Re: [PATCH 0/4] sched: Various reweight_entity() fixes

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Feb 18, 2026 at 10:37:19AM -0800, Shubhang Kaushik wrote:
> Hello Peter,
> 
> On Fri, 30 Jan 2026, Peter Zijlstra wrote:
> 
> > Two issues related to reweight_entity() were raised; poking at all that got me
> > these patches.
> > 
> > They're in queue.git/sched/core and I spend most of yesterday staring at traces
> > trying to find anything wrong. So far, so good.
> > 
> > Please test.
> > 
> > 
> 
> Analysis of commit bdba3187771c ("sched/fair: Use full weight to
> __calc_delta()") confirms it resolves a scaling bottleneck present in the
> v7.0-rc baseline (cee73b1e840c).
> 
> On Ampere Altra (80 cores) non-SMT SMP with CONFIG_NOHZ_FULL and
> CONFIG_SCHED_CLUSTER enabled, the baseline suffered from vruntime distortion
> that caused significant scheduling jitter, resulting in measurable
> performance cliffs under high parallel load.
> 
> Llama AIO for Inference Scaling : The baseline shows a performance cliffs at
> high thread counts, due to the scheduler miscalculating time slices for tail
> threads, which forced the entire synchronous inference batch to wait. The
> patch corrects the delta calculation for predictable scaling.
> 
> n_threads	Baseline (cee73b1e840c)	Patched (bdba3187771c)	Diff
> 51		8.46 t/s		12.12 t/s		+43.3%
> 52		17.75 t/s		22.26 t/s		+25.4%
> 64		12.40 t/s		14.66 t/s		+18.2%
> 
> Hackbench scaling in the baseline was non-linear indicating excessive
> scheduling overhead as process counts increased. By using the full weight
> for delta calculations, the patch reduces unnecessary migrations and
> improves message throughput.
> 
> Task Count	Baseline(sec)	Patched(sec)	Throughput
> 16 Processes	51.25s		44.09s		+16.2%
> 32 Processes	92.42s		78.09s		+18.3%
> 
> Tested-by: Shubhang Kaushik <shubhang@os.amperecomputing.com>

These are excellent results, thanks!

Let me finally get around to posting them.