Using Clear Linux As A Desktop Linux Distribution - It Works Well But With Some "Papercuts"

[ms178]

I wish Intel bundled Clear Linux with Intel MKL, VSML, IPP and Intel Compiler.
This would have been a killer move.

If I remember correctly a major hurdle is that ICC is proprietary. As they've open sourced the rest of the components you mentioned, ICC seems to be the last proprietary piece left standing. I'd welcome if they would open source it and use it more broadly in Clear Linux if it provides performance improvements. Even better would be if they incorporated the best pieces into GCC and Clang and would ditch it alltogether. I would certainly like to hear the reasoning why Intel spends the money to support three different C/C++ compiler teams.

[Royi]

I don't think they open sourced MKL, IPP or VSML. I read they might open source VSML but never heard about open sourcing the others.

For scientific code I think ICC will be much faster than GCC.
Its vectorization engine is the best out there .

[celrod]

I don't think they open sourced MKL, IPP or VSML. I read they might open source VSML but never heard about open sourcing the others.

For scientific code I think ICC will be much faster than GCC.
Its vectorization engine is the best out there .

Didn't they add VSML to glibc? See here. Here is the documented asm for logf16 for avx512, for example.
gcc and g++ already use it, and gfortran will soon / will already if you compile gcc trunk from source and follow these instructions. I tried it, and it works. And is very fast.

You can play around on godbolt. Here are a few (very easy to vectorize) examples, and see that they are indeed using zmm registers with the math calls. gcc "just works".
For Clang to vectorize it, you have to pass -fveclib=SVML, and then you'll also have to link SVML to access the function (otherwise you'll get a symbol look up error).

I don't think Phoronix's benchmarks do a good job highlighting this.
1. I don't know how much of the benchmark code is even written in a vectorizable manner. A lot of code isn't. Things like struct-of-arrays memory layout are often critical.
2. Necessary flags to allow vectorization. To get the vectorized transcendentals, you need the following on top of "O3": -fno-math-errno -funsafe-math-optimizations -ffinite-math-only. Note that you don't have to turn on associative-math, so IEEE requiring things like Kahan summation will still work. I was lazy and just used Ofast in the link.
3. For avx512, gcc needs "-mprefer-vector-width=512" on top of march=(something appropriate).


Because Intel processors are so much faster than AMD when it comes vectorized code (although, gen3 Ryzen will increase to two 256 bit fma units per core), increasing accessibility to optimized/vector libraries seems like a good strategy. It's one I'm certainly happy about, anyway.

I love SIMD. One day, I'll spend the time to see how much of my workloads can really be ported to a GPU, but I speculate tentatively "not much". So at the very least until then, the wider the CPU registers -- and the more code compatible with it -- the happier I am.

[skeetre]

I just happened to do a check for Clear Linux Nvidia and looks like they have some official guidance on installing the Nvidia binary drivers now.

https://clearlinux.org/documentation/clear-linux/tutorials/nvidia

I'm going to have to give this a real go as my 'desktop' os now.

[Michael]

I just happened to do a check for Clear Linux Nvidia and looks like they have some official guidance on installing the Nvidia binary drivers now.

https://clearlinux.org/documentation/clear-linux/tutorials/nvidia

I'm going to have to give this a real go as my 'desktop' os now.

Works out quite well, thanks for pointing it out - https://www.phoronix.com/scan.php?pa...x-nvidia&num=1

[skeetre]

I wanted to try it out yesterday, but I had to spend the night recovering a failed lvm partition on my new Fedora 30 server, and still playing around with my Odroid N-2. Nice to you see you were able to test it out so quickly. I've been doing checks every few weeks on the status of Nvidia binary drivers on Clear Linux and it never looked worth the hassle of trying until now.

[andrei_me]

Michael what about a 2020 follow up of this news? What changed from the desktop part since last year?

[marcdraco]

I just popped in here on the advice of Christine Hall (Bride of Linux).

Lack of CUDA is a killer for me - to a degree anyway. My rendering workstation is CUDA driven so that's a non-starter but my baby machine (an Lenovo X270 wtih a modest i7) is a whole different ball game. Clear is running from an external hard drive so I don't have to diddle around with partitioning the main drive and it makes it portable between the machines.

Anyway, the point is that on the laptop Clear absolutely smokes anything and everything else by a huge margin. So much so that it feels like I'm on a much more powerful machine.

I'm not a huge fan of Gnome 3 but I just need a shell so it's not in the way.

In Blender 2.81, I'm seeing performance increases of 200-300% compared to Windows 10 (which I'm stuck with for proprietary software). I wish I knew what strange voodoo was going on here. This is reminiscent of the days Microsoft hid a bunch of functions in the Windows stack so their own productivity software looked better than everyone else's.

It might still be a little early to promote this to the artistic community because my experience is limited to older, more modest machines but if this is the case across the board, then Clear is going to make a serious dent in everyone else.

My guess is they've benchmarked the kernel, found the bottlenecks and re-coded them in pure x86 asm (which is damn hard for most of us, particularly if it's more than a few lines). I'm not good enough at kernel programming to know but something is going and it's exposing a huge problem with the existing Linux distros - somewhere at the heart of everything is bloated. Probably as as result of developers adding more and more bits and no one actually going back to basics and fixing those problems. Maybe it's that they're using a better compiler?

For my single-use case though this is a winner for me - so far - so long as nothing breaks!

[sb56637]

Hey Michael, are you still running Clear as a daily driver? How's it working out for you on the long run? Any issues with updates? How about application availability?
I'm an openSUSE user, but I'm struggling with WebRTC and other videoconferencing apps just burning up my CPU, whereas Clear manages a *much* lower CPU load on the same hardware. So the magic sauce in Clear Linux really does make a tangible (even palpable!) difference in performance.