Hardware Prefetch for Atom E-Cores

[cyring]

In this Intel whitepaper 357930-001US we are reading new MSR dedicated to Atom sub-architecture within hybrid processors.

Feature	Register
LLC Streamer	0x1320
L2 NLP	0x1321
SELECTION	0x1323

Whereas the known MSR 0x1A4 is said to be available for both P-Cores and E-Cores

---

I'll appreciate if one can rdmsr the new MSR on E-Core CPU numbers.

Next, still on E-Cores, toggle the Enable bit using a wrmsr ; rdmsr modification sequence.

It would be also interesting to check the specs:

• try a read access of those Atom MSR on P-Cores
• verify the scope of writing: MSR should be module scope: flipping a MSR bit on an unique E-Core should be reflected on the others. But what is the module cluster formed with ?

---

Thanks for helping.

[cyring]

contributors

@gundami @Technologicat @justanerd @BugReporterZ @svmlegacy @kocoman1 @jowa2021 @Betaminos @huajian628 @vitaly-zdanevich @rushvora

Can your guys please help to investigate those registers on Hybrid architecture ?

[BugReporterZ]

I tried this.

#rdmsr -a 0x1321
CPU 19: 250122000001
CPU 18: 250122000001
CPU 17: 250122000001
CPU 16: 250122000001
rdmsr: CPU 15 cannot read MSR 0x00001321

#rdmsr -a 0x1323
CPU 19: 1f9cc00000000
CPU 18: 1f9cc00000000
CPU 17: 1f9cc00000000
CPU 16: 1f9cc00000000
rdmsr: CPU 15 cannot read MSR 0x00001323

===============================================

#rdmsr -a 0x1320
CPU 19: 100007e041000004
CPU 18: 100007e041000004
CPU 17: 100007e041000004
CPU 16: 100007e041000004
rdmsr: CPU 15 cannot read MSR 0x00001320

#wrmsr -p 19 0x1320 4
#rdmsr -a -x -0 0x1320
CPU 19: 0000000000000004
CPU 18: 0000000000000004
CPU 17: 0000000000000004
CPU 16: 0000000000000004
rdmsr: CPU 15 cannot read MSR 0x00001320

#wrmsr -p 19 0x1320 1152930164351434756
#rdmsr -a -x -0 0x1320
CPU 19: 100007e041000004
CPU 18: 100007e041000004
CPU 17: 100007e041000004
CPU 16: 100007e041000004
rdmsr: CPU 15 cannot read MSR 0x00001320

[cyring]

I tried this.

Excellent. I can code now. Thank you

[cyring]

• ROG STRIX Z790 Series BIOS Manual

Hardware Prefetcher
Allows you to enable or disable the MLC streamer prefetcher.
Configuration options: [Disabled] [Enabled]

Adjacent Cache Line Prefetch
Allows you to prefetch adjacent cache lines, reducing the DRAM loading time and improving 
the system performance.
Configuration options: [Disabled] [Enabled]

• SUPERMICRO C9Z790 USER MANUAL

  
  Hardware Prefetcher
  If set to Enabled, the hardware prefetcher will prefetch streams of data and instructions 
  from the main memory to the L2 cache to improve CPU performance. The options are 
  Disabled and Enabled.

Adjacent Cache Line Prefetch Select Enabled for the CPU to prefetch both cache lines for 128 bytes as comprised. Select Disabled for the CPU to prefetch both cache lines for 64 bytes. The options are Disabled and Enabled.

* ASRock Intel Z790 Motherboard
`<same as above>`

[cyring]

• MS03-6L0

  
  Enable LP [Global]
  Enables Logical processor (Software Method to Enable/Disable Logical Processor threads).
  Options available: ALL LPs, Single LP. Default setting is ALL LPs.

Hardware Prefetcher Select whether to enable the speculative prefetch unit of the processor. Options available: Enable, Disable. Default setting is Enable. L2 RF0 Prefetch Disable Options available: Enable, Disable. Default setting is Disable.

Adjacent Cache Prefetch When enabled, cache lines are fetched in pairs. When disabled, only the required cache line is fetched. Options available: Enable, Disable. Default setting is Enable.

DCU Streamer Prefetcher Enable/Disable DCU streamer prefetcher. Options available: Enable, Disable. Default setting is Enable.

DCU IP Prefetcher Enable/Disable DCU IP Prefetcher. Options available: Enable, Disable. Default setting is Enable.

* [BIOS for Dell PowerEdge Servers](https://infohub.delltechnologies.com/p/bios-settings-for-optimized-performance-on-next-generation-dell-poweredge-servers/)
* [Lenovo Processors UEFI_XEON_1ST_2ND](https://pubs.lenovo.com/uefi_xeon_1st_2nd/processors.html)
* [`HPC Cluster Tuning on
 3rd Generation Intel® Xeon`](https://www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=https://cdrdv2-public.intel.com/686419/HPC-Cluster-Tuning-Guide-on-3rd-Generation-Intel-Xeon-Scalable-Processors.pdf&ved=2ahUKEwi414W1upmDAxV4TaQEHWGVDqUQFnoECB8QAQ&usg=AOvVaw0rTxexBbnkrILDZM1PDzjb)
* [firmware-settings](https://github.com/microsoft/MSLab/blob/master/Docs/AzureStackHCI/02-AzSHCI-Deployment/01-hardware-configuration.md#firmware-settings)
* [Supermicro BIOS settings](https://github.com/cncf/cnf-testbed/issues/129#issuecomment-438191639)

[cyring]

The Intel Atom cores are placed in a group of four per module with private L1 caches for each core.

Each core has a set of L1 hardware prefetchers

The L2 prefetch block is shared for all cores in the module

trackers are shared between the cores

DCU – Data Caching Unit is the block that holds the L1 data cache.

[cyring]

+ @BugReporterZ : Hello,

Above is the last commit 0a0daea556fef89c6ff0249101cc51f794ab6e4c which adds DCU L1 NLP, a bit of MSR 0x1A4 unveiled in the Whitepaper.

---

Since commit cc5c3270c0ef699047b7e6d40dfe025fd71492c1 are added for E-Cores:

• L1 NPP
• L2 NLP
• LLC Streamer

Remark: Screenshots above are made from virtualization

[cyring]

In "Hardware LLC prefetch feature on 4th Gen Intel® Xeon® Scalable Processor (Codename Sapphire Rapids)" we are reading L3 prefetch disable at bit 42 of MSR 0x6D

Does it work on Desktop or Mobile processors ?

[cyring]

Pre-release in progress https://github.com/cyring/CoreFreq/discussions/472 If things are missing, please let me know.