Linux-PM Archive on lore.kernel.org
 help / color / Atom feed
WARNING: multiple messages refer to this Message-ID
From: Patrick Bellasi <patrick.bellasi@arm.com>
To: douglas.raillard@arm.com
Cc: linux-kernel@vger.kernel.org, linux-pm@vger.kernel.org,
	mingo@redhat.com, peterz@infradead.org, quentin.perret@arm.com,
	dietmar.eggemann@arm.com
Subject: Re: [RFC PATCH 6/7] sched/cpufreq: Improve sugov_cpu_is_busy accuracy
Date: Thu, 16 May 2019 13:55:52 +0100
Message-ID: <20190516125552.hol3rasllhveekxq@e110439-lin> (raw)
In-Reply-To: <20190508174301.4828-7-douglas.raillard@arm.com>

On 08-May 18:43, douglas.raillard@arm.com wrote:
> From: Douglas RAILLARD <douglas.raillard@arm.com>
> 
> Avoid assuming a CPU is busy when it has begun being idle before
> get_next_freq() is called. This is achieved by making sure the CPU will
> not be detected as busy by other CPUs whenever its utilization is
> decreasing.

If I understand it correctly, what you are after here is a "metric"
which tells you (in a shared performance domain) if a CPU has been
busy for a certain amount of time.
You do that by reworking the way idle_calls are accounted for the
sugov_update_single() case.

That approach could work but it looks a bit convoluted in the way it's
coded and it's also difficult to exclude there could be corner cases
with wired behaviors.
Isn't that why you "fix" the saved_idle_calls counter after all?

What about a different approach where we:

1. we annotate the timestamp a CPU wakes up from IDLE (last_wakeup_time)

2. we use that annotated last_wake_time and the rq->nr_running to
   define the "cpu is busy" heuristic.

Looking at a sibling CPU, I think we can end up with two main cases:

1. CPU's nr_running is == 0
   then we don't consider busy that CPU

2. CPU's nr_running is  > 0
   then the CPU is busy iff
      (current_time - last_wakeup_tim) >= busy_threshold

Notice that, when a CPU is active, its rq clock is periodically
updated, at least once per tick. Thus, provided a tick time is not too
long to declare busy a CPU, then the above logic should work.

Perhaps the busy_threshold can also be defined considering the PELT
dynamics and starting from an expected utilization increase converted
in time.

Could something like to above work or am I missing something?

> Signed-off-by: Douglas RAILLARD <douglas.raillard@arm.com>
> ---
>  kernel/sched/cpufreq_schedutil.c | 42 ++++++++++++++++++++++++++++----
>  1 file changed, 37 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> index a12b7e5bc028..ce4b90cafbb5 100644
> --- a/kernel/sched/cpufreq_schedutil.c
> +++ b/kernel/sched/cpufreq_schedutil.c
> @@ -62,6 +62,7 @@ struct sugov_cpu {
>  	/* The field below is for single-CPU policies only: */
>  #ifdef CONFIG_NO_HZ_COMMON
>  	unsigned long		saved_idle_calls;
> +	unsigned long		previous_util;
>  #endif
>  };
>  
> @@ -181,14 +182,35 @@ static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
>  	return ret;
>  }
>  
> -static void sugov_cpu_is_busy_update(struct sugov_cpu *sg_cpu)
> +static void sugov_cpu_is_busy_update(struct sugov_cpu *sg_cpu,
> +				     unsigned long util)
>  {
>  	unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
>  	sg_cpu->saved_idle_calls = idle_calls;
> +
> +	/*
> +	 * Make sure that this CPU will not be immediately considered as busy in
> +	 * cases where the CPU has already entered an idle state. In that case,
> +	 * the number of idle_calls will not vary anymore until it exits idle,
> +	 * which would lead sugov_cpu_is_busy() to say that this CPU is busy,
> +	 * because it has not (re)entered idle since the last time we looked at
> +	 * it.
> +	 * Assuming cpu0 and cpu1 are in the same policy, that will make sure
> +	 * this sequence of events leads to right cpu1 business status from
> +	 * get_next_freq(cpu=1)
> +	 * cpu0: [enter idle] -> [get_next_freq] -> [doing nothing] -> [wakeup]
> +	 * cpu1:                ...              -> [get_next_freq] ->   ...
> +	 */
> +	if (util <= sg_cpu->previous_util)
> +		sg_cpu->saved_idle_calls--;
> +
> +	sg_cpu->previous_util = util;
>  }
>  #else
>  static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
> -static void sugov_cpu_is_busy_update(struct sugov_cpu *sg_cpu) {}
> +static void sugov_cpu_is_busy_update(struct sugov_cpu *sg_cpu
> +				     unsigned long util)
> +{}
>  #endif /* CONFIG_NO_HZ_COMMON */
>  
>  /**
> @@ -507,10 +529,9 @@ static void sugov_update_single(struct update_util_data *hook, u64 time,
>  	if (!sugov_should_update_freq(sg_policy, time))
>  		return;
>  
> -	busy = sugov_cpu_is_busy(sg_cpu);
> -	sugov_cpu_is_busy_update(sg_cpu);
> -
>  	util = sugov_get_util(sg_cpu);
> +	busy = sugov_cpu_is_busy(sg_cpu);
> +	sugov_cpu_is_busy_update(sg_cpu, util);
>  	max = sg_cpu->max;
>  	util = sugov_iowait_apply(sg_cpu, time, util, max);
>  	next_f = get_next_freq(sg_policy, util, max);
> @@ -545,12 +566,15 @@ static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
>  	struct cpufreq_policy *policy = sg_policy->policy;
>  	unsigned long util = 0, max = 1;
>  	unsigned int j;
> +	unsigned long sg_cpu_util = 0;
>  
>  	for_each_cpu(j, policy->cpus) {
>  		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
>  		unsigned long j_util, j_max;
>  
>  		j_util = sugov_get_util(j_sg_cpu);
> +		if (j_sg_cpu == sg_cpu)
> +			sg_cpu_util = j_util;
>  		j_max = j_sg_cpu->max;
>  		j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
>  
> @@ -560,6 +584,14 @@ static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
>  		}
>  	}
>  
> +	/*
> +	 * Only update the business status if we are looking at the CPU for
> +	 * which a utilization change triggered a call to get_next_freq(). This
> +	 * way, we don't affect the "busy" status of CPUs that don't have any
> +	 * change in utilization.
> +	 */
> +	sugov_cpu_is_busy_update(sg_cpu, sg_cpu_util);
> +
>  	return get_next_freq(sg_policy, util, max);
>  }
>  
> -- 
> 2.21.0
> 

-- 
#include <best/regards.h>

Patrick Bellasi

From: Patrick Bellasi <patrick.bellasi@arm.com>
To: douglas.raillard@arm.com
Cc: linux-kernel@vger.kernel.org, linux-pm@vger.kernel.org,
	mingo@redhat.com, peterz@infradead.org, quentin.perret@arm.com,
	dietmar.eggemann@arm.com
Subject: Re: [RFC PATCH 6/7] sched/cpufreq: Improve sugov_cpu_is_busy accuracy
Date: Thu, 16 May 2019 13:55:52 +0100
Message-ID: <20190516125552.hol3rasllhveekxq@e110439-lin> (raw)
Message-ID: <20190516125552.K9R9UqhITWjeDAQRfTMh6Ls9hyXts1kWBjf7B3Or1q4@z> (raw)
In-Reply-To: <20190508174301.4828-7-douglas.raillard@arm.com>

On 08-May 18:43, douglas.raillard@arm.com wrote:
> From: Douglas RAILLARD <douglas.raillard@arm.com>
> 
> Avoid assuming a CPU is busy when it has begun being idle before
> get_next_freq() is called. This is achieved by making sure the CPU will
> not be detected as busy by other CPUs whenever its utilization is
> decreasing.

If I understand it correctly, what you are after here is a "metric"
which tells you (in a shared performance domain) if a CPU has been
busy for a certain amount of time.
You do that by reworking the way idle_calls are accounted for the
sugov_update_single() case.

That approach could work but it looks a bit convoluted in the way it's
coded and it's also difficult to exclude there could be corner cases
with wired behaviors.
Isn't that why you "fix" the saved_idle_calls counter after all?

What about a different approach where we:

1. we annotate the timestamp a CPU wakes up from IDLE (last_wakeup_time)

2. we use that annotated last_wake_time and the rq->nr_running to
   define the "cpu is busy" heuristic.

Looking at a sibling CPU, I think we can end up with two main cases:

1. CPU's nr_running is == 0
   then we don't consider busy that CPU

2. CPU's nr_running is  > 0
   then the CPU is busy iff
      (current_time - last_wakeup_tim) >= busy_threshold

Notice that, when a CPU is active, its rq clock is periodically
updated, at least once per tick. Thus, provided a tick time is not too
long to declare busy a CPU, then the above logic should work.

Perhaps the busy_threshold can also be defined considering the PELT
dynamics and starting from an expected utilization increase converted
in time.

Could something like to above work or am I missing something?

> Signed-off-by: Douglas RAILLARD <douglas.raillard@arm.com>
> ---
>  kernel/sched/cpufreq_schedutil.c | 42 ++++++++++++++++++++++++++++----
>  1 file changed, 37 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> index a12b7e5bc028..ce4b90cafbb5 100644
> --- a/kernel/sched/cpufreq_schedutil.c
> +++ b/kernel/sched/cpufreq_schedutil.c
> @@ -62,6 +62,7 @@ struct sugov_cpu {
>  	/* The field below is for single-CPU policies only: */
>  #ifdef CONFIG_NO_HZ_COMMON
>  	unsigned long		saved_idle_calls;
> +	unsigned long		previous_util;
>  #endif
>  };
>  
> @@ -181,14 +182,35 @@ static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
>  	return ret;
>  }
>  
> -static void sugov_cpu_is_busy_update(struct sugov_cpu *sg_cpu)
> +static void sugov_cpu_is_busy_update(struct sugov_cpu *sg_cpu,
> +				     unsigned long util)
>  {
>  	unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
>  	sg_cpu->saved_idle_calls = idle_calls;
> +
> +	/*
> +	 * Make sure that this CPU will not be immediately considered as busy in
> +	 * cases where the CPU has already entered an idle state. In that case,
> +	 * the number of idle_calls will not vary anymore until it exits idle,
> +	 * which would lead sugov_cpu_is_busy() to say that this CPU is busy,
> +	 * because it has not (re)entered idle since the last time we looked at
> +	 * it.
> +	 * Assuming cpu0 and cpu1 are in the same policy, that will make sure
> +	 * this sequence of events leads to right cpu1 business status from
> +	 * get_next_freq(cpu=1)
> +	 * cpu0: [enter idle] -> [get_next_freq] -> [doing nothing] -> [wakeup]
> +	 * cpu1:                ...              -> [get_next_freq] ->   ...
> +	 */
> +	if (util <= sg_cpu->previous_util)
> +		sg_cpu->saved_idle_calls--;
> +
> +	sg_cpu->previous_util = util;
>  }
>  #else
>  static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
> -static void sugov_cpu_is_busy_update(struct sugov_cpu *sg_cpu) {}
> +static void sugov_cpu_is_busy_update(struct sugov_cpu *sg_cpu
> +				     unsigned long util)
> +{}
>  #endif /* CONFIG_NO_HZ_COMMON */
>  
>  /**
> @@ -507,10 +529,9 @@ static void sugov_update_single(struct update_util_data *hook, u64 time,
>  	if (!sugov_should_update_freq(sg_policy, time))
>  		return;
>  
> -	busy = sugov_cpu_is_busy(sg_cpu);
> -	sugov_cpu_is_busy_update(sg_cpu);
> -
>  	util = sugov_get_util(sg_cpu);
> +	busy = sugov_cpu_is_busy(sg_cpu);
> +	sugov_cpu_is_busy_update(sg_cpu, util);
>  	max = sg_cpu->max;
>  	util = sugov_iowait_apply(sg_cpu, time, util, max);
>  	next_f = get_next_freq(sg_policy, util, max);
> @@ -545,12 +566,15 @@ static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
>  	struct cpufreq_policy *policy = sg_policy->policy;
>  	unsigned long util = 0, max = 1;
>  	unsigned int j;
> +	unsigned long sg_cpu_util = 0;
>  
>  	for_each_cpu(j, policy->cpus) {
>  		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
>  		unsigned long j_util, j_max;
>  
>  		j_util = sugov_get_util(j_sg_cpu);
> +		if (j_sg_cpu == sg_cpu)
> +			sg_cpu_util = j_util;
>  		j_max = j_sg_cpu->max;
>  		j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
>  
> @@ -560,6 +584,14 @@ static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
>  		}
>  	}
>  
> +	/*
> +	 * Only update the business status if we are looking at the CPU for
> +	 * which a utilization change triggered a call to get_next_freq(). This
> +	 * way, we don't affect the "busy" status of CPUs that don't have any
> +	 * change in utilization.
> +	 */
> +	sugov_cpu_is_busy_update(sg_cpu, sg_cpu_util);
> +
>  	return get_next_freq(sg_policy, util, max);
>  }
>  
> -- 
> 2.21.0
> 

-- 
#include <best/regards.h>

Patrick Bellasi

  parent reply index

Thread overview: 36+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-05-08 17:42 [RFC PATCH 0/7] sched/cpufreq: Make schedutil energy aware douglas.raillard
2019-05-08 17:42 ` douglas.raillard
2019-05-08 17:42 ` [RFC PATCH 1/7] PM: Introduce em_pd_get_higher_freq() douglas.raillard
2019-05-08 17:42   ` douglas.raillard
2019-05-16 12:42   ` Patrick Bellasi
2019-05-16 12:42     ` Patrick Bellasi
2019-05-16 13:01     ` Quentin Perret
2019-05-16 13:01       ` Quentin Perret
2019-05-16 13:22       ` Patrick Bellasi
2019-05-16 13:22         ` Patrick Bellasi
2019-06-19 16:08         ` Douglas Raillard
2019-06-20 13:04           ` Patrick Bellasi
2019-06-21 10:17             ` Quentin Perret
2019-06-21 10:22               ` Quentin Perret
2019-05-16 13:06     ` Douglas Raillard
2019-05-16 13:06       ` Douglas Raillard
2019-05-08 17:42 ` [RFC PATCH 2/7] sched/cpufreq: Attach perf domain to sugov policy douglas.raillard
2019-05-08 17:42   ` douglas.raillard
2019-05-08 17:42 ` [RFC PATCH 3/7] sched/cpufreq: Hook em_pd_get_higher_power() into get_next_freq() douglas.raillard
2019-05-08 17:42   ` douglas.raillard
2019-05-08 17:42 ` [RFC PATCH 4/7] sched/cpufreq: Move up sugov_cpu_is_busy() douglas.raillard
2019-05-08 17:42   ` douglas.raillard
2019-05-08 17:42 ` [RFC PATCH 5/7] sched/cpufreq: sugov_cpu_is_busy for shared policy douglas.raillard
2019-05-08 17:42   ` douglas.raillard
2019-05-08 17:43 ` [RFC PATCH 6/7] sched/cpufreq: Improve sugov_cpu_is_busy accuracy douglas.raillard
2019-05-08 17:43   ` douglas.raillard
2019-05-16 12:55   ` Patrick Bellasi [this message]
2019-05-16 12:55     ` Patrick Bellasi
2019-06-19 16:19     ` Douglas Raillard
2019-06-20 11:05       ` Patrick Bellasi
2019-05-08 17:43 ` [RFC PATCH 7/7] sched/cpufreq: Boost schedutil frequency ramp up douglas.raillard
2019-05-08 17:43   ` douglas.raillard
2019-05-13  7:12 ` [RFC PATCH 0/7] sched/cpufreq: Make schedutil energy aware Viresh Kumar
2019-05-13  7:12   ` Viresh Kumar
2019-05-13 13:52   ` Douglas Raillard
2019-05-13 13:52     ` Douglas Raillard

Reply instructions:

You may reply publically to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20190516125552.hol3rasllhveekxq@e110439-lin \
    --to=patrick.bellasi@arm.com \
    --cc=dietmar.eggemann@arm.com \
    --cc=douglas.raillard@arm.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-pm@vger.kernel.org \
    --cc=mingo@redhat.com \
    --cc=peterz@infradead.org \
    --cc=quentin.perret@arm.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Linux-PM Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/linux-pm/0 linux-pm/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 linux-pm linux-pm/ https://lore.kernel.org/linux-pm \
		linux-pm@vger.kernel.org linux-pm@archiver.kernel.org
	public-inbox-index linux-pm

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.linux-pm


AGPL code for this site: git clone https://public-inbox.org/ public-inbox