Re: [RFC][PATCH] memory controller per zone patches take 2 [9/10] per-zone-lru for memory cgroup

[Balbir Singh <balbir-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org>]

KAMEZAWA Hiroyuki wrote:
> On Mon, 19 Nov 2007 11:51:15 +0530
> Balbir Singh <balbir-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org> wrote:
>>> =
>>> /cgroup/group_A/group_A_1
>>>             .  /group_A_2
>>>                /group_A_3
>>> (LRU(s) will be used for maintaining parent/child groups.)
>>>
>> The LRU's will be shared, my vision is
>>
>> 		LRU
>> 		^ ^
>> 		| |
>> 	Mem-----+ +----Mem
>>
>>
>> That two or more mem_cgroup's can refer to the same LRU list and have
>> their own resource counters. This setup will be used in the case
>> of a hierarchy, so that a child can share memory with its parent
>> and have it's own limit.
>>
>> The mem_cgroup will basically then only contain a reference
>> to the LRU list.
>>
> Hmm, interesting. 
> 
> Then, 
>    group_A_1's usage + group_A_2's usage + group_A_3's usgae < group_A's limit.
>    group_A_1, group_A_2, group_A_3 has its own limit.

Yes that is correct

> In plan.
> 
> I wonder if we want rich control functions, we need "share" or "priority" among
> childs. (but maybe this will be complicated one.)
> 

That would nice and the end goal of providing this feature. We also need
to provide soft-limits (more complex) and guarantees (with the kernel
memory controller coming in - nice to have, but not necessary for now)

> Thank you for explanation.
> 
> Regards,
> -Kame
> 

Thanks for helping out the memory controller.

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