Received-SPF: pass (google.com: domain of linux-kernel+bounces-64003-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Date: Tue, 13 Feb 2024 17:45:26 +0000
In-Reply-To: <1e5ffefa-8c80-44b8-986f-ee574c3b3349@bytedance.com>
Precedence: bulk
Mime-Version: 1.0
References: <20240210-zswap-global-lru-v1-0-853473d7b0da@bytedance.com>
 <20240210-zswap-global-lru-v1-2-853473d7b0da@bytedance.com>
 <CAJD7tkYdEWrrQj+MFNpDR6FHc60czXcuawh3dd1rSF=QQm6+yg@mail.gmail.com> <1e5ffefa-8c80-44b8-986f-ee574c3b3349@bytedance.com>
Message-ID: <ZcuqtiIsAmDedOqK@google.com>
Subject: Re: [PATCH 2/2] mm/zswap: change zswap_pool kref to percpu_ref
From: Yosry Ahmed <yosryahmed@google.com>
To: Chengming Zhou <zhouchengming@bytedance.com>
Cc: Andrew Morton <akpm@linux-foundation.org>, Johannes Weiner <hannes@cmpxchg.org>, 
	Nhat Pham <nphamcs@gmail.com>, linux-mm@kvack.org, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: quoted-printable

On Tue, Feb 13, 2024 at 10:31:16PM +0800, Chengming Zhou wrote:
> On 2024/2/13 06:42, Yosry Ahmed wrote:
> > On Sun, Feb 11, 2024 at 5:58=E2=80=AFAM Chengming Zhou
> > <zhouchengming@bytedance.com> wrote:
> >>
> >> All zswap entries will take a reference of zswap_pool when
> >> zswap_store(), and drop it when free. Change it to use the
> >> percpu_ref is better for scalability performance.
> >>
> >> Testing kernel build in tmpfs with memory.max=3D2GB
> >> (zswap shrinker and writeback enabled with one 50GB swapfile).
> >>
> >>         mm-unstable  zswap-global-lru
> >> real    63.20        63.12
> >> user    1061.75      1062.95
> >> sys     268.74       264.44
> >=20
> > Are these numbers from a single run or the average of multiple runs?
>=20
> The average of 5 runs. And I just checked/compared each run result,
> the improvement is stable. So yes, it should be a real performance gain.
>=20
> > It just seems that the improvement is small, and percpu refcnt is
> > slightly less intuitive (and uses a bit more memory), so let's make
> > sure there is a real performance gain first.
>=20
> Right, percpu_ref use a bit more memory which should be ok for our use ca=
se,
> since we almost have only one zswap_pool to be using. The performance gai=
n is
> for zswap_store/load hotpath.
>=20
> >=20
> > It would also be useful to mention how many threads/CPUs are being used=
 here.
>=20
> My bad, the testing uses 32 threads on a 128 CPUs x86-64 machine.

Thanks for the clarification. Please include such details in the commit
message.