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

[Peter Zijlstra]

On Mon, Oct 06, 2025 at 05:23:46PM -0400, Chris Mason wrote:
> On 10/6/25 4:23 PM, Joseph Salisbury wrote:
> > Hi Chris,
> > 
> > During testing, we are seeing a ~6% performance regression with the 
> > upstream stable v6.12.43 kernel (And Oracle UEK 
> > 6.12.0-104.43.4.el9uek.x86_64 kernel) when running the Phoronix 
> > pts/apache benchmark with 100 concurrent requests [0].  The regression 
> > is seen with the following hardware:
> > 
> > PROCESSOR: Intel Xeon Platinum 8167M Core Count: 8 Thread Count: 16 
> > Extensions: SSE 4.2 + AVX512CD + AVX2 + AVX + RDRAND + FSGSBASE Cache 
> > Size: 16 MB Microcode: 0x1 Core Family: Cascade Lake
> > 
> > After performing a bisect, we found that the performance regression was 
> > introduced by the following commit:
> > 
> > Stable v6.12.43: fc4289233e4b ("sched/fair: Bump sd->max_newidle_lb_cost 
> > when newidle balance fails")
> > Mainline v6.17-rc1: 155213a2aed4 ("sched/fair: Bump 
> > sd->max_newidle_lb_cost when newidle balance fails")
> > 
> > Reverting this commit causes the performance regression to not exist.
> > 
> > I was hoping to get your feedback, since you are the patch author.  Do 
> > you think gathering any additional data will help diagnose this issue?
> 
> Hi everyone,
> 
> Peter, we've had a collection of regression reports based on this
> change, so it sounds like we need to make it less aggressive, or maybe
> we need to make the degrading of the cost number more aggressive?
> 
> Joe (and everyone else who has hit this), can I talk you into trying the
> drgn from
> https://lore.kernel.org/lkml/2fbf24bc-e895-40de-9ff6-5c18b74b4300@xxxxxxxx/
> 
> I'm curious if it degrades at all or just gets stuck up high.

Right, that decay is very slow compared to the bump we give it. A little
more data there might be very useful.

If this is indeed due to lack of decay, we could look at making
update_s[gd]_lb_stats(), which is part of the regular load-balance, and has
a wider view of things, boost decay when 'sum_nr_running > nr_cpus' or
something like that. Perhaps proportional to the difference
'sum_nr_running - nr_cpus', more extra tasks more decay.