Re: [ckrm-tech] [RFC 3/5] sched: Add CPU rate hard caps

[Balbir Singh <balbir@in.ibm.com>]

Peter Williams wrote:
> > > > The problem is that with per-task caps, if I have a resource group A
> > > > and I want to limit it to 10%, I need to limit each task in resource
> > > > group A to 10% (which makes resource groups not so useful). Is my
> > > > understanding correct?
> > >
> > >
> > > Well the general idea is correct but your maths is wrong.  You'd have 
> > > to give each of them a cap somewhere between 10% and 10% divided by 
> > > the number of tasks in group A.  Exactly where in that range would 
> > > vary depending on the CPU demand of each task and would need to be 
> > > adjusted dynamically (unless they were very boring tasks whose demands 
> > > were constant over time).
> >
> >
> > Hmm.. I thought my math was reasonable (but there is always so much to 
> > learn)
> > From your formula, if I have 1 task in group A, I need to provide it with
> > a cap of b/w 10 to 11%. For two tasks, I need to give them b/w 10 to 10.5%.
> > If I have a hundred, it needs to be b/w 10% and 10.01%
>
>
> Now your arithmetic is failing you.  According to my formula:
>
> 1. With one task in group A you give it 10% which is what you get when 
> you divide 10% by one.
>
> 2. With two tasks in group A you give them each somewhere between 5% 
> (which is 10% divided by 2) and 10%.  If they are equally busy you give 
> them each 5% and if they are not equally busy you give them you give 
> them larger caps.

Yes, I understand. I misinterpreted what you said earlier. I see you
clearly meant the range [cap_of_the_group/number_of_tasks, cap_of_the_group]

> Another, probably a better but more expensive, formula is to divide the 
> 10% between them in proportion to their demand.  Being careful not to 
> give any of them a zero cap, of course.  I.e. in the two task 10% case 
> they each get 5% if they are equally busy but if one is twice as busy as 
> the other it gets a 6.6% cap and the other gets 3.3% (approximately).

Yes, that makes a lot of sense

> Peter

Thanks for clarifying.

--

	Balbir Singh,
	Linux Technology Center,
	IBM Software Labs
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/