Received-SPF: pass (google.com: domain of linux-kernel+bounces-49229-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
From: Efly Young <yangyifei03@kuaishou.com>
To: <hannes@cmpxchg.org>
CC: <akpm@linux-foundation.org>, <android-mm@google.com>,
	<cgroups@vger.kernel.org>, <linux-kernel@vger.kernel.org>,
	<linux-mm@kvack.org>, <mhocko@kernel.org>, <mkoutny@suse.com>,
	<muchun.song@linux.dev>, <roman.gushchin@linux.dev>, <shakeelb@google.com>,
	<tjmercier@google.com>, <yangyifei03@kuaishou.com>, <yuzhao@google.com>
Subject: Re: [PATCH] mm: memcg: Use larger chunks for proactive reclaim
Date: Fri, 2 Feb 2024 13:02:47 +0800
Message-ID: <20240202050247.45167-1-yangyifei03@kuaishou.com>
In-Reply-To: <20240201153428.GA307226@cmpxchg.org>
References: <20240201153428.GA307226@cmpxchg.org>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset="y"
Content-Transfer-Encoding: 8bit

> On Thu, Feb 01, 2024 at 02:57:22PM +0100, Michal Koutný wrote:
> > Hello.
> > 
> > On Wed, Jan 31, 2024 at 04:24:41PM +0000, "T.J. Mercier" <tjmercier@google.com> wrote:
> > >     reclaimed = try_to_free_mem_cgroup_pages(memcg,
> > > -         min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
> > > +         max((nr_to_reclaim - nr_reclaimed) / 4,
> > > +             (nr_to_reclaim - nr_reclaimed) % 4),
> > 
> > The 1/4 factor looks like magic.
> 
> It's just cutting the work into quarters to balance throughput with
> goal accuracy. It's no more or less magic than DEF_PRIORITY being 12,
> or SWAP_CLUSTER_MAX being 32.
> 
> > Commit 0388536ac291 says:
> > | In theory, the amount of reclaimed would be in [request, 2 * request).
> 
> Looking at the code, I'm not quite sure if this can be read this
> literally. Efly might be able to elaborate, but we do a full loop of
> all nodes and cgroups in the tree before checking nr_to_reclaimed, and
> rely on priority level for granularity. So request size and complexity
> of the cgroup tree play a role. I don't know where the exact factor
> two would come from.

I'm sorry that this conclusion may be arbitrary. It might just only suit
for my case. In my case, I traced it loop twice every time before checking
nr_reclaimed, and it reclaimed less than my request size(1G) every time.
So I think the upper bound is 2 * request. But now it seems that this is
related to cgroup tree I constucted and my system status and my request
size(a relatively large chunk). So there are many influencing factors,
a specific upper bound is not accurate.

> IMO it's more accurate to phrase it like this:
> 
> Reclaim tries to balance nr_to_reclaim fidelity with fairness across
> nodes and cgroups over which the pages are spread. As such, the bigger
> the request, the bigger the absolute overreclaim error. Historic
> in-kernel users of reclaim have used fixed, small request batches to
> approach an appropriate reclaim rate over time. When we reclaim a user
> request of arbitrary size, use decaying batches to manage error while
> maintaining reasonable throughput.
> 
> > Doesn't this suggest 1/2 as a better option? (I didn't pursue the
> > theory.)
> 
> That was TJ's first suggestion as well, but as per above I suggested
> quartering as a safer option.
> 
> > Also IMO importantly, when nr_to_reclaim - nr_reclaimed is less than 8,
> > the formula gives arbitrary (unrelated to delta's magnitude) values.
> 
> try_to_free_mem_cgroup_pages() rounds up to SWAP_CLUSTER_MAX. So the
> error margin is much higher at the smaller end of requests anyway.
> But practically speaking, users care much less if you reclaim 32 pages
> when 16 were requested than if you reclaim 2G when 1G was requested.

Yes, I agreed completely that the bigger the request the bigger the
absolute overreclaim error. The focus now is the tradeoff between 
accurate reclaim and efficient reclaim. I think TJ's test is suggestive.