Making The AMD Ryzen Threadripper 3990X Run Even Faster - By Loading Up Intel's Clear Linux

[indepe]

The LTS Kernel is 4.x. The “native” kernel is at 5.5 since recently. (I”m reading the latest update also builds the nvidia driver again.)

EDIT: Actually there is also (or perhaps new) a LTS2019 kernel which comes with kernel 5.4.

[DCHart]

Pretty difficult actually, the patches you can see in their pkg repo is only a very tiny percentage of the modifications they make. Most of their optimizations are undocumented and many are hardcoded.

Yet it is fairly easy to recompile the RPM for a different distro. The patches and the configuration are all in the source RPM (I looked). The gcc optimizations are exported to the environment so we know what those are. However, it is a 4.x series kernel. Fedora, for example, is running 5.4.18. Mint is running 5.3.

[duby229]

What's with the Clear Linux push? Perhaps it is a legit distro besides it's special kernel optimizations, but how difficult is it to port those optimization to our kernels?

Pretty difficult actually, the patches you can see in their pkg repo is only a very tiny percentage of the modifications they make. Most of their optimizations are undocumented and many are hardcoded.

[deusexmachina]

What's with the Clear Linux push? Perhaps it is a legit distro besides it's special kernel optimizations, but how difficult is it to port those optimization to our kernels?

[DCHart]

I agree that Clear Linux is very fast. I have it loaded on an older i7 laptop. However, in its current form it is not usable. I can find no advantage of swupd over dnf or apt. However, there are disadvantages including the fact that you cannot search for a program by program description. Moreover, there is a great deal of missing software. For example, bind-utils is available but not bind. There is no id3v2 tagger, no mindlna or substitute, the KDE Partition manager is AWOL. This list goes on ad infinitum.

You could recompile Fedora RPMs. I checked the macros and they will put files in the right place (I think). The optimizations are exported to the environment. However that would be without an upgrade mechanism.

The bundle concept has no advantages. Fedora makes packages available through groups and individually which makes much more sense. Why install stuff that you do not want? As people move to SSDs they are getting by with less HD space.

Can anyone explain how CL works in the absence of an fstab file? How would one add, say, a USB drive by its UUID so that it is always in the same place?

And speaking of places, a lot of stuff is not where it belongs. I don't have the CL machine booted at the moment but if memory serves me correctly sudoers is in /usr/share/default/sudo/sudoers instead of /etc and that's just one example.

Ultimately why on earth did Intel feel the need to reinvent the wheel? The complement to RPMs used to be yum and is now dnf. It ain't broke but Intel "fixed" it. Similarly, long time Linux users (I am in my 20th year) expect certain things to be in certain places. Why make it difficult for people to transition from one distro to another?

It is fast but so what if you cannot use it?

[torsionbar28]

The results aren't surprising; just like Intel's compiler results in the fastest AMD binaries even though AMD isn't directly targeted, Intels distribution is the fastest on AMDs hardware, again, even if it isn't directly targeted. That highlights how well Intels optimizations are working across a variety of different architectures, and not relying on any special tricks to extract performance.

There was an article a few years ago where someone from AMD specifically said one of Zen's design goals was to perform well with existing (read: intel) optimizations rather than software developers having to design for Zen. Software optimization was one of the pain points with Bulldozer and the CMT design. It makes perfect sense then, that optimizations that target intel's chips also deliver results on Zen, as this was AMD's goal all along.

[gamerk2]

The results aren't surprising; just like Intel's compiler results in the fastest AMD binaries even though AMD isn't directly targeted, Intels distribution is the fastest on AMDs hardware, again, even if it isn't directly targeted. That highlights how well Intels optimizations are working across a variety of different architectures, and not relying on any special tricks to extract performance.

[CochainComplex]

1) The gap between first and last is very thin

This is the geometric mean if you have eg. Blender which is performing the same across all of the distros you will flatten the rather extremer cases like Octave or Video de/encoding. In my case "Octave" - makes a huhge difference if I have to wait 2h or 3,5h (which is more then 30%) until the result pops up.

3) Clear Linux may not be the right choice for your PC, as a lot of stuff have been stripped and this may render it wimp on certain desktop platforms (missing support for not-so-older CPUs, libraries, etc.). Also I don't know about its support and community (check the archwiki or how much support there's for CentOS or Ubuntu)

True, but if you have to use Python, Octave, PHP, redis, video encoders it is a astonishing gain.

I have to be mobile so I have a Workstation Laptop with the fastest mobile CPU available. Getting 30% more is an upgrade which I can not get without making use of this fast distro because there is simply no product on the market.
Besides sometimes 10% more overall performance can be easily a few hundred dollars if you buy a faster cpu (or system) instead.

Yes, if you don't have to rely on performance there is not much reason to use clearlinux.

p.s.: Besides for my NAS (amd 200ge) I'm using docker containers based on clear linux just because performance for free. There is not much of a difference between linuxserver/mariadb vs clearlinux/mariadb concerning the deployment.

[ms178]

I haven't had a great experience with the Steam flatpak at least, not only due to bugs which weren't in the native client but also the game performance was noticeably slower. I'd prefer native software anytime, at least for now.

I just want to remind people that some of the ClearLinux optimizations (Glibc dispatch for AVX2) aren't even in effect at the moment on AMD platforms. So there is still some performance potential which could be unlocked if AMD would help the Glibc community to test their hardware with this functionality, see: https://hub.packtpub.com/a-bug-found...nd-other-cpus/

[CochainComplex]

Maybe one should apply those optimizations to the flatpak runtimes?

I was thinking about this aswell - steam and co