Received-SPF: pass (google.com: domain of linux-kernel+bounces-205208-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.48.161 as permitted sender) client-ip=147.75.48.161;
Precedence: bulk
MIME-Version: 1.0
References: <20240606165303.431215-1-flintglass@gmail.com> <CAJD7tkZ=o3AN+4Cj5UBJv6zcrjPFW5T1_53iHB2qtShRRhKhbQ@mail.gmail.com>
In-Reply-To: <CAJD7tkZ=o3AN+4Cj5UBJv6zcrjPFW5T1_53iHB2qtShRRhKhbQ@mail.gmail.com>
From: Takero Funaki <flintglass@gmail.com>
Date: Fri, 7 Jun 2024 10:00:43 +0900
Message-ID: <CAPpoddc5ge0myLfN8burL9fEUjc0oaB0C8Yc3_J923hd_O9u4A@mail.gmail.com>
Subject: Re: [PATCH] mm: zswap: limit number of zpools based on CPU and RAM
To: Yosry Ahmed <yosryahmed@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>, Nhat Pham <nphamcs@gmail.com>, 
	Chengming Zhou <chengming.zhou@linux.dev>, Andrew Morton <akpm@linux-foundation.org>, linux-mm@kvack.org, 
	linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

2024=E5=B9=B46=E6=9C=887=E6=97=A5(=E9=87=91) 2:46 Yosry Ahmed <yosryahmed@g=
oogle.com>:

>
> There are a lot of magic numbers in this patch, and it seems like it's
> all based on theory. I don't object to making the number of zpools
> dynamic in some way, but unless we do it in a data-driven way where we
> understand the implications, I think the added complexity and
> inconsistency is not justified.
>
> For example, 2*CPU zpools is an overkill and will cause a lot of
> fragmentation. We use 32 zpools right now for machines with 100s of
> CPUs. I know that you are keeping 32 as the limit, but why 2*CPUs if
> nr_cpus <=3D 16?
>
> Also, the limitation based on memory size assumes that zsmalloc is the
> only allocator used by zswap, which is unfortunately not the case.
>
> The current implementation using 32 zpools all the time is not
> perfect, and I did write a patch to make it at least be min(nr_cpus,
> 32), but it is simple and it works. Complexity should be justified.
>

Thanks for your comments.
I agree the 2*cpu is too much. it was conservatively chosen assuming
1/2 contention while all cores are accessing zswap. Much smaller
factor or non-linear scale as your comments in the main thread would
be better.

I found your patch from the main thread.
One point I'm afraid, this hashing will fail if nr_zswap_zpools is 1
or is not rounded to order of 2. hash_ptr crashes when bit is 0.

> +   return entry->pool->zpools[hash_ptr(entry, ilog2(nr_zswap_zpools))];