Carlos R. Mafra | c24b513 | 2009-08-05 20:53:34 +0200 | [diff] [blame] | 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_direct [Tracepoint event] |
| 21 | kmem:mm_pagevec_free [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_pagevec_free -e kmem:mm_page_free_direct ./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_pagevec_free |
| 40 | 20934 kmem:mm_page_free_direct |
| 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_pagevec_free -e |
| 49 | kmem:mm_page_free_direct ./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_pagevec_free ( +- 0.934% ) |
| 61 | 18376 kmem:mm_page_free_direct ( +- 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_pagevec_free -e |
| 162 | kmem:mm_page_free_direct 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_pagevec_free |
| 169 | 254115 kmem:mm_page_free_direct |
| 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_pagevec_free -e |
| 178 | kmem:mm_page_free_direct 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_pagevec_free ( +- 4.820% ) |
| 185 | 25653 kmem:mm_page_free_direct ( +- 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 | |