Re: [PATCH v3 0/7] x86: Rewrite switch_to()

[Andy Lutomirski]

On Sat, Aug 13, 2016 at 11:45 AM, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
> * Brian Gerst <brgerst@xxxxxxxxx> wrote:
>
>> On Sat, Aug 13, 2016 at 1:16 PM, Linus Torvalds
>> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>> > On Sat, Aug 13, 2016 at 9:38 AM, Brian Gerst <brgerst@xxxxxxxxx> wrote:
>> >> This patch set simplifies the switch_to() code, by moving the stack switch
>> >> code out of line into an asm stub before calling __switch_to().  This ends
>> >> up being more readable, and using the C calling convention instead of
>> >> clobbering all registers improves code generation.  It also allows newly
>> >> forked processes to construct a special stack frame to seamlessly flow
>> >> to ret_from_fork, instead of using a test and branch, or an unbalanced
>> >> call/ret.
>> >
>> > Do you have performance numbers? Is it noticeable/measurable?
>>
>> How do I measure it?  The perf documentation isn't easy to understand.
>
> Something like this:
>
>   taskset 1 perf stat -a -e '{instructions,cycles}' --repeat 10 perf bench sched pipe
>
> ... will give a very good idea about the general impact of these changes on
> context switch overhead.
>

I will be quite surprised if you can measure any effect at all.  I've
never seen context switches take fewer than ~2k cycles, and on my
laptop, they take 8k-9k cycles.  The scheduler is really, really slow.

(Why doesn't that perf command show cycles per context switch?)

--Andy