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] [day] [month] [year] [list]
Message-ID: <2388de76-4827-46e3-a39f-4a19ddef9617@os.amperecomputing.com>
Date: Tue, 8 Jul 2025 21:27:02 +0800
From: Adam Li <adamli@...amperecomputing.com>
To: Peter Zijlstra <peterz@...radead.org>
Cc: mingo@...hat.com, juri.lelli@...hat.com, vincent.guittot@...aro.org,
 dietmar.eggemann@....com, rostedt@...dmis.org, bsegall@...gle.com,
 mgorman@...e.de, vschneid@...hat.com, cl@...ux.com,
 linux-kernel@...r.kernel.org, patches@...erecomputing.com
Subject: Re: [PATCH 0/2] Fix imbalance issue when balancing fork

On 7/4/2025 5:17 PM, Peter Zijlstra wrote:
> On Tue, Jul 01, 2025 at 02:45:47AM +0000, Adam Li wrote:
>> Load imbalance is observed when the workload frequently forks new threads.
>> Due to CPU affinity, the workload can run on CPU 0-7 in the first
>> group, and only on CPU 8-11 in the second group. CPU 12-15 are always idle.
>>
>> { 0 1 2 3 4 5 6 7 } {8 9 10 11 12 13 14 15}
>>   * * * * * * * *    * * *  *
>>
>> When looking for dst group for newly forked threads, in many times
>> update_sg_wakeup_stats() reports the second group has more idle CPUs
>> than the first group. The scheduler thinks the second group is less
>> busy. Then it selects least busy CPUs among CPU 8-11. So CPU 8-11 can be
>> crowded with newly forked threads, at the same time CPU 0-7 can be idle.
>>
>> The first patch 'Only update stats of allowed CPUs when looking for dst
>> group' *alone* can fix this imbalance issue.
>>
>> And I think the second patch also makes sense in this scenario. If group
>> weight includes CPUs a task cannot use, group classification can be
>> incorrect. Please comment.
>>
>> Adam Li (2):
>>   sched/fair: Only update stats of allowed CPUs when looking for dst
>>     group
>>   sched/fair: Only count group weight for allowed CPUs when looking for
>>     dst group
>>
>>  kernel/sched/fair.c | 6 ++++--
>>  1 file changed, 4 insertions(+), 2 deletions(-)
>>

Hi Peter,
> 
> Hurm... so the thing I noticed is that update_sg_wakeup_stats() and
> update_sg_lb_stats() are *very* similar.
> 
> Specifically, the first patch does something to wakeup_stats that
> lb_stats already does. While the second patch seems to do something that
> might also apply to lb_stats.
> 
Thanks for the idea. I am testing this.

> Is there no way we can unify this?
I will try to unify the common logic in update_sg_wakeup_stats() and
update_sg_lb_stats() into a single function. It seems not easy :).

Thanks,
-adam



Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ