Re: [PATCH] sched: minor optimizations in wake_affine and select_task_rq_fair

["Amit K. Arora" <aarora@linux.vnet.ibm.com>]

On Mon, Sep 29, 2008 at 10:09:41AM -0600, Chris Friesen wrote:
> Amit K. Arora wrote:
>> sched: Minor optimizations in wake_affine and select_task_rq_fair
>>
>> This patch does following:
>> o Reduces the number of arguments to wake_affine().
>
> At what point is it cheaper to pass items as args rather than recalculating 
> them?  If reducing the number of args is desirable, what about removing the 
> "this_cpu" and "prev_cpu" args and recalculating them in wake_affine()?

Thats a good question. Its kind of arguable and I wasn't sure if
everyone will be happy if I removed more arguments from wake_affine() than
what I did in my patch (because of the recalculations required).

wake_affine() currently has 11 arguments and I thought it may make sense
in reducing it to a sane number. For that I chose arguments which I
thought can be recalculated with minimum overhead (involves single struct
dereference, a simple per cpu variable and/or a simple arithmetic). And
one argument ("rq") which is being removed, isn't used at all in the
function!

Regarding the two variables you have mentioned, I didn't remove them as
args since I wasn't sure of "this_cpu" (which is nothing but
smp_processor_id()) as it is arch dependent, and calculating "prev_cpu"
involves two struct dereferences (((struct thread_info *)(task)->stack)->cpu).
And the calculation for other arguments (like, this_sd, load and this_load)
involves good amount of instructions.

If you disagree, what do you suggest we do here ?

Regards,
Amit Arora