topic Hi in Software Archive

Intel Xeon Phi Per Core Frequency

CPati2 — Fri, 25 Aug 2017 01:34:23 GMT

Hi All,

I have following system:

Operating System: CentOS Linux 7 (Core)
Kernel: Linux 3.10.0-514.10.2.el7.x86_64
Architecture: x86-64 Intel(R) Xeon Phi(TM) CPU 7210 @ 1.30GHz

Is the supported min and max frequency is 1GHz and 1.5GHz respectively?

I also observed that it's possible to set frequency of each core individually. I validated this by setting frequency of few cores to 1GHz and rest to default. Does that mean Xeon Phi has per core frequency DVFS?

Thanks.

The target frequency can be

McCalpinJohn — Fri, 25 Aug 2017 20:14:36 GMT

The target frequency can be set on a per core basis, but I believe that you will find that all of the active cores always run at the same frequency.

Hi John,

CPati2 — Fri, 25 Aug 2017 22:26:33 GMT

Hi John,

Does that mean, in terms of architecture there is unified frequency control for all cores, not per core dynamic clock and voltage scaling (DCVS)?

One thing I have observed is that setting mix and max frequency doesn't seem to have any effect on the current frequency. Below is the log of cpu0 and it shows that current frequency is 1.2 GHz even though I set max and min to 1GHz. Any idea why this may be happening?

ls /sys/devices/system/cpu/cpu0/cpufreq/

affected_cpus cpuinfo_transition_latency scaling_governor
cpuinfo_cur_freq related_cpus scaling_max_freq
cpuinfo_max_freq scaling_available_governors scaling_min_freq
cpuinfo_min_freq scaling_driver scaling_setspeed

sudo cat /sys/devices/system/cpu/cpu0/cpufreq/*
0
1219257
1500000
1000000
4294967295
0
performance powersave
intel_pstate
performance
1000000
1000000

Thanks.

Hi

aazue — Fri, 25 Aug 2017 23:33:41 GMT

Hi
Probably you must use (on demand) instead (performance)
for dynamically frequency change between (min) and (max) i don't use this hardware but
I think that it will be unable to him to work in his max frequency all the time
I don't know if management of dynamically range frequency of each cores are same on your
hardware,read subject information about the functionality of (governor)
if dynamically frequency is relative with his temperature maybe, you can compile last
source (freeipmi-1.5.7) for investigate more far.
Regards

Hi,

CPati2 — Fri, 25 Aug 2017 23:37:00 GMT

Hi,

The problem is that Xeon Phi doesn't have ondemand because it uses intel-pstate driver and not cpufreq. I can revert to cpufreq, but I don't know the reason for default intel-pstate and whether it's architecture dependent for Xeon Phi.

Refrence this blog post.

Thanks.

Hi

aazue — Sat, 26 Aug 2017 01:57:01 GMT

Hi
he seems that you have (powersave) that is equivalent (ondemand)
Maybe if revert to cpufreq, you will evaluate exactly if this change is more effective
with your hardware used for answer your expectations.
The difference is very variable with different model hardware tested.
Regards

Hi,

CPati2 — Sat, 26 Aug 2017 19:20:29 GMT

Hi,

I think what I am asking has nothing to do with which governor is in use. It's hardware and architecture based. So, whatever Intel desigend at the circuit level will not change by a software.

Meaning, per core DVFS if supported has to be at the architecture level. I don't know why sysfs is giving ambiguous data.

Thanks.

Hi

aazue — Sun, 27 Aug 2017 04:27:25 GMT

Hi
I don't share your reasoning ,it's always the kernel and the system that have the hand on all the
parameter and this management is above to all other elements ,if your request it's not propitious
for him, she will be rejected.

About your comment;
(So, whatever Intel designed at the circuit level will not change by a software.)
it's false, he can be temporary disabled or modulated by the system,and fortunately than its the case.
Here , now since two years and half we work for reduce consumption and improve performance of the system
I can confirm to you that taking account only reference based technical characteristic of hardware is
great error, the system make as that he want or can make only.
The system is not exclusively reserved in his conception to the Intel Hardware specificity unique
even if his quality is currently probably the better..
After several hundred approach tested that give no result really significant we understand now
where to operate,currently We obtain now a very very well results that are exceptional.
This improve is effective with hardware INTEL,AMD,ARM and PPC.
Currently we search how to solve without required recompile all the system differently,,with all process
that are situated in his level tree , this task seem not easy to make simple to solve.
I have find a solution that could be an shortcut but i am not yet certain it's operate correct.

About your second remarks:
If your order not answer correctly the problem is on side shared libraries where some process are locked.
You can test six dozen of models chipsets Intel different he will not change anything to this problem.
we referring to governor for evaluate approximately the part that are busy by the processors, but it's
not really an parameter very reliable.(more exactly ,how many time required for he decrease max to min)
I am not even certain that dynamic voltage scaling able to answer instantaneously or even change
progressively synchronized.

Regards

Hi

aazue — Mon, 28 Aug 2017 23:01:09 GMT

I have read speedily an part his source..
I think you don't query sufficient resources for he could change value to 1.5 by himself and
return to 1.3 after the hot point of your task ended.
It's not exactly as that you want but you see already if change of frequency (turbo) works.
Even if frequency is low your hardware extremely power (as the truck instead the car )

About multi-cores (after having submit to friends)
Even when you have standalone regulator each core the cost for driving each individually
would give result catastrophic. It's extremely complex, several problem exist .
Fortunately ,processors is only a part between the complete consumption of the hardware.

Regards