tools/perf/Documentation/examples.txt

   1
   2                 ------------------------------
   3                 ****** perf by examples ******
   4                 ------------------------------
   5
   6 [ From an e-mail by Ingo Molnar, http://lkml.org/lkml/2009/8/4/346 ]
   7
   8
   9 First, discovery/enumeration of available counters can be done via
  10 'perf list':
  11
  12 titan:~> perf list
  13   [...]
  14   kmem:kmalloc                             [Tracepoint event]
  15   kmem:kmem_cache_alloc                    [Tracepoint event]
  16   kmem:kmalloc_node                        [Tracepoint event]
  17   kmem:kmem_cache_alloc_node               [Tracepoint event]
  18   kmem:kfree                               [Tracepoint event]
  19   kmem:kmem_cache_free                     [Tracepoint event]
  20   kmem:mm_page_free                        [Tracepoint event]
  21   kmem:mm_page_free_batched                [Tracepoint event]
  22   kmem:mm_page_alloc                       [Tracepoint event]
  23   kmem:mm_page_alloc_zone_locked           [Tracepoint event]
  24   kmem:mm_page_pcpu_drain                  [Tracepoint event]
  25   kmem:mm_page_alloc_extfrag               [Tracepoint event]
  26
  27 Then any (or all) of the above event sources can be activated and
  28 measured. For example the page alloc/free properties of a 'hackbench
  29 run' are:
  30
  31  titan:~> perf stat -e kmem:mm_page_pcpu_drain -e kmem:mm_page_alloc
  32  -e kmem:mm_page_free_batched -e kmem:mm_page_free ./hackbench 10
  33  Time: 0.575
  34
  35  Performance counter stats for './hackbench 10':
  36
  37           13857  kmem:mm_page_pcpu_drain
  38           27576  kmem:mm_page_alloc
  39            6025  kmem:mm_page_free_batched
  40           20934  kmem:mm_page_free
  41
  42     0.613972165  seconds time elapsed
  43
  44 You can observe the statistical properties as well, by using the
  45 'repeat the workload N times' feature of perf stat:
  46
  47  titan:~> perf stat --repeat 5 -e kmem:mm_page_pcpu_drain -e
  48    kmem:mm_page_alloc -e kmem:mm_page_free_batched -e
  49    kmem:mm_page_free ./hackbench 10
  50  Time: 0.627
  51  Time: 0.644
  52  Time: 0.564
  53  Time: 0.559
  54  Time: 0.626
  55
  56  Performance counter stats for './hackbench 10' (5 runs):
  57
  58           12920  kmem:mm_page_pcpu_drain    ( +-   3.359% )
  59           25035  kmem:mm_page_alloc         ( +-   3.783% )
  60            6104  kmem:mm_page_free_batched  ( +-   0.934% )
  61           18376  kmem:mm_page_free          ( +-   4.941% )
  62
  63     0.643954516  seconds time elapsed   ( +-   2.363% )
  64
  65 Furthermore, these tracepoints can be used to sample the workload as
  66 well. For example the page allocations done by a 'git gc' can be
  67 captured the following way:
  68
  69  titan:~/git> perf record -f -e kmem:mm_page_alloc -c 1 ./git gc
  70  Counting objects: 1148, done.
  71  Delta compression using up to 2 threads.
  72  Compressing objects: 100% (450/450), done.
  73  Writing objects: 100% (1148/1148), done.
  74  Total 1148 (delta 690), reused 1148 (delta 690)
  75  [ perf record: Captured and wrote 0.267 MB perf.data (~11679 samples) ]
  76
  77 To check which functions generated page allocations:
  78
  79  titan:~/git> perf report
  80  # Samples: 10646
  81  #
  82  # Overhead          Command               Shared Object
  83  # ........  ...............  ..........................
  84  #
  85     23.57%       git-repack  /lib64/libc-2.5.so
  86     21.81%              git  /lib64/libc-2.5.so
  87     14.59%              git  ./git
  88     11.79%       git-repack  ./git
  89      7.12%              git  /lib64/ld-2.5.so
  90      3.16%       git-repack  /lib64/libpthread-2.5.so
  91      2.09%       git-repack  /bin/bash
  92      1.97%               rm  /lib64/libc-2.5.so
  93      1.39%               mv  /lib64/ld-2.5.so
  94      1.37%               mv  /lib64/libc-2.5.so
  95      1.12%       git-repack  /lib64/ld-2.5.so
  96      0.95%               rm  /lib64/ld-2.5.so
  97      0.90%  git-update-serv  /lib64/libc-2.5.so
  98      0.73%  git-update-serv  /lib64/ld-2.5.so
  99      0.68%             perf  /lib64/libpthread-2.5.so
 100      0.64%       git-repack  /usr/lib64/libz.so.1.2.3
 101
 102 Or to see it on a more finegrained level:
 103
 104 titan:~/git> perf report --sort comm,dso,symbol
 105 # Samples: 10646
 106 #
 107 # Overhead          Command               Shared Object  Symbol
 108 # ........  ...............  ..........................  ......
 109 #
 110      9.35%       git-repack  ./git                       [.] insert_obj_hash
 111      9.12%              git  ./git                       [.] insert_obj_hash
 112      7.31%              git  /lib64/libc-2.5.so          [.] memcpy
 113      6.34%       git-repack  /lib64/libc-2.5.so          [.] _int_malloc
 114      6.24%       git-repack  /lib64/libc-2.5.so          [.] memcpy
 115      5.82%       git-repack  /lib64/libc-2.5.so          [.] __GI___fork
 116      5.47%              git  /lib64/libc-2.5.so          [.] _int_malloc
 117      2.99%              git  /lib64/libc-2.5.so          [.] memset
 118
 119 Furthermore, call-graph sampling can be done too, of page
 120 allocations - to see precisely what kind of page allocations there
 121 are:
 122
 123  titan:~/git> perf record -f -g -e kmem:mm_page_alloc -c 1 ./git gc
 124  Counting objects: 1148, done.
 125  Delta compression using up to 2 threads.
 126  Compressing objects: 100% (450/450), done.
 127  Writing objects: 100% (1148/1148), done.
 128  Total 1148 (delta 690), reused 1148 (delta 690)
 129  [ perf record: Captured and wrote 0.963 MB perf.data (~42069 samples) ]
 130
 131  titan:~/git> perf report -g
 132  # Samples: 10686
 133  #
 134  # Overhead          Command               Shared Object
 135  # ........  ...............  ..........................
 136  #
 137     23.25%       git-repack  /lib64/libc-2.5.so
 138                 |
 139                 |--50.00%-- _int_free
 140                 |
 141                 |--37.50%-- __GI___fork
 142                 |          make_child
 143                 |
 144                 |--12.50%-- ptmalloc_unlock_all2
 145                 |          make_child
 146                 |
 147                  --6.25%-- __GI_strcpy
 148     21.61%              git  /lib64/libc-2.5.so
 149                 |
 150                 |--30.00%-- __GI_read
 151                 |          |
 152                 |           --83.33%-- git_config_from_file
 153                 |                     git_config
 154                 |                     |
 155    [...]
 156
 157 Or you can observe the whole system's page allocations for 10
 158 seconds:
 159
 160 titan:~/git> perf stat -a -e kmem:mm_page_pcpu_drain -e
 161 kmem:mm_page_alloc -e kmem:mm_page_free_batched -e
 162 kmem:mm_page_free sleep 10
 163
 164  Performance counter stats for 'sleep 10':
 165
 166          171585  kmem:mm_page_pcpu_drain
 167          322114  kmem:mm_page_alloc
 168           73623  kmem:mm_page_free_batched
 169          254115  kmem:mm_page_free
 170
 171    10.000591410  seconds time elapsed
 172
 173 Or observe how fluctuating the page allocations are, via statistical
 174 analysis done over ten 1-second intervals:
 175
 176  titan:~/git> perf stat --repeat 10 -a -e kmem:mm_page_pcpu_drain -e
 177    kmem:mm_page_alloc -e kmem:mm_page_free_batched -e
 178    kmem:mm_page_free sleep 1
 179
 180  Performance counter stats for 'sleep 1' (10 runs):
 181
 182           17254  kmem:mm_page_pcpu_drain    ( +-   3.709% )
 183           34394  kmem:mm_page_alloc         ( +-   4.617% )
 184            7509  kmem:mm_page_free_batched  ( +-   4.820% )
 185           25653  kmem:mm_page_free          ( +-   3.672% )
 186
 187     1.058135029  seconds time elapsed   ( +-   3.089% )
 188
 189 Or you can annotate the recorded 'git gc' run on a per symbol basis
 190 and check which instructions/source-code generated page allocations:
 191
 192  titan:~/git> perf annotate __GI___fork
 193  ------------------------------------------------
 194   Percent |      Source code & Disassembly of libc-2.5.so
 195  ------------------------------------------------
 196           :
 197           :
 198           :      Disassembly of section .plt:
 199           :      Disassembly of section .text:
 200           :
 201           :      00000031a2e95560 <__fork>:
 202  [...]
 203      0.00 :        31a2e95602:   b8 38 00 00 00          mov    $0x38,%eax
 204      0.00 :        31a2e95607:   0f 05                   syscall
 205     83.42 :        31a2e95609:   48 3d 00 f0 ff ff       cmp    $0xfffffffffffff000,%rax
 206      0.00 :        31a2e9560f:   0f 87 4d 01 00 00       ja     31a2e95762 <__fork+0x202>
 207      0.00 :        31a2e95615:   85 c0                   test   %eax,%eax
 208
 209 ( this shows that 83.42% of __GI___fork's page allocations come from
 210   the 0x38 system call it performs. )
 211
 212 etc. etc. - a lot more is possible. I could list a dozen of
 213 other different usecases straight away - neither of which is
 214 possible via /proc/vmstat.
 215
 216 /proc/vmstat is not in the same league really, in terms of
 217 expressive power of system analysis and performance
 218 analysis.
 219
 220 All that the above results needed were those new tracepoints
 221 in include/tracing/events/kmem.h.
 222
 223         Ingo
 224
 225