Benchmarks Of Arch Linux's Zen Kernel Flavor

[creative]

Gentoo Ricers FTW!!!!111

eh... no

[svenh]

@Michael: I like the comparison graphs for 2 setups. I would suggest a short first line that explains what follows (as equivalent to "More is better") because I always have to think twice what the graph bars mean. The worst result of the two is the baseline, and the percentage is the relative disadvantage of the setup (in the current column), i.e. the empty column setup is N % better than the column with the bar?

[Weasel]

Compiling the kernel with -O4 like many Gentoo users have been doing for years would have surely given them better results.

Oh, wait, -O4 doesn't exist and works as -O3 and -O3 in its turn often doesn't give any benefits but makes your binaries a lot bigger.

Well, time to dig into `man gcc` and find something no one has been using before. -ffast-math for the win! Oh wait, it's unsafe to use.

Exercises with uber-optimizations will probably never cease. Maybe 'cause people don't write compilers while believing there's some hidden untapped potential only for them.

-ffast-math can make a huge difference. And it's "unsafe" only if the program is really designed around exact IEEE fp semantics.

99% of programmers use floats just to calculate "real number" stuff, they don't even know what a exponent/significand is, so it's mostly safe.

For the kernel though, it won't make a difference since it doesn't use floating point to begin with.

[FireBurn]

As a gentoo user I'd like to point out that unless you're editing Kernel make files, there's only -O2 or -Os options for compiling the kernel, I do however use that patches that allow -march=native from https://github.com/graysky2/kernel_gcc_patch which has been benchmarked

[ckonte]

I use Zen for audio recording and I'm quite happy with it, there's a big difference in latency.

[theriddick]

Been using Linux-Zen for a while, pretty happy with it, also includes fsync changes.

[Michael]

Some numbers for Mesa git comming. Are ready but must be sent and/or edited.
As teaser:
~ +20% for OpenGL (radeonsi, Polaris 20)
and size down over a third

@Michael:
I'll going over Mesa Devel List but CC you and you can resubmit if you like.

Sure can CC to make sure I see, thanks.

[nuetzel]

Profile Guided Optimizations are what make Clear so much faster than other distributions. They are well worth the effort.

Some numbers for Mesa git comming. Are ready but must be sent and/or edited.
As teaser:
~ +20% for OpenGL (radeonsi, Polaris 20)
and size down over a third

@Michael:
I'll going over Mesa Devel List but CC you and you can resubmit if you like.

[Mark Rose]

Well, there are some cases where there are significant gains to be had by choosing specific compiler optimizations - but just adding -O3 usually doesn't do all that much. I recall a paper about using genetic algorithms to choose compiler flags from about 20 years ago but I don't think it caught on. Probably not worth the effort except in very rare cirumstances.

Profile Guided Optimizations are what make Clear so much faster than other distributions. They are well worth the effort.

[Archprogrammer]

Exercises with uber-optimizations will probably never cease. Maybe 'cause people don't write compilers while believing there's some hidden untapped potential only for them.

Well, there are some cases where there are significant gains to be had by choosing specific compiler optimizations - but just adding -O3 usually doesn't do all that much. I recall a paper about using genetic algorithms to choose compiler flags from about 20 years ago but I don't think it caught on. Probably not worth the effort except in very rare cirumstances.