Re: [PATCH v2] sched/fair: bump sd->max_newidle_lb_cost when newidle balance fails

[Chen, Yu C]

On 7/16/2025 7:25 PM, Peter Zijlstra wrote:
> On Tue, Jul 15, 2025 at 06:08:43PM +0800, Chen, Yu C wrote:
> > On 7/15/2025 3:08 PM, kernel test robot wrote:
> > > Hello,
> > >
> > > kernel test robot noticed a 22.9% regression of unixbench.throughput on:
> > >
> > >
> > > commit: ac34cb39e8aea9915ec2f4e08c979eb2ed1d7561 ("[PATCH v2] sched/fair: bump sd->max_newidle_lb_cost when newidle balance fails")
> > > url: https://github.com/intel-lab-lkp/linux/commits/Chris-Mason/sched-fair-bump-sd-max_newidle_lb_cost-when-newidle-balance-fails/20250626-224805
> > > base: https://git.kernel.org/cgit/linux/kernel/git/tip/tip.git 5bc34be478d09c4d16009e665e020ad0fcd0deea
> > > patch link: https://lore.kernel.org/all/20250626144017.1510594-2-clm@xxxxxx/
> > > patch subject: [PATCH v2] sched/fair: bump sd->max_newidle_lb_cost when newidle balance fails
> > >
> > > testcase: unixbench
> > > config: x86_64-rhel-9.4
> > > compiler: gcc-12
> > > test machine: 64 threads 2 sockets Intel(R) Xeon(R) Gold 6346 CPU @ 3.10GHz (Ice Lake) with 256G memory
> > > parameters:
> > >
> > > 	runtime: 300s
> > > 	nr_task: 100%
> > > 	test: shell1
> > > 	cpufreq_governor: performance
> > ...
> >
> > > commit:
> > >     5bc34be478 ("sched/core: Reorganize cgroup bandwidth control interface file writes")
> > >     ac34cb39e8 ("sched/fair: bump sd->max_newidle_lb_cost when newidle balance fails")
> > >
> > > 5bc34be478d09c4d ac34cb39e8aea9915ec2f4e08c9
> > > ---------------- ---------------------------
> > >            %stddev     %change         %stddev
> > >                \          |                \
> > ...
> >
> > >        40.37           +16.9       57.24        mpstat.cpu.all.idle%
> >
> > This commit inhibits the newidle balance.
>
> When not successful. So when newidle balance is not succeeding to pull
> tasks, it is backing off and doing less of it.
>
> > It seems that some workloads
> > do not like newlyidle balance, like schbench, which is short duration
> > task. While other workloads want the newidle balance to pull at its best
> > effort, like unixbench shell test case.
> > Just wonder if we can check the sched domain's average utilization to
> > decide how hard we should trigger the newly idle balance, or can we check
> > the overutilized flag to decide whether we should launch the
> > new idle balance, something I was thinking of:
>
> Looking at the actual util signal might be interesting, but as Chris
> already noted, overutilized isn't the right thing to look at. Simply
> taking rq->cfs.avg.util_avg might be more useful. Very high util and
> failure to pull might indicate new-idle just isn't very important /
> effective. While low util and failure might mean we should try harder.
>
> Other things to look at:
>
>   - if the sysctl_sched_migration_cost limit isn't artificially limiting
>     actual scanning costs. Eg. very large domains might perhaps have
>     costs that are genuinely larger than that somewhat random number.
>
>   - if despite the apparent failure to pull, we do already have something
>     to run (eg. wakeups).
>
>   - if the 3/2 backoff is perhaps too aggressive vs the 1% per second
>     decay.

Thanks for the suggestions, let me try to reproduce this issue locally
to see what is the proper way to get it addressed.


thanks,
Chenyu