linux-perf-users.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Milian Wolff <milian.wolff@kdab.com>
To: shiny.sebastian@intel.com
Cc: linux-perf-users@vger.kernel.org
Subject: Re: Perf report --percentage
Date: Wed, 08 Jun 2016 09:34:58 +0200	[thread overview]
Message-ID: <1936512.XMEGsyMYkf@agathebauer> (raw)
In-Reply-To: <6076A35470B38B408600AA0E3B4A31E767471A9E@ORSMSX114.amr.corp.intel.com>

[-- Attachment #1: Type: text/plain, Size: 4075 bytes --]

On Dienstag, 7. Juni 2016 19:21:15 CEST Sebastian, Shiny wrote:
> Hi,
>  
> I work with Intel and am running perf tool within Linux VMs on Windows
> Hyper-V.  With kernel 4.6 and corresponding perf version, percentages of
> call stacks at deeper levels  do not seem to add up to 100% or parent value
> (26.97%): example below. I have tried the -percentage flag with both
> 'absolute' and 'relative' values and they don't seem to change the output
> at all. Is this a bug or something wrong with my perf flags ?
> 
> Commands used:
> #perf record  -a -g -c 10000 sleep 10
> #perf report -i cpu3.data.old -C 1 --no-children
> 
> -   27.48%  fio          [kernel.kallsyms]  [k] __softirqentry_text_start
>      __softirqentry_text_start
>    - irq_exit
>       - 26.97% smp_call_function_single_interrupt
>          - call_function_single_interrupt
>             + 4.21% __blk_run_queue
>             + 4.09% hv_ringbuffer_write
>             + 1.83% 0x99c
>             + 1.78% __blockdev_direct_IO
>             + 0.96% do_blockdev_direct_IO
>             + 0.95% generic_make_request
>             + 0.86% blk_queue_bio
>             + 0.70% iov_iter_get_pages
>             + 0.64% md_make_request
>             + 0.55% do_io_submit
>               0.51% entry_SYSCALL_64_fastpath
>       + 0.51% hyperv_vector_handler

Can you try to look at the output of

perf report -i cpu3.data.old -C 1 --no-children -g graph,0

i.e. potentially there are more entries with a cost of 0.5% or less, which by 
default falls below the threshold.

That said, I just tested it and it doesn't work for me either:

tmp$ perf record --call-graph dwarf kwrite
[ perf record: Woken up 65 times to write data ]
[ perf record: Captured and wrote 16.308 MB perf.data (2032 samples) ]
tmp$ perf report -g graph,0 --no-children

-    9.47%  kwrite           ld-2.23.so                     [.] do_lookup_x
   - 2.21% do_lookup_x
      - _dl_lookup_symbol_x
         + 1.42% _dl_relocate_object
         + 0.74% _dl_fixup
+    7.06%  kwrite           ld-2.23.so                     [.] strcmp 

It's also bogus when I look at it with --children:

-   10.76%     9.47%  kwrite           ld-2.23.so           [.] do_lookup_x
   + 1.54% _start
   + 1.34% do_lookup_x
   + 0.61% _dl_runtime_resolve_avx
-    7.06%     7.06%  kwrite           ld-2.23.so           [.] strcmp
   + 0.42% _start
   + 0.22% _dl_runtime_resolve_avx

So someone else with more insight should answer whether this output can be 
explained somehow.

> Older version: Perf version 4.2 on kernel 4.2 - seems to normalize them to
> 100%.  27.69%  swapper          [kernel.kallsyms]            [k] intel_idle
> 
>                     ---intel_idle
> 
>                        |--99.84%-- cpuidle_enter_state
>                        |
>                        |          cpuidle_enter
>                        |          call_cpuidle
>                        |          cpu_startup_entry
>                        |          
>                        |          |--97.27%-- start_secondary
>                        |          |
>                        |           --2.73%-- rest_init
>                        |           
>                        |                     start_kernel
>                        |                     x86_64_start_reservations
>                        |                     x86_64_start_kernel
> 
>                         --0.16%-- [...]

This output can be reproduced on newer perf by passing `-g fractal` to `perf 
report`. I still get the issues as shown above though with a newer perf:

-    9.47%  kwrite           ld-2.23.so                     [.] do_lookup_x
   + 23.31% do_lookup_x
-    7.06%  kwrite           ld-2.23.so                     [.] strcmp
   + 9.06% strcmp
+    5.71%  kwrite           ld-2.23.so                     [.] check_match

I.e. ~74% and 91% of the samples don't seen to have a proper backtrace 
attribution?

Cheers

-- 
Milian Wolff | milian.wolff@kdab.com | Software Engineer
KDAB (Deutschland) GmbH&Co KG, a KDAB Group company
Tel: +49-30-521325470
KDAB - The Qt Experts

[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 5903 bytes --]

  reply	other threads:[~2016-06-08  7:35 UTC|newest]

Thread overview: 9+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2016-06-07 19:21 Perf report --percentage Sebastian, Shiny
2016-06-08  7:34 ` Milian Wolff [this message]
2016-06-08 16:21   ` Sebastian, Shiny
2016-06-11  8:31     ` Namhyung Kim
2016-06-11 11:53       ` Milian Wolff
2016-06-12 13:55         ` Namhyung Kim
2016-06-13 11:01           ` Milian Wolff
2016-06-09 15:13   ` Arnaldo Carvalho de Melo
2016-06-10  3:14     ` Namhyung Kim

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1936512.XMEGsyMYkf@agathebauer \
    --to=milian.wolff@kdab.com \
    --cc=linux-perf-users@vger.kernel.org \
    --cc=shiny.sebastian@intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).