How to install perf ( kernel performance monitor ) on Ubuntu , CentOS

[ install perf on Ubuntu 12.04 ]

# tail -1 /etc/lsb-release DISTRIB_DESCRIPTION="Ubuntu 12.04 LTS" # uname -ri 3.2.0-25-generic x86_64

install

# apt-get install linux-tools -y

monitor a chromium-browser process.

fly chromium-browser via perf like this

on GUI , issue the following command

# perf record chromium-browser

nnn, there appears a waning message.. , kernel pointer ??

Kernel address maps (/proc/{kallsyms,modules}) were restricted.                                                                         Check /proc/sys/kernel/kptr_restrict before running 'perf record'.

to solve this:

# cat /proc/sys/kernel/kptr_restrict 1 # echo 0 > /proc/sys/kernel/kptr_restrict # cat /proc/sys/kernel/kptr_restrict 0 # cat /proc/kallsyms  | head -10 0000000000000000 D irq_stack_union 0000000000000000 D __per_cpu_start 0000000000004000 D gdt_page 0000000000005000 d exception_stacks 000000000000b000 d tlb_vector_offset 000000000000b080 d cpu_loops_per_jiffy 000000000000b0c0 D xen_vcpu_info 000000000000b100 D xen_vcpu 000000000000b108 d idt_desc 000000000000b118 d xen_cr0_value

try again.

Okay , there are no warning messages.

check the result file

$ perf report Events: 3K cycles 73.37%  chromium-browse  [unknown]                [.] 0x7f613588d718  2.80%  chromium-browse  [kernel.kallsyms]        [k] native_write_msr_safe  1.25%  chromium-browse  [kernel.kallsyms]        [k] page_fault  1.02%  chromium-browse  libfreebl3.so            [.] 0x27d5e <snip>

you can specify events you’d like to collect.

# perf list List of pre-defined events (to be used in -e):  cpu-cycles OR cycles                               [Hardware event]  stalled-cycles-frontend OR idle-cycles-frontend    [Hardware event]  stalled-cycles-backend OR idle-cycles-backend      [Hardware event]  instructions                                       [Hardware event]  cache-references                                   [Hardware event]  cache-misses                                       [Hardware event]  branch-instructions OR branches                    [Hardware event]  branch-misses                                      [Hardware event]  bus-cycles                                         [Hardware event]  cpu-clock                                          [Software event]  task-clock                                         [Software event]  page-faults OR faults                              [Software event]  minor-faults                                       [Software event]  major-faults                                       [Software event]  context-switches OR cs                             [Software event]  cpu-migrations OR migrations                       [Software event]

$ perf record -e cpu-cycles -e cpu-clock -e context-switches chromium-browser

$ perf report

perf stat

$ perf stat chromium-browser Performance counter stats for 'chromium-browser':       2349.438123 task-clock                #    0.211 CPUs utilized                       6,912 context-switches          #    0.003 M/sec                                 635 CPU-migrations            #    0.000 M/sec                              87,726 page-faults               #    0.037 M/sec                       4,824,461,601 cycles                    #    2.053 GHz                     [83.95%]     2,375,916,048 stalled-cycles-frontend   #   49.25% frontend cycles idle    [82.50%]     1,467,114,781 stalled-cycles-backend    #   30.41% backend  cycles idle    [67.95%]     5,208,469,775 instructions              #    1.08  insns per cycle                                                     #    0.46  stalled cycles per insn [83.95%]     1,091,803,683 branches                  #  464.708 M/sec                   [82.98%]        31,296,254 branch-misses             #    2.87% of all branches         [83.44%]      11.149857740 seconds time elapsed

[ install perf on CentOS6 ]

# cat /etc/centos-release CentOS release 6.2 (Final) # uname -ri 2.6.32-220.17.1.el6.x86_64 x86_64 # yum install -y perf

perf : Performance monitoring for the Linux kernel

- about perf

https://perf.wiki.kernel.org/index.php/Tutorial

[root@fc17 ~]# cat /etc/fedora-release Fedora release 17 (Beefy Miracle) [root@fc17 ~]# uname -ri 3.4.0-1.fc17.x86_64 x86_64

install perf via yum

[root@fc17 ~]# yum install perf -y

[ example usage ]

# perf top

you can see which processes affect linux kernel performance.

you could also see the detail of specific process info.

select a process -> type enter




here’s an output of map details.




- perf record and perf repot

record kernel performance into a file

[root@fc17 ~]# perf record -a sleep 5 [ perf record: Woken up 5 times to write data ] [ perf record: Captured and wrote 2.234 MB perf.data (~97596 samples) ]

info is written into perf.data file
.
read that data
“perf report” command reads info from perf.data file.

# perf report Events: 14K cycles                                                               12.32%          swapper  [kernel.kallsyms]                   [k] native_safe_ha  5.25%              dig  ld-2.15.so                          [.] do_lookup_x  2.75%              dig  ld-2.15.so                          [.] strcmp  2.37%           chrome  libflashplayer.so                   [.] 0x00000000003d  1.75%              dig  [kernel.kallsyms]                   [k] clear_page_c  1.66%              dig  [kernel.kallsyms]                   [k] page_fault  1.49%              dig  libc-2.15.so                        [.] __memset_x86_6  1.38%              dig  ld-2.15.so                          [.] _dl_lookup_sym  1.32%   gnome-terminal  libvte2_90.so.9.3200.1              [.] 0x000000000003  1.20%              dig  [kernel.kallsyms]                   [k] unmap_single_v  1.13%              dig  ld-2.15.so                          [.] _dl_relocate_o  0.97%      gnome-shell  r600_dri.so                         [.] 0x000000000005

# perf report --sort=cpu Events: 14K cycles                                                               26.35%  3 25.43%  0 24.50%  1 23.73%  2

- monitor a specific process

check PID you want to monitor

[root@fc17 ~]# ps aux | grep unbound | grep -v grep unbound   1772  0.1  0.1  51596  9884 ?        Ss   00:23   0:03 unbound -c /etc/unbound/unbound.conf

[root@fc17 ~]# perf stat -p 1772 sleep 5 Performance counter stats for process id '1772':       2619.300788 task-clock                #    0.524 CPUs utilized           [100.00%]             2,294 context-switches          #    0.876 K/sec                   [100.00%]               147 CPU-migrations            #    0.056 K/sec                   [100.00%]               252 page-faults               #    0.096 K/sec                       2,104,000,503 cycles                    #    0.803 GHz                     [49.81%]       576,211,571 stalled-cycles-frontend   #   27.39% frontend cycles idle    [49.55%]       702,855,782 stalled-cycles-backend    #   33.41% backend  cycles idle    [49.55%]     1,512,642,754 instructions              #    0.72  insns per cycle                                                     #    0.46  stalled cycles per insn [50.19%]       296,171,521 branches                  #  113.073 M/sec                   [50.45%]        23,773,996 branch-misses             #    8.03% of all branches         [50.45%]       5.001789912 seconds time elapsed

interesting tool.