Re: [PATCH v4] time/tick-sched: idle load balancing when nohz_full cpu becomes idle.

[Frederic Weisbecker]

On Thu, May 16, 2024 at 10:49:11AM +0200, Peter Zijlstra wrote:
> On Thu, May 16, 2024 at 09:20:08AM +0100, Yun Levi wrote:
> > > None of that HK nonsense is relevant. The NOHZ_FULL nonsense implies
> > > single CPU partitions, and *that* should be avoiding any and all
> > > load-balancing.
> > 
> > Do you mean.. tick_nohz_full cpu (non-HK-ticked cpu) shouldn't belong
> > to any sched_domain?
> 
> AFAIK NOHZ_FULL still hard relies on the isolcpus garbage, so yeah, it
> should be all single cpu partitions, which don't have a domain.
> 
> (this really should migrate to use cpusets partitions)
> 
> > > If there still is, that's a bug, but that's not related to HK goo.
> > >
> > > As such, I don't think the HK_TYPE_SCHED check in
> > > nohz_balance_enter_idle() actually makes sense, the on_null_omain()
> > > check a little below that should already take care of things, no?
> > 
> > IIUC,
> > currently, whether cpu belongs on domain or null is determined by
> > HK_DOMAIN_FLAGS
> 
> No! you can create NULL domains without any of the HK nonsense. Both
> isolcpus and cpusets can create single CPU partitions.
> 
> > However, when "nohz_full=" is used, it still on HK_DOMAIN, so it
> > belongs to sched_domain
> > so, it couldn't be filtered out by on_null_domain().
> > 
> > unless "isolcpus=domain" or "isolcpus={cpu_list}", it's on null domain.
> > with "isolcpus=tick", it participates sched_domain.
> 
> Frederic ?!? You can use nohz_full without isolcpus? That makes no
> sense. If you do that you get to keep the pieces.

I fear you can yes, even though most users combine it with isolcpus. I
know, that interface is terrible but it dates from times when we weren't
sure about all the potential usecases of nohz_full. There was a possibility
that HPC could just want to reduce ticks without all the hard and costly
isolation around. But all the usecases I have witnessed so far in ten years
involved wanting 0 noise after all...