From:   Valentin Schneider <vschneid@redhat.com>
To:     tmmdh <921889738@qq.com>, hupu <921889738@qq.com>,
        peterz@infradead.org
Cc:     bristot@redhat.com, bsegall@google.com, dietmar.eggemann@arm.com,
        hupu@oppo.com, ionela.voinescu@arm.com, juri.lelli@redhat.com,
        len.brown@intel.com, linux-kernel@vger.kernel.org, mgorman@suse.de,
        rafael.j.wysocki@intel.com, ravi.v.shankar@intel.com,
        ricardo.neri-calderon@linux.intel.com, ricardo.neri@intel.com,
        rostedt@goodmis.org, srinivas.pandruvada@linux.intel.com,
        tim.c.chen@intel.com, tim.c.chen@linux.intel.com,
        vincent.guittot@linaro.org, x86@kernel.org
Subject: Re: [PATCH v3 06/10] sched/fair: Use the prefer_sibling flag of the
 current sched domain
In-Reply-To: <tencent_6C38D389245FD03C6E1312999FEDD394F606@qq.com>
References: <xhsmh1qmkr7pa.mognet@vschneid.remote.csb>
 <tencent_6C38D389245FD03C6E1312999FEDD394F606@qq.com>
Date:   Tue, 21 Feb 2023 18:15:34 +0000
Message-ID: <xhsmhttzeq3zt.mognet@vschneid.remote.csb>
MIME-Version: 1.0
Content-Type: text/plain
Precedence: bulk

On 21/02/23 17:49, tmmdh wrote:
> From: Valentin Schneider <vschneid@redhat.com>
> But I am still confused by Peter's description about prefer_sibling in the last email, linked at
> https://lore.kernel.org/all/Y+Z2b%2FOtZDk9cT53@hirez.programming.kicks-ass.net/
>> this morning -- conceptually, prefer siblings says you want to try
>> sibling domains before filling up your current domain.
>
> Why should we try sibling domains before filling up your current domain? Why does Peter think that sibling domains is better than current domain.
> My understanding about this problem is described as follows, but I am not
> sure if it is correct. I think the sibling domain is a domain lower than
> the current level. Just like SMT is the sibling of MC, while DIE is the
> sibling of NUMA.

That's the wrong way around; going up (or down) the sched_domain hierarchy is
done via parent (or child) pointers. Sibling means going sideways (i.e. the
same topology level but viewed from a different CPU)

> Is it because the cpus covered by sibling domains share more resources (such as cache), which can improve the performance of task running?
>

If sibling domains are in same cache hierarchy then spreading tasks between
them can improve overall performance. That doesn't work with NUMA or
big.LITTLE domains, so we don't have the flag on those.