On 15 December 2015 at 09:50, Luca Abeni [off-list ref] wrote:

On 12/15/2015 05:59 AM, Vincent Guittot wrote:
[...]

quoted

quoted

quoted

quoted

So I don't think this is right. AFAICT this projects the WCET as the
amount of time actually used by DL. This will, under many
circumstances, vastly overestimate the amount of time actually
spend on it. Therefore unduly pessimisme the fair capacity of this
CPU.

I agree that if the WCET is far from reality, we will underestimate
available capacity for CFS. Have you got some use case in mind which
overestimates the WCET ?
If we can't rely on this parameters to evaluate the amount of capacity
used by deadline scheduler on a core, this will imply that we can't
also use it for requesting capacity to cpufreq and we should fallback
on a monitoring mechanism which reacts to a change instead of
anticipating it.

I think a more "theoretically sound" approach would be to track the
_active_ utilisation (informally speaking, the sum of the utilisations
of the tasks that are actually active on a core - the exact definition
of "active" is the trick here).

The point is that we probably need 2 definitions of "active" tasks.

Ok; thanks for clarifying. I do not know much about the remaining capacity
used by CFS; however, from what you write I guess CFS really need an
"average"
utilisation (while frequency scaling needs the active utilisation).

yes. this patch is only about the "average" utilization

So, I suspect you really need to track 2 different things.
From a quick look at the code that is currently in mainline, it seems to
me that it does a reasonable thing for tracking the remaining capacity
used by CFS...

quoted

The 1st one would be used to scale the frequency. From a power saving
point of view, it have to reflect the minimum frequency needed at the
current time to handle all works without missing deadline.

Right. And it can be computed as shown in the GRUB-PA paper I mentioned
in a previous mail (that is, by tracking the active utilisation, as done
by my patches).

I fully trust you on that part.

quoted

This one
should be updated quite often with the wake up and the sleep of tasks
as well as the throttling.

Strictly speaking, the active utilisation must be updated when a task
wakes up and when a task sleeps/terminates (but when a task
sleeps/terminates
you cannot decrease the active utilisation immediately: you have to wait
some time because the task might already have used part of its "future
utilisation").
The active utilisation must not be updated when a task is throttled: a
task is throttled when its current runtime is 0, so it already used all
of its utilisation for the current period (think about two tasks with
runtime=50ms and period 100ms: they consume 100% of the time on a CPU,
and when the first task consumed all of its runtime, you cannot decrease
the active utilisation).

 I haven't read the paper you pointed in the previous email but it's
on my todo list. Does the GRUB-PA take into account the frequency
transition when selecting the best frequency ?

quoted

The 2nd definition is used to compute the remaining capacity for the
CFS scheduler. This one doesn't need to be updated at each wake/sleep
of a deadline task but should reflect the capacity used by deadline in
a larger time scale. The latter will be used by the CFS scheduler  at
the periodic load balance pace

Ok, so as I wrote above this really looks like an average utilisation.
My impression (but I do not know the CFS code too much) is that the mainline
kernel is currently doing the right thing to compute it, so maybe there is
no
need to change the current code in this regard.
If the current code is not acceptable for some reason, an alternative would
be to measure the active utilisation for frequency scaling, and then apply a
low-pass filter to it for CFS.


                                Luca

quoted

quoted

As done, for example, here:
https://github.com/lucabe72/linux-reclaiming/tree/track-utilisation-v2
(in particular, see

https://github.com/lucabe72/linux-reclaiming/commit/49fc786a1c453148625f064fa38ea538470df55b
)
I understand this approach might look too complex... But I think it is
much less pessimistic while still being "safe".
If there is something that I can do to make that code more acceptable,
let me know.


                         Luca