Re: /proc/profile
From: Tommy Reynolds (TommyReynolds_at_yahoo.com)
Date: 04/30/05
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: Sat, 30 Apr 2005 09:11:32 -0500
On Wed, 27 Apr 2005 09:45:02 -0700, Jaggu wrote:
> I tried to profile kernel, by using readprofile I got
> following data.I dont know how to analyze data, document says the third
> column is normalized load of the procedure..calculated as ratio between
> number of clock ticks and length of procedure...
>
> root@ml310:/# readprofile
> 4 _stext 0.1000
> 4722 _nmask_and_or_msr 147.5625
> 1 copy_page 0.0069
> 1 do_page_fault 0.0009
> 1 memset 0.0109
Think of the third column this way:
1) A counter is incremented on every clock tick. The counter is chosen
based on which kernel subroutine is being executed. (This is not
really how it is implemented, but it gets the idea across.)
2) Now, we expect that a large subroutine takes longer to execute than a
smaller subroutine. If the small subroutine gets lots and lots of counts,
then that subroutine is contributing more than its share to the system
performance load.
3) So if you sort the output in descending order, you can see which
kernel routines are contributing most to the overall system load.
Concentrate on optimizing those routines and you can make the most
improvement on the system.
How do we tell how big subroutines are? Hints from ld(1).
Cheers
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|
