |                                                                  |
   60000 ++                                                                 |
         |                                                                  |
   50000 ++                                                                 |
         |                                                                  |
   40000 ++                                                                 |
         | O O                                                              |

30000 O+ O O O OO O O O O OO O O O OO O O O OO O O O O OO O O O OO O O
| O |
20000 ++-----------------------------------------------------------------+