[PATCH v6 5/6] sched: replace capacity_factor by usage

Wanpeng Li kernellwp at gmail.com
Sat Nov 22 16:22:48 PST 2014

Hi Vincent,
On 10/3/14, 8:50 PM, Vincent Guittot wrote:
> On 3 October 2014 11:35, Morten Rasmussen <morten.rasmussen at arm.com> wrote:
>> On Fri, Oct 03, 2014 at 08:24:23AM +0100, Vincent Guittot wrote:
>>> On 2 October 2014 18:57, Morten Rasmussen <morten.rasmussen at arm.com> wrote:
>>>> On Tue, Sep 23, 2014 at 05:08:04PM +0100, Vincent Guittot wrote:
>>>>> Below are two examples to illustrate the problem that this patch solves:
>>>>> 1 - capacity_factor makes the assumption that max capacity of a CPU is
>>>>> SCHED_CAPACITY_SCALE and the load of a thread is always is
>>>>> SCHED_LOAD_SCALE. It compares the output of these figures with the sum
>>>>> of nr_running to decide if a group is overloaded or not.
>>>>> But if the default capacity of a CPU is less than SCHED_CAPACITY_SCALE
>>>>> (640 as an example), a group of 3 CPUS will have a max capacity_factor
>>>>> of 2 ( div_round_closest(3x640/1024) = 2) which means that it will be
>>>>> seen as overloaded if we have only one task per CPU.
>>>> I did some testing on TC2 which has the setup you describe above on the
>>>> A7 cluster when the clock-frequency property is set in DT. The two A15s
>>>> have max capacities above 1024. When using sysbench with five threads I
>>>> still get three tasks on the two A15s and two tasks on the three A7
>>>> leaving one cpu idle (on average).
>>>> Using cpu utilization (usage) does correctly identify the A15 cluster as
>>>> overloaded and it even gets as far as selecting the A15 running two
>>>> tasks as the source cpu in load_balance(). However, load_balance() bails
>>>> out without pulling the task due to calculate_imbalance() returning a
>>>> zero imbalance. calculate_imbalance() bails out just before the hunk you
>>>> changed due to comparison of the sched_group avg_loads. sgs->avg_load is
>>>> basically the sum of load in the group divided by its capacity. Since
>>>> load isn't scaled the avg_load of the overloaded A15 group is slightly
>>>> _lower_ than the partially idle A7 group. Hence calculate_imbalance()
>>>> bails out, which isn't what we want.
>>>> I think we need to have a closer look at the imbalance calculation code
>>>> and any other users of sgs->avg_load to get rid of all code making
>>>> assumptions about cpu capacity.
>>> We already had this discussion during the review of a previous version
>>> https://lkml.org/lkml/2014/6/3/422
>>> and my answer has not changed; This patch is necessary to solve the 1
>>> task per CPU issue of the HMP system but is not enough. I have a patch
>>> for solving the imbalance calculation issue and i have planned to send
>>> it once this patch will be in a good shape for being accepted by
>>> Peter.
>>> I don't want to mix this patch and the next one because there are
>>> addressing different problem: this one is how evaluate the capacity of
>>> a system and detect when a group is overloaded and the next one will
>>> handle the case when the balance of the system can't rely on the
>>> average load figures of the group because we have a significant
>>> capacity difference between groups. Not that i haven't specifically
>>> mentioned HMP for the last patch because SMP system can also take
>>> advantage of it
>> You do mention 'big' and 'little' cores in your commit message and quote
>> example numbers with are identical to the cpu capacities for TC2. That
> By last patch, i mean the patch about imbalance that i haven't sent
> yet, but it's not related with this patch
>> lead me to believe that this patch would address the issues we see for
>> HMP systems. But that is clearly wrong. I would suggest that you drop
> This patch addresses one issue: correctly detect how much capacity we
> have and correctly detect when the group is overloaded; HMP system

What's the meaning of "HMP system"?

Wanpeng Li

> clearly falls in this category but not only.
> This is the only purpose of this patch and not the "1 task per CPU
> issue" that you mentioned.
> The second patch is about correctly balance the tasks on system where
> the capacity of a group is significantly less than another group. It
> has nothing to do in capacity computation and overload detection but
> it will use these corrected/new metrics to make a better choice.
> Then, there is the "1 task per CPU issue on HMP" that you mentioned
> and this latter will be solved thanks to these 2 patchsets but this is
> not the only/main target of these patchsets so i don't want to reduce
> them into: "solve an HMP issue"
>> mentioning big and little cores and stick to only describing cpu
>> capacity reductions due to rt tasks and irq to avoid any confusion about
>> the purpose of the patch. Maybe explicitly say that non-default cpu
>> capacities (capacity_orig) isn't addressed yet.
> But they are addressed by this patchset. you just mixed various goal
> together (see above)
>> I think the two problems you describe are very much related. This patch
>> set is half the solution of the HMP balancing problem. We just need the
>> last bit for avg_load and then we can add scale-invariance on top.
> i don't see the link between scale invariance and a bad load-balancing
> choice. The fact that the current load balancer puts more tasks than
> CPUs in a group on HMP system should not influence or break the scale
> invariance load tracking. Isn't it ?
> Now, i could send the other patchset but i'm afraid that this will
> generate more confusion than help in the process review.
> Regards,
> Vincent
>> IMHO, it would be good to have all the bits and pieces for cpu capacity
>> scaling and scaling of per-entity load-tracking on the table so we fit
>> things together. We only have patches for parts of the solution posted
>> so far.
>> Morten
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo at vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/

More information about the linux-arm-kernel mailing list