Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:5 as permitted sender) client-ip=2620:137:e000::3:5;
MIME-Version: 1.0
References: <20231115154946.3933808-1-dhowells@redhat.com> <20231115154946.3933808-9-dhowells@redhat.com>
In-Reply-To: <20231115154946.3933808-9-dhowells@redhat.com>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Wed, 15 Nov 2023 11:28:22 -0500
Message-ID: <CAHk-=wjytv+Gy-Ra0rhLCAW_120BvnzLC63tfkkZVXzGgD3_+w@mail.gmail.com>
Subject: Re: [PATCH v3 08/10] iov_iter: Add benchmarking kunit tests
To:     David Howells <dhowells@redhat.com>
Cc:     Christian Brauner <christian@brauner.io>,
        Jens Axboe <axboe@kernel.dk>,
        Al Viro <viro@zeniv.linux.org.uk>,
        Christoph Hellwig <hch@lst.de>,
        David Laight <David.Laight@aculab.com>,
        Matthew Wilcox <willy@infradead.org>,
        Brendan Higgins <brendanhiggins@google.com>,
        David Gow <davidgow@google.com>, linux-fsdevel@vger.kernel.org,
        linux-block@vger.kernel.org, linux-mm@kvack.org,
        netdev@vger.kernel.org, linux-kselftest@vger.kernel.org,
        kunit-dev@googlegroups.com, linux-kernel@vger.kernel.org,
        Christian Brauner <brauner@kernel.org>,
        David Hildenbrand <david@redhat.com>,
        John Hubbard <jhubbard@nvidia.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, 15 Nov 2023 at 10:50, David Howells <dhowells@redhat.com> wrote:
>
> Add kunit tests to benchmark 256MiB copies to a KVEC iterator, a BVEC
> iterator, an XARRAY iterator and to a loop that allocates 256-page BVECs
> and fills them in (similar to a maximal bio struct being set up).

I see *zero* advantage of doing this in the kernel as opposed to doing
this benchmarking in user space.

If you cannot see the performance difference due to some user space
interface costs, then the performance difference doesn't matter.

Yes, some of the cases may be harder to trigger than others.
iov_iter_xarray() isn't as common an op as ubuf/iovec/etc, but that
either means that it doesn't matter enough, or that maybe some more
filesystems could be taught to use it for splice or whatever.

Particularly for something like different versions of memcpy(), this
whole benchmarking would want

 (a) profiles

 (b) be run on many different machines

 (c) be run repeatedly to get some idea of variance

and all of those only get *harder* to do with Kunit tests. In user
space? Just run the damn binary (ok, to get profiles you then have to
make sure you have the proper permission setup to get the kernel
profiles too, but a

   echo 1 > /proc/sys/kernel/perf_event_paranoid

as root will do that for you without you having to then do the actual
profiling run as root)

                Linus