Re: [patch 0/6] mm: memcg: statistics implementation cleanups

[KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>]

(2012/05/15 20:03), Johannes Weiner wrote:

> On Tue, May 15, 2012 at 09:19:33AM +0900, KAMEZAWA Hiroyuki wrote:
>> (2012/05/15 3:00), Johannes Weiner wrote:
>>
>>> Before piling more things (reclaim stats) on top of the current mess,
>>> I thought it'd be better to clean up a bit.
>>>
>>> The biggest change is printing statistics directly from live counters,
>>> it has always been annoying to declare a new counter in two separate
>>> enums and corresponding name string arrays.  After this series we are
>>> down to one of each.
>>>
>>>  mm/memcontrol.c |  223 +++++++++++++++++------------------------------
>>>  1 file changed, 82 insertions(+), 141 deletions(-)
>>
>> to all 1-6. Thank you.
>>
>> Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
> 
> Thanks!
> 
>> One excuse for my old implementation of mem_cgroup_get_total_stat(),
>> which is fixed in patch 6, is that I thought it's better to touch all counters
>> in a cachineline at once and avoiding long distance for-each loop.
>>
>> What number of performance difference with some big hierarchy(100+children) tree ?
>> (But I agree your code is cleaner. I'm just curious.)
> 
> I set up a parental group with hierarchy enabled, then created 512
> children and did a 4-job kernel bench in one of them.  Every 0.1
> seconds, I read the stats of the parent, which requires reading each
> stat/event/lru item from 512 groups before moving to the next one:
> 
>                         512stats-vanilla        512stats-patched
> Walltime (s)            62.61 (  +0.00%)        62.88 (  +0.43%)
> Walltime (stddev)        0.17 (  +0.00%)         0.14 (  -3.17%)
> 
> That should be acceptable, I think.
> 
> 


Yes, thank you. 
Thanks,
-Kame