[Devel] Re: [RFC][PATCH] Per container statistics

[Balbir Singh]

Paul Menage wrote:
> Hi Balbir,
> 
> On 5/14/07, Balbir Singh <balbir at linux.vnet.ibm.com> wrote:
>>
>> This patch implements per container statistics infrastructure and re-uses
>> code from the taskstats interface. A new set of container operations are
>> registered with commands and attributes. It should be very easy to
>> extend per container statistics, by adding members to the containerstats
>> structure.
> 
> Sorry for the delay in looking at this. (I've been travelling a bit).
> 
> The basic idea of being able to get stats on a per-container basis
> seems good, but I've got some suggestions on the API/implementation:
> 
> - saving a mount pointer in the containerfsroot structure won't work
> because a hierarchy (superblock) can be mounted in more than one
> place, or even in zero places (if you unmount a hierarchy with active
> containers, the superblock and the containers stay active). Also it
> might be possible to move a mounted container hierarchy via mount
> --move, although I've not actually tried that.

Yes, Good catch!

> 
> - a cleaner way to pass in a container id would be to pass a file
> descriptor on a container directory. The dentry associated with this
> fd would unambiguously identify the hierarchy and the container, so
> then even if we didn't maintain a per-container task likst, the
> for_each_thread() loop would involve just a single comparison per task
> to see if the task was in the desired container.

I thought about this approach, but did not implement the code this way
because a system could have thousands of containers and expecting a
statistics application to open a file descriptor each time for each
container will turn out to be an expensive operation

overhead =  2 syscalls (open + close) * number of containers * frequency
            of stats collections

I guess for now this is a good choice

> 
> - if we're trying to integrate this with taskstats, then it would be
> nice to support retrieving all the other taskstats values (where they
> make sense) on a per-container basis (in the same way that we can
> currently request them on a per-thread or a per-process basis). i.e.
> don't create a new container_taskstats structure, but instead augment
> the current taskstats structure, and allow the user to retrieve the
> aggregate of that for the entire container. Similarly, being able to
> get taskstats notifications based on the container memberships of a
> task might be nice.
> 

The reason why I did not do so (augmenting container stats) is that
we send out an event on every exit and that might be heavy for the
container. The current implementation does not support the push
model (where data is pushed from the kernel), but that should be
easy to support for container events (as and when we find a useful
event to support).



> Paul




-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL