Re: [RFC][PATCH] another swap controller for cgroup

[Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org>]

On 2008/05/15 17:56 +0900, YAMAMOTO Takashi wrote:
>>>  > If so, why is this better
>>>  > than charging for actual swap usage?
>>>
>>>  its behaviour is more determinstic and it uses less memory.
>>>  (than nishimura-san's one, which charges for actual swap usage.)
>>>

Consuming more memory cannot be helped for my controller...

>> Using less memory is good, but maybe not worth it if the result isn't so useful.
>>
>> I'd say that it's less deterministic than nishimura-san's controller -
>> with his you just need to know how much swap is in use (which you can
>> tell by observing the app on a real system) but with yours you also
>> have to know whether there are any processes sharing anon pages (but
>> not mms).
> 
> deterministic in the sense that, even when two or more processes
> from different cgroups are sharing a page, both of them, rather than
> only unlucky one, are always charged.
> 

I'm not sure whether this behavior itself is good or bad,
but I think it's not good idea to make memory controller,
which charges only one process for a shared page,
and swap controller behave differently.
I think it will be confusing for users. At least,
I would feel it strange.

> another related advantage is that it's possible to move charges
> quite precisely when moving a task among cgroups.
> 

Moving charges is one of future todo of my controller.
But, as you say, it won't be so precise as yours.


Thanks,
Daisuke Nishimura.