The Fastest Linux OS For AMD Ryzen Zen 3? It's Still Intel Clear Linux

[phoronix]

Phoronix: The Fastest Linux OS For AMD Ryzen Zen 3? It's Still Intel Clear Linux

As we have shown with prior AMD Ryzen and EPYC processors, the Linux distribution generally offering the fastest out-of-the-box performance is Intel's Clear Linux platform. Given there aren't many other distributions as aggressively optimizing their default package set and engaging in features like AutoFDO, PGO, and various out-of-tree patches in the name of modern Intel x86_64 Linux performance -- and in turn, AMD performance benefits as well -- Clear Linux really shines with modern hardware. Testing of the latest Clear Linux with a Ryzen 9 5900X continues to delivering promising performance compared to the likes of Fedora, openSUSE, Manjaro, Debian, and Ubuntu.

http://www.phoronix.com/vr.php?view=29791

[vladpetric]

Wow, this is amazing, thank you Michael!

A few months ago someone told me about issues when attempting qemu/KVM on top of Clear Linux / Ryzen (my current setup is a clear linux/kvm/haswell xeon).

Any more recent thoughts?

[OnikenX]

it would be could to see manjaro or arch with the zen kernel, or a comparation between kernels like xanmod, zen and stock

[OnikenX]

*cool

[f0rmat]

I know that this request may be a bridge too far, but given how well AMD does on an Intel optimized system, it might be interesting to see how AMD does on an AMD optimized Gentoo. But that is a lot of work.

[perpetually high]

All my homies run ClearLinux.

This is great stuff, thanks Michael. These are serious gains to be had. 15% is nothing to sneeze at for some compiler flags and kernel patches/tweaks.

I don't run ClearLinux but I do use their kernel patches and C/CXX flags with -march and my Haswell system flies on Ubuntu 20.04 LTS.

[Vistaus]

I know that this request may be a bridge too far, but given how well AMD does on an Intel optimized system, it might be interesting to see how AMD does on an AMD optimized Gentoo. But that is a lot of work.

It might also be interesting to see what AMD could do with Linux, i.e. a Clear Linux competitor from AMD.

[f0rmat]

It might also be interesting to see what AMD could do with Linux, i.e. a Clear Linux competitor from AMD.

That is a very interesting point..

bridgman What do you think?

[duby229]

I know that this request may be a bridge too far, but given how well AMD does on an Intel optimized system, it might be interesting to see how AMD does on an AMD optimized Gentoo. But that is a lot of work.

Clear linux does a whole lot more than just compiler flags tweaks. They make a bunch of modifications. It would be impossible to duplicate all that Clear linux does because quite a lot of it is (or at least was) hardcoded.

EDIT: I haven't followed it in quite some time though, so I don't know if they ever got to distributing patches for all the modifications they make? They didn't used to. Many of the modifications to non-GPL projects they just simply didn't make patches available. Like I said I don't know if they ever corrected that.

EDIT: Which is one of the reasons I think the GPL is a superior license. All of their kernel modifications are available and always have been because the GPL requires it. But to duplicate Clear linux in full at least used to be impossible.

[Zan Lynx]

Some of the biggest things and hardest to replicate are the automatic flag detection and the PGO.

Flag detection and bringing in the optimized code on demand is required to have a distro that boots on everything. Otherwise, like some people's Gentoo builds, you can only run it on specific system types.

PGO requires two compile runs. But the hardest part is building a useful profile. It has to exercise the code paths you want to optimize. For example, just opening Firefox and loading a web page does what? Things specific to that one page. You need to run the CSS transistion code, various JS compile types, WebGL, video playback, and who knows what else. Many things that are easy to forget.

For PGO libraries with test suites it is easy to run the tests to build the profile. Problem is, and what I ran into with a C++ library of mine, the test sets tend to exercise uncommon and error path code more than mainline. This gives your PGO profile an unrealistic view of how the library is used. So you need to build a test subset that builds an accurate profile of the common success paths.

Etc.