Intel Continues To Demonstrate The Importance Of Software Optimizations: Clear Linux + Xeon Max Benchmarks

[brucethemoose]

Does anyone have actual experience of deploying Clear Linux on servers. We have large shared servers in an academic environment, used by students to run their research. These machines run Ubuntu and get upgraded every 2-4 years, using LTS (and overall do not require much maintenance).

Having more performance is enticing, but a rolling release doesn't seem ideal in that case. Although presumably servers are the target of Intel, not desktop. How well does it work? Is it a constant upgrade hell? Does it break often (including locally compiled software)? How is the availability of drivers (such as NVIDIA GPU drivers*) and software (e.g., pyTorch)? Anyone has given Clear Linux a try outside of benchmarks?

-----------
* No need to start a discussion of the evil ways of NVIDIA. Whether I/we like it or not, it remains the path of least resistance at this point in time to get AI/ML done, and thru purchasing department.

Clear Linux does not support Nvidia by default. There are some community scripts, and it might work OK if you set it up once and never update... But otherwise it can be a pain to maintain.

Other than that, I love the distro, but I dunno about general stability either. Clear Linux seems to be tragically underused out in the wild :/

Personally I use CachyOS (an Arch distro) on my desktop for ML performance. Arch/CachyOS Nvidia support is excellent, and the performance is excellent compared to other 3090 desktops running the same projects. But I tend to link python venvs to native packages using the `--system-site-packages` flag to get the most out of the optimized system packages.

[ms178]

Michael Speaking of Intel and software optimizations, could you ask your Intel contacts about the APO feature for 14th Gen CPUs and if we get to see it on Linux, too? As I understand it, it ties into Thread Director and scheduling. I wonder if that feature is even implementable or how it would integrate into the Linux scheduling.

[geearf]

Michael it'd be nice to see if with a few tweaks you could go from a much slower distro to almost Clear's level, the governor being one of them as you suggested in the first bench.
I think showing that Clear is faster is not that new anymore, but showing us how easy it'd be to (almost) match it would be great!

[Michael]

Michael it'd be nice to see if with a few tweaks you could go from a much slower distro to almost Clear's level, the governor being one of them as you suggested in the first bench.
I think showing that Clear is faster is not that new anymore, but showing us how easy it'd be to (almost) match it would be great!

I've already done that in a prior article showing how making changes to Ubuntu can come closer to Clear but still not matching it all the way across the board.

[geearf]

I've already done that in a prior article showing how making changes to Ubuntu can come closer to Clear but still not matching it all the way across the board.

Oh I am sorry I don't remember. Please share the link or the title so I can find it.
Thank you!

edit: actually, are you talking about this one https://www.phoronix.com/review/ubuntu-clear-tweaks ? Do you expect little to have change on Clear since? Thank you! Maybe next time you could try that against CachyOS since they have more aggressive build flags.

[Michael]

Oh I am sorry I don't remember. Please share the link or the title so I can find it.
Thank you!

edit: actually, are you talking about this one https://www.phoronix.com/review/ubuntu-clear-tweaks ? Do you expect little to have change on Clear since? Thank you! Maybe next time you could try that against CachyOS since they have more aggressive build flags.

https://www.phoronix.com/review/ubuntu1810-fast-clear

[geearf]

https://www.phoronix.com/review/ubuntu1810-fast-clear

Thank you!
The docker one is really close, while it may not win half of the tests, number wise it's good enough I'd say.
Having seen that, now I want to know what part of the compiled packages mattered, was it the autoFDO, the flags, etc...

[aaupov]

Michael

> While Intel hasn't had any big breakthroughs to announce around their in-house Clear Linux distribution recently

There has been an indication that ClearLinux developers are looking at BOLT or alternatives:

https://github.com/clearlinux/distribution/issues/2996

[Michael]

Michael

> While Intel hasn't had any big breakthroughs to announce around their in-house Clear Linux distribution recently

There has been an indication that ClearLinux developers are looking at BOLT or alternatives:

https://github.com/clearlinux/distribution/issues/2996

Thanks, checking out.