Re: [PATCH v5 11/11] sched: introduce cgroup file stat_percpu

[Glauber Costa <glommer@parallels.com>]

On 01/10/2013 01:27 AM, Glauber Costa wrote:
> On 01/10/2013 01:17 AM, Andrew Morton wrote:
>> On Thu, 10 Jan 2013 01:10:02 +0400
>> Glauber Costa <glommer@parallels.com> wrote:
>>
>>> The main advantage I see in this approach, is that there is way less
>>> data to be written using a header. Although your way works, it means we
>>> will write the strings "nice", "system", etc. #cpu times. Quite a waste.
>>
>> Yes, overhead can be a significant issue with this type of interface. 
>> But we already incurred a massive overhead by using a human-readable
>> ascii interface.  If performance is an issue, perhaps the whole thing
>> should be grafted onto taskstats instead.  Or create a new
>> taskstats-like thing.
> 
> I think this would be a little alienish in the already alien world of
> cgroups.
> 
> However, I was not so much talking about plain performance overhead as
> measurable in miliseconds-to-parse, but rather just alluding to the fact
> that we would be writing the same set of strings multiple times when a
> header would do just fine.
> 
> This is the same method used for instance by slabinfo.
> 
>>
>> btw, a more typical interface would be
>>
>> cat /.../cpu0
>> nice:nn
>> system:nn
>> irq:nn
>>
> 
> Well, yes. But welcome to cgroups: directories have a meaning, so the
> only way to organize stuff is with plain files in the current hierarchy
> is by filling it with files. As many files as we have cpus.
> 
> At this point you are certain to miss all the other files present in the
> directory.
> 
> 
>> - the traditional one-per-line name:value tuples.  But I'd assumed that
>> having a file per CPU would be aawkward.
>>
> Indeed.
> 

Andrew,

Given my arguments above, which interface would you prefer for me to
settle down? I still don't see any problems with the header, specially
given the fact that it exists precisely to allow fields to come and go
if needed.