Re: [CFS Bandwidth Control v4 3/7] sched: throttle cfs_rq entitieswhich exceed their local quota

From: Paul Turner
Date: Thu Feb 24 2011 - 22:59:55 EST

On Thu, Feb 24, 2011 at 9:20 AM, Peter Zijlstra <a.p.zijlstra@xxxxxxxxx> wrote:
> On Thu, 2011-02-24 at 22:09 +0530, Bharata B Rao wrote:
>> On Thu, Feb 24, 2011 at 04:52:53PM +0100, Peter Zijlstra wrote:
>> > On Thu, 2011-02-24 at 21:15 +0530, Bharata B Rao wrote:
>> > > While I admit that our load balancing semantics wrt thorttled entities are
>> > > not consistent (we don't allow pulling of tasks directly from throttled
>> > > cfs_rqs, while allow pulling of tasks from a throttled hierarchy as in the
>> > > above case), I am beginning to think if it works out to be advantageous.
>> > > Is there a chance that the task gets to run on other CPU where the hierarchy
>> > > isn't throttled since runtime is still available ?
>> >
>> > Possible yes, but the load-balancer doesn't know about that, not should
>> > it (its complicated, and broken, enough, no need to add more cruft to
>> > it).
>> >
>> > I'm starting to think you all should just toss all this and start over,
>> > its just too smelly.
>> Hmm... You have brought up 3 concerns:
>> 1. Hierarchy semantics
>> If you look at the heirarchy semantics we currently have while ignoring the
>> load balancer interactions for a moment, I guess what we have is a reasonable
>> one.
>> - Only group entities are throttled
>> - Throttled entities are taken off the runqueue and hence they never
>>   get picked up for scheduling.
>> - New or child entites are queued up to the throttled entities and not
>>   further up. As I said in another thread, having the tree intact and correct
>>   underneath the throttled entity allows us to rebuild the hierarchy during
>>   unthrottling with least amount of effort.
> It also gets you into all that load-balancer mess, and I'm not going to
> let you off lightly there.

I think the example was a little cuckoo. As you say, it's dequeued
and invisible to the load balancer.

The special case of block->wakeup->throttle->put only exists for the
current task which is ineligible for non-active load-balance anyway.

>> - Group entities in a hierarchy are throttled independent of each other based
>>   on their bandwidth specification.
> That's missing out quite a few details.. for one there is no mention of
> hierarchical implication of/constraints on bandwidth, can children have
> more bandwidth than their parent (I hope not).

I wasn't planning to enforce it since I believe there is value in
non-conformant constraints:


- I have some application that I want to limit to 3 cpus
I have a 2 workers in that application, across a period I would like
those workers to use a maximum of say 2.5 cpus each (suppose they
serve some sort of co-processor request per user and we want to
prevent a single user eating our entire limit and starving out
everything else).

The goal in this case is not preventing over-subscription, but
ensuring that some part threads is not allowed to blow our entire
quota, while not destroying the (relatively) work-conserving aspect of
its performance in general.

The above occurs sufficiently often that at the very least I think
conformance checking would have to be gated by a sysctl so that this
use case is still enabled.

- There's also the case of "I want to manage a newly abusive user,
being smart I've given his hierarchy a unique root so that I can
constrain them."
A non-conformant constraint avoids the adversarial problem of having
to find and bring all of their set (possibly maliciously large) limits
within the global limit I want to impose upon them.

My viewpoint was that if some idiot wants to set up such a tree
(unintentionally) it's their own damn fault but I suppose we should at
least give them a safety :) I'll add it.

>> 2. Handling of throttled entities by load balancer
>> This definetely needs to improve and be more consistent. We can work on this.
> Feh, improve is being nice about it, it needs a complete overhaul, the
> current situation is a cobbled together leaky mess.

I think as long as the higher level semantics are correct and
throttling happens /sanely/ this is a non-issue.

>> 3. per-cgroup vs global period specification
>> I thought per-cgroup specification would be most flexible and hence started
>> out with that. This would allow groups/workloads/VMs to define their
>> own bandwidth rate.
> Most flexible yes, most 'interesting' too, now if you consider running a
> child task is also running the parent entity and therefore you're
> consuming bandwidth up the entire hierarchy, what happens when the
> parent has a much larger period than the child?
> In that case your child doesn't get ran while the parent is throttled,
> and the child's period is violated.

There are definitely cases where this is both valid and useful. I
think gating conformancy allows for both (especially if it defaults to

>> Let us know if you have other design concerns besides these.
> Yeah, that weird time accounting muck, bandwidth should decrease on
> usage and incremented on replenishment, this gets you 0 as the natural
> boundary between credit and debt, no need to keep two variables.

Yes, agreed! Fixing :)

> Also, the above just about covers all the patch set does, isn't that
> enough justification to throw the thing out and start over?
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at