On Thu, Oct 13, 2016 at 9:04 AM, Patrick Bellasi patrick.bellasi@arm.com wrote:
On 13-Oct 08:50, Andres Oportus wrote:
On Thu, Oct 13, 2016 at 6:43 AM, Leo Yan leo.yan@linaro.org wrote:
On Thu, Oct 13, 2016 at 03:18:16PM +0200, Vincent Guittot wrote:
On 13 October 2016 at 15:05, Patrick Bellasi <
patrick.bellasi@arm.com>
wrote:
On 10-Oct 16:35, Leo Yan wrote:
Add extra two performance optimization methods by setting sysfs
nodes:
Method 1: set sched_migration_cost_ns to 0:
By default sched_migration_cost_ns = 50000, scheduler calls
That's 50us...
In fact default value is 500000 = 500us not 50000 = 50us
Sorry, should be 500us.
can_migrate_task() to check if tasks are cache hot or not and it compares sched_migration_cost_ns to avoid migrate tasks
frequently.
This introduces side effects to easily pack tasks on the same
one
CPU
and introduce latency to spread tasks within multi-cores,
especially
if we think energy aware scheduling is easily to pack tasks on
single
CPU. So after task packing on one CPU with high utilization, we
can
easily spread out tasks after we set sched_migration_cost_ns to
... dunno how exactly this metric is used by the scheduler but, according to its name and you explanation, it seems that in the use-case you are targeting, tasks needs to me migrated more often
than
50us. Is that the case?
main advantage is that there are sysfs entry for that so it can be tuned for each platform
If set this value to 0, the most benefit I can see is if there have idle CPU and there have two tasks are runnable on another CPU, it can give more chance to migrate one of the runnable tasks onto the idle CPU immediately.
The currently available mechanism to have less latency and spread tasks
is
to set the prefer_idle flag (spreads tasks in the corresponding cgroup as long as there are idle cores).
That's what AOSP's kernels use for the wakeup path.
wouldn't placement in the wakeup path possibly place a task in an idle cpu to begin with and get the same or higher perf improvement compared to moving the task as part of load balancing?
Is this set in this experiment? Isn't load balancing supposed to be more about throughput and hence "slower" to kick in moving tasks as needed?
What Leo is addressing here is idle load balance, when a CPU is going to become idle we would like to pull tasks from CPUs which have many "as soon as possible".
Is seems that based on his experiments, the "migration cost" is impacting on the movement of some tasks by introducing latencies. By tuning the migration cost value (actually setting it to 0) _some_ benchmarks have been measured to get an uplift on performances.
What we need to understand (at fist instance) is how much "generic" (i.e. platform and workload independent) is the proposed solution.
I agree, I would think that migration cost tuning could improve load balancing behavior but setting it to 0 effectively saying that there is no cost to task movement as part of load balancing seems incorrect. I'm wondering if this is a scenario that could be improved/tuned say in the wakeup path rather than assuming no migration cost.
As a general comment, I can understand that an hardcoded 50us value could be not generic at all, however: is there any indication on
how
to properly dimension this value for a specific target?
Maybe a specific set of synthetics experiments can be used to
figure
out what is the best value to be used. In that case we should probably report in the documentation how to measure and tune experimentally this value instead of just
changing an
hardcoded value with another one.
Method 2: set busy_factor to 1:
This decreases load balance inteval time, so it will give more
chance
for active load balance for migration running task from little
core
to
big core.
Same reasoning as before, how can be sure that the value you are proposing (ie. busy_factor=1) it is really generic enough?
Method 1 can improve prominent performance on one big.LITTLE
system
(which has CA53x4 + CA72x4 cores), from the Geekbench testing
result
the
score can improve performance ~5%.
Tested method 1 with Geekbench on the ARM Juno R2 board for
multi-thread
case, the score can be improved from 2348 to 2368, so can improve performance ~0.84%.
Am I correct on assuming that potentially different values can
give us
even better performance but we tried and tested only the two values you are proposing?
Yes, I only tried these two values. Just like Patrick suggested, the methodology is more important rather than hard-coded value.
For the 1st test, the root cause was that tasks was hot on a CPU and can't be selected to migrate on other CPU because of its hotness so decreasing the sched_migration_cost_ns directly reduce the hotness period during which a task can't migrate on another CPU
That being said, i'm not sure that this should be put in the Documentation/scheduler/sched-energy.tx
I think for EAS, these two paramters are more important than traditional SMP load balance. Because EAS will have more chance to pack tasks onto single CPU or into one cluster, so we need utilize the existed machenism to spread out these tasks, so sched_migration_cost_ns and busy_factors are two things we can rely on.
Moreover, do we have any measure of the impact on energy
consumption
for the proposed value?
From one member's platform, have not observed power impaction for energy consumption. I will try to use video playback case on Juno to generate out more power data.
Tested method 2 on Juno as well, but it has very minor performance boosting.
That seems to support the idea that what you are proposing are
values
"optimal" only for performance on a specific platform. Isn't it?
Yes.
Signed-off-by: Leo Yan leo.yan@linaro.org
Documentation/scheduler/sched-energy.txt | 24
++++++++++++++++++++++++
1 file changed, 24 insertions(+)
diff --git a/Documentation/scheduler/sched-energy.txt
b/Documentation/scheduler/sched-energy.txt
index dab2f90..c0e62fe 100644 --- a/Documentation/scheduler/sched-energy.txt +++ b/Documentation/scheduler/sched-energy.txt @@ -360,3 +360,27 @@ of the cpu from idle/busy power of the shared
resources. The cpu can be tricked
into different per-cpu idle states by disabling the other states.
Based on
various combinations of measurements with specific cpus busy and
disabling
idle-states it is possible to extrapolate the idle-state power.
+Performance tunning method +==========================
+Below setting may impact heavily for performance tunning:
+echo 0 > /proc/sys/kernel/sched_migration_cost_ns
+After set sched_migration_cost_ns to 0, it is helpful to spread
tasks within
+the big cluster. Otherwise when scheduler executes load balance,
it
calls
+can_migrate_task() to check if tasks are cache hot or not and it
compares
+sched_migration_cost_ns to avoid migrate tasks frequently. This
introduce side
+effect to easily pack tasks on the same one CPU and introduce
latency to
+spread tasks within multi-cores, especially if we think about
energy
awared
+scheduling to pack tasks on single CPU.
+echo 1 > /proc/sys/kernel/sched_domain/cpuX/domain0/busy_factor +echo 1 > /proc/sys/kernel/sched_domain/cpuX/domain1/busy_factor
+After set busy_factor to 1, it decreases load balance inteval
time.
So if we
+take min_interval = 8, that means we permit the load balance
interval =
+busy_factor * min_interval = 8ms. So this will shorten task
migration latency,
+especially if we want to migrate a running task from little core
to
big core
+to trigger active load balance.
1.9.1
-- #include <best/regards.h>
Patrick Bellasi
eas-dev mailing list eas-dev@lists.linaro.org https://lists.linaro.org/mailman/listinfo/eas-dev
eas-dev mailing list eas-dev@lists.linaro.org https://lists.linaro.org/mailman/listinfo/eas-dev
-- #include <best/regards.h>
Patrick Bellasi