Re: [External] : Re: [REGRESSION][v6.17-rc1]sched/fair: Bump sd->max_newidle_lb_cost when newidle balance fails

[Dietmar Eggemann <dietmar.eggemann@arm.com>]

On 30.10.25 10:30, Peter Zijlstra wrote:
> 
> Hi Adam,
> 
> On Thu, Oct 30, 2025 at 03:22:34PM +0800, Adam Li wrote:
> 
>> Commit 155213a2aed4 brings ~6% regression for SpecJBB critical-jOPS on
>> AmpereOne server. 
>>
>> Baseline: v6.17 kernel + patch. Compared with baseline:
>> NI_TARGET	+7%
>> NI_SPARE	-3%
>> NI_RUNNABLE	-4%
>> NI_RANDOM	-3%
>>
>> So NI_TARGET reverts the performance regression.
>> The other options brings more regression for SpecJBB.
> 
> Excellent, how does NI_TARGET+NI_RANDOM work for you? Like Chris, I
> found that the schbench workload was really suffering from doing too
> many idle balance attempts that weren't really working out.
> 
> NI_RANDOM reduces the number of newidle balances based on the success
> ratio. Eg. if you have successful balancing 25% of the time, it will do
> 1 in 4 balances and count a successful balance as 4 to keep the
> accounting straight.
> 
> So workloads with a high success rate will keep newidle balance calls
> frequent, while workloads with a low success rate (like schbench) will
> greatly reduce the number of calls.

Results with OLTP 'hammerdb - mysqld' on Arm64 VMs

				NOPM		P50 latency

6.18-rc3			baseline	baseline

6.18-rc3-revert-155213a2aed4	+12%		-8.5%

6.18-rc3 NI_TARGET		+12%		-8.2%

6.18-rc3 NI_SPARE		+-0%		+-0%

6.18-rc3 NI_RUNNABLE		-1%		+1%

6.18-rc3 NI_RANDOM	        -1%		+1%

Looks like only NI_TARGET reverts the regression here.