Hi all,
This patchset changes slightly the calculation of target frequency to
eliminate the deadband effect (explained in patch 2 changelog) that it
seems to slow down the CPU in low and medium loads.
Patch 1 introduces a new relation (RELATION_C) for the next frequency
selection, which chooses the closest frequency to target.
Patch 2 is the actual change to ondemand governor.
You may find graphs with the 'deadband' effect and benchmark results:
https://docs.google.com/spreadsheets/d/16kDBh5lyc6YvBnoS1hUa1t2O38z0xrWvaEj5XtJ8auw/edit#gid=2072493052
Thanks
Stratos Karafotis (2):
cpufreq: Introduce new relation for freq selection
cpufreq: ondemand: Eliminate the deadband effect in low loads
drivers/cpufreq/cpufreq_ondemand.c | 11 +++++++----
drivers/cpufreq/freq_table.c | 12 +++++++++++-
include/linux/cpufreq.h | 1 +
3 files changed, 19 insertions(+), 5 deletions(-)
--
1.9.3
Introduce CPUFREQ_RELATION_C for frequency selection.
It selects the frequency with the minimum euclidean distance to target.
In case of equal distance between 2 frequencies, it will select the
greater frequency.
Signed-off-by: Stratos Karafotis <[email protected]>
---
drivers/cpufreq/freq_table.c | 12 +++++++++++-
include/linux/cpufreq.h | 1 +
2 files changed, 12 insertions(+), 1 deletion(-)
diff --git a/drivers/cpufreq/freq_table.c b/drivers/cpufreq/freq_table.c
index 1632981..df14766 100644
--- a/drivers/cpufreq/freq_table.c
+++ b/drivers/cpufreq/freq_table.c
@@ -117,7 +117,7 @@ int cpufreq_frequency_table_target(struct cpufreq_policy *policy,
.frequency = 0,
};
struct cpufreq_frequency_table *pos;
- unsigned int freq, i = 0;
+ unsigned int freq, diff, i = 0;
pr_debug("request for target %u kHz (relation: %u) for cpu %u\n",
target_freq, relation, policy->cpu);
@@ -127,6 +127,7 @@ int cpufreq_frequency_table_target(struct cpufreq_policy *policy,
suboptimal.frequency = ~0;
break;
case CPUFREQ_RELATION_L:
+ case CPUFREQ_RELATION_C:
optimal.frequency = ~0;
break;
}
@@ -168,6 +169,15 @@ int cpufreq_frequency_table_target(struct cpufreq_policy *policy,
}
}
break;
+ case CPUFREQ_RELATION_C:
+ diff = abs(freq - target_freq);
+ if (diff < optimal.frequency ||
+ (diff == optimal.frequency &&
+ freq > table[optimal.driver_data].frequency)) {
+ optimal.frequency = diff;
+ optimal.driver_data = i;
+ }
+ break;
}
}
if (optimal.driver_data > i) {
diff --git a/include/linux/cpufreq.h b/include/linux/cpufreq.h
index ec4112d..00fad91 100644
--- a/include/linux/cpufreq.h
+++ b/include/linux/cpufreq.h
@@ -176,6 +176,7 @@ static inline void disable_cpufreq(void) { }
#define CPUFREQ_RELATION_L 0 /* lowest frequency at or above target */
#define CPUFREQ_RELATION_H 1 /* highest frequency below or at target */
+#define CPUFREQ_RELATION_C 2 /* closest frequency to target */
struct freq_attr {
struct attribute attr;
--
1.9.3
Currently, ondemand calculates the target frequency proportional to load
using the formula:
Target frequency = C * load
where C = policy->cpuinfo.max_freq / 100
Though, in many cases, the minimum available frequency is pretty high and
the above calculation introduces a dead band from load 0 to
100 * policy->cpuinfo.min_freq / policy->cpuinfo.max_freq where the target
frequency is always calculated to less than policy->cpuinfo.min_freq and
the minimum frequency is selected.
For example: on Intel i7-3770 @ 3.4GHz the policy->cpuinfo.min_freq = 1600000
and the policy->cpuinfo.max_freq = 3400000 (without turbo). Thus, the CPU
starts to scale up at a load above 47.
On quad core 1500MHz Krait the policy->cpuinfo.min_freq = 384000
and the policy->cpuinfo.max_freq = 1512000. Thus, the CPU starts to scale
at load above 25.
Change the calculation of target frequency to eliminate the above effect using
the formula:
Target frequency = A + B * load
where A = policy->cpuinfo.min_freq and
B = (policy->cpuinfo.max_freq - policy->cpuinfo->min_freq) / 100
This will map load values 0 to 100 linearly to cpuinfo.min_freq to
cpuinfo.max_freq.
Also, use the CPUFREQ_RELATION_C in __cpufreq_driver_target to select the
closest frequency in frequency_table. This is necessary to avoid selection
of minimum frequency only when load equals to 0. It will also help for selection
of frequencies using a more 'fair' criterion.
Tables below show the difference in selected frequency for specific values
of load without and with this patch. On Intel i7-3770 @ 3.40GHz:
Without With
Load Target Selected Target Selected
0 0 1600000 1600000 1600000
5 170050 1600000 1690050 1700000
10 340100 1600000 1780100 1700000
15 510150 1600000 1870150 1900000
20 680200 1600000 1960200 2000000
25 850250 1600000 2050250 2100000
30 1020300 1600000 2140300 2100000
35 1190350 1600000 2230350 2200000
40 1360400 1600000 2320400 2400000
45 1530450 1600000 2410450 2400000
50 1700500 1900000 2500500 2500000
55 1870550 1900000 2590550 2600000
60 2040600 2100000 2680600 2600000
65 2210650 2400000 2770650 2800000
70 2380700 2400000 2860700 2800000
75 2550750 2600000 2950750 3000000
80 2720800 2800000 3040800 3000000
85 2890850 2900000 3130850 3100000
90 3060900 3100000 3220900 3300000
95 3230950 3300000 3310950 3300000
100 3401000 3401000 3401000 3401000
On ARM quad core 1500MHz Krait:
Without With
Load Target Selected Target Selected
0 0 384000 384000 384000
5 75600 384000 440400 486000
10 151200 384000 496800 486000
15 226800 384000 553200 594000
20 302400 384000 609600 594000
25 378000 384000 666000 702000
30 453600 486000 722400 702000
35 529200 594000 778800 810000
40 604800 702000 835200 810000
45 680400 702000 891600 918000
50 756000 810000 948000 918000
55 831600 918000 1004400 1026000
60 907200 918000 1060800 1026000
65 982800 1026000 1117200 1134000
70 1058400 1134000 1173600 1134000
75 1134000 1134000 1230000 1242000
80 1209600 1242000 1286400 1242000
85 1285200 1350000 1342800 1350000
90 1360800 1458000 1399200 1350000
95 1436400 1458000 1455600 1458000
100 1512000 1512000 1512000 1512000
Tested on Intel i7-3770 CPU @ 3.40GHz and on ARM quad core 1500MHz Krait
(Android smartphone).
Benchmarks on Intel i7 shows a performance improvement on low and medium
work loads with lower power consumption. Specifics:
Phoronix Linux Kernel Compilation 3.1:
Time: -0.40%, energy: -0.07%
Phoronix Apache:
Time: -4.98%, energy: -2.35%
Phoronix FFMPEG:
Time: -6.29%, energy: -4.02%
Also, running mp3 decoding (very low load) shows no differences with and
without this patch.
Signed-off-by: Stratos Karafotis <[email protected]>
---
drivers/cpufreq/cpufreq_ondemand.c | 11 +++++++----
1 file changed, 7 insertions(+), 4 deletions(-)
diff --git a/drivers/cpufreq/cpufreq_ondemand.c b/drivers/cpufreq/cpufreq_ondemand.c
index 18d4091..ad3f38f 100644
--- a/drivers/cpufreq/cpufreq_ondemand.c
+++ b/drivers/cpufreq/cpufreq_ondemand.c
@@ -170,21 +170,24 @@ static void od_check_cpu(int cpu, unsigned int load)
dbs_freq_increase(policy, policy->max);
} else {
/* Calculate the next frequency proportional to load */
- unsigned int freq_next;
- freq_next = load * policy->cpuinfo.max_freq / 100;
+ unsigned int freq_next, min_f, max_f;
+
+ min_f = policy->cpuinfo.min_freq;
+ max_f = policy->cpuinfo.max_freq;
+ freq_next = min_f + load * (max_f - min_f) / 100;
/* No longer fully busy, reset rate_mult */
dbs_info->rate_mult = 1;
if (!od_tuners->powersave_bias) {
__cpufreq_driver_target(policy, freq_next,
- CPUFREQ_RELATION_L);
+ CPUFREQ_RELATION_C);
return;
}
freq_next = od_ops.powersave_bias_target(policy, freq_next,
CPUFREQ_RELATION_L);
- __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
+ __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
}
}
--
1.9.3
Hi!
> Tested on Intel i7-3770 CPU @ 3.40GHz and on ARM quad core 1500MHz Krait
> (Android smartphone).
> Benchmarks on Intel i7 shows a performance improvement on low and medium
> work loads with lower power consumption. Specifics:
>
> Phoronix Linux Kernel Compilation 3.1:
> Time: -0.40%, energy: -0.07%
> Phoronix Apache:
> Time: -4.98%, energy: -2.35%
> Phoronix FFMPEG:
> Time: -6.29%, energy: -4.02%
Hmm. Intel i7 should be race-to-idle machine. So basically rule like
if (load > 0) go to max frequency else go to lowest frequency would do
the right thing in your test, right?
So... should we do that, or do we need better benchmark?
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
Hi Pavel!
On 11/07/2014 07:57 μμ, Pavel Machek wrote:
> Hi!
>
>> Tested on Intel i7-3770 CPU @ 3.40GHz and on ARM quad core 1500MHz Krait
>> (Android smartphone).
>> Benchmarks on Intel i7 shows a performance improvement on low and medium
>> work loads with lower power consumption. Specifics:
>>
>> Phoronix Linux Kernel Compilation 3.1:
>> Time: -0.40%, energy: -0.07%
>> Phoronix Apache:
>> Time: -4.98%, energy: -2.35%
>> Phoronix FFMPEG:
>> Time: -6.29%, energy: -4.02%
>
> Hmm. Intel i7 should be race-to-idle machine. So basically rule like
> if (load > 0) go to max frequency else go to lowest frequency would do
> the right thing in your test, right?
I don't think that "if (load > 0) go to max" will work even on i7.
For low load this will have impact on energy consumption.
On my tests, a simple mp3 decoding (very low load on my machine) have no
difference with and without this patch.
> So... should we do that, or do we need better benchmark?
I'm sorry. I'm not sure I understood exactly what do you mean by "better
benchmark".
Of course, we should do as many benchmarks as we can.
I usually do these 5 sets of benchmarks on my i7 that IMHO give a good
indication about the changes in different CPU loads.
1) Linux kernel compilation (about 85% busy CPU)
2) Apache (about 32% busy CPU)
3) ffmpeg (about 24% busy CPU)
4) mp3 decoding (about 0.3% CPU)
5) Idle system (about 0.06% CPU)
The patch was also tested on a Android smartphone (kernel 3.4). The kernel
distributed to 1000+ users.
Unfortunately I have no benchmarks, but no regressions reported on
consumption. Actually, there reports for better performance and
lower power consumption, but of course we can't rely on these reports. :)
Thanks for your comments!
Stratos
On Fri 2014-07-11 20:29:57, Stratos Karafotis wrote:
> Hi Pavel!
>
> On 11/07/2014 07:57 μμ, Pavel Machek wrote:
> > Hi!
> >
> >> Tested on Intel i7-3770 CPU @ 3.40GHz and on ARM quad core 1500MHz Krait
> >> (Android smartphone).
> >> Benchmarks on Intel i7 shows a performance improvement on low and medium
> >> work loads with lower power consumption. Specifics:
> >>
> >> Phoronix Linux Kernel Compilation 3.1:
> >> Time: -0.40%, energy: -0.07%
> >> Phoronix Apache:
> >> Time: -4.98%, energy: -2.35%
> >> Phoronix FFMPEG:
> >> Time: -6.29%, energy: -4.02%
> >
> > Hmm. Intel i7 should be race-to-idle machine. So basically rule like
> > if (load > 0) go to max frequency else go to lowest frequency would do
> > the right thing in your test, right?
>
> I don't think that "if (load > 0) go to max" will work even on i7.
> For low load this will have impact on energy consumption.
Are you sure? CPU frequency should not matter on idle CPU.
(Can you try to modify your code and rerun for example the apache
test?)
> > So... should we do that, or do we need better benchmark?
>
> I'm sorry. I'm not sure I understood exactly what do you mean by "better
> benchmark".
I believe that any increase of frequency in frequency will make the
benchmarks you qouted better (on i7). Actually, you can probably just
select performance governor...?
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
On 11/07/2014 09:34 μμ, Pavel Machek wrote:
> On Fri 2014-07-11 20:29:57, Stratos Karafotis wrote:
>> Hi Pavel!
>>
>> On 11/07/2014 07:57 μμ, Pavel Machek wrote:
>>> Hi!
>>>
>>>> Tested on Intel i7-3770 CPU @ 3.40GHz and on ARM quad core 1500MHz Krait
>>>> (Android smartphone).
>>>> Benchmarks on Intel i7 shows a performance improvement on low and medium
>>>> work loads with lower power consumption. Specifics:
>>>>
>>>> Phoronix Linux Kernel Compilation 3.1:
>>>> Time: -0.40%, energy: -0.07%
>>>> Phoronix Apache:
>>>> Time: -4.98%, energy: -2.35%
>>>> Phoronix FFMPEG:
>>>> Time: -6.29%, energy: -4.02%
>>>
>>> Hmm. Intel i7 should be race-to-idle machine. So basically rule like
>>> if (load > 0) go to max frequency else go to lowest frequency would do
>>> the right thing in your test, right?
>>
>> I don't think that "if (load > 0) go to max" will work even on i7.
>> For low load this will have impact on energy consumption.
>
> Are you sure? CPU frequency should not matter on idle CPU.
Even on a totally idle CPU there will be a small impact because of leakage
current (thanks to Dirk Brandewie for this info).
This simple test on a nearly idle system shows this:
[root@albert cpufreq]# for CPUFREQ in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do [ -f $CPUFREQ ] || continue; echo -n performance > $CPUFREQ; done
[root@albert cpufreq]# /home/stratosk/kernels/linux-pm/tools/power/x86/turbostat/turbostat -J sleep 20
Core CPU Avg_MHz %Busy Bzy_MHz TSC_MHz SMI CPU%c1 CPU%c3 CPU%c6 CPU%c7 CoreTmp PkgTmp Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 Pkg_J Cor_J GFX_J time
- - 2 0.06 2712 3392 0 0.30 0.00 99.63 0.00 34 34 8.09 0.00 81.94 0.00 380.41 14.51 1.64 20.00
0 0 0 0.02 1891 3392 0 0.09 0.00 99.88 0.00 34 34 8.09 0.00 81.94 0.00 380.41 14.51 1.64 20.00
0 4 1 0.04 3006 3392 0 0.07
1 1 1 0.04 2501 3392 0 0.62 0.00 99.33 0.00 34
1 5 0 0.01 2346 3392 0 0.66
2 2 0 0.01 1996 3392 0 0.44 0.00 99.55 0.00 34
2 6 4 0.18 2278 3392 0 0.26
3 3 5 0.15 3449 3392 0 0.07 0.01 99.77 0.00 34
3 7 0 0.01 1839 3392 0 0.21
20.000899 sec
[root@albert cpufreq]# ^C
[root@albert cpufreq]# for CPUFREQ in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do [ -f $CPUFREQ ] || continue; echo -n ondemand > $CPUFREQ; done
[root@albert cpufreq]# /home/stratosk/kernels/linux-pm/tools/power/x86/turbostat/turbostat -J sleep 20
Core CPU Avg_MHz %Busy Bzy_MHz TSC_MHz SMI CPU%c1 CPU%c3 CPU%c6 CPU%c7 CoreTmp PkgTmp Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 Pkg_J Cor_J GFX_J time
- - 2 0.09 1693 3392 0 0.35 0.01 99.55 0.00 35 36 8.33 0.00 84.31 0.00 377.68 12.23 1.15 20.00
0 0 1 0.08 1603 3392 0 0.13 0.00 99.79 0.00 35 36 8.33 0.00 84.31 0.00 377.68 12.23 1.15 20.00
0 4 1 0.08 1646 3392 0 0.13
1 1 1 0.06 1647 3392 0 0.66 0.00 99.28 0.00 35
1 5 0 0.01 1611 3392 0 0.71
2 2 0 0.02 1617 3392 0 0.50 0.02 99.46 0.00 35
2 6 4 0.22 1764 3392 0 0.30
3 3 4 0.25 1701 3392 0 0.07 0.00 99.68 0.00 35
3 7 0 0.01 1602 3392 0 0.31
20.001580 sec
So, for low loads the impact will be higher.
This is the reason that the intel_pstate driver don't use 'performance'
and try to request a low P state when there is no load.
> (Can you try to modify your code and rerun for example the apache
> test?)
Yes, I can do the apache test if the above example is not enough.
>>> So... should we do that, or do we need better benchmark?
>>
>> I'm sorry. I'm not sure I understood exactly what do you mean by "better
>> benchmark".
>
> I believe that any increase of frequency in frequency will make the
> benchmarks you qouted better (on i7). Actually, you can probably just
> select performance governor...?
Maybe in benchmarks where the CPU load is high. But definitely not, in mp3
decoding and idle system test.
The point is, as you mentioned, more tests and of course on other CPUs.
Unfortunately, I can test only on i7 and krait as mentioned in changelog.
I will happily run any test you would like for more info.
Thanks,
Stratos
On 2014.07.30 10:00 Stratos Karafotis wrote:
> This patchset changes slightly the calculation of target frequency to
> eliminate the deadband effect (explained in patch 2 changelog) that it
> seems to slow down the CPU in low and medium loads.
>
> Patch 1 introduces a new relation (RELATION_C) for the next frequency
> selection, which chooses the closest frequency to target.
>
> Patch 2 is the actual change to ondemand governor.
> You may find graphs with the 'deadband' effect and benchmark results:
> https://docs.google.com/spreadsheets/d/16kDBh5lyc6YvBnoS1hUa1t2O38z0xrWvaEj5XtJ8auw/edit#gid=2072493052
I did the same benchmark tests before (without) and after (with) this patch set on my i7-2600K system.
I added the results, which are similar to Stratos', under a new "benchmark" tab on the spreadsheet.
... Doug
Hi Doug,
On 12/07/2014 06:45 μμ, Doug Smythies wrote:
>
> On 2014.07.30 10:00 Stratos Karafotis wrote:
>
>> This patchset changes slightly the calculation of target frequency to
>> eliminate the deadband effect (explained in patch 2 changelog) that it
>> seems to slow down the CPU in low and medium loads.
>>
>> Patch 1 introduces a new relation (RELATION_C) for the next frequency
>> selection, which chooses the closest frequency to target.
>>
>> Patch 2 is the actual change to ondemand governor.
>
>> You may find graphs with the 'deadband' effect and benchmark results:
>> https://docs.google.com/spreadsheets/d/16kDBh5lyc6YvBnoS1hUa1t2O38z0xrWvaEj5XtJ8auw/edit#gid=2072493052
>
> I did the same benchmark tests before (without) and after (with) this patch set on my i7-2600K system.
> I added the results, which are similar to Stratos', under a new "benchmark" tab on the spreadsheet.
Thank you very much for your benchmarks!
Stratos
Hi!
> >>>> Tested on Intel i7-3770 CPU @ 3.40GHz and on ARM quad core 1500MHz Krait
> >>>> (Android smartphone).
> >>>> Benchmarks on Intel i7 shows a performance improvement on low and medium
> >>>> work loads with lower power consumption. Specifics:
> >>>>
> >>>> Phoronix Linux Kernel Compilation 3.1:
> >>>> Time: -0.40%, energy: -0.07%
> >>>> Phoronix Apache:
> >>>> Time: -4.98%, energy: -2.35%
> >>>> Phoronix FFMPEG:
> >>>> Time: -6.29%, energy: -4.02%
> >>>
> >>> Hmm. Intel i7 should be race-to-idle machine. So basically rule like
> >>> if (load > 0) go to max frequency else go to lowest frequency would do
> >>> the right thing in your test, right?
> >>
> >> I don't think that "if (load > 0) go to max" will work even on i7.
> >> For low load this will have impact on energy consumption.
> >
> > Are you sure? CPU frequency should not matter on idle CPU.
>
> Even on a totally idle CPU there will be a small impact because of leakage
> current (thanks to Dirk Brandewie for this info).
Are you sure? IIRC Intel cpus will automatically lower CPU frequency
in deep C states..
> This simple test on a nearly idle system shows this:
>
> [root@albert cpufreq]# for CPUFREQ in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do [ -f $CPUFREQ ] || continue; echo -n performance > $CPUFREQ; done
> [root@albert cpufreq]# /home/stratosk/kernels/linux-pm/tools/power/x86/turbostat/turbostat -J sleep 20
> Core CPU Avg_MHz %Busy Bzy_MHz TSC_MHz SMI CPU%c1 CPU%c3 CPU%c6 CPU%c7 CoreTmp PkgTmp Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 Pkg_J Cor_J GFX_J time
> - - 2 0.06 2712 3392 0 0.30 0.00 99.63 0.00 34 34 8.09 0.00 81.94 0.00 380.41 14.51 1.64 20.00
> 0 0 0 0.02 1891 3392 0 0.09 0.00 99.88 0.00 34 34 8.09 0.00 81.94 0.00 380.41 14.51 1.64 20.00
> 0 4 1 0.04 3006 3392 0 0.07
> 1 1 1 0.04 2501 3392 0 0.62 0.00 99.33 0.00 34
> 1 5 0 0.01 2346 3392 0 0.66
> 2 2 0 0.01 1996 3392 0 0.44 0.00 99.55 0.00 34
> 2 6 4 0.18 2278 3392 0 0.26
> 3 3 5 0.15 3449 3392 0 0.07 0.01 99.77 0.00 34
> 3 7 0 0.01 1839 3392 0 0.21
> 20.000899 sec
> [root@albert cpufreq]# ^C
> [root@albert cpufreq]# for CPUFREQ in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do [ -f $CPUFREQ ] || continue; echo -n ondemand > $CPUFREQ; done
> [root@albert cpufreq]# /home/stratosk/kernels/linux-pm/tools/power/x86/turbostat/turbostat -J sleep 20
> Core CPU Avg_MHz %Busy Bzy_MHz TSC_MHz SMI CPU%c1 CPU%c3 CPU%c6 CPU%c7 CoreTmp PkgTmp Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 Pkg_J Cor_J GFX_J time
> - - 2 0.09 1693 3392 0 0.35 0.01 99.55 0.00 35 36 8.33 0.00 84.31 0.00 377.68 12.23 1.15 20.00
> 0 0 1 0.08 1603 3392 0 0.13 0.00 99.79 0.00 35 36 8.33 0.00 84.31 0.00 377.68 12.23 1.15 20.00
> 0 4 1 0.08 1646 3392 0 0.13
> 1 1 1 0.06 1647 3392 0 0.66 0.00 99.28 0.00 35
> 1 5 0 0.01 1611 3392 0 0.71
> 2 2 0 0.02 1617 3392 0 0.50 0.02 99.46 0.00 35
> 2 6 4 0.22 1764 3392 0 0.30
> 3 3 4 0.25 1701 3392 0 0.07 0.00 99.68 0.00 35
> 3 7 0 0.01 1602 3392 0 0.31
> 20.001580 sec
>
>
> So, for low loads the impact will be higher.
So it seems ondemand saves cca 1% of energy?
Best regards,
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
On 21/07/2014 12:51 πμ, Pavel Machek wrote:
> Hi!
>
>>>>>> Tested on Intel i7-3770 CPU @ 3.40GHz and on ARM quad core 1500MHz Krait
>>>>>> (Android smartphone).
>>>>>> Benchmarks on Intel i7 shows a performance improvement on low and medium
>>>>>> work loads with lower power consumption. Specifics:
>>>>>>
>>>>>> Phoronix Linux Kernel Compilation 3.1:
>>>>>> Time: -0.40%, energy: -0.07%
>>>>>> Phoronix Apache:
>>>>>> Time: -4.98%, energy: -2.35%
>>>>>> Phoronix FFMPEG:
>>>>>> Time: -6.29%, energy: -4.02%
>>>>>
>>>>> Hmm. Intel i7 should be race-to-idle machine. So basically rule like
>>>>> if (load > 0) go to max frequency else go to lowest frequency would do
>>>>> the right thing in your test, right?
>>>>
>>>> I don't think that "if (load > 0) go to max" will work even on i7.
>>>> For low load this will have impact on energy consumption.
>>>
>>> Are you sure? CPU frequency should not matter on idle CPU.
>>
>> Even on a totally idle CPU there will be a small impact because of leakage
>> current (thanks to Dirk Brandewie for this info).
>
> Are you sure? IIRC Intel cpus will automatically lower CPU frequency
> in deep C states..
I'm sorry. I don't know further details about the leakage current
in deeper C states.
>> This simple test on a nearly idle system shows this:
>>
>> [root@albert cpufreq]# for CPUFREQ in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do [ -f $CPUFREQ ] || continue; echo -n performance > $CPUFREQ; done
>> [root@albert cpufreq]# /home/stratosk/kernels/linux-pm/tools/power/x86/turbostat/turbostat -J sleep 20
>> Core CPU Avg_MHz %Busy Bzy_MHz TSC_MHz SMI CPU%c1 CPU%c3 CPU%c6 CPU%c7 CoreTmp PkgTmp Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 Pkg_J Cor_J GFX_J time
>> - - 2 0.06 2712 3392 0 0.30 0.00 99.63 0.00 34 34 8.09 0.00 81.94 0.00 380.41 14.51 1.64 20.00
>> 0 0 0 0.02 1891 3392 0 0.09 0.00 99.88 0.00 34 34 8.09 0.00 81.94 0.00 380.41 14.51 1.64 20.00
>> 0 4 1 0.04 3006 3392 0 0.07
>> 1 1 1 0.04 2501 3392 0 0.62 0.00 99.33 0.00 34
>> 1 5 0 0.01 2346 3392 0 0.66
>> 2 2 0 0.01 1996 3392 0 0.44 0.00 99.55 0.00 34
>> 2 6 4 0.18 2278 3392 0 0.26
>> 3 3 5 0.15 3449 3392 0 0.07 0.01 99.77 0.00 34
>> 3 7 0 0.01 1839 3392 0 0.21
>> 20.000899 sec
>> [root@albert cpufreq]# ^C
>> [root@albert cpufreq]# for CPUFREQ in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor; do [ -f $CPUFREQ ] || continue; echo -n ondemand > $CPUFREQ; done
>> [root@albert cpufreq]# /home/stratosk/kernels/linux-pm/tools/power/x86/turbostat/turbostat -J sleep 20
>> Core CPU Avg_MHz %Busy Bzy_MHz TSC_MHz SMI CPU%c1 CPU%c3 CPU%c6 CPU%c7 CoreTmp PkgTmp Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 Pkg_J Cor_J GFX_J time
>> - - 2 0.09 1693 3392 0 0.35 0.01 99.55 0.00 35 36 8.33 0.00 84.31 0.00 377.68 12.23 1.15 20.00
>> 0 0 1 0.08 1603 3392 0 0.13 0.00 99.79 0.00 35 36 8.33 0.00 84.31 0.00 377.68 12.23 1.15 20.00
>> 0 4 1 0.08 1646 3392 0 0.13
>> 1 1 1 0.06 1647 3392 0 0.66 0.00 99.28 0.00 35
>> 1 5 0 0.01 1611 3392 0 0.71
>> 2 2 0 0.02 1617 3392 0 0.50 0.02 99.46 0.00 35
>> 2 6 4 0.22 1764 3392 0 0.30
>> 3 3 4 0.25 1701 3392 0 0.07 0.00 99.68 0.00 35
>> 3 7 0 0.01 1602 3392 0 0.31
>> 20.001580 sec
>>
>>
>> So, for low loads the impact will be higher.
>
> So it seems ondemand saves cca 1% of energy?
Yes, in this small test, on my nearly "idle" system.
Stratos
On Monday, June 30, 2014 07:59:32 PM Stratos Karafotis wrote:
> Hi all,
>
> This patchset changes slightly the calculation of target frequency to
> eliminate the deadband effect (explained in patch 2 changelog) that it
> seems to slow down the CPU in low and medium loads.
>
> Patch 1 introduces a new relation (RELATION_C) for the next frequency
> selection, which chooses the closest frequency to target.
>
> Patch 2 is the actual change to ondemand governor.
>
> You may find graphs with the 'deadband' effect and benchmark results:
> https://docs.google.com/spreadsheets/d/16kDBh5lyc6YvBnoS1hUa1t2O38z0xrWvaEj5XtJ8auw/edit#gid=2072493052
>
> Thanks
>
>
> Stratos Karafotis (2):
> cpufreq: Introduce new relation for freq selection
> cpufreq: ondemand: Eliminate the deadband effect in low loads
>
> drivers/cpufreq/cpufreq_ondemand.c | 11 +++++++----
> drivers/cpufreq/freq_table.c | 12 +++++++++++-
> include/linux/cpufreq.h | 1 +
> 3 files changed, 19 insertions(+), 5 deletions(-)
OK, I've queued up these two patches for 3.17. We'll see if anyone sees
any problems related to them. Thanks!
--
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.
On 07/23/2014 02:50 AM, Rafael J. Wysocki wrote:
> On Monday, June 30, 2014 07:59:32 PM Stratos Karafotis wrote:
>> Hi all,
>>
>> This patchset changes slightly the calculation of target frequency to
>> eliminate the deadband effect (explained in patch 2 changelog) that it
>> seems to slow down the CPU in low and medium loads.
>>
>> Patch 1 introduces a new relation (RELATION_C) for the next frequency
>> selection, which chooses the closest frequency to target.
>>
>> Patch 2 is the actual change to ondemand governor.
>>
>> You may find graphs with the 'deadband' effect and benchmark results:
>> https://docs.google.com/spreadsheets/d/16kDBh5lyc6YvBnoS1hUa1t2O38z0xrWvaEj5XtJ8auw/edit#gid=2072493052
>>
>> Thanks
>>
>>
>> Stratos Karafotis (2):
>> cpufreq: Introduce new relation for freq selection
>> cpufreq: ondemand: Eliminate the deadband effect in low loads
>>
>> drivers/cpufreq/cpufreq_ondemand.c | 11 +++++++----
>> drivers/cpufreq/freq_table.c | 12 +++++++++++-
>> include/linux/cpufreq.h | 1 +
>> 3 files changed, 19 insertions(+), 5 deletions(-)
>
> OK, I've queued up these two patches for 3.17. We'll see if anyone sees
> any problems related to them. Thanks!
>
Thank you!
Stratos