Re: [patch V3] percpu_counter: scalability works

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

On Mon, 2011-05-16 at 22:22 +0800, Eric Dumazet wrote:
> Le lundi 16 mai 2011 à 11:35 +0200, Eric Dumazet a écrit :
> 
> > Given that vm_committed has one percent resolution need
> > (sysctl_overcommit_ratio is expressed with percent resolution), it
> > should be used with an appropriate batch value, something like :
> > 
> > vm_committed_as_batch = max(percpu_counter_batch,
> > 			    total_ram_pages/(num_possible_cpus()*100));
> > 
> 
> 
> Funny thing with vm_committed_as is we dont even read its value with
> default vm configuration 
> 
> (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS or OVERCOMMIT_GUESS)
> 
> [ In this case, we read it only for /proc/meminfo output ]
> 
> Ideally, we could dynamically change vm_committed_as_batch when
> sysctl_overcommit_memory or other param is changed. This would need a
> mechanism to ask all cpus to transfert/clear their local s32 into global
> fbc->count [when lowering vm_committed_as_batch]
I actually posted something like this before
http://marc.info/?l=linux-mm&m=130144785326028&w=2
but this could affect /proc/meminfo read.

> Another idea would be to use an atomic when manipulating the percpu s32,
> so that __percpu_counter_sum() is able to make this operation itself :
> At the end of __percpu_counter_sum(), fbc->count would be the final
> result, and all s32 would be zero, unless some cpus called _add()
> meanwhile.
don't understand it. But if concurrent _add can introduce deviation,
this is good.

I'm still interesting in improving percpu_counter itself. If we can
improve it, why we don't? My patch doesn't slow down anything for all
tests. Why didn't you ever look at it?

Thanks,
Shaohua