lkml.org 
[lkml]   [2011]   [Sep]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: CFS Bandwidth Control - Test results of cgroups tasks pinned vs unpinnede
On 09/13/11 11:23, Peter Zijlstra wrote:
> On Tue, 2011-09-13 at 23:31 +0530, Srivatsa Vaddagiri wrote:
>> * Peter Zijlstra<a.p.zijlstra@chello.nl> [2011-09-13 16:19:39]:
>>
>>>> Booting with "nohz=off" also helps significantly.
>>>>
>>>> With nohz=on, average idle time (over 1 min) is 10.3%
>>>> With nohz=off, average idle time (over 1 min) is 3.9%

I think more compelling here is that it looks like nohz load-balance
needs more love.

>>>
>>> So we should put the cpufreq/idle governor into the nohz/idle path, it
>>> already tries to predict the idle duration in order to pick a C state,
>>> that same prediction should be used to determine if stopping the tick is
>>> worth it.
>>
>> Hmm ..I tried performance governor and found that it slightly increases
>> idle time.
>>
>> With nohz=off&& ondemand governor, idle time = 4%
>> With nohz=off&& performance governor on all cpus, idle time = 6%
>>
>> I can't see obvious reasons for that ..afaict bandwidth capping should
>> be independent of frequency (i.e task gets capped by "used" time,
>> irrespective of frequency at which it was "using" the cpu)?
>
> That's not what I said.. what I said is that the nohz code should also
> use the idle time prognosis.. disabling the tick is a costly operation,
> doing it only to have to undo it costs time, and will be accounted to
> idle time, hence your improvement with nohz=off.
>

Enabling Venki's CONFIG_IRQ_TIME_ACCOUNTING=y would discount to provide
a definitive answer here yes?

- Paul



\
 
 \ /
  Last update: 2011-09-16 10:17    [W:0.188 / U:0.284 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site