Module vs Kernel main performacne

Abu Rasheda rcpilot2010 at gmail.com
Wed May 30 20:17:11 EDT 2012


On Wed, May 30, 2012 at 2:44 PM, Abu Rasheda <rcpilot2010 at gmail.com> wrote:
> I did another experiment.
>
> Wrote a stand alone module and user program which does ioctl and pass
> buffer to kernel module.
>
> User program passes a buffer through ioctl and kernel module does
> kmalloc on it and calls copy_from_user, kfree and return. Test program
> send 120 gigabyte data to module.
>
> If I pass 1k buffer per call, I get
>
> 115,396,349,819 instructions              #    0.90  insns per cycle
>      [95.00%]
>
> as I increase size of buffer, insns per cycle keep decreasing. Here is the data:
>
>    1k 0.90  insns per cycle
>    8k 0.43  insns per cycle
>  43k 0.18  insns per cycle
> 100k 0.08  insns per cycle
>
> Showing that cop_from_user is more efficient when copy data is small,
> why it is so ?

Did another experiment:

User program sending 43k and allocating 43k after entering ioctl and
copy_from_user smaller portion in each call to copy_from_user:
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
copy_from_user  0.25k at a time 0.56  insns per cycle
copy_from_user  0.50k at a time 0.42  insns per cycle
copy_from_user  1.00k at a time 0.36  insns per cycle
copy_from_user  2.00k at a time 0.29  insns per cycle
copy_from_user  3.00k at a time 0.26  insns per cycle
copy_from_user  4.00k at a time 0.23  insns per cycle
copy_from_user  8.00k at a time 0.21  insns per cycle
copy_from_user 16.00k at a time 0.19  insns per cycle


User program sending 43k, allocating smaller chunk and sending that
chunk to call to copy_from_user:
--------------------------------------------------------------------------------------------------------------------------------------------------------------
Allocated 0.25k and copy_from_user  0.25k at a time 1.04 insns per cycle
Allocated 0.50k and copy_from_user  0.50k at a time 0.90 insns per cycle
Allocated 1.00k and copy_from_user  1.00k at a time 0.79 insns per cycle
Allocated 2.00k and copy_from_user  2.00k at a time 0.67 insns per cycle
Allocated 4.00k and copy_from_user  4.00k at a time 0.53 insns per cycle
Allocated 8.00k and copy_from_user  8.00k at a time 0.42 insns per cycle
Allocated 16.00k and copy_from_user 16.00k at a time 0.33 insns per cycle
Allocated 32.00k and copy_from_user 32.00k at a time 0.22 insns per cycle



More information about the Kernelnewbies mailing list