update_load_avg() is the heaviest caller of the cpufreq utilization callbacks right now and it runs too frequently from various parts of the fair scheduler.
It may not be optimal to update the CPU frequency that frequently though.
We are more concerned about the events when the CPU utilization changes sharply and that mostly happens when a task gets attached/detached to/from a runqueue.
Perhaps we aren't that concerned about the case when the utilization changes gradually (small delta) due the time a task gets to run and updating the CPU frequency at the tick rate should be good enough in that case.
Sometimes update_load_avg() gets called right before migrating a task to a new CPU and that can steal the chance of updating the frequency when we actually needed it (i.e. after util_avg is updated with the contributions of the migrated task). This can happen because the cpufreq governors change the frequency only once in a fixed period of time (i.e. rate_limit_us for the schedutil governor).
The fair scheduler already calls the utilization hooks from attach/detach paths and that takes care of the most important case we are concerned about.
This patch relocates the call to utilization hook from update_cfs_rq_load_avg() to task_tick_fair().
Testing:
Platform: Skylake, Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz. Quad core (8 threads).
Distribution: Ubutnu 16.04 with cmdline intel_pstate=passive. Based over 4.12-rc3.
All the below tests were run 3 times with and without this change over 4.12-rc3 and the average of all the runs is shown here (RESULT). The number of times the schedutil hook is called during the tests is also present in the table (UPDATE-COUNT).
Bench:
- pipe-test
Command: perf bench sched pipe
4.12-rc3 4.12-rc3 + this patch Change -------- --------------------- ----------
RESULT 4.21 3.68 - 12.53 %
UPDATE-COUNT 5405 886 - 83 %
- Hackbench
Command: sync; sleep 1; for i in {1..20}; do sleep 2; \ hackbench -P -g 1 | grep Time; done
4.12-rc3 4.12-rc3 + this patch Change -------- --------------------- ----------
RESULT 0.00925 0.00853 - 7.75 %
UPDATE-COUNT 10342 1925 - 81 %
- ebizzy
Command: ebizzy -t N -S 20
4.12-rc3 4.12-rc3 + this patch Change -------- --------------------- ----------
RESULT (N=8) 270221 247841 - 8 % RESULT (N=16) 252692 249839 - 1.1 % RESULT (N=24) 251151 254114 + 1.1 % RESULT (N=32) 249274 250115 + 0.3 %
UPDATE-COUNT 41964 40279 - 4 %
- cyclictest (with latency)
Command: cyclictest -q -t N -i 10000 --latency=1000000 -D 20
4.12-rc3 4.12-rc3 + this patch Change -------- --------------------- ----------
RESULT (N=8) 326 332 + 1.75 % RESULT (N=16) 297 328 + 10 % RESULT (N=24) 288 317 + 9.9 % RESULT (N=32) 268 313 + 17 %
UPDATE-COUNT 28949 313 - 98.6 %
Some more investigation (with ftrace) was done to check the cause of regression in this case (as the below test doesn't have the same results without latency).
o The number of time we went into a cpuidle state and came out of it is almost same with both the images. o The average time between sched_waking and sched_switch traces for the cyclictest threads is 19.8 us with rc3 and 13.5 us with this patch. And so the response to the threads is actually better with this patch. o Need more suggestions to verify why the regression is there.
- cyclictest (without latency)
Command: cyclictest -q -t N -i 10000 --latency=0 -D 20
4.12-rc3 4.12-rc3 + this patch Change -------- --------------------- ----------
RESULT (N=8) 16.34 8.87 - 45.78 % RESULT (N=16) 13.91 6.41 - 54 % RESULT (N=24) 12.42 9.1 - 26.77 % RESULT (N=32) 13 3.34 - 74 %
UPDATE-COUNT 26000 350 - 98.6 %
Note: Most of the MAX values with rc3 were between 10-30 and that with this patch were between 2-20. There were few random high values (100-400) though with both the images and are discarded to get the averages.
Summary:
o The number of times cpufreq utilization callbacks were called (UPDATE-COUNT) during the tests has come down heavily (81-96 %), except for ebizzy. And that's a lot of code removed from scheduler's hot path.
o pipe-test and hackbench have shown a clear improvement with the patch.
o ebizzy has shown a regression (specially at N=8). Though during some individual tests, rc3 also reported low numbers as compared to the kernel with this patch.
o cyclictest with 0 latency (i.e. no deeper idle states) shows clear improvement, while with latency=1000000 there is considerable regression.
Signed-off-by: Viresh Kumar viresh.kumar@linaro.org
--- I have also created a callgraph of where exactly are cpufreq hooks getting called from within the fair scheduler: https://yuml.me/dcc6d554. --- kernel/sched/fair.c | 13 ++++--------- 1 file changed, 4 insertions(+), 9 deletions(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 47a0c552c77b..62fe58b68e43 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -3276,7 +3276,7 @@ static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) { struct sched_avg *sa = &cfs_rq->avg; - int decayed, removed_load = 0, removed_util = 0; + int decayed, removed_load = 0;
if (atomic_long_read(&cfs_rq->removed_load_avg)) { s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0); @@ -3290,7 +3290,6 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0); sub_positive(&sa->util_avg, r); sub_positive(&sa->util_sum, r * LOAD_AVG_MAX); - removed_util = 1; set_tg_cfs_propagate(cfs_rq); }
@@ -3301,9 +3300,6 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) cfs_rq->load_last_update_time_copy = sa->last_update_time; #endif
- if (update_freq && (decayed || removed_util)) - cfs_rq_util_change(cfs_rq); - return decayed || removed_load; }
@@ -3481,10 +3477,7 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) #define UPDATE_TG 0x0 #define SKIP_AGE_LOAD 0x0
-static inline void update_load_avg(struct sched_entity *se, int not_used1) -{ - cpufreq_update_util(rq_of(cfs_rq_of(se)), 0); -} +static inline void update_load_avg(struct sched_entity *se, int not_used1) {}
static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} @@ -9021,6 +9014,8 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) entity_tick(cfs_rq, se, queued); }
+ cpufreq_update_util(rq, 0); + if (static_branch_unlikely(&sched_numa_balancing)) task_tick_numa(rq, curr); }
On Thu, Jun 01, 2017 at 05:04:27PM +0530, Viresh Kumar wrote:
This patch relocates the call to utilization hook from update_cfs_rq_load_avg() to task_tick_fair().
That's not right. Consider hardware where 'setting' the DVFS is a 'cheap' MSR write, doing that once every 10ms (HZ=100) is absurd.
We spoke about this problem in Pisa, the proposed solution was having each driver provide a cost metric and the generic code doing a max filter over the window constructed from that cost metric.
+ Patrick,
On 01-06-17, 14:22, Peter Zijlstra wrote:
On Thu, Jun 01, 2017 at 05:04:27PM +0530, Viresh Kumar wrote:
This patch relocates the call to utilization hook from update_cfs_rq_load_avg() to task_tick_fair().
That's not right. Consider hardware where 'setting' the DVFS is a 'cheap' MSR write, doing that once every 10ms (HZ=100) is absurd.
Yeah, that may be too much for such a platforms. Actually we (/me & Vincent) were worried about the current location of the utilization update hooks and believed that they are getting called way too often. But yeah, this patch optimized it way too much.
One of the goals of this patch was to avoid doing small OPP updates from update_load_avg() which can potentially block significant utilization changes (and hence big OPP changes) while a task is attached or detached, etc.
We spoke about this problem in Pisa, the proposed solution was having each driver provide a cost metric and the generic code doing a max filter over the window constructed from that cost metric.
So we want to compensate for the lost opportunities (due to rate_limit_us window) by changing the OPP based on what has happened in the previous rate_limit_us window. I am not sure how will that help.
Case 1: A periodic RT task runs for a small time in the rate_limit_us window and the timing is such that we (almost) never go to the max OPP because of rate_limit_us window.
Wouldn't a better solution towards such a case is what Patrick [1] proposed earlier (i.e. ignore rate_limit_us for RT/DL tasks), as we will run at high OPP when we really needed it the most.
Case 2: A high utilization periodic CFS task runs for short duration and keeps on migrating to other CPUs. We miss the opportunity to update the OPP based on this tasks utilization because of rate_limit_us window and by the time we update the OPP again, this task is already migrated and so the utilization is low again.
If the task has already migrated, why should we increase the OPP on assumption that this task will come back on this CPU? There are enough chances that the selected (higher) OPP will not be utilized by the current load on the CPU.
Also if this CFS tasks runs once every 2 (or more) ticks on the same CPU, then we are back to the same problem again.
1 2 3 4 |---------|---------|---------|---------|
T T
1,2,3,4 are representing the events on which we try to update the OPP and are placed rate_limit_us distance apart. And the task T happens to run between 1-2 and 3-4. We will not change the frequency until the event 2 in this case as rate_limit_us window isn't over yet. We go to higher OPP on 2 (which is really wasted for the current loads) because T happened in the last window. On 3 we come back to the OPP proportional to the current load. And the next time T runs again, we are still stuck on the low OPP. So instead of fixing it, we made it worse by wasting power unnecessarily.
Is there any case I am missing that you are concerned about ?
On Wed, Jun 07, 2017 at 05:36:55PM +0530, Viresh Kumar wrote:
- Patrick,
On 01-06-17, 14:22, Peter Zijlstra wrote:
On Thu, Jun 01, 2017 at 05:04:27PM +0530, Viresh Kumar wrote:
This patch relocates the call to utilization hook from update_cfs_rq_load_avg() to task_tick_fair().
That's not right. Consider hardware where 'setting' the DVFS is a 'cheap' MSR write, doing that once every 10ms (HZ=100) is absurd.
Yeah, that may be too much for such a platforms. Actually we (/me & Vincent) were worried about the current location of the utilization update hooks and believed that they are getting called way too often. But yeah, this patch optimized it way too much.
One of the goals of this patch was to avoid doing small OPP updates from update_load_avg() which can potentially block significant utilization changes (and hence big OPP changes) while a task is attached or detached, etc.
To me that sounds like you want to apply a more clever filter to the utilization updates than a simple rate limiter as Peter suggests below. IMHO, it would be better to feed schedutil with all the available information and improve the filtering policy there instead of trying to hack the policy tweaking the input data.
We spoke about this problem in Pisa, the proposed solution was having each driver provide a cost metric and the generic code doing a max filter over the window constructed from that cost metric.
Maybe it is possible to somehow let the rate at which we allow OPP changes depend on the size of the 'error' delta between the current OPP and what we need. So radical changes causes OPP changes immediately, and small corrections have to wait longer?
On Wed, Jun 7, 2017 at 5:43 PM, Morten Rasmussen morten.rasmussen@arm.com wrote:
On Wed, Jun 07, 2017 at 05:36:55PM +0530, Viresh Kumar wrote:
- Patrick,
On 01-06-17, 14:22, Peter Zijlstra wrote:
On Thu, Jun 01, 2017 at 05:04:27PM +0530, Viresh Kumar wrote:
This patch relocates the call to utilization hook from update_cfs_rq_load_avg() to task_tick_fair().
That's not right. Consider hardware where 'setting' the DVFS is a 'cheap' MSR write, doing that once every 10ms (HZ=100) is absurd.
Yeah, that may be too much for such a platforms. Actually we (/me & Vincent) were worried about the current location of the utilization update hooks and believed that they are getting called way too often. But yeah, this patch optimized it way too much.
One of the goals of this patch was to avoid doing small OPP updates from update_load_avg() which can potentially block significant utilization changes (and hence big OPP changes) while a task is attached or detached, etc.
To me that sounds like you want to apply a more clever filter to the utilization updates than a simple rate limiter as Peter suggests below. IMHO, it would be better to feed schedutil with all the available information and improve the filtering policy there instead of trying to hack the policy tweaking the input data.
Agreed.
Unless the tweaked input data would be used somewhere else too, that is.
We spoke about this problem in Pisa, the proposed solution was having each driver provide a cost metric and the generic code doing a max filter over the window constructed from that cost metric.
Maybe it is possible to somehow let the rate at which we allow OPP changes depend on the size of the 'error' delta between the current OPP and what we need. So radical changes causes OPP changes immediately, and small corrections have to wait longer?
That sounds reasonable to me.
Thanks, Rafael
linaro-kernel@lists.linaro.org
-
Morten Rasmussen -
Peter Zijlstra -
Rafael J. Wysocki -
Viresh Kumar