Re: [patch V3] percpu_counter: scalability works

[Shaohua Li <shaohua.li@intel.com>]

On Mon, 2011-05-16 at 15:44 +0800, Eric Dumazet wrote:
> Le lundi 16 mai 2011 à 15:15 +0800, Shaohua Li a écrit :
> 
> > I can, but you can't prevent me to optimize percpu_counter.
> > 
> 
> Well, I have the right to say you're wrong.
sure, but please give a reason.

> Your last patch is not good, sorry. 
> Please take the time to read it
> again and fix obvious problems.
what kind of obvious problems?

> And also give us numbers if one process
> does the mmap()/munmap() loop, before and after your patch.
I did a stress test with one thread

while {
__percpu_counter_add(+count)
__percpu_counter_add(-count)
}
the loop do 10000000 times.
in _add fast path (no locking hold):
before my patch:
real    0m0.133s
user    0m0.000s
sys     0m0.124s
after:
real    0m0.129s
user    0m0.000s
sys     0m0.120s
the difference is variation.

in _add slow path (locking hold):
before my patch:
real    0m0.374s
user    0m0.000s
sys     0m0.372s
after:
real    0m0.245s
user    0m0.000s
sys     0m0.020s

My patch actually makes _add faster, because we removed spin_lock.

> A percpu_counter is already a beast as is, you're suggesting to double
> its size, for a pathological case.
> 
> Its absolutely not clear to me why vm_committed_as is using the default
> percpu_counter_batch. 
> 
> By the way could you make sure percpu_counter_batch has the right value
> on your 24 cpus machine ?
> 
> Your 128Mbyte mmap threshold sounds like percpu_counter_batch=32 instead
> of 48
let's not argue the batch size anymore. If we can make percpu_counter
faster, why we don't (even your patch mentioned this).

Thanks,
Shaohua