Re: [PATCH 1/9] sched/balancing: Switch the 'DEFINE_SPINLOCK(balancing)' spinlock into an 'atomic_t sched_balance_running' flag

From: Ingo Molnar <mingo@kernel.org>
To: Shrikanth Hegde <sshegde@linux.ibm.com>
Cc: linux-kernel@vger.kernel.org,
	Peter Zijlstra <peterz@infradead.org>,
	Vincent Guittot <vincent.guittot@linaro.org>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Valentin Schneider <vschneid@redhat.com>
Subject: Re: [PATCH 1/9] sched/balancing: Switch the 'DEFINE_SPINLOCK(balancing)' spinlock into an 'atomic_t sched_balance_running' flag
Date: Fri, 8 Mar 2024 12:23:10 +0100	[thread overview]
Message-ID: <Zer1Hkxh/UMxs3xs@gmail.com> (raw)
In-Reply-To: <bf612672-f7c3-4585-ac31-e02a1ebf614c@linux.ibm.com>

* Shrikanth Hegde <sshegde@linux.ibm.com> wrote:

> system is at 75% load		<-- 	25.6% contention
> 113K probe:rebalance_domains_L37
> 84K probe:rebalance_domains_L55
> 
> 87
> system is at 100% load		<--	87.5% contention.
> 64K probe:rebalance_domains_L37
> 8K probe:rebalance_domains_L55
> 
> 
> A few reasons for contentions could be: 
>
> 1. idle load balance is running and some other cpu is becoming idle, and 
>    tries newidle_balance.
>
> 2. when system is busy, every CPU would do busy balancing, it would 
>    contend for the lock. It will not do balance as should_we_balance says 
>    this CPU need not balance. It bails out and release the lock.

Thanks, these measurements are really useful!

Would it be possible to disambiguate these two cases?

I think we should probably do something about this contention on this large 
system: especially if #2 'no work to be done' bailout is the common case.

Thanks,

	Ingo