Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 13 Feb 2020 08:53:48 -0500
From:   Tejun Heo <tj@kernel.org>
To:     Michal Hocko <mhocko@kernel.org>
Cc:     Johannes Weiner <hannes@cmpxchg.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        Roman Gushchin <guro@fb.com>, linux-mm@kvack.org,
        cgroups@vger.kernel.org, linux-kernel@vger.kernel.org,
        kernel-team@fb.com
Subject: Re: [PATCH v2 3/3] mm: memcontrol: recursive memory.low protection
Message-ID: <20200213135348.GF88887@mtj.thefacebook.com>
References: <20191219200718.15696-1-hannes@cmpxchg.org>
 <20191219200718.15696-4-hannes@cmpxchg.org>
 <20200130170020.GZ24244@dhcp22.suse.cz>
 <20200203215201.GD6380@cmpxchg.org>
 <20200211164753.GQ10636@dhcp22.suse.cz>
 <20200212170826.GC180867@cmpxchg.org>
 <20200213074049.GA31689@dhcp22.suse.cz>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200213074049.GA31689@dhcp22.suse.cz>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Hello, Michal.

On Thu, Feb 13, 2020 at 08:40:49AM +0100, Michal Hocko wrote:
> So how are we going to deal with hierarchies where the actual workload
> of interest is a leaf deeper in the hierarchy and the upper levels of
> the hierarchy are shared between unrelated workloads?  Look at how
> systemd organizes system into cgroups for example (slices vs. scopes)
> and say you want to add a protection to a single user or a service.

The above scenario where descendants of a cgroup behave unrelated to
each other sound plausible in theory but doesn't really work in
practice because memory management is closely tied with IO and all IO
control mechanisms are strictly hierarchical and arbitrate
level-by-level.

A workload's memory footprint can't be protected without protecting
its IOs and a descendant cgroup's IOs can't be protected without
protecting its ancestors, so implementing that in memory controller in
isolation, while doable, won't serve many practical purposes. It just
ends up creating cgroups which are punished on memory while greedly
burning up IOs.

The same pattern for CPU control, so for any practical configuration,
the hierarchy layout has to follow the resource distribution hierarchy
of the system - it's what the whole thing is for after all. The
existing memory.min/low semantics is mostly from failing to recognize
them being closely intertwined with IO and resembling weight based
control mechanisms, and rather blindly copying memory.high/max
behaviors, for which, btw, individual configurations may make sense.

Thanks.

-- 
tejun