Message-ID: <ZCa8KB7Os9yJWz3P@pc636>
Date:   Fri, 31 Mar 2023 12:55:36 +0200
From:   Uladzislau Rezki <urezki@...il.com>
To:     "Paul E. McKenney" <paulmck@...nel.org>
Cc:     Uladzislau Rezki <urezki@...il.com>,
        Joel Fernandes <joel@...lfernandes.org>,
        "Zhang, Qiang1" <qiang1.zhang@...el.com>,
        "Zhuo, Qiuxu" <qiuxu.zhuo@...el.com>, RCU <rcu@...r.kernel.org>,
        quic_neeraju@...cinc.com, Boqun Feng <boqun.feng@...il.com>,
        LKML <linux-kernel@...r.kernel.org>,
        Oleksiy Avramchenko <oleksiy.avramchenko@...y.com>,
        Steven Rostedt <rostedt@...dmis.org>,
        Frederic Weisbecker <frederic@...nel.org>
Subject: Re: [PATCH 1/1] Reduce synchronize_rcu() waiting time

On Thu, Mar 30, 2023 at 02:16:36PM -0700, Paul E. McKenney wrote:
> On Thu, Mar 30, 2023 at 09:18:44PM +0200, Uladzislau Rezki wrote:
> > On Thu, Mar 30, 2023 at 11:58:41AM -0700, Paul E. McKenney wrote:
> > > On Thu, Mar 30, 2023 at 05:43:15PM +0200, Uladzislau Rezki wrote:
> > > > On Thu, Mar 30, 2023 at 03:09:33PM +0000, Joel Fernandes wrote:
> > > > > On Tue, Mar 28, 2023 at 08:26:13AM -0700, Paul E. McKenney wrote:
> > > > > > On Mon, Mar 27, 2023 at 10:29:31PM -0400, Joel Fernandes wrote:
> > > > > > > Hello,
> > > > > > > 
> > > > > > > > On Mar 27, 2023, at 9:06 PM, Paul E. McKenney <paulmck@...nel.org> wrote:
> > > > > > > > 
> > > > > > > > On Mon, Mar 27, 2023 at 11:21:23AM +0000, Zhang, Qiang1 wrote:
> > > > > > > >>>> From: Uladzislau Rezki (Sony) <urezki@...il.com>
> > > > > > > >>>> Sent: Tuesday, March 21, 2023 6:28 PM
> > > > > > > >>>> [...]
> > > > > > > >>>> Subject: [PATCH 1/1] Reduce synchronize_rcu() waiting time
> > > > > > > >>>> 
> > > > > > > >>>> A call to a synchronize_rcu() can be expensive from time point of view.
> > > > > > > >>>> Different workloads can be affected by this especially the ones which use this
> > > > > > > >>>> API in its time critical sections.
> > > > > > > >>>> 
> > > > > > > >>> 
> > > > > > > >>> This is interesting and meaningful research. ;-)
> > > > > > > >>> 
> > > > > > > >>>> For example in case of NOCB scenario the wakeme_after_rcu() callback
> > > > > > > >>>> invocation depends on where in a nocb-list it is located. Below is an example
> > > > > > > >>>> when it was the last out of ~3600 callbacks:
> > > > > > > >>> 
> > > > > > > >> 
> > > > > > > >> 
> > > > > > > >> 
> > > > > > > >> Can it be implemented separately as follows?  it seems that the code is simpler
> > > > > > > >> (only personal opinion)  😊.
> > > > > > > >> 
> > > > > > > >> But I didn't test whether this reduce synchronize_rcu() waiting time
> > > > > > > >> 
> > > > > > > >> +static void rcu_poll_wait_gp(struct rcu_tasks *rtp)
> > > > > > > >> +{
> > > > > > > >> +       unsigned long gp_snap;
> > > > > > > >> +
> > > > > > > >> +       gp_snap = start_poll_synchronize_rcu();
> > > > > > > >> +       while (!poll_state_synchronize_rcu(gp_snap))
> > > > > > > >> +               schedule_timeout_idle(1);
> > > > > > > > 
> > > > > > > > I could be wrong, but my guess is that the guys working with
> > > > > > > > battery-powered devices are not going to be very happy with this loop.
> > > > > > > > 
> > > > > > > > All those wakeups by all tasks waiting for a grace period end up
> > > > > > > > consuming a surprisingly large amount of energy.
> > > > > > > 
> > > > > > > Is that really the common case? On the general topic of wake-ups:
> > > > > > > Most of the time there should be only one
> > > > > > > task waiting synchronously on a GP to end. If that is
> > > > > > > true, then it feels like waking
> > > > > > > up nocb Kthreads which indirectly wake other threads is doing more work than usual?
> > > > > > 
> > > > > > A good question, and the number of outstanding synchronize_rcu()
> > > > > > calls will of course be limited by the number of tasks in the system.
> > > > > > But I myself have raised the ire of battery-powered embedded folks with
> > > > > > a rather small number of wakeups, so...
> > > > > 
> > > > > But unless I am missing something, even if there is single synchronize_rcu(),
> > > > > you have a flurry of potential wakeups right now, instead of the bare minimum
> > > > > I think. I have not measured how many wake ups, but I'd love to when I get
> > > > > time. Maybe Vlad has some numbers.
> > > > > 
> > > > I will measure and have a look at wake-ups. But, what we have for now is
> > > > if there are two callers of synchronize_rcu() on different CPUs, i guess
> > > > two nocb-kthreads have to handle it, thus two nocb-kthreads have to be
> > > > awaken to do the work. This patch needs only one wake-up to serve all
> > > > users.
> > > 
> > > One wakeup per synchronize_rcu(), right?
> > > 
> > The gp-kthread wake-ups only one work, in its turn a worker wake-ups all
> > registered users of synchronize_rcu() for which a gp was passed. How many
> > users of synchonize_rcu() awaken by one worker depends on how many were
> > registered before initiating a new GP by the gp-kthread.
> > 
> > > > Anyway, i will provide some data and analysis of it.
> > > 
> > > Looking forward to seeing it!
> > > 
> > Good. I will switch fully on it soon. I need to sort out some perf.
> > issues at work.
> 
> And if you are looking for reduced wakeups instead of lower latency for
> synchronize_rcu(), I could see where the extra workqueue wakeup might
> be a problem for you.
> 
> Assuming that this is all default-off, you could keep a count of the
> number of required wakeups for each grace period (indexed as usual by
> the bottom few bits of the grace-period counter without the low-order
> state bits), and do the wakeups directly from the grace-period kthread
> if there are not all that many of them.
> 
At least if there is only one user of synchronize_rcu(), we can wake
it directly, i mean to invoke comlete() from the gp-kthread. I think
we should split such parts into different patches.

>
> Except that, given that workqueues try hard to make the handler be on the
> same CPU as the one that did the corresponding schedule_work() invocation,
> it is not clear that this particular wakeup is really costing you enough
> to notice.  (That CPU is not idle, after all.)  But there is nothing
> quite like measuring the actual energy consumption on real hardware!
> 
AFAICR the schedule_work() wants to use current CPU, indeed. For few
users it might be OK to comlete() them directly. Energy wise, +1 wake-up
of our worker to kick all users, if not "direct" option. I do not think
energy is a problem here.

--
Uladzislau Rezki

Powered by blists - more mailing lists