Received-SPF: pass (google.com: domain of linux-kernel+bounces-30344-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Date: Thu, 18 Jan 2024 11:16:01 -0500
From: Johannes Weiner <hannes@cmpxchg.org>
To: Yosry Ahmed <yosryahmed@google.com>
Cc: Nhat Pham <nphamcs@gmail.com>,
	Ronald Monthero <debug.penguin32@gmail.com>, sjenning@redhat.com,
	ddstreet@ieee.org, vitaly.wool@konsulko.com,
	akpm@linux-foundation.org, chrisl@kernel.org, linux-mm@kvack.org,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH] mm/zswap: Improve with alloc_workqueue() call
Message-ID: <20240118161601.GJ939255@cmpxchg.org>
References: <CAKEwX=NLe-N6dLvOVErPSL3Vfw6wqHgcUBQoNRLeWkN6chdvLQ@mail.gmail.com>
 <20240116133145.12454-1-debug.penguin32@gmail.com>
 <CAKEwX=PjraCg_NjP4Tnkbv8uqnVw8yJGh-mbuZC02Gp6HMcDBw@mail.gmail.com>
 <CAJD7tkb_uC_K7+C3GjVqg1rDRCmUkbHcEw950CkUHG66yokbcg@mail.gmail.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <CAJD7tkb_uC_K7+C3GjVqg1rDRCmUkbHcEw950CkUHG66yokbcg@mail.gmail.com>

On Wed, Jan 17, 2024 at 11:30:50AM -0800, Yosry Ahmed wrote:
> On Wed, Jan 17, 2024 at 11:14 AM Nhat Pham <nphamcs@gmail.com> wrote:
> >
> > On Tue, Jan 16, 2024 at 5:32 AM Ronald Monthero
> > <debug.penguin32@gmail.com> wrote:
> >
> > + Johannes and Yosry
> >
> > >
> > > The core-api create_workqueue is deprecated, this patch replaces
> > > the create_workqueue with alloc_workqueue. The previous
> > > implementation workqueue of zswap was a bounded workqueue, this
> > > patch uses alloc_workqueue() to create an unbounded workqueue.
> > > The WQ_UNBOUND attribute is desirable making the workqueue
> > > not localized to a specific cpu so that the scheduler is free
> > > to exercise improvisations in any demanding scenarios for
> > > offloading cpu time slices for workqueues.
> >
> > nit: extra space between paragraph would be nice.
> >
> > > For example if any other workqueues of the same primary cpu
> > > had to be served which are WQ_HIGHPRI and WQ_CPU_INTENSIVE.
> > > Also Unbound workqueue happens to be more efficient
> > > in a system during memory pressure scenarios in comparison
> > >  to a bounded workqueue.
> > >
> > > shrink_wq = alloc_workqueue("zswap-shrink",
> > >                      WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
> > >
> > > Overall the change suggested in this patch should be
> > > seamless and does not alter the existing behavior,
> > > other than the improvisation to be an unbounded workqueue.
> > >
> > > Signed-off-by: Ronald Monthero <debug.penguin32@gmail.com>
> > > ---
> > >  mm/zswap.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > index 74411dfdad92..64dbe3e944a2 100644
> > > --- a/mm/zswap.c
> > > +++ b/mm/zswap.c
> > > @@ -1620,7 +1620,8 @@ static int zswap_setup(void)
> > >                 zswap_enabled = false;
> > >         }
> > >
> > > -       shrink_wq = create_workqueue("zswap-shrink");
> > > +       shrink_wq = alloc_workqueue("zswap-shrink",
> > > +                       WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
> >
> > Have you benchmarked this to check if there is any regression, just to
> > be safe? With an unbounded workqueue, you're gaining scheduling
> > flexibility at the cost of cache locality. My intuition is that it
> > doesn't matter too much here, but you should probably double check by
> > stress testing - run some workload with a relatively small zswap pool
> > limit (i.e heavy global writeback), and see if there is any difference
> > in performance.
> 
> I also think this shouldn't make a large difference. The global
> shrinking work is already expensive, and I imagine that it exhausts
> the caches anyway by iterating memcgs. A performance smoketest would
> be reassuring for sure, but I believe it won't make a difference.

The LRU inherently makes the shrinker work on the oldest and coldest
entries, so I doubt we benefit a lot from cache locality there.

What could make a difference though is the increased concurrency by
switching max_active from 1 to 0. This could cause a higher rate of
shrinker runs, which might increase lock contention and reclaim
volume. That part would be good to double check with the shrinker
benchmarks.

> > On a different note, I wonder if it would help to perform synchronous
> > reclaim here instead. With our current design, the zswap store failure
> > (due to global limit hit) would leave the incoming page going to swap
> > instead, creating an LRU inversion. Not sure if that's ideal.
> 
> The global shrink path keeps reclaiming until zswap can accept again
> (by default, that means reclaiming 10% of the total limit). I think
> this is too expensive to be done synchronously.

That thresholding code is a bit weird right now.

It wakes the shrinker and rejects at the same time. We're guaranteed
to see rejections, even if the shrinker has no trouble flushing some
entries a split second later.

It would make more sense to wake the shrinker at e.g. 95% full and
have it run until 90%.

But with that in place we also *should* do synchronous reclaim once we
hit 100%. Just enough to make room for the store. This is important to
catch the case where reclaim rate exceeds swapout rate. Rejecting and
going to swap means the reclaimer will be throttled down to IO rate
anyway, and the app latency isn't any worse. But this way we keep the
pipeline alive, and keep swapping out the oldest zswap entries,
instead of rejecting and swapping what would be the hottest ones.