cpufreq: governor: Be friendly towards latency-sensitive bursty workloads

Cpufreq governors like the ondemand governor calculate the load on the CPU
periodically by employing deferrable timers. A deferrable timer won't fire
if the CPU is completely idle (and there are no other timers to be run), in
order to avoid unnecessary wakeups and thus save CPU power.

However, the load calculation logic is agnostic to all this, and this can
lead to the problem described below.

Time (ms)               CPU 1

100                Task-A running

110                Governor's timer fires, finds load as 100% in the last
                   10ms interval and increases the CPU frequency.

110.5              Task-A running

120		   Governor's timer fires, finds load as 100% in the last
		   10ms interval and increases the CPU frequency.

125		   Task-A went to sleep. With nothing else to do, CPU 1
		   went completely idle.

200		   Task-A woke up and started running again.

200.5		   Governor's deferred timer (which was originally programmed
		   to fire at time 130) fires now. It calculates load for the
		   time period 120 to 200.5, and finds the load is almost zero.
		   Hence it decreases the CPU frequency to the minimum.

210		   Governor's timer fires, finds load as 100% in the last
		   10ms interval and increases the CPU frequency.

So, after the workload woke up and started running, the frequency was suddenly
dropped to absolute minimum, and after that, there was an unnecessary delay of
10ms (sampling period) to increase the CPU frequency back to a reasonable value.
And this pattern repeats for every wake-up-from-cpu-idle for that workload.
This can be quite undesirable for latency- or response-time sensitive bursty
workloads. So we need to fix the governor's logic to detect such wake-up-from-
cpu-idle scenarios and start the workload at a reasonably high CPU frequency.

One extreme solution would be to fake a load of 100% in such scenarios. But
that might lead to undesirable side-effects such as frequency spikes (which
might also need voltage changes) especially if the previous frequency happened
to be very low.

We just want to avoid the stupidity of dropping down the frequency to a minimum
and then enduring a needless (and long) delay before ramping it up back again.
So, let us simply carry forward the previous load - that is, let us just pretend
that the 'load' for the current time-window is the same as the load for the
previous window. That way, the frequency and voltage will continue to be set
to whatever values they were set at previously. This means that bursty workloads
will get a chance to influence the CPU frequency at which they wake up from
cpu-idle, based on their past execution history. Thus, they might be able to
avoid suffering from slow wakeups and long response-times.

However, we should take care not to over-do this. For example, such a "copy
previous load" logic will benefit cases like this: (where # represents busy
and . represents idle)

##########.........#########.........###########...........##########........

but it will be detrimental in cases like the one shown below, because it will
retain the high frequency (copied from the previous interval) even in a mostly
idle system:

##########.........#.................#.....................#...............

(i.e., the workload finished and the remaining tasks are such that their busy
periods are smaller than the sampling interval, which causes the timer to
always get deferred. So, this will make the copy-previous-load logic copy
the initial high load to subsequent idle periods over and over again, thus
keeping the frequency high unnecessarily).

So, we modify this copy-previous-load logic such that it is used only once
upon every wakeup-from-idle. Thus if we have 2 consecutive idle periods, the
previous load won't get blindly copied over; cpufreq will freshly evaluate the
load in the second idle interval, thus ensuring that the system comes back to
its normal state.

[ The right way to solve this whole problem is to teach the CPU frequency
governors to also track load on a per-task basis, not just a per-CPU basis,
and then use both the data sources intelligently to set the appropriate
frequency on the CPUs. But that involves redesigning the cpufreq subsystem,
so this patch should make the situation bearable until then. ]

Experimental results:
+-------------------+

I ran a modified version of ebizzy (called 'sleeping-ebizzy') that sleeps in
between its execution such that its total utilization can be a user-defined
value, say 10% or 20% (higher the utilization specified, lesser the amount of
sleeps injected). This ebizzy was run with a single-thread, tied to CPU 8.

Behavior observed with tracing (sample taken from 40% utilization runs):
------------------------------------------------------------------------

Without patch:
~~~~~~~~~~~~~~
kworker/8:2-12137  416.335742: cpu_frequency: state=2061000 cpu_id=8
kworker/8:2-12137  416.335744: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy
      <...>-40753  416.345741: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2
kworker/8:2-12137  416.345744: cpu_frequency: state=4123000 cpu_id=8
kworker/8:2-12137  416.345746: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy
      <...>-40753  416.355738: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2
<snip>  ---------------------------------------------------------------------  <snip>
      <...>-40753  416.402202: sched_switch: prev_comm=ebizzy ==> next_comm=swapper/8
     <idle>-0      416.502130: sched_switch: prev_comm=swapper/8 ==> next_comm=ebizzy
      <...>-40753  416.505738: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2
kworker/8:2-12137  416.505739: cpu_frequency: state=2061000 cpu_id=8
kworker/8:2-12137  416.505741: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy
      <...>-40753  416.515739: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2
kworker/8:2-12137  416.515742: cpu_frequency: state=4123000 cpu_id=8
kworker/8:2-12137  416.515744: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy

Observation: Ebizzy went idle at 416.402202, and started running again at
416.502130. But cpufreq noticed the long idle period, and dropped the frequency
at 416.505739, only to increase it back again at 416.515742, realizing that the
workload is in-fact CPU bound. Thus ebizzy needlessly ran at the lowest frequency
for almost 13 milliseconds (almost 1 full sample period), and this pattern
repeats on every sleep-wakeup. This could hurt latency-sensitive workloads quite
a lot.

With patch:
~~~~~~~~~~~

kworker/8:2-29802  464.832535: cpu_frequency: state=2061000 cpu_id=8
<snip>  ---------------------------------------------------------------------  <snip>
kworker/8:2-29802  464.962538: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy
      <...>-40738  464.972533: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2
kworker/8:2-29802  464.972536: cpu_frequency: state=4123000 cpu_id=8
kworker/8:2-29802  464.972538: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy
      <...>-40738  464.982531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2
<snip>  ---------------------------------------------------------------------  <snip>
kworker/8:2-29802  465.022533: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy
      <...>-40738  465.032531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2
kworker/8:2-29802  465.032532: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy
      <...>-40738  465.035797: sched_switch: prev_comm=ebizzy ==> next_comm=swapper/8
     <idle>-0      465.240178: sched_switch: prev_comm=swapper/8 ==> next_comm=ebizzy
      <...>-40738  465.242533: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2
kworker/8:2-29802  465.242535: sched_switch: prev_comm=kworker/8:2 ==> next_comm=ebizzy
      <...>-40738  465.252531: sched_switch: prev_comm=ebizzy ==> next_comm=kworker/8:2

Observation: Ebizzy went idle at 465.035797, and started running again at
465.240178. Since ebizzy was the only real workload running on this CPU,
cpufreq retained the frequency at 4.1Ghz throughout the run of ebizzy, no
matter how many times ebizzy slept and woke-up in-between. Thus, ebizzy
got the 10ms worth of 4.1 Ghz benefit during every sleep-wakeup (as compared
to the run without the patch) and this boost gave a modest improvement in total
throughput, as shown below.

Sleeping-ebizzy records-per-second:
-----------------------------------

Utilization  Without patch  With patch  Difference (Absolute and % values)
    10%         274767        277046        +  2279 (+0.829%)
    20%         543429        553484        + 10055 (+1.850%)
    40%        1090744       1107959        + 17215 (+1.578%)
    60%        1634908       1662018        + 27110 (+1.658%)

A rudimentary and somewhat approximately latency-sensitive workload such as
sleeping-ebizzy itself showed a consistent, noticeable performance improvement
with this patch. Hence, workloads that are truly latency-sensitive will benefit
quite a bit from this change. Moreover, this is an overall win-win since this
patch does not hurt power-savings at all (because, this patch does not reduce
the idle time or idle residency; and the high frequency of the CPU when it goes
to cpu-idle does not affect/hurt the power-savings of deep idle states).

Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

1 files changed, 55 insertions, 3 deletions

diff --git a/drivers/cpufreq/cpufreq_governor.c b/drivers/cpufreq/cpufreq_governor.c
index e1c6433b16e0..9004450863be 100644
--- a/drivers/cpufreq/cpufreq_governor.c
+++ b/drivers/cpufreq/cpufreq_governor.c
@@ -36,14 +36,29 @@ void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
 	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 	struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
 	struct cpufreq_policy *policy;
+	unsigned int sampling_rate;
 	unsigned int max_load = 0;
 	unsigned int ignore_nice;
 	unsigned int j;
 
-	if (dbs_data->cdata->governor == GOV_ONDEMAND)
+	if (dbs_data->cdata->governor == GOV_ONDEMAND) {
+		struct od_cpu_dbs_info_s *od_dbs_info =
+				dbs_data->cdata->get_cpu_dbs_info_s(cpu);
+
+		/*
+		 * Sometimes, the ondemand governor uses an additional
+		 * multiplier to give long delays. So apply this multiplier to
+		 * the 'sampling_rate', so as to keep the wake-up-from-idle
+		 * detection logic a bit conservative.
+		 */
+		sampling_rate = od_tuners->sampling_rate;
+		sampling_rate *= od_dbs_info->rate_mult;
+
 		ignore_nice = od_tuners->ignore_nice_load;
-	else
+	} else {
+		sampling_rate = cs_tuners->sampling_rate;
 		ignore_nice = cs_tuners->ignore_nice_load;
+	}
 
 	policy = cdbs->cur_policy;
 
@@ -96,7 +111,36 @@ void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
 		if (unlikely(!wall_time || wall_time < idle_time))
 			continue;
 
-		load = 100 * (wall_time - idle_time) / wall_time;
+		/*
+		 * If the CPU had gone completely idle, and a task just woke up
+		 * on this CPU now, it would be unfair to calculate 'load' the
+		 * usual way for this elapsed time-window, because it will show
+		 * near-zero load, irrespective of how CPU intensive that task
+		 * actually is. This is undesirable for latency-sensitive bursty
+		 * workloads.
+		 *
+		 * To avoid this, we reuse the 'load' from the previous
+		 * time-window and give this task a chance to start with a
+		 * reasonably high CPU frequency. (However, we shouldn't over-do
+		 * this copy, lest we get stuck at a high load (high frequency)
+		 * for too long, even when the current system load has actually
+		 * dropped down. So we perform the copy only once, upon the
+		 * first wake-up from idle.)
+		 *
+		 * Detecting this situation is easy: the governor's deferrable
+		 * timer would not have fired during CPU-idle periods. Hence
+		 * an unusually large 'wall_time' (as compared to the sampling
+		 * rate) indicates this scenario.
+		 */
+		if (unlikely(wall_time > (2 * sampling_rate)) &&
+						j_cdbs->copy_prev_load) {
+			load = j_cdbs->prev_load;
+			j_cdbs->copy_prev_load = false;
+		} else {
+			load = 100 * (wall_time - idle_time) / wall_time;
+			j_cdbs->prev_load = load;
+			j_cdbs->copy_prev_load = true;
+		}
 
 		if (load > max_load)
 			max_load = load;
@@ -318,11 +362,19 @@ int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 		for_each_cpu(j, policy->cpus) {
 			struct cpu_dbs_common_info *j_cdbs =
 				dbs_data->cdata->get_cpu_cdbs(j);
+			unsigned int prev_load;
 
 			j_cdbs->cpu = j;
 			j_cdbs->cur_policy = policy;
 			j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
 					       &j_cdbs->prev_cpu_wall, io_busy);
+
+			prev_load = (unsigned int)
+				(j_cdbs->prev_cpu_wall - j_cdbs->prev_cpu_idle);
+			j_cdbs->prev_load = 100 * prev_load /
+					(unsigned int) j_cdbs->prev_cpu_wall;
+			j_cdbs->copy_prev_load = true;
+
 			if (ignore_nice)
 				j_cdbs->prev_cpu_nice =
 					kcpustat_cpu(j).cpustat[CPUTIME_NICE];