Re: [PATCH v3] perf sched timehist: Add pre-migration wait time option

[Madadi Vineeth Reddy]

Hi Tim Chen,

On 04/10/24 03:34, Tim Chen wrote:
> On Wed, 2024-10-02 at 22:09 +0530, Madadi Vineeth Reddy wrote:
>> pre-migration wait time is the time that a task unnecessarily spends
>> on the runqueue of a CPU but doesn't get switched-in there. In terms
>> of tracepoints, it is the time between sched:sched_wakeup and
>> sched:sched_migrate_task.
>>
>> Let's say a task woke up on CPU2, then it got migrated to CPU4 and
>> then it's switched-in to CPU4. So, here pre-migration wait time is
>> time that it was waiting on runqueue of CPU2 after it is woken up.
>>
>> The general pattern for pre-migration to occur is:
>> sched:sched_wakeup
>> sched:sched_migrate_task
>> sched:sched_switch
> 
> If a task migrate from CPU A to CPU B, but is unlucky that
> someone took CPU B and it has to wait. Then it is yet again migrated to CPU C.
> Do you only compute pre-mig time as 
> t_sched_migrate_task_CPU_A - t_sched_migrate_task_CPU_A?
> 
> The task also spend some pre-mig time on CPU_B that probably
> should be included.  And that time is when it migrates to B
> till it is migrated away from B.  Do you take that into
> account?
> 

You are right. Currently the time spent on CPU B is not accounted. I will
consider the latest migrate time and that should fix it.

>>
>> The sched:sched_waking event is used to capture the wakeup time,
>> as it aligns with the existing code and only introduces a negligible
>> time difference.
>>
>> pre-migrations are generally not useful and it increases migrations.
> 
> Not sure long pre-migration itself leads to migrations.  But rather migrations
> is a function of the aggressiveness of the idle load balancer.

You're right. pre-migration wait time itself doesn't directly cause migrations,
but it is a useful metric for understanding the overall impact on task scheduling,
especially in the context of the idle load balancer as you mentioned. I agree
that this is most relevant for scenarios where newly idle CPUs pull tasks from
other CPUs.

By capturing the pre-migration wait time, the intention is to evaluate how
efficiently tasks are balanced and whether there are opportunities for improvement
in the wakeup and load-balancing paths.

Thank you for the feedback.

Thanks,
Madadi Vineeth Reddy

> 
>> This metric would be helpful in testing patches mainly related to wakeup
>> and load-balancer code paths 
>>
> 
> I think it is most relevant for the idle load balancer code paths, where
> newly idle CPU pull task from other CPUs to run.
> 
> Are there cases where you have used this metric to tune the scheduler?
> 
> Thanks.
> 
> Tim 
> 
>> as better wakeup logic would choose an
>> optimal CPU where task would be switched-in and thereby reducing pre-
>> migrations.
>>
>> The sample output(s) when -P or --pre-migrations is used:
>> =================
>>            time    cpu  task name                       wait time  sch delay   run time  pre-mig time
>>                         [tid/pid]                          (msec)     (msec)     (msec)     (msec)
>> --------------- ------  ------------------------------  ---------  ---------  ---------  ---------
>>    38456.720806 [0001]  schbench[28634/28574]               4.917      4.768      1.004      0.000
>>    38456.720810 [0001]  rcu_preempt[18]                     3.919      0.003      0.004      0.000
>>    38456.721800 [0006]  schbench[28779/28574]              23.465     23.465      1.999      0.000
>>    38456.722800 [0002]  schbench[28773/28574]              60.371     60.237      3.955     60.197
>>    38456.722806 [0001]  schbench[28634/28574]               0.004      0.004      1.996      0.000
>>    38456.722811 [0001]  rcu_preempt[18]                     1.996      0.005      0.005      0.000
>>    38456.723800 [0000]  schbench[28833/28574]               4.000      4.000      3.999      0.000
>>    38456.723800 [0004]  schbench[28762/28574]              42.951     42.839      3.999     39.867
>>    38456.723802 [0007]  schbench[28812/28574]              43.947     43.817      3.999     40.866
>>    38456.723804 [0001]  schbench[28587/28574]               7.935      7.822      0.993      0.000
>>
>> Signed-off-by: Madadi Vineeth Reddy <vineethr@xxxxxxxxxxxxx>
>>
>> ---
>> Changes in v3:
>> - Use the sched:sched_waking event to calculate the wakeup time. (Namhyung Kim)
>> - Rebase against perf-tools-next commit 80f192724e31 ("perf tests: Add more
>>   topdown events regroup tests")
>>
>> Changes in v2:
>> - Use timehist_sched_wakeup_event() to get the sched_wakeup time. (Namhyung Kim)
>> - Rebase against perf-tools-next commit b38c49d8296b ("perf/test: Speed up test
>>   case perf annotate basic tests")
>>
>> Signed-off-by: Madadi Vineeth Reddy <vineethr@xxxxxxxxxxxxx>
>> ---
>>  tools/perf/Documentation/perf-sched.txt |  8 +++
>>  tools/perf/builtin-sched.c              | 83 ++++++++++++++++---------
>>  2 files changed, 60 insertions(+), 31 deletions(-)
>>
>> diff --git a/tools/perf/Documentation/perf-sched.txt b/tools/perf/Documentation/perf-sched.txt
>> index 3db64954a267..6dbbddb6464d 100644
>> --- a/tools/perf/Documentation/perf-sched.txt
>> +++ b/tools/perf/Documentation/perf-sched.txt
>> @@ -221,6 +221,14 @@ OPTIONS for 'perf sched timehist'
>>  	priorities are specified with -: 120-129. A combination of both can also be
>>  	provided: 0,120-129.
>>  
>> +-P::
>> +--pre-migrations::
>> +	Show pre-migration wait time. pre-migration wait time is the time spent
>> +	by a task waiting on a runqueue but not getting the chance to run there
>> +	and is migrated to a different runqueue where it is finally run. This
>> +	time between sched_wakeup and migrate_task is the pre-migration wait
>> +	time.
>> +
>>  OPTIONS for 'perf sched replay'
>>  ------------------------------
>>  
>> diff --git a/tools/perf/builtin-sched.c b/tools/perf/builtin-sched.c
>> index fdf979aaf275..e6540803ced6 100644
>> --- a/tools/perf/builtin-sched.c
>> +++ b/tools/perf/builtin-sched.c
>> @@ -225,6 +225,7 @@ struct perf_sched {
>>  	bool		show_wakeups;
>>  	bool		show_next;
>>  	bool		show_migrations;
>> +	bool		pre_migrations;
>>  	bool		show_state;
>>  	bool		show_prio;
>>  	u64		skipped_samples;
>> @@ -244,7 +245,9 @@ struct thread_runtime {
>>  	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
>>  	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
>>  	u64 dt_delay;       /* time between wakeup and sched-in */
>> +	u64 dt_pre_mig;     /* time between migration and wakeup */
>>  	u64 ready_to_run;   /* time of wakeup */
>> +	u64 migrated;	    /* time when a thread is migrated */
>>  
>>  	struct stats run_stats;
>>  	u64 total_run_time;
>> @@ -252,6 +255,7 @@ struct thread_runtime {
>>  	u64 total_iowait_time;
>>  	u64 total_preempt_time;
>>  	u64 total_delay_time;
>> +	u64 total_pre_mig_time;
>>  
>>  	char last_state;
>>  
>> @@ -2073,14 +2077,15 @@ static void timehist_header(struct perf_sched *sched)
>>  		printf(" ");
>>  	}
>>  
>> -	if (sched->show_prio) {
>> -		printf(" %-*s  %-*s  %9s  %9s  %9s",
>> -		       comm_width, "task name", MAX_PRIO_STR_LEN, "prio",
>> -		       "wait time", "sch delay", "run time");
>> -	} else {
>> -		printf(" %-*s  %9s  %9s  %9s", comm_width,
>> -		       "task name", "wait time", "sch delay", "run time");
>> -	}
>> +	printf(" %-*s", comm_width, "task name");
>> +
>> +	if (sched->show_prio)
>> +		printf("  %-*s", MAX_PRIO_STR_LEN, "prio");
>> +
>> +	printf("  %9s  %9s  %9s", "wait time", "sch delay", "run time");
>> +
>> +	if (sched->pre_migrations)
>> +		printf("  %9s", "pre-mig time");
>>  
>>  	if (sched->show_state)
>>  		printf("  %s", "state");
>> @@ -2095,17 +2100,15 @@ static void timehist_header(struct perf_sched *sched)
>>  	if (sched->show_cpu_visual)
>>  		printf(" %*s ", ncpus, "");
>>  
>> -	if (sched->show_prio) {
>> -		printf(" %-*s  %-*s  %9s  %9s  %9s",
>> -		       comm_width, "[tid/pid]", MAX_PRIO_STR_LEN, "",
>> -		       "(msec)", "(msec)", "(msec)");
>> -	} else {
>> -		printf(" %-*s  %9s  %9s  %9s", comm_width,
>> -		       "[tid/pid]", "(msec)", "(msec)", "(msec)");
>> -	}
>> +	printf(" %-*s", comm_width, "[tid/pid]");
>>  
>> -	if (sched->show_state)
>> -		printf("  %5s", "");
>> +	if (sched->show_prio)
>> +		printf("  %-*s", MAX_PRIO_STR_LEN, "");
>> +
>> +	printf("  %9s  %9s  %9s", "(msec)", "(msec)", "(msec)");
>> +
>> +	if (sched->pre_migrations)
>> +		printf("  %9s", "(msec)");
>>  
>>  	printf("\n");
>>  
>> @@ -2117,15 +2120,15 @@ static void timehist_header(struct perf_sched *sched)
>>  	if (sched->show_cpu_visual)
>>  		printf(" %.*s ", ncpus, graph_dotted_line);
>>  
>> -	if (sched->show_prio) {
>> -		printf(" %.*s  %.*s  %.9s  %.9s  %.9s",
>> -		       comm_width, graph_dotted_line, MAX_PRIO_STR_LEN, graph_dotted_line,
>> -		       graph_dotted_line, graph_dotted_line, graph_dotted_line);
>> -	} else {
>> -		printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
>> -		       graph_dotted_line, graph_dotted_line, graph_dotted_line,
>> -		       graph_dotted_line);
>> -	}
>> +	printf(" %.*s", comm_width, graph_dotted_line);
>> +
>> +	if (sched->show_prio)
>> +		printf("  %.*s", MAX_PRIO_STR_LEN, graph_dotted_line);
>> +
>> +	printf("  %.9s  %.9s  %.9s", graph_dotted_line, graph_dotted_line, graph_dotted_line);
>> +
>> +	if (sched->pre_migrations)
>> +		printf("  %.9s", graph_dotted_line);
>>  
>>  	if (sched->show_state)
>>  		printf("  %.5s", graph_dotted_line);
>> @@ -2180,6 +2183,8 @@ static void timehist_print_sample(struct perf_sched *sched,
>>  
>>  	print_sched_time(tr->dt_delay, 6);
>>  	print_sched_time(tr->dt_run, 6);
>> +	if (sched->pre_migrations)
>> +		print_sched_time(tr->dt_pre_mig, 6);
>>  
>>  	if (sched->show_state)
>>  		printf(" %5c ", thread__tid(thread) == 0 ? 'I' : state);
>> @@ -2239,6 +2244,7 @@ static void timehist_update_runtime_stats(struct thread_runtime *r,
>>  	r->dt_iowait  = 0;
>>  	r->dt_preempt = 0;
>>  	r->dt_run     = 0;
>> +	r->dt_pre_mig = 0;
>>  
>>  	if (tprev) {
>>  		r->dt_run = t - tprev;
>> @@ -2247,6 +2253,11 @@ static void timehist_update_runtime_stats(struct thread_runtime *r,
>>  				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
>>  			else
>>  				r->dt_delay = tprev - r->ready_to_run;
>> +
>> +			if (r->ready_to_run && r->migrated) {
>> +				if ((r->migrated > r->ready_to_run) && (r->migrated < tprev))
>> +					r->dt_pre_mig = r->migrated - r->ready_to_run;
>> +			}
>>  		}
>>  
>>  		if (r->last_time > tprev)
>> @@ -2270,6 +2281,7 @@ static void timehist_update_runtime_stats(struct thread_runtime *r,
>>  	r->total_sleep_time   += r->dt_sleep;
>>  	r->total_iowait_time  += r->dt_iowait;
>>  	r->total_preempt_time += r->dt_preempt;
>> +	r->total_pre_mig_time += r->dt_pre_mig;
>>  }
>>  
>>  static bool is_idle_sample(struct perf_sample *sample,
>> @@ -2684,8 +2696,14 @@ static int timehist_migrate_task_event(const struct perf_tool *tool,
>>  
>>  	tr->migrations++;
>>  
>> +	if (tr->migrated == 0)
>> +		tr->migrated = sample->time;
>> +
>>  	/* show migrations if requested */
>> -	timehist_print_migration_event(sched, evsel, sample, machine, thread);
>> +	if (sched->show_migrations) {
>> +		timehist_print_migration_event(sched, evsel, sample,
>> +							machine, thread);
>> +	}
>>  
>>  	return 0;
>>  }
>> @@ -2836,11 +2854,13 @@ static int timehist_sched_change_event(const struct perf_tool *tool,
>>  		/* last state is used to determine where to account wait time */
>>  		tr->last_state = state;
>>  
>> -		/* sched out event for task so reset ready to run time */
>> +		/* sched out event for task so reset ready to run time and migrated time */
>>  		if (state == 'R')
>>  			tr->ready_to_run = t;
>>  		else
>>  			tr->ready_to_run = 0;
>> +
>> +		tr->migrated = 0;
>>  	}
>>  
>>  	evsel__save_time(evsel, sample->time, sample->cpu);
>> @@ -3280,8 +3300,8 @@ static int perf_sched__timehist(struct perf_sched *sched)
>>  		goto out;
>>  	}
>>  
>> -	if (sched->show_migrations &&
>> -	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
>> +	if ((sched->show_migrations || sched->pre_migrations) &&
>> +		perf_session__set_tracepoints_handlers(session, migrate_handlers))
>>  		goto out;
>>  
>>  	/* pre-allocate struct for per-CPU idle stats */
>> @@ -3823,6 +3843,7 @@ int cmd_sched(int argc, const char **argv)
>>  	OPT_BOOLEAN(0, "show-prio", &sched.show_prio, "Show task priority"),
>>  	OPT_STRING(0, "prio", &sched.prio_str, "prio",
>>  		   "analyze events only for given task priority(ies)"),
>> +	OPT_BOOLEAN('P', "pre-migrations", &sched.pre_migrations, "Show pre-migration wait time"),
>>  	OPT_PARENT(sched_options)
>>  	};
>>  
>