Re: Revert for cgroups CPU accounting subsystem patch

From: Balbir Singh <balbir@linux.vnet.ibm.com>
To: Paul Menage <menage@google.com>
Cc: containers@lists.linux-foundation.org, vatsa@linux.vnet.ibm.com,
	LKML <linux-kernel@vger.kernel.org>, Ingo Molnar <mingo@elte.hu>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Andrew Morton <akpm@linux-foundation.org>
Subject: Re: Revert for cgroups CPU accounting subsystem patch
Date: Tue, 13 Nov 2007 12:59:59 +0530	[thread overview]
Message-ID: <47395277.1060006@linux.vnet.ibm.com> (raw)
In-Reply-To: <6599ad830711122310nf7530cfs5ef1fea061b1252c@mail.gmail.com>

Paul Menage wrote:
> On Nov 12, 2007 11:00 PM, Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
>> Right now, one of the limitations of the CPU controller is that
>> the moment you create another control group, the bandwidth gets
>> divided by the default number of shares. We can't create groups
>> just for monitoring.
> 
> Could we get around this with, say, a flag that always treats a CFS
> schedulable entity as having a weight equal to the number of runnable
> tasks in it? So CPU bandwidth would be shared between groups in
> proportion to the number of runnable tasks, which would distribute the
> cycles approximately equivalently to them all being separate
> schedulable entities.
> 

I think it's a good hack, but not sure about the complexity to implement
the code. I worry that if the number of tasks increase (say run into
thousands for one or more groups and a few groups have just a few
tasks), we'll lose out on accuracy.

>> cpu_acct fills this gap.
> 
> Agreed, but not in the right way IMO.
> 

I think we already have the code, we need to make it more useful and
reusable.

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