About crystal frequency slow down in Xeon (non skylake)

[templex]

I saw a kernel path:  Fix erroneous TSC rate on Skylake Xeon  

In this patch, it mentioned that crystal frequency was a bit slower than expect on Skylake Xeon.

I wonder is this issue only would happen on Skylake Xeon or not?

Because on Cascade Lake Xeon, I haven't find the slow down.

I'm using TSC as the source of timestamp, so it's real important for me to figure it out.

Thanks!

---------

Additional Information:

My question actually comes from <Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3> 18.7.3 Determining the Processor Base Frequency, there is a Table 18-85. Nominal Core Crystal Clock Frequency

Processor Families/Processor Number Series 1	Nominal CoreCrystalClockFrequency
Intel® Xeon® Processor Scalable Family with CPUID signature 06_55H.	25 MHz
6th and 7th generation Intel® Core™ processors and Intel® Xeon® W ProcessorFamily.	24 MHz
Next Generation Intel® Atom™ processors based on Goldmont Microarchitecturewith CPUID signature 06_5CH (does notinclude Intel Xeon processors).	19.2 MHz

Xeon Scalable has CPUID signature 06_55H, but they may have different microarchitecture, e.g. Skylake, Cascade Lake. 

For Skylake Xeon, there is comment in kernel patch:  

All SKX subject the crystal to an EMI reduction circuit that
reduces its actual frequency by (approximately) -0.25%

I wonder is there the same issue in further microarchitecture? If there is, which means the  Nominal Core Crystal Clock Frequency of Xeon Scalable in manual is not accurate. I can't use it as timestamp source.

[Emeth_O_Intel]

Hello templex,

Thank you for contacting Intel Xeon Community Support.

I was reviewing your thread and this issue occurred on the kernel version 4.9 with some Intel Kabylake and SkylakeX CPUs Models due to the fact the parameters added in the kernel path (code arch/x86/kernel/tsc.c) affects the TSC frequency. However, the developers already fix the inconvenience.

      switch (boot_cpu_data.x86_model) {

      case INTEL_FAM6_SKYLAKE_MOBILE:

      case INTEL_FAM6_SKYLAKE_DESKTOP:

      case INTEL_FAM6_KABYLAKE_MOBILE:

      case INTEL_FAM6_KABYLAKE_DESKTOP:

         crystal_khz = 24000;   /* 24.0 MHz */

         break;

      case INTEL_FAM6_SKYLAKE_X:

         crystal_khz = 25000;   /* 25.0 MHz */

         break;

      case INTEL_FAM6_ATOM_GOLDMONT:

         crystal_khz = 19200;   /* 19.2 MHz */

         break;

So, far we have not seen any unusual activity related to the Intel Cascade Lake Xeon Models regarding TSC frequency.

Of course, it is very important to have the kernel up to date if possible in order to have the most recent updates and fixes on the systems.

References:

https://github.com/torvalds/linux/blob/16f73eb02d7e1765ccab3d2018e0bd98eb93d973/arch/x86/kernel/tsc.c

https://source.puri.sm/dorota.czaplejewicz/linux-emcraft/-/commit/6baf3d61821f5b38f27b4e9f044ad4d1e8f3d14f?view=inline&w=1

https://lore.kernel.org/patchwork/patch/875647/

Emeth O.

Intel® Server Specialist

[templex]

Thank you for your details explanation about the TSC frequency in kernel.

My question actually comes from <Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3> 18.7.3 Determining the Processor Base Frequency, there is a Table 18-85. Nominal Core Crystal Clock Frequency

Processor Families/Processor Number Series 1	Nominal Core Crystal ClockFrequency
Intel® Xeon® Processor Scalable Family with CPUID signature 06_55H.	25 MHz
6th and 7th generation Intel® Core™ processors and Intel® Xeon® W Processor Family.	24 MHz
Next Generation Intel® Atom™ processors based on Goldmont Microarchitecture with CPUID signature 06_5CH (does notinclude Intel Xeon processors).	19.2 MHz

Xeon Scalable has CPUID signature 06_55H, but they may have different microarchitecture, e.g. Skylake, Cascade Lake. 

For Skylake Xeon, there is comment in kernel patch:  

All SKX subject the crystal to an EMI reduction circuit that
reduces its actual frequency by (approximately) -0.25%

I wonder is there the same issue in further microarchitecture? If there is, which means the  Nominal Core Crystal Clock Frequency of Xeon Scalable in manual is not accurate. I can't use it as timestamp source.

[Emeth_O_Intel]

Hello Templex,

Thank you for your prompt reply.

Our best recommendation will be to create a new question in the following forum about Software Community: https://community.intel.com/t5/Software-Tuning-Performance/bd-p/software-tuning-perf-optimization.

This matter should be handled by a different team, and I am very sure they will be more than happy to assist you and provide you the most accurate information.

Have a wonderful day.

Regards,

Emeth O.

Intel Server Specialist.