[Bug: 21.5-b28] more detailed gprof profile, performance on aix,
cpu hog, tprof, --without-error-checking --without-kkcc
Thomas Mittelstaedt
T.Mittelstaedt at cadenas.de
Thu Mar 13 09:55:07 EDT 2008
I don't want to litter this list with profile data, but this one should
be complete. The previous one
I posted, did not show the full listing.
Hope this helps. The configuration of xemacs was
--without-error-checking' '--without-kkcc'
Thomas Mittelstaedt schrieb:
> Tried out the following:
>
> ./configure '--without-error-checking'
> '--with-late-packages=/usr/local/lib/xemacs' '--without-kkcc'
>
> and xemacs still starts using lots of cpu but then goes down to 4-11%
> range and stays there.
> tprof can also generate micro-profiles of the individual functions. If
> somebody is interested I can send a tar
> of these profiles.
>
>
>
> Thomas Mittelstaedt schrieb:
>
>> With xemacs configured without error-checking, I still have 30% + cpu
>> usage, but interactive usage is
>> significantly better.
>>
>> Total % For All Processes (/usr/local/bin/xemacs-21.5-b28) = 27,48
>>
>> Subroutine %
>> Source
>> ========== ======
>> ======
>> .add_ichar_rune_1 1,47
>> redisplay.c
>> .extent_list_locate 1,22
>> extents.c
>> .re_match_2_internal 1,21
>> regex.c
>> .assq_no_quit 1,02
>> fns.c
>> .execute_optimized_program 0,89
>> bytecode.c
>> .create_text_block 0,76
>> redisplay.c
>> .display_table_entry 0,75
>> glyphs.c
>> .kkcc_marking 0,73
>> gc.c
>> .BYTE_BUF_BYTE_ADDRESS_NO_VERIFY 0,53
>> inline.c
>> .compare_runes 0,53
>> redisplay-output.c
>> .Fcons 0,42
>> alloc.c
>> .noseeum_cons 0,35
>> alloc.c
>> .kkcc_gc_stack_push_lisp_object_1 0,35
>> gc.c
>> .recompute_funcall_allocation_flag 0,33
>> alloc.c
>> .readchar 0,33
>> lread.c
>> .Fmemq 0,33
>> fns.c
>> .generate_fstring_runes 0,32
>> redisplay.c
>> .extent_list_at 0,32
>> extents.c
>> .specifier_get_inst_list 0,30
>> specifier.c
>> .free_cons 0,29
>> alloc.c
>> .kkcc_gc_stack_push_1 0,28
>> gc.c
>> .find_symbol_value_1 0,28
>> symbols.c
>> .recompute_need_to_garbage_collect 0,26
>> gc.c
>> .BYTE_BUF_BYTE_ADDRESS 0,26
>> inline.c
>> .extent_ancestor 0,22
>> inline.c
>> .store_symval_forwarding 0,22
>> symbols.c
>> .bytebpos_to_charbpos 0,19
>> inline.c
>> .unbind_to_1 0,19
>> eval.c
>> .Fset 0,19
>> symbols.c
>> .Ffuncall 0,18
>> eval.c
>> .extent_fragment_update 0,18
>> extents.c
>> .unbind_to_hairy 0,18
>> eval.c
>> .do_symval_forwarding 0,18
>> symbols.c
>> .kkcc_bt_push 0,18
>> gc.c
>> .no_conversion_convert 0,17
>> file-coding.c
>> .alloc_managed_lcrecord 0,17
>> alloc.c
>> .offset_post_relocation 0,17
>> regex.c
>> .free_managed_lcrecord 0,17
>> alloc.c
>> .malloc 0,16
>> gmalloc.c
>> .lispdesc_indirect_count 0,16
>> inline.c
>> .symbol_is_constant 0,15
>> symbols.c
>> .add_ichar_rune 0,15
>> redisplay.c
>> .compare_display_blocks 0,15
>> redisplay-output.c
>> .reject_constant_symbols 0,15
>> symbols.c
>> .specifier_instance_from_inst_list 0,14
>> specifier.c
>> .specifier_instance_1 0,14
>> specifier.c
>>
>> Thomas Mittelstaedt schrieb:
>>
>>
>>> The previous output was from an image, which I had instrumented for
>>> gprof. Here is
>>> the normal one:
>>> User
>>> Process %
>>> =============
>>> ======
>>> /usr/local/bin/xemacs-21.5-b28.bak
>>> 48,33
>>> /usr/local/bin/bash
>>> 0,01
>>>
>>> Profile: /usr/local/bin/xemacs-21.5-b28.bak
>>>
>>> Total % For All Processes (/usr/local/bin/xemacs-21.5-b28.bak) = 48,33
>>>
>>> Subroutine %
>>> Source
>>> ========== ======
>>> ======
>>> .error_check_cons 3,09
>>> inline.c
>>> .re_match_2_internal 2,36
>>> regex.c
>>> .extent_list_locate 2,23
>>> extents.c
>>> .error_check_lstream 2,16
>>> inline.c
>>> .Dynarr_verify_1 1,93
>>> inline.c
>>> .error_check_buffer 1,25
>>> inline.c
>>> .error_check_string 0,86
>>> inline.c
>>> .kkcc_marking 0,82
>>> gc.c
>>> .add_ichar_rune_1 0,77
>>> redisplay.c
>>> .deadbeef_memory 0,75
>>> alloc.c
>>> .free_cons 0,65
>>> alloc.c
>>> .readchar 0,64
>>> lread.c
>>> .assq_no_quit 0,63
>>> fns.c
>>> .Dynarr_verify_mod_1 0,62
>>> inline.c
>>> .error_check_image_instance 0,55
>>> inline.c
>>> .execute_optimized_program 0,55
>>> bytecode.c
>>> .error_check_symbol 0,54
>>> inline.c
>>> .noseeum_cons 0,53
>>> alloc.c
>>> .BYTE_BUF_BYTE_ADDRESS_NO_VERIFY 0,52
>>> inline.c
>>> .offset_post_relocation 0,46
>>> regex.c
>>> .extent_list_at 0,45
>>> extents.c
>>> .wrap_record_1 0,45
>>> inline.c
>>> .BYTE_BUF_BYTE_ADDRESS 0,45
>>> inline.c
>>> .recompute_funcall_allocation_flag 0,44
>>> alloc.c
>>> .mark_object_maybe_checking_free_1 0,41
>>> gc.c
>>> .compare_runes 0,40
>>> redisplay-output.c
>>> .display_table_entry 0,40
>>> glyphs.c
>>> .kkcc_gc_stack_push_lisp_object_1 0,40
>>> gc.c
>>> .alloc_managed_lcrecord 0,37
>>> alloc.c
>>> .create_text_block 0,36
>>> redisplay.c
>>> .Fset 0,35
>>> symbols.c
>>> .error_check_vector 0,31
>>> inline.c
>>> .error_check_window 0,30
>>> inline.c
>>> .generate_fstring_runes 0,30
>>> redisplay.c
>>> .no_conversion_convert 0,30
>>> file-coding.c
>>> .kkcc_gc_stack_push_1 0,29
>>> gc.c
>>> .Fcons 0,28
>>> alloc.c
>>> .recompute_need_to_garbage_collect 0,27
>>> gc.c
>>> .re_search_2 0,26
>>> regex.c
>>> .specifier_get_inst_list 0,26
>>> specifier.c
>>> .unbind_to_hairy 0,26
>>> eval.c
>>> .unbind_to_1 0,26
>>> eval.c
>>>
>>>
>>>
>>> Thomas Mittelstaedt schrieb:
>>>
>>>
>>>
>>>> I think, I found something interesting using tprof, which I ran like the
>>>> following:
>>>>
>>>> tprof -us -p xemacs-21.5-b28 -M /localbuild/xemacs-21.5.28-latest/src/
>>>> -x /usr/local/bin/xemacs-21.5-b28
>>>>
>>>> User
>>>> Process %
>>>> =============
>>>> ======
>>>> /usr/local/bin/xemacs-21.5-b28
>>>> 60,65
>>>>
>>>> Profile: /usr/local/bin/xemacs-21.5-b28
>>>>
>>>> Total % For All Processes (/usr/local/bin/xemacs-21.5-b28) = 60,65
>>>>
>>>> Subroutine %
>>>> Source
>>>> ========== ======
>>>> ======
>>>> .__mcount 22,40
>>>> /lib/libc/__mcount.c
>>>> .qincrement 3,02
>>>> /lib/libc/__mcount.c
>>>> .error_check_lstream 2,70
>>>> inline.c
>>>> .kkcc_marking 1,68
>>>> gc.c
>>>> .readchar 1,39
>>>> lread.c
>>>> .error_check_cons 1,21
>>>> inline.c
>>>> .extent_list_locate 1,10
>>>> extents.c
>>>> .error_check_string 1,00
>>>> inline.c
>>>> .no_conversion_convert 0,81
>>>> file-coding.c
>>>> .__stack_pointer 0,79 _stk.s
>>>>
>>>>
>>>>
>>>>
>>>> Thomas Mittelstaedt schrieb:
>>>>
>>>>
>>>>
>>>>
>>>>> Using trace and curt, I produced another performance output, which is
>>>>> attached.
>>>>> Maybe it helps.
>>>>>
>>>>> Thomas Mittelstaedt schrieb:
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> Okay, I have attached the truss output with the start:
>>>>>>
>>>>>>
>>>>>>
>>>>>> Jerry James schrieb:
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>> Thomas,
>>>>>>>
>>>>>>> Thanks for going to the effort of collecting this profiling data. It
>>>>>>> gives us some interesting information.
>>>>>>>
>>>>>>> 2008/3/11 Thomas Mittelstaedt <T.Mittelstaedt at cadenas.de>:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>> ngranularity: Jedes gefundene Beispiel deckt 4 Byte ab. Zeit: 97,87 Sekunden
>>>>>>>>
>>>>>>>> % kumulativ selbst selbst gesamt
>>>>>>>> Zeit Sekunden Sekunden Aufrufe ms/Aufruf ms/Aufruf Name
>>>>>>>> 0,2 0,21 0,21 .__is_wctype_sb [1]
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>> [snip]
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>> 0,0 1,00 0,01 .xtCreate [48]
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>> So the top 48 time hogs together consume a grand total of 1.00 seconds
>>>>>>> out of 97.87 seconds of execution. Your other profile was similar:
>>>>>>> out of 56.55 seconds of execution, the top 43 collectively consumed
>>>>>>> 0.73 seconds of execution. This indicates that the performance
>>>>>>> problems you are seeing are not occurring in the profiled (i.e.,
>>>>>>> XEmacs) code. We appear to either be triggering a performance bug in
>>>>>>> some library, or using some library in a suboptimal way. In either
>>>>>>> case, this is as far as a profiler is going to take us.
>>>>>>>
>>>>>>> I'm still curious about the large number of kioctl calls in the truss
>>>>>>> output you showed us earlier. Do you know why that output did not
>>>>>>> start with the launching of the XEmacs executable? Can you get us
>>>>>>> truss output that does start from launch? I want to know what file
>>>>>>> descriptor is the target of all that manipulation. It may be a red
>>>>>>> herring, but I think it's worth checking out.
>>>>>>>
>>>>>>> Thank you,
>>>>>>>
>>>>>>>
>>>>>>> ------------------------------------------------------------------------
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> XEmacs-Beta mailing list
>>>>>>> XEmacs-Beta at xemacs.org
>>>>>>> http://calypso.tux.org/cgi-bin/mailman/listinfo/xemacs-beta
>>>>>>>
>>>>>>>
>>>>>>> ------------------------------------------------------------------------
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> XEmacs-Beta mailing list
>>>>>>> XEmacs-Beta at xemacs.org
>>>>>>> http://calypso.tux.org/cgi-bin/mailman/listinfo/xemacs-beta
>>>>>>>
>>>>>>> ------------------------------------------------------------------------
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> XEmacs-Beta mailing list
>>>>>>> XEmacs-Beta at xemacs.org
>>>>>>> http://calypso.tux.org/cgi-bin/mailman/listinfo/xemacs-beta
>>>>>>>
>>>>>>> ------------------------------------------------------------------------
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> XEmacs-Beta mailing list
>>>>>>> XEmacs-Beta at xemacs.org
>>>>>>> http://calypso.tux.org/cgi-bin/mailman/listinfo/xemacs-beta
>>>>>>>
>>>>>>> ------------------------------------------------------------------------
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> XEmacs-Beta mailing list
>>>>>>> XEmacs-Beta at xemacs.org
>>>>>>> http://calypso.tux.org/cgi-bin/mailman/listinfo/xemacs-beta
>>>>>>>
>
> _______________________________________________
> XEmacs-Beta mailing list
> XEmacs-Beta at xemacs.org
> http://calypso.tux.org/cgi-bin/mailman/listinfo/xemacs-beta
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: gprof.out.gz
Type: application/x-gzip
Size: 174446 bytes
Desc: not available
Url : http://lists.xemacs.org/pipermail/xemacs-beta/attachments/20080313/c0f3b974/gprof.out-0001.bin
More information about the XEmacs-Beta
mailing list
