linux-kernel - Re: [PATCH v3 0/3] sched/fair: Capacity aware wakeup rework

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <9b6d0ea2-cf5e-ed89-6c08-6f43c345265f@arm.com>
Date:   Tue, 28 Jan 2020 11:34:21 +0000
From:   Valentin Schneider <valentin.schneider@....com>
To:     Dietmar Eggemann <dietmar.eggemann@....com>,
        linux-kernel@...r.kernel.org
Cc:     mingo@...hat.com, peterz@...radead.org, vincent.guittot@...aro.org,
        morten.rasmussen@....com, qperret@...gle.com,
        adharmap@...eaurora.org
Subject: Re: [PATCH v3 0/3] sched/fair: Capacity aware wakeup rework

On 28/01/2020 09:24, Dietmar Eggemann wrote:
> On 26/01/2020 21:09, Valentin Schneider wrote:
> 
> [...]
> 
>> v2 -> v3
>> --------
>> o Added missing sync_entity_load_avg() (Quentin)
>> o Added fallback CPU selection (maximize capacity)
>> o Added special case for CPU hogs: task_fits_capacity() will always return 'false'
>>   for tasks that are simply too big, due to the margin.
> 
> v3 fixes the Geekbench multicore regression I saw on Pixel4 (Android 10,
> Android Common Kernel v4.14 based, Snapdragon 855) running v1.
> 
> I changed the Pixel4 kernel a bit (PELT instead WALT, mainline
> select_idle_sibling() instead the csctate aware one), mainline
> task_fits_capacity()) for base, v1 & v3.
> 
> Since it's not mainline kernel the results have to be taken with a pinch
> of salt but they probably show that the new condition:
> 
> if (rq->cpu_capacity_orig == READ_ONCE(rq->rd->max_cpu_capacity) && ...
>     return cpu;
> 
> has an effect when dealing with tasks with util_est values > 0.8 * 1024.
> 

Thanks for going through that hassle! I'm quite pleased with the results
of that change - if you also look at e.g. the stddev for sysbench on Juno,
runs are much more consistent now.