MIME-Version: 1.0
References: <20230202012348.885402-1-nhuck@google.com> <alpine.LRH.2.21.2302021052580.21238@file01.intranet.prod.int.rdu2.redhat.com>
In-Reply-To: <alpine.LRH.2.21.2302021052580.21238@file01.intranet.prod.int.rdu2.redhat.com>
From:   Nathan Huckleberry <nhuck@google.com>
Date:   Tue, 7 Feb 2023 14:54:21 -0800
Message-ID: <CAJkfWY5d+FqGb_3PMZbMWYdNW4ED4CqiuJzMQFog_GLLEtZiBg@mail.gmail.com>
Subject: Re: [dm-devel] [PATCH] dm-verity: Remove WQ_UNBOUND.
To:     Mikulas Patocka <mpatocka@redhat.com>
Cc:     Mike Snitzer <snitzer@kernel.org>, linux-kernel@vger.kernel.org,
        Eric Biggers <ebiggers@kernel.org>, dm-devel@redhat.com,
        Sami Tolvanen <samitolvanen@google.com>,
        Alasdair Kergon <agk@redhat.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Feb 7, 2023 at 12:16 AM Mikulas Patocka <mpatocka@redhat.com> wrote:
>
>
>
> On Wed, 1 Feb 2023, Nathan Huckleberry wrote:
>
> > Setting WQ_UNBOUND increases scheduler latency on ARM64.  This is likely
> > due to the asymmetric architecture of ARM64 processors.
> >
> > I've been unable to reproduce the results that claim WQ_UNBOUND gives a
> > performance boost on x86-64.
> >
> > This flag is causing performance issues for multiple subsystems within
> > Android.  Notably, the same slowdown exists for decompression with
> > EROFS.
> >
> > | open-prebuilt-camera  | WQ_UNBOUND | ~WQ_UNBOUND   |
> > |-----------------------|------------|---------------|
> > | verity wait time (us) | 11746      | 119 (-98%)    |
> > | erofs wait time (us)  | 357805     | 174205 (-51%) |
> >
> > | sha256 ramdisk random read | WQ_UNBOUND    | ~WQ_UNBOUND |
> > |----------------------------|-----------=---|-------------|
> > | arm64 (accelerated)        | bw=42.4MiB/s  | bw=212MiB/s |
> > | arm64 (generic)            | bw=16.5MiB/s  | bw=48MiB/s  |
> > | x86_64 (generic)           | bw=233MiB/s   | bw=230MiB/s |
> >
> > Cc: Sami Tolvanen <samitolvanen@google.com>
> > Cc: Eric Biggers <ebiggers@kernel.org>
> > Signed-off-by: Nathan Huckleberry <nhuck@google.com>
> > ---
> >  drivers/md/dm-verity-target.c | 4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/md/dm-verity-target.c b/drivers/md/dm-verity-target.c
> > index ccf5b852fbf7..020fd2341025 100644
> > --- a/drivers/md/dm-verity-target.c
> > +++ b/drivers/md/dm-verity-target.c
> > @@ -1399,8 +1399,8 @@ static int verity_ctr(struct dm_target *ti, unsigned argc, char **argv)
> >               goto bad;
> >       }
> >
> > -     /* WQ_UNBOUND greatly improves performance when running on ramdisk */
> > -     wq_flags = WQ_MEM_RECLAIM | WQ_UNBOUND;
> > +     wq_flags = WQ_MEM_RECLAIM;
> > +
> >       /*
> >        * Using WQ_HIGHPRI improves throughput and completion latency by
> >        * reducing wait times when reading from a dm-verity device.
>
> Hi
>
> If you remove WQ_UNBOUND, you should also change the last argument of
> alloc_workqueue from num_online_cpus() to either 0 or 1. Try both 0 and 1
> and tell us which performs better.

They look roughly the same on ARM64. There's a slight advantage for
using 1, but they're too close to conclusively say one is better.

Thanks,
Huck

>
> Mikulas
>