Re: [PATCH RFC] sched/fair: Avoid unnecessary IPIs for ILB

[Ingo Molnar]

* Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx> wrote:

> From: Vineeth Pillai <vineethrp@xxxxxxxxxx>
> 
> Whenever a CPU stops its tick, it now requires another idle CPU to handle the
> balancing for it because it can't perform its own periodic load balancing.
> This means it might need to update 'nohz.next_balance' to 'rq->next_balance' if
> the upcoming nohz-idle load balancing is too distant in the future. This update
> process is done by triggering an ILB, as the general ILB handler
> (_nohz_idle_balance) that manages regular nohz balancing also refreshes
> 'nohz.next_balance' by looking at the 'rq->next_balance' of all other idle CPUs
> and selecting the smallest value.
> 
> Triggering this ILB can be achieved by setting the NOHZ_NEXT_KICK flag. This
> primarily results in the ILB handler updating 'nohz.next_balance' while
> possibly not doing any load balancing at all. However, sending an IPI merely to
> refresh 'nohz.next_balance' seems excessive, and there ought to be a more
> efficient method to update 'nohz.next_balance' from the local CPU.
> 
> Fortunately, there already exists a mechanism to directly invoke the ILB
> handler (_nohz_idle_balance) without initiating an IPI. It's accomplished by
> setting the NOHZ_NEWILB_KICK flag. This flag is set during regular "newly idle"
> balancing and solely exists to update a CPU's blocked load if it couldn't pull
> more tasks during regular "newly idle balancing" - and it does so without
> having to send any IPIs. Once the flag is set, the ILB handler is called
> directly from do_idle()-> nohz_run_idle_balance(). While its goal is to update
> the blocked load without an IPI, in our situation, we aim to refresh
> 'nohz.next_balance' without an IPI but we can piggy back on this.
> 
> So in this patch, we reuse this mechanism by also setting the NOHZ_NEXT_KICK to
> indicate nohz.next_balance needs an update via this direct call shortcut. Note
> that we set this flag without knowledge that the tick is about to be stopped,
> because at the point we do it, we have no way of knowing that. However we do
> know that the CPU is about to enter idle. In our testing, the reduction in IPIs
> is well worth updating nohz.next_balance a few more times.
> 
> Also just to note, without this patch we observe the following pattern:
> 
> 1. A CPU is about to stop its tick.
> 2. It sets nohz.needs_update to 1.
> 3. It then stops its tick and goes idle.
> 4. The scheduler tick on another CPU checks this flag and decides an ILB kick is needed.
> 5. The ILB CPU ends up being the one that just stopped its tick!
> 6. This results in an IPI to the tick-stopped CPU which ends up waking it up
>    and disturbing it!
> 
> Testing shows a considerable reduction in IPIs when doing this:
> 
> Running "cyclictest -i 100 -d 100 --latency=1000 -t -m" on a 4vcpu VM
> the IPI call count profiled over 10s period is as follows:
> without fix: ~10500
> with fix: ~1000
> 
> Fixes: 7fd7a9e0caba ("sched/fair: Trigger nohz.next_balance updates when a CPU goes NOHZ-idle")
> 
> [ Joel: wrote commit messages, collaborated on fix, helped reproduce issue etc. ]
> 
> Cc: Suleiman Souhlal <suleiman@xxxxxxxxxx>
> Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
> Cc: Hsin Yi <hsinyi@xxxxxxxxxx>
> Cc: Frederic Weisbecker <frederic@xxxxxxxxxx>
> Cc: Paul E. McKenney <paulmck@xxxxxxxxxx>
> Signed-off-by: Vineeth Pillai <vineethrp@xxxxxxxxxx>
> Co-developed-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> ---
>  kernel/sched/fair.c | 21 ++++++++++++++-------
>  1 file changed, 14 insertions(+), 7 deletions(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index cb225921bbca..2ece55f32782 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -11786,13 +11786,12 @@ void nohz_balance_enter_idle(int cpu)
>  	/*
>  	 * Ensures that if nohz_idle_balance() fails to observe our
>  	 * @idle_cpus_mask store, it must observe the @has_blocked
> -	 * and @needs_update stores.
> +	 * stores.
>  	 */
>  	smp_mb__after_atomic();
>  
>  	set_cpu_sd_state_idle(cpu);
>  
> -	WRITE_ONCE(nohz.needs_update, 1);
>  out:
>  	/*
>  	 * Each time a cpu enter idle, we assume that it has blocked load and
> @@ -11945,21 +11944,25 @@ static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
>  }
>  
>  /*
> - * Check if we need to run the ILB for updating blocked load before entering
> - * idle state.
> + * Check if we need to run the ILB for updating blocked load and/or updating
> + * nohz.next_balance before entering idle state.
>   */
>  void nohz_run_idle_balance(int cpu)
>  {
>  	unsigned int flags;
>  
> -	flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
> +	flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK | NOHZ_NEXT_KICK, nohz_flags(cpu));
> +
> +	if (!flags)
> +		return;
>  
>  	/*
>  	 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
>  	 * (ie NOHZ_STATS_KICK set) and will do the same.
>  	 */
> -	if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
> -		_nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
> +	if ((flags == (flags & (NOHZ_NEXT_KICK | NOHZ_NEWILB_KICK))) &&
> +	    !need_resched())
> +		_nohz_idle_balance(cpu_rq(cpu), flags);
>  }
>  
>  static void nohz_newidle_balance(struct rq *this_rq)
> @@ -11977,6 +11980,10 @@ static void nohz_newidle_balance(struct rq *this_rq)
>  	if (this_rq->avg_idle < sysctl_sched_migration_cost)
>  		return;
>  
> +	/* If rq->next_balance before nohz.next_balance, trigger ILB */
> +	if (time_before(this_rq->next_balance, READ_ONCE(nohz.next_balance)))
> +		atomic_or(NOHZ_NEXT_KICK, nohz_flags(this_cpu));
> +
>  	/* Don't need to update blocked load of idle CPUs*/
>  	if (!READ_ONCE(nohz.has_blocked) ||
>  	    time_before(jiffies, READ_ONCE(nohz.next_blocked)))

Ok, judging by your IPI reduction numbers this is definitely an 
optimization we want to do.

The patch does make _nohz_idle_balance() run more parallel, as previously 
it would be generally run by the first-idle CPU in nohz.idle_cpus_mask (at 
least for next_balance updates), but I think it's still SMP-safe, as all 
key data structure updates are already rq-locked AFAICS.

One thing I noticed is that we now use nohz.needs_update only in a single 
remaining case, when _nohz_idle_balance() "self-defers":

                /*
                 * If this CPU gets work to do, stop the load balancing
                 * work being done for other CPUs. Next load
                 * balancing owner will pick it up.
                 */
                if (need_resched()) {
                        if (flags & NOHZ_STATS_KICK)
                                has_blocked_load = true;
                        if (flags & NOHZ_NEXT_KICK)
                                WRITE_ONCE(nohz.needs_update, 1);
                        goto abort;
                }

Getting a need-resched flag set on this CPU is a pretty dubious reason to 
skip an ILB run IMO, and we could do entirely without that complication, 
allowing us to remove the nohz.needs_update flag handling logic altogether?

If we do that then the !need_resched() flag needs to go from 
nohz_run_idle_balance() too:

        /*
         * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
         * (ie NOHZ_STATS_KICK set) and will do the same.
         */
        if ((flags == (flags & (NOHZ_NEXT_KICK | NOHZ_NEWILB_KICK))) &&
            !need_resched())
                _nohz_idle_balance(cpu_rq(cpu), flags);

... if I'm reading the code right that is.

Thanks,

	Ingo