What should I do if I want to get Perf samples and Read Performance Counter at the same time?

What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[Peipei Wang]

Hi all,

Here is my issue, looking forward to the suggestions from you guys. I
have a java process and I can get the process id. I want to get the
information about the event cache-misses for this process at the
sampling rate 100(For this, I use perf record -e cache-misses -c 100
-p pid) . And I also want to know how many cache-misses events occurs
for each second(For this I use watch perf stat -e cache-misses -p pid
sleep 1)

Any one has ideas about how to get both of them?
Thanks for all your help!


Yours,
Wang Peipei

Re: What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[David Ahern]

On 8/28/13 11:21 AM, Peipei Wang wrote:
> Hi all,
>
> Here is my issue, looking forward to the suggestions from you guys. I
> have a java process and I can get the process id. I want to get the
> information about the event cache-misses for this process at the
> sampling rate 100(For this, I use perf record -e cache-misses -c 100 -p
> pid) . And I also want to know how many cache-misses events occurs for
> each second(For this I use watch perf stat -e cache-misses -p pid sleep 1)
>
> Any one has ideas about how to get both of them?

I believe you will need to write your own perf command to do that

David

Re: What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[Mikhail Kulemin]

There is a possibility to use perf script with python or perl.

1) Record needed events with perf record

[mih@home perf]$ perf record -e cache-misses ls
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.013 MB perf.data (~556 samples) ]

2) You can see actual info about events using 'perf script'

[mih@home perf]$ perf script
# ========
# captured on: Thu Aug 29 08:33:20 2013

<...some info ommited...>

# cmdline : /usr/bin/perf record -e cache-misses ls
# event : name = cache-misses, type = 0, config = 0x3, config1 = 0x0, config2 =
# HEADER_CPU_TOPOLOGY info available, use -I to display
# HEADER_NUMA_TOPOLOGY info available, use -I to display
# pmu mappings: cpu = 4, software = 1, tracepoint = 2, uncore_cbox_0 = 6, uncore
# ========
#
           :2717  2717  1148.761687: cache-misses:  ffffffff8107cc0d task_tgid_n
           :2717  2717  1148.761690: cache-misses:  ffffffff8107cc0d task_tgid_n
              ls  2717  1148.761693: cache-misses:  ffffffff8111e2f6 __perf_even
              ls  2717  1148.761696: cache-misses:  ffffffff8106e410 flush_signa
              ls  2717  1148.761702: cache-misses:  ffffffff812f1867 clear_page_
              ls  2717  1148.761834: cache-misses:        3a5bc17f70 open64 (/us
              ls  2717  1148.762713: cache-misses:  ffffffff8139c103 do_output_c
(END)

3) You can see time of each sample but can not see number of events in
each sample. By default there can be different number of events in
each sample. To get such info you can use custom script. First of all
generate basic template

[mih@home perf]$ perf script -g py
generated Python script: perf-script.py

This script handled only tracepoint events. (For more information
about scripting for perf read "man perf-script"). To handle hw events
you can use function process_event(). This function is not well
documented but rather simple.  There is one difficulty - you need
parse C structure of perf_sample to  get info about time and number of
events. I use scruct python module and use perf_struct defenition (for
example from http://lxr.linux.no/#linux+v3.10.9/tools/perf/util/event.h#L83
)

My perf-script.py :

import os
import sys
import struct

sys.path.append(os.environ['PERF_EXEC_PATH'] + \
    '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')

from perf_trace_context import *
from Core import *


def trace_begin():
    print "in trace_begin"

def trace_end():
    print "in trace_end"

def process_event(param_dict):
        event_attr = param_dict["attr"]
        sample     = param_dict["sample"]
        raw_buf    = param_dict["raw_buf"]
        comm       = param_dict["comm"]
        name       = param_dict["ev_name"]

        if (param_dict.has_key("symbol")):
                symbol = param_dict["symbol"]
        else:
                symbol = "Unknown_symbol"

        # Extract counts value from perf_sample struct (Dirty.
Addition to API neeeded.)
        (counts,) = struct.unpack("=Q",sample[48:56])
        (time,) = struct.unpack("=Q",sample[16:24])

        print name, counts, time, symbol, comm


4) Run your custom script perf script -s perf-script.py
[mih@home perf]$ perf script -s perf-script.py
in trace_begin
cache-misses 1 1148761687085 task_tgid_nr_ns :2717
cache-misses 1 1148761690924 task_tgid_nr_ns :2717
cache-misses 9 1148761693240 __perf_event__output_id_sample ls
cache-misses 121 1148761696003 flush_signal_handlers ls
cache-misses 1467 1148761702371 clear_page_c_e ls
cache-misses 8503 1148761834505 open64 ls
cache-misses 8469 1148762713882 do_output_char ls
in trace_end

Time in nanosec - third field
number of events in sample second field

5) If you use -c option when running perf record script will report
only 1 event perf sample (but actial number of events per sample is
determine by -c parameter)

Mikhail

2013/8/29 David Ahern <dsahern@gmail.com>:
> On 8/28/13 11:21 AM, Peipei Wang wrote:
>>
>> Hi all,
>>
>> Here is my issue, looking forward to the suggestions from you guys. I
>> have a java process and I can get the process id. I want to get the
>> information about the event cache-misses for this process at the
>> sampling rate 100(For this, I use perf record -e cache-misses -c 100 -p
>> pid) . And I also want to know how many cache-misses events occurs for
>> each second(For this I use watch perf stat -e cache-misses -p pid sleep 1)
>>
>> Any one has ideas about how to get both of them?
>
>
> I believe you will need to write your own perf command to do that
>
> David
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-perf-users"
> in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[Mikhail Kulemin]

Sorry, I didn`t mention that I described perf behaviour for latest
stable kernel 3.10.9.
For my examples I use perf on Fedora 19 with latest kernel available.

What disrtib are you use? Can you update it to latest version?

As far as I understand you are working with kernel 3.2.0 and perf
corresponding this version of kernel has different API for python.
I quickly look at sources
(http://lxr.linux.no/#linux+v3.2/tools/perf/util/scripting-engines/trace-event-python.c)
and could not find mention about process_event() function. It looks
like hw events handled with trace_unhandled() function as another
kinds of events. So you should use another script for tracing with
this version of kernel.

According scripting support: It looks like you need to recompile perf
package (or package that has perf utility) with scripting enable. You
need to check how package maintainer compile perf in you distribution.
For example in Fedora perf located in package perf when compiling
kernel (because perf sources located in kernel tree).
Perf compiled with make command:
make -C tools/perf -s V=1 WERROR=0 NO_LIBUNWIND=1
HAVE_CPLUS_DEMANGLE=1 NO_GTK2=1 NO_LIBNUMA=1 NO_STRLCPY=1 pref

Check such build scripts for your distribution, maybe Python scripting
disabled in build time (possible).


2013/8/29 Peipei Wang <wangpeipei.90@gmail.com>:
> Hi Mikhail,
>
> I got stuck when enabling Python scripting in perf. It tells me to install
> python and recompile perf after that. However, even after I removed the
> linux-tools at first, reinstalled python-dev amd reinstalled linux-tools,
> the perf script -g py still shows messages as follows:
>
> perf script -g py
> Python scripting not supported.  Install libpython and rebuild perf to
> enable it
> For example:
>   # apt-get install python-dev (ubuntu)
>   # yum install python-devel (Fedora)
>   etc.
> # ========
> # captured on: Thu Aug 29 11:16:01 2013
> # hostname : sif.thho.net
> # os release : 3.2.0-43-generic
> # perf version : 3.2.42
> # arch : x86_64
> # nrcpus online : 4
> # nrcpus avail : 4
> # cpudesc : Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
> # cpuid : GenuineIntel,6,42,7
> # total memory : 8133464 kB
> # cmdline : /usr/bin/perf_3.2.0-43 record -e cache-misses ls
> # event : name = cache-misses, type = 0, config = 0x3, config1 = 0x0,
> config2 =
> # HEADER_CPU_TOPOLOGY info available, use -I to display
> # HEADER_NUMA_TOPOLOGY info available, use -I to display
> # ========
>
>
> Best wishes.
> Yours,
> Wang Peipei
>
>
> On Thu, Aug 29, 2013 at 12:59 AM, Mikhail Kulemin <mihkulemin@gmail.com>
> wrote:
>>
>> There is a possibility to use perf script with python or perl.
>>
>> 1) Record needed events with perf record
>>
>> [mih@home perf]$ perf record -e cache-misses ls
>> [ perf record: Woken up 1 times to write data ]
>> [ perf record: Captured and wrote 0.013 MB perf.data (~556 samples) ]
>>
>> 2) You can see actual info about events using 'perf script'
>>
>> [mih@home perf]$ perf script
>> # ========
>> # captured on: Thu Aug 29 08:33:20 2013
>>
>> <...some info ommited...>
>>
>> # cmdline : /usr/bin/perf record -e cache-misses ls
>> # event : name = cache-misses, type = 0, config = 0x3, config1 = 0x0,
>> config2 =
>> # HEADER_CPU_TOPOLOGY info available, use -I to display
>> # HEADER_NUMA_TOPOLOGY info available, use -I to display
>> # pmu mappings: cpu = 4, software = 1, tracepoint = 2, uncore_cbox_0 = 6,
>> uncore
>> # ========
>> #
>>            :2717  2717  1148.761687: cache-misses:  ffffffff8107cc0d
>> task_tgid_n
>>            :2717  2717  1148.761690: cache-misses:  ffffffff8107cc0d
>> task_tgid_n
>>               ls  2717  1148.761693: cache-misses:  ffffffff8111e2f6
>> __perf_even
>>               ls  2717  1148.761696: cache-misses:  ffffffff8106e410
>> flush_signa
>>               ls  2717  1148.761702: cache-misses:  ffffffff812f1867
>> clear_page_
>>               ls  2717  1148.761834: cache-misses:        3a5bc17f70
>> open64 (/us
>>               ls  2717  1148.762713: cache-misses:  ffffffff8139c103
>> do_output_c
>> (END)
>>
>> 3) You can see time of each sample but can not see number of events in
>> each sample. By default there can be different number of events in
>> each sample. To get such info you can use custom script. First of all
>> generate basic template
>>
>> [mih@home perf]$ perf script -g py
>> generated Python script: perf-script.py
>>
>> This script handled only tracepoint events. (For more information
>> about scripting for perf read "man perf-script"). To handle hw events
>> you can use function process_event(). This function is not well
>> documented but rather simple.  There is one difficulty - you need
>> parse C structure of perf_sample to  get info about time and number of
>> events. I use scruct python module and use perf_struct defenition (for
>> example from
>> http://lxr.linux.no/#linux+v3.10.9/tools/perf/util/event.h#L83
>> )
>>
>> My perf-script.py :
>>
>> import os
>> import sys
>> import struct
>>
>> sys.path.append(os.environ['PERF_EXEC_PATH'] + \
>>     '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
>>
>> from perf_trace_context import *
>> from Core import *
>>
>>
>> def trace_begin():
>>     print "in trace_begin"
>>
>> def trace_end():
>>     print "in trace_end"
>>
>> def process_event(param_dict):
>>         event_attr = param_dict["attr"]
>>         sample     = param_dict["sample"]
>>         raw_buf    = param_dict["raw_buf"]
>>         comm       = param_dict["comm"]
>>         name       = param_dict["ev_name"]
>>
>>         if (param_dict.has_key("symbol")):
>>                 symbol = param_dict["symbol"]
>>         else:
>>                 symbol = "Unknown_symbol"
>>
>>         # Extract counts value from perf_sample struct (Dirty.
>> Addition to API neeeded.)
>>         (counts,) = struct.unpack("=Q",sample[48:56])
>>         (time,) = struct.unpack("=Q",sample[16:24])
>>
>>         print name, counts, time, symbol, comm
>>
>>
>> 4) Run your custom script perf script -s perf-script.py
>> [mih@home perf]$ perf script -s perf-script.py
>> in trace_begin
>> cache-misses 1 1148761687085 task_tgid_nr_ns :2717
>> cache-misses 1 1148761690924 task_tgid_nr_ns :2717
>> cache-misses 9 1148761693240 __perf_event__output_id_sample ls
>> cache-misses 121 1148761696003 flush_signal_handlers ls
>> cache-misses 1467 1148761702371 clear_page_c_e ls
>> cache-misses 8503 1148761834505 open64 ls
>> cache-misses 8469 1148762713882 do_output_char ls
>> in trace_end
>>
>> Time in nanosec - third field
>> number of events in sample second field
>>
>> 5) If you use -c option when running perf record script will report
>> only 1 event perf sample (but actial number of events per sample is
>> determine by -c parameter)
>>
>> Mikhail
>>
>> 2013/8/29 David Ahern <dsahern@gmail.com>:
>> > On 8/28/13 11:21 AM, Peipei Wang wrote:
>> >>
>> >> Hi all,
>> >>
>> >> Here is my issue, looking forward to the suggestions from you guys. I
>> >> have a java process and I can get the process id. I want to get the
>> >> information about the event cache-misses for this process at the
>> >> sampling rate 100(For this, I use perf record -e cache-misses -c 100 -p
>> >> pid) . And I also want to know how many cache-misses events occurs for
>> >> each second(For this I use watch perf stat -e cache-misses -p pid sleep
>> >> 1)
>> >>
>> >> Any one has ideas about how to get both of them?
>> >
>> >
>> > I believe you will need to write your own perf command to do that
>> >
>> > David
>> >
>> >
>> > --
>> > To unsubscribe from this list: send the line "unsubscribe
>> > linux-perf-users"
>> > in
>> > the body of a message to majordomo@vger.kernel.org
>> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
>
>

Re: What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[Mikhail Kulemin]

You can ignote message about Android. As far as I understand form
output you should see compiled executable of perf in tools/perf
directory. Please check it.

2013/8/30 Peipei Wang <wangpeipei.90@gmail.com>:
> Hi Mikhail,
>
> After installing all these dependency libraries, still get the following
> error. I have no idea about the whole thing of "Android/api-level.h".
>
> peipei@sif:~/linux-perf$ sudo make -C tools/perf -s V=1 WERROR=0
> NO_LIBUNWIND=1 HAVE_CPLUS_DEMANGLE=1 NO_GTK2=1 NO_LIBNUMA=1 NO_STRLCPY=1
>     CHK -fstack-protector-all
>     CHK -Wstack-protector
>     CHK -Wvolatile-register-var
>     CHK -D_FORTIFY_SOURCE=2
>     CHK bionic
> <stdin>:1:31: fatal error: android/api-level.h: No such file or directory
> compilation terminated.
>     CHK libelf
>     CHK libdw
>     CHK -DLIBELF_MMAP
>     CHK libaudit
>     CHK libslang
>     CHK perl
>     CHK python
>     CHK python version
>     CHK -DHAVE_ON_EXIT
>     CHK -DBACKTRACE_SUPPORT
>     * new build flags or prefix
>
> Best wishes.
> Yours,
> Wang Peipei
>
>
> On Thu, Aug 29, 2013 at 11:40 PM, Mikhail Kulemin <mihkulemin@gmail.com>
> wrote:
>>
>> About perf version: yes, you need to use perf version corresponding to
>> the linux kernel version.
>>
>> According perf compilation: you need to install smth like
>> "binutils-dev" package to resolve dependency form library and try to
>> compile perf
>>
>> 2013/8/30 Peipei Wang <wangpeipei.90@gmail.com>:
>> > Hi Mikhail,
>> >
>> > I get a git clone from
>> > https://git.kernel.org/cgit/linux/kernel/git/namhyung/linux-perf.git/,
>> > then
>> > I checkout the tag v3.10, install libperl-dev, and use exactly the
>> > command
>> > you give to me, but there are still a few problems:
>> >
>> > peipei@sif:~/linux-perf$ sudo make -C tools/perf -s V=1 WERROR=0
>> > NO_LIBUNWIND=1 HAVE_CPLUS_DEMANGLE=1 NO_GTK2=1 NO_LIBNUMA=1 NO_STRLCPY=1
>> >     CHK -fstack-protector-all
>> >     CHK -Wstack-protector
>> >     CHK -Wvolatile-register-var
>> >     CHK -D_FORTIFY_SOURCE=2
>> >     CHK bionic
>> > <stdin>:1:31: fatal error: android/api-level.h: No such file or
>> > directory
>> > compilation terminated.
>> >     CHK libelf
>> >     CHK libdw
>> > <stdin>:1:19: fatal error: dwarf.h: No such file or directory
>> > compilation terminated.
>> > Makefile:603: No libdw.h found or old libdw.h found or elfutils is older
>> > than 0.138, disables dwarf support. Please install new
>> > elfutils-devel/libdw-dev
>> >     CHK -DLIBELF_MMAP
>> >     CHK libaudit
>> > <stdin>:1:22: fatal error: libaudit.h: No such file or directory
>> > compilation terminated.
>> > Makefile:681: No libaudit.h found, disables 'trace' tool, please install
>> > audit-libs-devel or libaudit-dev
>> >     CHK libslang
>> >     CHK perl
>> >     CHK python
>> >     CHK python version
>> >     CHK -DHAVE_ON_EXIT
>> >     CHK -DBACKTRACE_SUPPORT
>> >     * new build flags or prefix
>> > /usr/bin/ld: cannot find -liberty
>> > collect2: ld returned 1 exit status
>> > make: *** [perf] Error 1
>> >
>> > What should I do next??
>> >
>> >
>> > Best wishes.
>> > Yours,
>> > Wang Peipei
>> >
>> >
>> > On Thu, Aug 29, 2013 at 9:14 PM, Peipei Wang <wangpeipei.90@gmail.com>
>> > wrote:
>> >>
>> >> Hi Mikhail,
>> >>
>> >> I am using ubuntu 12.04 and I just update my kernel version to the same
>> >> as
>> >> yours:3.10.9.Then I got problems about which perf version I should use.
>> >> The
>> >> error is like this:
>> >>
>> >> peipei@sif:~/linux_install$ perf script
>> >> perf_3.10.9-031009 not found
>> >> You may need to install linux-tools-3.10.9-031009
>> >>
>> >>
>> >> What version do you use?? I could not found a corresponding version of
>> >> linux-tools on ubuntu repository, neither on google results. Please how
>> >> I
>> >> can get an appropriate perf version to use.
>> >>
>> >>
>> >> Best wishes.
>> >> Yours,
>> >> Wang Peipei
>> >>
>> >>
>> >> On Thu, Aug 29, 2013 at 2:58 PM, Mikhail Kulemin <mihkulemin@gmail.com>
>> >> wrote:
>> >>>
>> >>> Sorry, I didn`t mention that I described perf behaviour for latest
>> >>> stable kernel 3.10.9.
>> >>> For my examples I use perf on Fedora 19 with latest kernel available.
>> >>>
>> >>> What disrtib are you use? Can you update it to latest version?
>> >>>
>> >>> As far as I understand you are working with kernel 3.2.0 and perf
>> >>> corresponding this version of kernel has different API for python.
>> >>> I quickly look at sources
>> >>>
>> >>>
>> >>> (http://lxr.linux.no/#linux+v3.2/tools/perf/util/scripting-engines/trace-event-python.c)
>> >>> and could not find mention about process_event() function. It looks
>> >>> like hw events handled with trace_unhandled() function as another
>> >>> kinds of events. So you should use another script for tracing with
>> >>> this version of kernel.
>> >>>
>> >>> According scripting support: It looks like you need to recompile perf
>> >>> package (or package that has perf utility) with scripting enable. You
>> >>> need to check how package maintainer compile perf in you distribution.
>> >>> For example in Fedora perf located in package perf when compiling
>> >>> kernel (because perf sources located in kernel tree).
>> >>> Perf compiled with make command:
>> >>> make -C tools/perf -s V=1 WERROR=0 NO_LIBUNWIND=1
>> >>> HAVE_CPLUS_DEMANGLE=1 NO_GTK2=1 NO_LIBNUMA=1 NO_STRLCPY=1 pref
>> >>>
>> >>> Check such build scripts for your distribution, maybe Python scripting
>> >>> disabled in build time (possible).
>> >>>
>> >>>
>> >>> 2013/8/29 Peipei Wang <wangpeipei.90@gmail.com>:
>> >>> > Hi Mikhail,
>> >>> >
>> >>> > I got stuck when enabling Python scripting in perf. It tells me to
>> >>> > install
>> >>> > python and recompile perf after that. However, even after I removed
>> >>> > the
>> >>> > linux-tools at first, reinstalled python-dev amd reinstalled
>> >>> > linux-tools,
>> >>> > the perf script -g py still shows messages as follows:
>> >>> >
>> >>> > perf script -g py
>> >>> > Python scripting not supported.  Install libpython and rebuild perf
>> >>> > to
>> >>> > enable it
>> >>> > For example:
>> >>> >   # apt-get install python-dev (ubuntu)
>> >>> >   # yum install python-devel (Fedora)
>> >>> >   etc.
>> >>> > # ========
>> >>> > # captured on: Thu Aug 29 11:16:01 2013
>> >>> > # hostname : sif.thho.net
>> >>> > # os release : 3.2.0-43-generic
>> >>> > # perf version : 3.2.42
>> >>> > # arch : x86_64
>> >>> > # nrcpus online : 4
>> >>> > # nrcpus avail : 4
>> >>> > # cpudesc : Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
>> >>> > # cpuid : GenuineIntel,6,42,7
>> >>> > # total memory : 8133464 kB
>> >>> > # cmdline : /usr/bin/perf_3.2.0-43 record -e cache-misses ls
>> >>> > # event : name = cache-misses, type = 0, config = 0x3, config1 =
>> >>> > 0x0,
>> >>> > config2 =
>> >>> > # HEADER_CPU_TOPOLOGY info available, use -I to display
>> >>> > # HEADER_NUMA_TOPOLOGY info available, use -I to display
>> >>> > # ========
>> >>> >
>> >>> >
>> >>> > Best wishes.
>> >>> > Yours,
>> >>> > Wang Peipei
>> >>> >
>> >>> >
>> >>> > On Thu, Aug 29, 2013 at 12:59 AM, Mikhail Kulemin
>> >>> > <mihkulemin@gmail.com>
>> >>> > wrote:
>> >>> >>
>> >>> >> There is a possibility to use perf script with python or perl.
>> >>> >>
>> >>> >> 1) Record needed events with perf record
>> >>> >>
>> >>> >> [mih@home perf]$ perf record -e cache-misses ls
>> >>> >> [ perf record: Woken up 1 times to write data ]
>> >>> >> [ perf record: Captured and wrote 0.013 MB perf.data (~556 samples)
>> >>> >> ]
>> >>> >>
>> >>> >> 2) You can see actual info about events using 'perf script'
>> >>> >>
>> >>> >> [mih@home perf]$ perf script
>> >>> >> # ========
>> >>> >> # captured on: Thu Aug 29 08:33:20 2013
>> >>> >>
>> >>> >> <...some info ommited...>
>> >>> >>
>> >>> >> # cmdline : /usr/bin/perf record -e cache-misses ls
>> >>> >> # event : name = cache-misses, type = 0, config = 0x3, config1 =
>> >>> >> 0x0,
>> >>> >> config2 =
>> >>> >> # HEADER_CPU_TOPOLOGY info available, use -I to display
>> >>> >> # HEADER_NUMA_TOPOLOGY info available, use -I to display
>> >>> >> # pmu mappings: cpu = 4, software = 1, tracepoint = 2,
>> >>> >> uncore_cbox_0 =
>> >>> >> 6,
>> >>> >> uncore
>> >>> >> # ========
>> >>> >> #
>> >>> >>            :2717  2717  1148.761687: cache-misses:
>> >>> >> ffffffff8107cc0d
>> >>> >> task_tgid_n
>> >>> >>            :2717  2717  1148.761690: cache-misses:
>> >>> >> ffffffff8107cc0d
>> >>> >> task_tgid_n
>> >>> >>               ls  2717  1148.761693: cache-misses:
>> >>> >> ffffffff8111e2f6
>> >>> >> __perf_even
>> >>> >>               ls  2717  1148.761696: cache-misses:
>> >>> >> ffffffff8106e410
>> >>> >> flush_signa
>> >>> >>               ls  2717  1148.761702: cache-misses:
>> >>> >> ffffffff812f1867
>> >>> >> clear_page_
>> >>> >>               ls  2717  1148.761834: cache-misses:
>> >>> >> 3a5bc17f70
>> >>> >> open64 (/us
>> >>> >>               ls  2717  1148.762713: cache-misses:
>> >>> >> ffffffff8139c103
>> >>> >> do_output_c
>> >>> >> (END)
>> >>> >>
>> >>> >> 3) You can see time of each sample but can not see number of events
>> >>> >> in
>> >>> >> each sample. By default there can be different number of events in
>> >>> >> each sample. To get such info you can use custom script. First of
>> >>> >> all
>> >>> >> generate basic template
>> >>> >>
>> >>> >> [mih@home perf]$ perf script -g py
>> >>> >> generated Python script: perf-script.py
>> >>> >>
>> >>> >> This script handled only tracepoint events. (For more information
>> >>> >> about scripting for perf read "man perf-script"). To handle hw
>> >>> >> events
>> >>> >> you can use function process_event(). This function is not well
>> >>> >> documented but rather simple.  There is one difficulty - you need
>> >>> >> parse C structure of perf_sample to  get info about time and number
>> >>> >> of
>> >>> >> events. I use scruct python module and use perf_struct defenition
>> >>> >> (for
>> >>> >> example from
>> >>> >> http://lxr.linux.no/#linux+v3.10.9/tools/perf/util/event.h#L83
>> >>> >> )
>> >>> >>
>> >>> >> My perf-script.py :
>> >>> >>
>> >>> >> import os
>> >>> >> import sys
>> >>> >> import struct
>> >>> >>
>> >>> >> sys.path.append(os.environ['PERF_EXEC_PATH'] + \
>> >>> >>     '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
>> >>> >>
>> >>> >> from perf_trace_context import *
>> >>> >> from Core import *
>> >>> >>
>> >>> >>
>> >>> >> def trace_begin():
>> >>> >>     print "in trace_begin"
>> >>> >>
>> >>> >> def trace_end():
>> >>> >>     print "in trace_end"
>> >>> >>
>> >>> >> def process_event(param_dict):
>> >>> >>         event_attr = param_dict["attr"]
>> >>> >>         sample     = param_dict["sample"]
>> >>> >>         raw_buf    = param_dict["raw_buf"]
>> >>> >>         comm       = param_dict["comm"]
>> >>> >>         name       = param_dict["ev_name"]
>> >>> >>
>> >>> >>         if (param_dict.has_key("symbol")):
>> >>> >>                 symbol = param_dict["symbol"]
>> >>> >>         else:
>> >>> >>                 symbol = "Unknown_symbol"
>> >>> >>
>> >>> >>         # Extract counts value from perf_sample struct (Dirty.
>> >>> >> Addition to API neeeded.)
>> >>> >>         (counts,) = struct.unpack("=Q",sample[48:56])
>> >>> >>         (time,) = struct.unpack("=Q",sample[16:24])
>> >>> >>
>> >>> >>         print name, counts, time, symbol, comm
>> >>> >>
>> >>> >>
>> >>> >> 4) Run your custom script perf script -s perf-script.py
>> >>> >> [mih@home perf]$ perf script -s perf-script.py
>> >>> >> in trace_begin
>> >>> >> cache-misses 1 1148761687085 task_tgid_nr_ns :2717
>> >>> >> cache-misses 1 1148761690924 task_tgid_nr_ns :2717
>> >>> >> cache-misses 9 1148761693240 __perf_event__output_id_sample ls
>> >>> >> cache-misses 121 1148761696003 flush_signal_handlers ls
>> >>> >> cache-misses 1467 1148761702371 clear_page_c_e ls
>> >>> >> cache-misses 8503 1148761834505 open64 ls
>> >>> >> cache-misses 8469 1148762713882 do_output_char ls
>> >>> >> in trace_end
>> >>> >>
>> >>> >> Time in nanosec - third field
>> >>> >> number of events in sample second field
>> >>> >>
>> >>> >> 5) If you use -c option when running perf record script will report
>> >>> >> only 1 event perf sample (but actial number of events per sample is
>> >>> >> determine by -c parameter)
>> >>> >>
>> >>> >> Mikhail
>> >>> >>
>> >>> >> 2013/8/29 David Ahern <dsahern@gmail.com>:
>> >>> >> > On 8/28/13 11:21 AM, Peipei Wang wrote:
>> >>> >> >>
>> >>> >> >> Hi all,
>> >>> >> >>
>> >>> >> >> Here is my issue, looking forward to the suggestions from you
>> >>> >> >> guys.
>> >>> >> >> I
>> >>> >> >> have a java process and I can get the process id. I want to get
>> >>> >> >> the
>> >>> >> >> information about the event cache-misses for this process at the
>> >>> >> >> sampling rate 100(For this, I use perf record -e cache-misses -c
>> >>> >> >> 100 -p
>> >>> >> >> pid) . And I also want to know how many cache-misses events
>> >>> >> >> occurs
>> >>> >> >> for
>> >>> >> >> each second(For this I use watch perf stat -e cache-misses -p
>> >>> >> >> pid
>> >>> >> >> sleep
>> >>> >> >> 1)
>> >>> >> >>
>> >>> >> >> Any one has ideas about how to get both of them?
>> >>> >> >
>> >>> >> >
>> >>> >> > I believe you will need to write your own perf command to do that
>> >>> >> >
>> >>> >> > David
>> >>> >> >
>> >>> >> >
>> >>> >> > --
>> >>> >> > To unsubscribe from this list: send the line "unsubscribe
>> >>> >> > linux-perf-users"
>> >>> >> > in
>> >>> >> > the body of a message to majordomo@vger.kernel.org
>> >>> >> > More majordomo info at
>> >>> >> > http://vger.kernel.org/majordomo-info.html
>> >>> >
>> >>> >
>> >>
>> >>
>> >
>
>

Re: What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[David Ahern]

On 8/29/13 10:28 PM, Mikhail Kulemin wrote:
> You can ignote message about Android. As far as I understand form
> output you should see compiled executable of perf in tools/perf
> directory. Please check it.

Yes, you can ignore that message during the CHK - it's just probing for 
build options.

David

Re: What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[Mikhail Kulemin]

You should fix perf-script.py

replace   sys.path.append(os.environ['PERF_EXEC_PATH'] +
'/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
with

sys.path.append(os.getcwd()+"/tools/perf/scripts/python/
Perf-Trace-Util/lib/Perf/Trace")

this string instruct python interpreter where to find modules to load.


Please check that "perf script" command works correctly (output all
sample records from perf.data file)

Mikhail.

2013/8/30 Peipei Wang <wangpeipei.90@gmail.com>:
> Hi Mikhail,
>
> Yes, you are right. I could use perf now. But I get errors again running
> "perf script -s perf-script.py". Here it is:
> perf script -s perf-script.py
> Traceback (most recent call last):
>   File "perf-script.py", line 18, in <module>
>     from Core import *
> ImportError: No module named Core
>
> I do have the module Core.py in the directory of
> tools/perf/scripts/python/Perf-Trace-Util/lib/Perf/Trace
>
> Is there anything else I should pay attention to when using this?
>
> Best wishes.
> Yours,
> Wang Peipei
>
>
> On Fri, Aug 30, 2013 at 9:33 AM, David Ahern <dsahern@gmail.com> wrote:
>>
>> On 8/29/13 10:28 PM, Mikhail Kulemin wrote:
>>>
>>> You can ignote message about Android. As far as I understand form
>>> output you should see compiled executable of perf in tools/perf
>>> directory. Please check it.
>>
>>
>> Yes, you can ignore that message during the CHK - it's just probing for
>> build options.
>>
>> David
>
>

Re: What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[David Ahern]

On 8/30/13 11:15 AM, Mikhail Kulemin wrote:
> You should fix perf-script.py
>
> replace   sys.path.append(os.environ['PERF_EXEC_PATH'] +
> '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
> with

better to set/export PERF_EXEC_PATH before running the command.

David


>
> sys.path.append(os.getcwd()+"/tools/perf/scripts/python/
> Perf-Trace-Util/lib/Perf/Trace")
>
> this string instruct python interpreter where to find modules to load.
>
>
> Please check that "perf script" command works correctly (output all
> sample records from perf.data file)

Re: What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[Mikhail Kulemin]

Indeed, It is preferable. I`m miss it.

2013/8/30 David Ahern <dsahern@gmail.com>:
> On 8/30/13 11:15 AM, Mikhail Kulemin wrote:
>>
>> You should fix perf-script.py
>>
>> replace   sys.path.append(os.environ['PERF_EXEC_PATH'] +
>> '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
>> with
>
>
> better to set/export PERF_EXEC_PATH before running the command.
>
> David
>
>
>
>>
>> sys.path.append(os.getcwd()+"/tools/perf/scripts/python/
>> Perf-Trace-Util/lib/Perf/Trace")
>>
>> this string instruct python interpreter where to find modules to load.
>>
>>
>> Please check that "perf script" command works correctly (output all
>> sample records from perf.data file)
>
>

Re: What should I do if I want to get Perf samples and Read Performance Counter at the same time?

[Peipei Wang]

Thanks David.

That was I did at first. You know I compiled perf from git repository
and it is actually not an official version. So I copy all stuffs in
the directory of perf to
/usr/src/linux-header3-3.10.9-031009/tools/perf  and make a soft link
to /usr/bin/perf.

I read the perf-script.py, actually, and set the environment variable
PERF_EXEC_PATH in /etc/profile :
                        export PERF_EXEC_PATH=
/usr/src/linux-header3-3.10.9-031009/tools/perf

But it doesn't work. But after I use the absolute directory, it seems
to work now.  But setting absolute directory still doesn't work on
Redhat 6.0, even if I replaced the total directory of perf.

Here it is. I get a Segmentation fault here.

[pwang7@bn20-171 ~]$ perf script -s perf-script.py
# ========
# captured on: Fri Aug 30 14:00:01 2013
# hostname : bn20-171.dcs.mcnc.org
# os release : 2.6.32-71.el6.x86_64
# perf version : 2.6.32-358.11.1.el6.x86_64.debug
# arch : x86_64
# nrcpus online : 24
# nrcpus avail : 24
# cpudesc : Intel(R) Xeon(R) CPU E5645 @ 2.40GHz
# cpuid : GenuineIntel,6,44,2
# total memory : 99061748 kB
# cmdline : /usr/bin/perf record -e cache-misses ls
# event : name = cache-misses, type = 0, config = 0x3, config1 = 0x0,
config2 = 0x0, excl_usr = 0, excl_kern = 0, excl_host = 0, excl_guest
= 0, precise_ip = 0, id = { 1797, 1798, 1799, 1800, 1801, 1802, 1803,
1804, 1805, 1806, 1807, 1808, 1809, 1810, 1811, 1812, 1813, 1814,
1815, 1816, 1817, 1818, 1819, 1820 }
# HEADER_CPU_TOPOLOGY info available, use -I to display
# HEADER_NUMA_TOPOLOGY info available, use -I to display
# ========
#
in trace_begin
Segmentation fault (core dumped)
Best wishes.
Yours,
Wang Peipei


On Fri, Aug 30, 2013 at 2:01 PM, David Ahern <dsahern@gmail.com> wrote:
> On 8/30/13 11:15 AM, Mikhail Kulemin wrote:
>>
>> You should fix perf-script.py
>>
>> replace   sys.path.append(os.environ['PERF_EXEC_PATH'] +
>> '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
>> with
>
>
> better to set/export PERF_EXEC_PATH before running the command.
>
> David
>
>
>
>>
>> sys.path.append(os.getcwd()+"/tools/perf/scripts/python/
>> Perf-Trace-Util/lib/Perf/Trace")
>>
>> this string instruct python interpreter where to find modules to load.
>>
>>
>> Please check that "perf script" command works correctly (output all
>> sample records from perf.data file)
>
>