Re: sched: Improve load balancing in the presence of idle CPUs

[Preeti U Murthy]

On 04/02/2015 11:29 AM, Jason Low wrote:
> On Wed, 2015-04-01 at 18:04 +0100, Morten Rasmussen wrote:
>> On Wed, Apr 01, 2015 at 07:49:56AM +0100, Preeti U Murthy wrote:
> 
>>> I am sorry I don't quite get this. Can you please elaborate?
>>
>> I think the scenario is that we are in nohz_idle_balance() and decide to
>> bail out because we have pulled some tasks, but before leaving
>> nohz_idle_balance() we want to check if more balancing is necessary
>> using nohz_kick_needed() and potentially kick somebody to continue.
> 
> Also, below is an example patch.
> 
> (Without the conversion to idle_cpu(), the check for rq->idle_balance
> would not be accurate anymore)
> 
> ---
>  kernel/sched/fair.c |   17 ++++++++++-------
>  1 files changed, 10 insertions(+), 7 deletions(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index fdae26e..7749a14 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -7620,6 +7620,8 @@ out:
>  }
> 
>  #ifdef CONFIG_NO_HZ_COMMON
> +static inline bool nohz_kick_needed(struct rq *rq);
> +
>  /*
>   * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
>   * rebalancing for all the cpus for whom scheduler ticks are stopped.
> @@ -7629,6 +7631,7 @@ static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
>  	int this_cpu = this_rq->cpu;
>  	struct rq *rq;
>  	int balance_cpu;
> +	bool done_balancing = false;
> 
>  	if (idle != CPU_IDLE ||
>  	    !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
> @@ -7644,7 +7647,7 @@ static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
>  		 * balancing owner will pick it up.
>  		 */
>  		if (need_resched())
> -			break;
> +			goto end;
> 
>  		rq = cpu_rq(balance_cpu);
> 
> @@ -7663,9 +7666,12 @@ static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
>  		if (time_after(this_rq->next_balance, rq->next_balance))
>  			this_rq->next_balance = rq->next_balance;
>  	}
> +	done_balancing = true;
>  	nohz.next_balance = this_rq->next_balance;
>  end:
>  	clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
> +	if (!done_balancing && nohz_kick_needed(this_rq))
> +		nohz_balancer_kick();
>  }
> 
>  /*
> @@ -7687,7 +7693,7 @@ static inline bool nohz_kick_needed(struct rq *rq)
>  	int nr_busy, cpu = rq->cpu;
>  	bool kick = false;
> 
> -	if (unlikely(rq->idle_balance))
> +	if (unlikely(idle_cpu(cpu)))
>  		return false;
> 
>         /*
> @@ -7757,16 +7763,13 @@ static void run_rebalance_domains(struct softirq_action *h)
>  	enum cpu_idle_type idle = this_rq->idle_balance ?
>  						CPU_IDLE : CPU_NOT_IDLE;
> 
> +	rebalance_domains(this_rq, idle);
>  	/*
>  	 * If this cpu has a pending nohz_balance_kick, then do the
>  	 * balancing on behalf of the other idle cpus whose ticks are
> -	 * stopped. Do nohz_idle_balance *before* rebalance_domains to
> -	 * give the idle cpus a chance to load balance. Else we may
> -	 * load balance only within the local sched_domain hierarchy
> -	 * and abort nohz_idle_balance altogether if we pull some load.
> +	 * stopped.
>  	 */
>  	nohz_idle_balance(this_rq, idle);
> -	rebalance_domains(this_rq, idle);
>  }
> 
>  /*
> 

Solution 1: As exists in the mainline
Solution 2: nohz_idle_balance(); rebalance_domains() on the ILB CPU
Solution 3: Above patch.

I observe that Solution 3 is not as aggressive in spreading load as
Solution 2. With Solution 2, the load gets spread within the first 3-4
seconds, while with Solution3, the load gets spread within the first 6-7
seconds. I think this is because, the above patch decides to further
nohz_idle_load_balance() based on the load on the current ILB CPU which
has most likely pulled just one task. This will abort further load
balancing. However, Solution 3 is certainly better at spreading load
than Solution 1.

Wrt IPIs, I see that Solution 2 results in increase in the number of
IPIs by around 2% over Solution 3, probably for the same reason that
Morten pointed out.

Regards
Preeti U Murthy

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/