Xeon Phi - Power And Clock Gating

[CPati2]

Hi All,

Does Xeon Phi Kights Landing has power gating? I believe it doesn't have clock gating. But, wondering if change number of active/online cores will add to power?

As per my experiments, running a workload on specific core while turning offline cores online doesn't add to total power consumption?

Thanks.

[McCalpinJohn]

Linux will automatically put idle cores into low-power states.  This is the job of the "cpuidle" subsystem.  In recent versions of Linux, the requested idle state is a function of how busy the system is and how long it will be before the next high-resolution timer expiration on each core.   With the default Linux configuration, on a system that has been idle for more than a few seconds, most of the cores should drop into the lowest power state (Core C6).

I don't see any reason to expect that cores that are taken offline in the BIOS would use less power than cores that are in the Core C6 state.

[CPati2]

Hi John,

Thank you.

In Linux, it's possible to put cores "offline" by writing either 1 (ON) or 0 (OFF) to sysfs: /sys/devices/system/cpu/cpuX/online (X = 0 to 255).

With Intel P-State driver I am not able to change frequency knobs to see effect of power consumption based on cores + frequency. I was wondering if putting cores to offline mode using above method really turns core off physically (it seems it will go to lowest power state)?

By this I can understand the effect of active cores on total power consumption.

Thanks.

[McCalpinJohn]

Core that are not running anything will not be active.  As I said above, I don't know of any mechanism that the OS could use to lower the power of a core other than putting in C6 state, so I don't expect any power differences between cores that are idle and cores that are "offline".

If you want to test the effect of active cores on power consumption, all you need to do is run on different numbers of cores.  The OS will ensure that unused cores are put in low-power states.  The OS makes the request and the hardware decides exactly what to do.

Be sure to pay attention to the tile location of the active and idle cores.  The hardware may be able to save more energy (i.e., turn more stuff off) if two cores are idle in one tile than if one core is inactive in each of two tiles.

If you want to study the effect of frequency on the power consumption of active cores, you should boot with "intel_pstate=disable" and use the "acpi_cpufreq" driver to control frequency.  My measurements suggest that KNL has only one frequency domain -- all active cores will run at exactly the same frequency.   Inactive cores will have their core clocks disabled, so the frequency does not matter (except for the cache, which remains active in most core C states).  I have not been able to find more detailed public documentation on the C states in Knights Landing.

The PKG RAPL domain in KNL also includes the MCDRAM.  MCDRAM power consumption will vary under load (of course), but may also vary when not being used, depending on what power-savings functionality the MCDRAM supports and whether the hardware decides to use it.   (I have not looked at this specifically for MCDRAM, but for DDR4, the power consumption varies by quite a bit on "idle" systems as the hardware enables and disables various power-saving modes.)