[PATCH] cpufreq: CPPC: Resolve the large frequency discrepancy from cpuinfo_cur_freq
lihuisong (C)
lihuisong at huawei.com
Tue Jan 9 23:09:48 PST 2024
Hi Ionela,
在 2024/1/8 22:03, Ionela Voinescu 写道:
> Hi,
>
> On Friday 05 Jan 2024 at 15:04:47 (+0800), lihuisong (C) wrote:
>> Hi Vanshi,
>>
>> 在 2024/1/5 8:48, Vanshidhar Konda 写道:
>>> On Thu, Jan 04, 2024 at 05:36:51PM +0800, lihuisong (C) wrote:
>>>> 在 2024/1/4 1:53, Ionela Voinescu 写道:
>>>>> Hi,
>>>>>
>>>>> On Tuesday 12 Dec 2023 at 15:26:17 (+0800), Huisong Li wrote:
>>>>>> Many developers found that the cpu current frequency is greater than
>>>>>> the maximum frequency of the platform, please see [1], [2] and [3].
>>>>>>
>>>>>> In the scenarios with high memory access pressure, the patch [1] has
>>>>>> proved the significant latency of cpc_read() which is used to obtain
>>>>>> delivered and reference performance counter cause an absurd frequency.
>>>>>> The sampling interval for this counters is very critical and
>>>>>> is expected
>>>>>> to be equal. However, the different latency of cpc_read() has a direct
>>>>>> impact on their sampling interval.
>>>>>>
>>>>> Would this [1] alternative solution work for you?
>>>> It would work for me AFAICS.
>>>> Because the "arch_freq_scale" is also from AMU core and constant
>>>> counter, and read together.
>>>> But, from their discuss line, it seems that there are some tricky
>>>> points to clarify or consider.
>>> I think the changes in [1] would work better when CPUs may be idle. With
>>> this
>>> patch we would have to wake any core that is in idle state to read the
>>> AMU
>>> counters. Worst case, if core 0 is trying to read the CPU frequency of
>>> all
>>> cores, it may need to wake up all the other cores to read the AMU
>>> counters.
>> From the approach in [1], if all CPUs (one or more cores) under one policy
>> are idle, they still cannot be obtained the CPU frequency, right?
>> In this case, the [1] API will return 0 and have to back to call
>> cpufreq_driver->get() for cpuinfo_cur_freq.
>> Then we still need to face the issue this patch mentioned.
> With the implementation at [1], arch_freq_get_on_cpu() will not return 0
> for idle CPUs and the get() callback will not be called to wake up the
> CPUs.
Right, arch_freq_get_on_cpu() will not return 0 for idle CPUs.
However, for no-housekeeping CPUs, it will return 0 and have to call
get() callback, right?
>
> Worst case, arch_freq_get_on_cpu() will return a frequency based on the
> AMU counter values obtained on the last tick on that CPU. But if that CPU
> is not a housekeeping CPU, a housekeeping CPU in the same policy will be
> selected, as it would have had a more recent tick, and therefore a more
> recent frequency value for the domain.
But this frequency is from the last tick,
this last tick is probably a long time ago and it doesn't update
'arch_freq_scale' for some reasons like CPU dile.
In addition, I'm not sure if there is possible that
amu_scale_freq_tick() is executed delayed under high stress case.
It also have an impact on the accuracy of the cpu frequency we query.
>
> I understand that the frequency returned here will not be up to date,
> but there's no proper frequency feedback for an idle CPU. If one only
> wakes up a CPU to sample counters, before the CPU goes back to sleep,
> the obtained frequency feedback is meaningless.
>
>>> For systems with 128 cores or more, this could be very expensive and
>>> happen
>>> very frequently.
>>>
>>> AFAICS, the approach in [1] would avoid this cost.
>> But the CPU frequency is just an average value for the last tick period
>> instead of the current one the CPU actually runs at.
>> In addition, there are some conditions to use 'arch_freq_scale' in this
>> approach.
> What are the conditions you are referring to?
It depends on the housekeeping CPUs.
>
>> So I'm not sure if this approach can entirely cover the frequency
>> discrepancy issue.
> Unfortunately there is no perfect frequency feedback. By the time you
> observe/use the value of scaling_cur_freq/cpuinfo_cur_freq, the frequency
> of the CPU might have already changed. Therefore, an average value might
> be a better indication of the recent performance level of a CPU.
An average value for CPU frequency is ok. It may be better if it has not
any delaying.
The original implementation for cpuinfo_cur_freq can more reflect their
meaning in the user-guide [1]. The user-guide said:
"cpuinfo_cur_freq : Current frequency of the CPU as obtained from the
hardware, in KHz.
This is the frequency the CPU actually runs at."
[1]https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
>
> Would you be able to test [1] on your platform and usecase?
I has tested it on my platform (CPU number: 64, SMT: off and CPU base
frequency: 2.7GHz).
Accoding to the testing result,
1> I found that patch [1] and [2] cannot cover the no housekeeping CPUs.
They still have to face the large frequency discrepancy issue my patch
mentioned.
2> Additionally, the frequency value of all CPUs are almost the same by
using the 'arch_freq_scale' factor way. I'm not sure if it is ok.
The patch [1] has been modified silightly as below:
-->
@@ -1756,7 +1756,10 @@ static unsigned int
cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
{
unsigned int new_freq;
- new_freq = cpufreq_driver->get(policy->cpu);
+ new_freq = arch_freq_get_on_cpu(policy->cpu);
+ if (!new_freq)
+ new_freq = cpufreq_driver->get(policy->cpu);
+
if (!new_freq)
return 0;
And the result is as follows:
*case 1:**No setting the nohz_full and cpufreq use performance governor*
*--> Step1: *read 'cpuinfo_cur_freq' in no pressure
0: 2699264 2: 2699264 4: 2699264 6: 2699264
8: 2696628 10: 2696628 12: 2696628 14: 2699264
16: 2699264 18: 2696628 20: 2699264 22: 2696628
24: 2699264 26: 2696628 28: 2699264 30: 2696628
32: 2696628 34: 2696628 36: 2696628 38: 2696628
40: 2699264 42: 2699264 44: 2696628 46: 2696628
48: 2696628 50: 2699264 52: 2699264 54: 2696628
56: 2696628 58: 2696628 60: 2696628 62: 2696628
64: 2696628 66: 2699264 68: 2696628 70: 2696628
72: 2699264 74: 2696628 76: 2696628 78: 2699264
80: 2696628 82: 2696628 84: 2699264 86: 2696628
88: 2696628 90: 2696628 92: 2696628 94: 2699264
96: 2696628 98: 2699264 100: 2699264 102: 2696628
104: 2699264 106: 2699264 108: 2699264 110: 2696628
112: 2699264 114: 2699264 116: 2699264 118: 2699264
120: 2696628 122: 2699264 124: 2696628 126: 2699264
Note: the frequency of all CPUs are almost the same.
*--> Step 2: *read 'cpuinfo_cur_freq' in the high memory access pressure.
0: 2696628 2: 2696628 4: 2696628 6: 2696628
8: 2696628 10: 2696628 12: 2696628 14: 2696628
16: 2696628 18: 2696628 20: 2696628 22: 2696628
24: 2696628 26: 2696628 28: 2696628 30: 2696628
32: 2696628 34: 2696628 36: 2696628 38: 2696628
40: 2696628 42: 2696628 44: 2696628 46: 2696628
48: 2696628 50: 2696628 52: 2696628 54: 2696628
56: 2696628 58: 2696628 60: 2696628 62: 2696628
64: 2696628 66: 2696628 68: 2696628 70: 2696628
72: 2696628 74: 2696628 76: 2696628 78: 2696628
80: 2696628 82: 2696628 84: 2696628 86: 2696628
88: 2696628 90: 2696628 92: 2696628 94: 2696628
96: 2696628 98: 2696628 100: 2696628 102: 2696628
104: 2696628 106: 2696628 108: 2696628 110: 2696628
112: 2696628 114: 2696628 116: 2696628 118: 2696628
120: 2696628 122: 2696628 124: 2696628 126: 2696628
*Case 2: setting nohz_full and cpufreq use ondemand governor*
There is "isolcpus=1-10,41-50 nohz_full=1-10,41-50 rcu_nocbs=1-10,41-50"
in /proc/cmdline.
*--> Step 1: *setting ondemand governor to all policy and query
'cpuinfo_cur_freq' in no pressure case.
And the frequency of CPUs all are about 400MHz.
*--> Step 2:* read 'cpuinfo_cur_freq' in the high memory access pressure.
The high memory access pressure is from the command: "stress-ng -c 64
--cpu-load 100% --taskset 0-63"
The result:
0: 2696628 1: 400000 2: 400000 3: 400909
4: 400000 5: 400000 6: 400000 7: 400000
8: 400000 9: 400000 10: 400600 11: 2696628
12: 2696628 13: 2696628 14: 2696628 15: 2696628
16: 2696628 17: 2696628 18: 2696628 19: 2696628
20: 2696628 21: 2696628 22: 2696628 23: 2696628
24: 2696628 25: 2696628 26: 2696628 27: 2696628
28: 2696628 29: 2696628 30: 2696628 31: 2696628
32: 2696628 33: 2696628 34: 2696628 35: 2696628
36: 2696628 37: 2696628 38: 2696628 39: 2696628
40: 2696628 41: 400000 42: 400000 43: 400000
44: 400000 45: 398847 46: 400000 47: 400000
48: 400000 49: 400000 50: 400000 51: 2696628
52: 2696628 53: 2696628 54: 2696628 55: 2696628
56: 2696628 57: 2696628 58: 2696628 59: 2696628
60: 2696628 61: 2696628 62: 2696628 63: 2699264
Note:
(1) The frequency of 1-10 and 41-50 CPUs work on the lowest frequency.
It turned out that nohz full was already work.
I guess that stress-ng cannot use the CPU in the range of nohz full.
Because the CPU frequency will be increased to 2.7G by binding CPU
to other application.
(2) The frequency of the nohz full core is calculated by get() callback
according to ftrace.
[1] https://lore.kernel.org/lkml/20230418113459.12860-7-sumitg@nvidia.com/
[2]
https://lore.kernel.org/lkml/20231127160838.1403404-3-beata.michalska@arm.com/
>
> Many thanks,
> Ionela.
>
>> /Huisong
>>
>>>>> [1] https://lore.kernel.org/lkml/20231127160838.1403404-1-beata.michalska@arm.com/
>>>>>
>>>>> Thanks,
>>>>> Ionela.
>>>>>
>>>>>> This patch adds a interface, cpc_read_arch_counters_on_cpu, to read
>>>>>> delivered and reference performance counter together. According to my
>>>>>> test[4], the discrepancy of cpu current frequency in the
>>>>>> scenarios with
>>>>>> high memory access pressure is lower than 0.2% by stress-ng
>>>>>> application.
>>>>>>
>>>>>> [1] https://lore.kernel.org/all/20231025093847.3740104-4-zengheng4@huawei.com/
>>>>>> [2] https://lore.kernel.org/all/20230328193846.8757-1-yang@os.amperecomputing.com/
>>>>>> [3]
>>>>>> https://lore.kernel.org/all/20230418113459.12860-7-sumitg@nvidia.com/
>>>>>>
>>>>>> [4] My local test:
>>>>>> The testing platform enable SMT and include 128 logical CPU in total,
>>>>>> and CPU base frequency is 2.7GHz. Reading "cpuinfo_cur_freq" for each
>>>>>> physical core on platform during the high memory access pressure from
>>>>>> stress-ng, and the output is as follows:
>>>>>> 0: 2699133 2: 2699942 4: 2698189 6: 2704347
>>>>>> 8: 2704009 10: 2696277 12: 2702016 14: 2701388
>>>>>> 16: 2700358 18: 2696741 20: 2700091 22: 2700122
>>>>>> 24: 2701713 26: 2702025 28: 2699816 30: 2700121
>>>>>> 32: 2700000 34: 2699788 36: 2698884 38: 2699109
>>>>>> 40: 2704494 42: 2698350 44: 2699997 46: 2701023
>>>>>> 48: 2703448 50: 2699501 52: 2700000 54: 2699999
>>>>>> 56: 2702645 58: 2696923 60: 2697718 62: 2700547
>>>>>> 64: 2700313 66: 2700000 68: 2699904 70: 2699259
>>>>>> 72: 2699511 74: 2700644 76: 2702201 78: 2700000
>>>>>> 80: 2700776 82: 2700364 84: 2702674 86: 2700255
>>>>>> 88: 2699886 90: 2700359 92: 2699662 94: 2696188
>>>>>> 96: 2705454 98: 2699260 100: 2701097 102: 2699630
>>>>>> 104: 2700463 106: 2698408 108: 2697766 110: 2701181
>>>>>> 112: 2699166 114: 2701804 116: 2701907 118: 2701973
>>>>>> 120: 2699584 122: 2700474 124: 2700768 126: 2701963
>>>>>>
>>>>>> Signed-off-by: Huisong Li <lihuisong at huawei.com>
>>>>>> ---
>>>>>>
[snip]
> .
More information about the linux-arm-kernel
mailing list