CachyOS Optimizing More Packages With PGO For Up To ~10% Better Performance

[phoronix]

Phoronix: CachyOS Begins Optimizing More Packages With PGO For Up To ~10% Better Performance

The Arch Linux based CachyOS operating system that is known for pursuing aggressive performance while still delivering a nice Linux desktop experience is out with its September 2024 release...

https://www.phoronix.com/news/CachyOS-September-2024

[zamazan4ik]

Great to see that more and more engineers are trying to optimize software with PGO. I hope one day this optimization also will be used for more packages in more "regular" distros.

[edxposed]

A 10% improvement is not very high for clang. A really highly optimized clang can usually be 50% faster than distro clang and 100% faster than distro gcc.

[intelfx]

A 10% improvement is not very high for clang. A really highly optimized clang can usually be 50% faster than distro clang and 100% faster than distro gcc.

Do you have anything more substantial on that than an anecdotal claim?

[V1tol]

A 10% improvement is not very high for clang. A really highly optimized clang can usually be 50% faster than distro clang and 100% faster than distro gcc.

I hope you are not working in IT anywhere.

[ahrs]

CachyOS has also been transitioning from Zlib to Zlib-ng

I wonder if they have a proper fix for Chromium and QtWebengine? I made this switch on my Gentoo system after seeing Fedora do so but Chromium and QtWebengine have build failures now due to their minizip usage (minizip-ng isn't completely compatible with minizip).

Fedora used bundled/vendored libs to get around this but Gentoo won't do that, they think it's an upstream issue with minizip-ng (and to be fair they're right). On the bright side this made me uninstall a bunch of crap and now I run Chromium or anything that needs WebEngine in Flatpaks.

[ptr1337]

A 10% improvement is not very high for clang. A really highly optimized clang can usually be 50% faster than distro clang and 100% faster than distro gcc.

Yes, we know, but for a Dynamic/Shared compiled Clang this is still quite good. We are already happy to optimize that much the dynamic clang.
Here you can find the Results of the Dynamic Clang:
https://github.com/CachyOS/CachyOS-PKGBUILDS/pull/329

Here with a static build llvm/clang:
https://github.com/ptr1337/llvm-bolt...ter/results.md

The difference is massive, but it is what it is. We follow archlinux with the packaging and that will stay like this.

Thanks for posting Michael

[ptr1337]

I wonder if they have a proper fix for Chromium and QtWebengine? I made this switch on my Gentoo system after seeing Fedora do so but Chromium and QtWebengine have build failures now due to their minizip usage (minizip-ng isn't completely compatible with minizip).

Fedora used bundled/vendored libs to get around this but Gentoo won't do that, they think it's an upstream issue with minizip-ng (and to be fair they're right). On the bright side this made me uninstall a bunch of crap and now I run Chromium or anything that needs WebEngine in Flatpaks.

We have currently just replaced with the zlib-ng-compat layer, but on Archlinux we will push likely in the near future also zlib-ng as replacement and then supported packages will use the native zlib-ng.

[ptr1337]

Great to see that more and more engineers are trying to optimize software with PGO. I hope one day this optimization also will be used for more packages in more "regular" distros.

Thanks for your collection related to PGO. We have looked quite much into this and took also some workloads for you, to create these PKGBUILDs.
Really helped us much and we want to work on more packages for this.

Many distributions do not like to use PGO, because it does not work well with reproduciblity. The only way would be, to "precreate" PGO profiles, upload them into the source directory and after that use htem

[edxposed]

Yes, we know, but for a Dynamic/Shared compiled Clang this is still quite good. We are already happy to optimize that much the dynamic clang.
Here you can find the Results of the Dynamic Clang:
https://github.com/CachyOS/CachyOS-PKGBUILDS/pull/329

Here with a static build llvm/clang:
https://github.com/ptr1337/llvm-bolt...ter/results.md

The difference is massive, but it is what it is. We follow archlinux with the packaging and that will stay like this.

Thanks for posting Michael

This would be a shame, as a reliable fast compiler would be valuable for CI use cases. We currently have to maintain our own clang builds.