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

[Johannes Weiner <hannes@cmpxchg.org>]

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@jp.fujitsu.com>

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.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>