Date:   Thu, 05 Sep 2019 12:18:55 +0100
From:   Patrick Bellasi <patrick.bellasi@....com>
To:     Peter Zijlstra <peterz@...radead.org>
Cc:     Subhra Mazumdar <subhra.mazumdar@...cle.com>,
        linux-kernel@...r.kernel.org, mingo@...hat.com, tglx@...utronix.de,
        steven.sistare@...cle.com, dhaval.giani@...cle.com,
        daniel.lezcano@...aro.org, vincent.guittot@...aro.org,
        viresh.kumar@...aro.org, tim.c.chen@...ux.intel.com,
        mgorman@...hsingularity.net, parth@...ux.ibm.com
Subject: Re: [RFC PATCH 1/9] sched,cgroup: Add interface for latency-nice


On Thu, Sep 05, 2019 at 11:46:16 +0100, Peter Zijlstra wrote...

> On Thu, Sep 05, 2019 at 10:45:27AM +0100, Patrick Bellasi wrote:
>
>> > From just reading the above, I would expect it to have the range
>> > [-20,19] just like normal nice. Apparently this is not so.
>> 
>> Regarding the range for the latency-nice values, I guess we have two
>> options:
>> 
>>   - [-20..19], which makes it similar to priorities
>>   downside: we quite likely end up with a kernel space representation
>>   which does not match the user-space one, e.g. look at
>>   task_struct::prio.
>> 
>>   - [0..1024], which makes it more similar to a "percentage"
>> 
>> Being latency-nice a new concept, we are not constrained by POSIX and
>> IMHO the [0..1024] scale is a better fit.
>> 
>> That will translate into:
>> 
>>   latency-nice=0 : default (current mainline) behaviour, all "biasing"
>>   policies are disabled and we wakeup up as fast as possible
>> 
>>   latency-nice=1024 : maximum niceness, where for example we can imaging
>>   to turn switch a CFS task to be SCHED_IDLE?
>
> There's a few things wrong there; I really feel that if we call it nice,
> it should be like nice. Otherwise we should call it latency-bias and not
> have the association with nice to confuse people.
>
> Secondly; the default should be in the middle of the range. Naturally
> this would be a signed range like nice [-(x+1),x] for some x. but if you
> want [0,1024], then the default really should be 512, but personally I
> like 0 better as a default, in which case we need negative numbers.
>
> This is important because we want to be able to bias towards less
> importance to (tail) latency as well as more importantance to (tail)
> latency.
>
> Specifically, Oracle wants to sacrifice (some) latency for throughput.
> Facebook OTOH seems to want to sacrifice (some) throughput for latency.

Right, we have this dualism to deal with and current mainline behaviour
is somehow in the middle.

BTW, the FB requirement is the same we have in Android.
We want some CFS tasks to have very small latency and a low chance
to be preempted by the wake-up of less-important "background" tasks.

I'm not totally against the usage of a signed range, but I'm thinking
that since we are introducing a new (non POSIX) concept we can get the
chance to make it more human friendly.

Give the two extremes above, would not be much simpler and intuitive to
have 0 implementing the FB/Android (no latency) case and 1024 the
(max latency) Oracle case?

Moreover, we will never match completely the nice semantic, give that
a 1 nice unit has a proper math meaning, isn't something like 10% CPU
usage change for each step?

For latency-nice instead we will likely base our biasing strategies on
some predefined (maybe system-wide configurable) const thresholds.

Could changing the name to "latency-tolerance" break the tie by marking
its difference wrt prior/nice levels? AFAIR, that was also the original
proposal [1] by PaulT during the OSPM discussion.

Best,
Patrick

[1] https://youtu.be/oz43thSFqmk?t=1302

-- 
#include <best/regards.h>

Patrick Bellasi

Powered by blists - more mailing lists