CentOS Stream & Clear Linux Achieve Greater Performance On 4th Gen Xeon Scalable Sapphire Rapids, EPYC Genoa

[phoronix]

Phoronix: CentOS Stream & Clear Linux Achieve Greater Performance On 4th Gen Xeon Scalable Sapphire Rapids, EPYC Genoa

As part of other ongoing performance tests of Intel 4th Gen Xeon Scalable "Sapphire Rapids" testing, I was curious to see how the more well-tuned Linux distributions are performing with the flagship Xeon Platinum 8490H processors relative to the common Ubuntu 22.04 LTS release. Here are those benchmark numbers alongside AMD's flagship Genoa server platform with two EPYC 9654 processors.

https://www.phoronix.com/review/centos-clear-spr

[dkokron]

What is the BIOS "Determinism" setting for these experiments (Performance vs Power)?

[Michael]

What is the BIOS "Determinism" setting for these experiments (Performance vs Power)?

You can always assume the BIOS defaults unless otherwise noted in the article, so performance.

[dkokron]

I have measured significant (~5%) performance improvement for HPL and another HPC application with the determinism set to Power. Have you done any testing with this setting?

[dkokron]

Disregard last question. I found what I need here https://www.phoronix.com/review/amd-...hmarks#page-10

[Linuxxx]

Thanks for the excellent article, Michael!

Appreciate the inclusion of the performance governor results for Ubuntu, which allows it to match CentOS Stream.

The remaining difference can be attributed to CentOS using newer software, always enabling transparent hugepages & not shipping "irqbalance" by default.

Theoretically, this should settle the debate whether Ubuntu is objectively a bloated & slow OS or not, but given the "special" individuals among the audience of Phoronix, I already know falsifications like this will continue, no matter the facts...

[jorgepl]

Thanks for the excellent article, Michael!

Appreciate the inclusion of the performance governor results for Ubuntu, which allows it to match CentOS Stream.

The remaining difference can be attributed to CentOS using newer software, always enabling transparent hugepages & not shipping "irqbalance" by default.

Theoretically, this should settle the debate whether Ubuntu is objectively a bloated & slow OS or not, but given the "special" individuals among the audience of Phoronix, I already know falsifications like this will continue, no matter the facts...

This doesn't settle anything. You cannot test whether a system is bloated or not on a cutting edge Xeon processor and 1TB of RAM memory.

OBVIOUSLY Ubuntu is not bloated there. How could it? Even Windows 11 would run amazingly well there.

The bloat is felt on a Celeron processor and 4GB of ram. There you can tell. And with snaps, you'll soon realize Ubuntu is bloated (because each snap consumes just by the virtue of being installed -not executed- around 4mb according to my own tests on Virtual Machine, and according to some comments in Launchpad by Canonical engineers).

[sabian2008]

I have measured significant (~5%) performance improvement for HPL and another HPC application with the determinism set to Power. Have you done any testing with this setting?

What processor? On old Intels I wouldn't be surprised. HPL would probably want to use AVX512 and with the performance governor it would hit thermal throttling more easily.

I'm happy Intel seems to have finally pulled off a decent AVX512 implementation.

[dkokron]

What processor? On old Intels I wouldn't be surprised. HPL would probably want to use AVX512 and with the performance governor it would hit thermal throttling more easily.

I'm happy Intel seems to have finally pulled off a decent AVX512 implementation.

Dual socket EPYC 7402

[darkbasic]

I think that a 20% average performance improvement can no longer be ignored by other distros.