Currently, the margin of cpu frequency raising and cpu overutilized are
hard-coded as 25% (1280/1024). Make the margin tunable
to control the aggressive for placement and frequency control. Such as
for power tuning framework could adjust smaller margin to slow down
frequency raising speed and let task stay in smaller cpu.
For light loading scenarios, like beach buggy blitz and messaging apps,
the app threads are moved big core with 25% margin and causing
unnecessary power.
With 0% capacity margin (1024/1024), the app threads could be kept in
little core and deliver better power results without any fps drop.
capacity margin 0% 10% 20% 30%
current current current current
Fps (mA) Fps (mA) Fps (mA) Fps (mA)
Beach buggy blitz 60 198.164 60 203.211 60 209.984 60 213.374
Yahoo browser 60 232.301 59.97 237.52 59.95 248.213 60 262.809
Change-Id: Iba48c556ed1b73c9a2699e9e809bc7d9333dc004
Signed-off-by: YT Chang <[email protected]>
---
include/linux/sched/cpufreq.h | 19 +++++++++++++++++++
include/linux/sched/sysctl.h | 1 +
include/linux/sysctl.h | 1 +
kernel/sched/fair.c | 4 +++-
kernel/sysctl.c | 15 +++++++++++++++
5 files changed, 39 insertions(+), 1 deletion(-)
diff --git a/include/linux/sched/cpufreq.h b/include/linux/sched/cpufreq.h
index 6205578..8a6c23a1 100644
--- a/include/linux/sched/cpufreq.h
+++ b/include/linux/sched/cpufreq.h
@@ -23,6 +23,23 @@ void cpufreq_add_update_util_hook(int cpu, struct update_util_data *data,
void cpufreq_remove_update_util_hook(int cpu);
bool cpufreq_this_cpu_can_update(struct cpufreq_policy *policy);
+#ifdef CONFIG_SMP
+extern unsigned int sysctl_sched_capacity_margin;
+
+static inline unsigned long map_util_freq(unsigned long util,
+ unsigned long freq, unsigned long cap)
+{
+ freq = freq * util / cap;
+ freq = freq * sysctl_sched_capacity_margin / SCHED_CAPACITY_SCALE;
+
+ return freq;
+}
+
+static inline unsigned long map_util_perf(unsigned long util)
+{
+ return util * sysctl_sched_capacity_margin / SCHED_CAPACITY_SCALE;
+}
+#else
static inline unsigned long map_util_freq(unsigned long util,
unsigned long freq, unsigned long cap)
{
@@ -33,6 +50,8 @@ static inline unsigned long map_util_perf(unsigned long util)
{
return util + (util >> 2);
}
+#endif
+
#endif /* CONFIG_CPU_FREQ */
#endif /* _LINUX_SCHED_CPUFREQ_H */
diff --git a/include/linux/sched/sysctl.h b/include/linux/sched/sysctl.h
index db2c0f3..5dee024 100644
--- a/include/linux/sched/sysctl.h
+++ b/include/linux/sched/sysctl.h
@@ -10,6 +10,7 @@
#ifdef CONFIG_SMP
extern unsigned int sysctl_hung_task_all_cpu_backtrace;
+extern unsigned int sysctl_sched_capacity_margin;
#else
#define sysctl_hung_task_all_cpu_backtrace 0
#endif /* CONFIG_SMP */
diff --git a/include/linux/sysctl.h b/include/linux/sysctl.h
index d99ca99..af6d70f 100644
--- a/include/linux/sysctl.h
+++ b/include/linux/sysctl.h
@@ -41,6 +41,7 @@
#define SYSCTL_ZERO ((void *)&sysctl_vals[0])
#define SYSCTL_ONE ((void *)&sysctl_vals[1])
#define SYSCTL_INT_MAX ((void *)&sysctl_vals[2])
+#define SCHED_CAPACITY_MARGIN_MIN 1024
extern const int sysctl_vals[];
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 20aa234..609b431 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -111,7 +111,9 @@ int __weak arch_asym_cpu_priority(int cpu)
*
* (default: ~20%)
*/
-#define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
+unsigned int sysctl_sched_capacity_margin = 1280;
+EXPORT_SYMBOL_GPL(sysctl_sched_capacity_margin);
+#define fits_capacity(cap, max) ((cap) * sysctl_sched_capacity_margin < (max) * 1024)
/*
* The margin used when comparing CPU capacities.
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index 14edf84..d6d2b84 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -127,6 +127,11 @@
static int six_hundred_forty_kb = 640 * 1024;
#endif
+/* this is needed for the proc of sysctl_sched_capacity_margin */
+#ifdef CONFIG_SMP
+static int min_sched_capacity_margin = 1024;
+#endif /* CONFIG_SMP */
+
/* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
@@ -1716,6 +1721,16 @@ int proc_do_static_key(struct ctl_table *table, int write,
.mode = 0644,
.proc_handler = proc_dointvec,
},
+#ifdef CONFIG_SMP
+ {
+ .procname = "sched_capcity_margin",
+ .data = &sysctl_sched_capacity_margin,
+ .maxlen = sizeof(unsigned int),
+ .mode = 0644,
+ .proc_handler = proc_dointvec_minmax,
+ .extra1 = &min_sched_capacity_margin,
+ },
+#endif
#ifdef CONFIG_SCHEDSTATS
{
.procname = "sched_schedstats",
--
1.9.1
On Wed, 16 Jun 2021 at 17:06, YT Chang <[email protected]> wrote:
>
> Currently, the margin of cpu frequency raising and cpu overutilized are
> hard-coded as 25% (1280/1024). Make the margin tunable
cpu_overutilized is 20% not 25%. Even if they light looks similar
these 2 margins are differents
> to control the aggressive for placement and frequency control. Such as
> for power tuning framework could adjust smaller margin to slow down
> frequency raising speed and let task stay in smaller cpu.
>
> For light loading scenarios, like beach buggy blitz and messaging apps,
> the app threads are moved big core with 25% margin and causing
> unnecessary power.
> With 0% capacity margin (1024/1024), the app threads could be kept in
> little core and deliver better power results without any fps drop.
>
> capacity margin 0% 10% 20% 30%
> current current current current
> Fps (mA) Fps (mA) Fps (mA) Fps (mA)
> Beach buggy blitz 60 198.164 60 203.211 60 209.984 60 213.374
> Yahoo browser 60 232.301 59.97 237.52 59.95 248.213 60 262.809
Would be good to know the impact of each part:
Changing the +25% in cpufreq governor
Changing the 20% margin to detect overloaded CPU
Also, IIUC your description, the current values are ok with some UCs
but not with others like the 2 aboves. Have you evaluated whether it
was not your power model that was not accurate ?
>
> Change-Id: Iba48c556ed1b73c9a2699e9e809bc7d9333dc004
> Signed-off-by: YT Chang <[email protected]>
> ---
> include/linux/sched/cpufreq.h | 19 +++++++++++++++++++
> include/linux/sched/sysctl.h | 1 +
> include/linux/sysctl.h | 1 +
> kernel/sched/fair.c | 4 +++-
> kernel/sysctl.c | 15 +++++++++++++++
> 5 files changed, 39 insertions(+), 1 deletion(-)
>
> diff --git a/include/linux/sched/cpufreq.h b/include/linux/sched/cpufreq.h
> index 6205578..8a6c23a1 100644
> --- a/include/linux/sched/cpufreq.h
> +++ b/include/linux/sched/cpufreq.h
> @@ -23,6 +23,23 @@ void cpufreq_add_update_util_hook(int cpu, struct update_util_data *data,
> void cpufreq_remove_update_util_hook(int cpu);
> bool cpufreq_this_cpu_can_update(struct cpufreq_policy *policy);
>
> +#ifdef CONFIG_SMP
> +extern unsigned int sysctl_sched_capacity_margin;
> +
> +static inline unsigned long map_util_freq(unsigned long util,
> + unsigned long freq, unsigned long cap)
> +{
> + freq = freq * util / cap;
> + freq = freq * sysctl_sched_capacity_margin / SCHED_CAPACITY_SCALE;
> +
> + return freq;
> +}
> +
> +static inline unsigned long map_util_perf(unsigned long util)
> +{
> + return util * sysctl_sched_capacity_margin / SCHED_CAPACITY_SCALE;
> +}
> +#else
> static inline unsigned long map_util_freq(unsigned long util,
> unsigned long freq, unsigned long cap)
> {
> @@ -33,6 +50,8 @@ static inline unsigned long map_util_perf(unsigned long util)
> {
> return util + (util >> 2);
> }
> +#endif
> +
> #endif /* CONFIG_CPU_FREQ */
>
> #endif /* _LINUX_SCHED_CPUFREQ_H */
> diff --git a/include/linux/sched/sysctl.h b/include/linux/sched/sysctl.h
> index db2c0f3..5dee024 100644
> --- a/include/linux/sched/sysctl.h
> +++ b/include/linux/sched/sysctl.h
> @@ -10,6 +10,7 @@
>
> #ifdef CONFIG_SMP
> extern unsigned int sysctl_hung_task_all_cpu_backtrace;
> +extern unsigned int sysctl_sched_capacity_margin;
> #else
> #define sysctl_hung_task_all_cpu_backtrace 0
> #endif /* CONFIG_SMP */
> diff --git a/include/linux/sysctl.h b/include/linux/sysctl.h
> index d99ca99..af6d70f 100644
> --- a/include/linux/sysctl.h
> +++ b/include/linux/sysctl.h
> @@ -41,6 +41,7 @@
> #define SYSCTL_ZERO ((void *)&sysctl_vals[0])
> #define SYSCTL_ONE ((void *)&sysctl_vals[1])
> #define SYSCTL_INT_MAX ((void *)&sysctl_vals[2])
> +#define SCHED_CAPACITY_MARGIN_MIN 1024
>
> extern const int sysctl_vals[];
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 20aa234..609b431 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -111,7 +111,9 @@ int __weak arch_asym_cpu_priority(int cpu)
> *
> * (default: ~20%)
> */
> -#define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
> +unsigned int sysctl_sched_capacity_margin = 1280;
> +EXPORT_SYMBOL_GPL(sysctl_sched_capacity_margin);
> +#define fits_capacity(cap, max) ((cap) * sysctl_sched_capacity_margin < (max) * 1024)
>
> /*
> * The margin used when comparing CPU capacities.
> diff --git a/kernel/sysctl.c b/kernel/sysctl.c
> index 14edf84..d6d2b84 100644
> --- a/kernel/sysctl.c
> +++ b/kernel/sysctl.c
> @@ -127,6 +127,11 @@
> static int six_hundred_forty_kb = 640 * 1024;
> #endif
>
> +/* this is needed for the proc of sysctl_sched_capacity_margin */
> +#ifdef CONFIG_SMP
> +static int min_sched_capacity_margin = 1024;
> +#endif /* CONFIG_SMP */
> +
> /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
> static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
>
> @@ -1716,6 +1721,16 @@ int proc_do_static_key(struct ctl_table *table, int write,
> .mode = 0644,
> .proc_handler = proc_dointvec,
> },
> +#ifdef CONFIG_SMP
> + {
> + .procname = "sched_capcity_margin",
> + .data = &sysctl_sched_capacity_margin,
> + .maxlen = sizeof(unsigned int),
> + .mode = 0644,
> + .proc_handler = proc_dointvec_minmax,
> + .extra1 = &min_sched_capacity_margin,
> + },
> +#endif
> #ifdef CONFIG_SCHEDSTATS
> {
> .procname = "sched_schedstats",
> --
> 1.9.1
>
Hi YT Chang
Thanks for the patch.
On 06/16/21 23:05, YT Chang wrote:
> Currently, the margin of cpu frequency raising and cpu overutilized are
> hard-coded as 25% (1280/1024). Make the margin tunable
The way I see cpu overutilized is that we check if we're above the 80% range.
> to control the aggressive for placement and frequency control. Such as
> for power tuning framework could adjust smaller margin to slow down
> frequency raising speed and let task stay in smaller cpu.
>
> For light loading scenarios, like beach buggy blitz and messaging apps,
> the app threads are moved big core with 25% margin and causing
> unnecessary power.
> With 0% capacity margin (1024/1024), the app threads could be kept in
> little core and deliver better power results without any fps drop.
>
> capacity margin 0% 10% 20% 30%
> current current current current
> Fps (mA) Fps (mA) Fps (mA) Fps (mA)
> Beach buggy blitz 60 198.164 60 203.211 60 209.984 60 213.374
> Yahoo browser 60 232.301 59.97 237.52 59.95 248.213 60 262.809
>
> Change-Id: Iba48c556ed1b73c9a2699e9e809bc7d9333dc004
> Signed-off-by: YT Chang <[email protected]>
> ---
We are aware of the cpu overutilized value not being adequate on some modern
platforms. But I haven't considered or seen any issues with the frequency one.
So the latter is an interesting one.
I like your patch, but sadly I can't agree with it too.
The dilemma is that there are several options forward based on what we've seen
vendors do/want:
1. Modify the margin to be small for high end SoC and larger for lower
end ones. Which is what your patch allows.
2. Some vendors have a per cluster (perf domain) value. So within the
same SoC different margins are used for each capacity level.
3. Some vendors have asymmetric margin. A margin to move up and a
different margin to go down.
We're still not sure which approach is the best way forward.
Your patch allows 1, but if it turned out options 2 or 3 are better; the ABI
will make it hard to change.
Have you considered all these options? Do you have any data to help support
1 is enough for the range of platforms you work with at least?
We were considering also whether we can have a smarter logic to automagically
set a better value for the platform, but no concrete suggestions yet.
So while I agree the current margin value of one size fits all is no longer
suitable. But the variation of hardware and the possible approaches we could
take need more careful thinking and consideration before committing to an ABI.
This patch is a good start for this discussion :)
Thanks
--
Qais Yousef