Re: [PATCH v2 14/14] Change CPUACCT to default n

[Glauber Costa <glommer@parallels.com>]

On 11/17/2011 12:58 AM, Balbir Singh wrote:
> On Thu, Nov 17, 2011 at 8:19 AM, Glauber Costa<glommer@parallels.com>  wrote:
>> On 11/16/2011 09:52 PM, KAMEZAWA Hiroyuki wrote:
>>>
>>> On Wed, 16 Nov 2011 15:51:27 +0530
>>> Balbir Singh<bsingharora@gmail.com>    wrote:
>>>
>>>>>
>>>>> On the other hand, I don't think much discussion remains for cpuacct,
>>>>> everyone's pretty unanimous in that they'd like to see it deprecated.
>>>>> By splitting this up we can close out that quickly while we figure out
>>>>> the
>>>>> best way to resolve the above.
>>>>>
>>>>
>>>> I'd give it a thumbs up, if we can create sched groups and provide
>>>> accounting without control - like we can for the memory cgroup today.
>>>>
>>>
>>> Isn't it possible ?
>>>
>>> Thanks,
>>> -Kame
>>>
>> I must say I don't really understand what exactly you propose, and how it is
>> different from what we have today.
>>
>> My take is that you are talking about a single cgroup in which you can have
>> the functionality of both cpuacct and cpu, but surrounded by knobs that
>> allows you to turn them off individually.
>>
>> Am I right?
>>
>
> No here is what I am asking for
>
> I don't want CPU control, just accounting, so I create the following groups
>
>                            a
>                         /    \
>                        V    V
>                        b     c
>
> Today, with the cpu controller, the moment I create a, b and c, they
> get default shares and if I put tasks, their b/w is decided by the
> shares, what if I don't want control, but I want to account for their
> time only?
>
> Balbir

I think that if this really a requirement, cpuacct should stay. I was 
working under the assumption that it was not really an important case - 
so thanks for the clarification. Peter and Paul can chime in here, but I 
think that this requirement poses constraints to the cpu cgroup and 
consequently the scheduler - both in its current incarnation and in what 
come in the future - that may not be acceptable. What I am concerned 
about is that it might mandate the scheduler to always test whether or 
not the grouping has a scheduling effect or not - and then walk the 
group if it is not, etc. In a summary, if we can or cannot bundle 
processes together for scheduling purposes, we'll likely need separate 
data structures anyway.

A lot of the code I wrote can be reused to at least make it faster in 
the case in which only the root is mounted - for cpuacct.stat at least.

However, the big question remains: The most expensive operation for 
cpuacct also seem to be the most important, cpuusage, which was a big 
part of the motivation to bundle them all together. Maybe then Paul's 
co-mounting idea starts to make sense, but it will still be quite slow 
for your usage, in which the groups are clearly different.

I think the best I can come up with right now, is to base my work on 
cpuacct - I am fine with that, and it was actually how my first version 
looked like - and then think about a way to make cpuusage faster later...