Ikey Doherty's Serpent OS Continues Building Up Its Rust Infrastructure

[ermo]

ermo What's the current status about a x86-64-v3 or x86-64-v2 as baseline? There was some back and forth on this. Also sunnyflunk criticized CachyOS in a blog post for doing whole repo builds with x86-64-v3 while his own data didn't back up his claims as it showed a net improvement overall. Was there more testing in the meantime and are there any news to share on this topic?

No news to share yet, no, other than -v2 being the current default profile.

We're still mulling over how to approach this properly.

[skeevy420]

Any idea what is the plan then ?

Step 1: Rewrite it in Rust
Step 2: ?????
Step 3: Profit

[sunnyflunk]

Also sunnyflunk criticized CachyOS in a blog post for doing whole repo builds with x86-64-v3 while his own data didn't back up his claims as it showed a net improvement overall. Was there more testing in the meantime and are there any news to share on this topic?

What you have stated is not correct. My critique was on the statement "if x86-64-v3 is detected it will automatically use the optimized packages, which yields more than 10% performance improvement.". The statement is wildly misleading as my data showed that x86-64-v3 could reduce performance of packages and in most instances led to only a small improvement.

The 10% number came from testing a handful of packages that would likely be benefited by AVX2 and taking an average or median I can't remember which. x86-64-v3 is beneficial to many packages, but is wrong is that you cannot do is claim this result broadly across your packages.

What my data showed was that -O3 was much more beneficial than x86-64-v3 for the tested packages. The -O2 setting likely needs some work in gcc as there was significant advantages by using -O3.

Also I'm not affiliated with Serpent OS in any way and have not been for over a year now. So any tweaks that I had previously added or blogged about you can't assume are still present as some have been removed.

[Estranged1906]

Why is SerpentOS using Gnome? Why not Budgie?

[ms178]

What you have stated is not correct. My critique was on the statement "if x86-64-v3 is detected it will automatically use the optimized packages, which yields more than 10% performance improvement."

Fair enough, but I think this statement on the CachyOS website wasn't meant as a scientific accurate statement in the first place but more so as a rule of thumb. You made a case of "false advertising" out of it, which is misleading, too.

The statement is wildly misleading as my data showed that x86-64-v3 could reduce performance of packages and in most instances led to only a small improvement.

The 10% number came from testing a handful of packages that would likely be benefited by AVX2 and taking an average or median I can't remember which. x86-64-v3 is beneficial to many packages, but is wrong is that you cannot do is claim this result broadly across your packages.

What my data showed was that -O3 was much more beneficial than x86-64-v3 for the tested packages. The -O2 setting likely needs some work in gcc as there was significant advantages by using -O3.

I have commented in more detail in the reddit thread on your article, I didn't get any response to the valid points I've raised there to this date. And to be honest, producing headlines of your own which were based on the testing of a very small sample size of packages, isn't fair to the CachyOS project at all. Quoting my reddit post:

"Your blog post leaves a lot to discuss and to interpret though and I have a problem with your choice of your testing platform. In the end, the Phoronix testing and even your own numbers have proven that CachyOS can deliver benefits over other Arch-based distributions overall. I think it is fair to say that CachyOS is more than just "marketing hype" and delivers improvements in areas that profit from using these newer instruction sets, but also from other optimizations made by the project.

While the wording of the passage you cited could be improved, its meaning was taken out of context as the project makes it also very clear that the advertised improvements don't come from using x86-64-v3 alone. The project offers a highly patched Kernel, ananicy-cpp with a decent set of rules, carries some Clear Linux patches for several packages with them, offers optimized KDE Plasma packages and delivers a more customized and performance-oriented experience than other Arch-derivatives while staying compatible with its repos.

While you mentioned that some repos weren't even build with x86-64-v3, it might be important for some to mention multilib and community explicitly. Wheras the community repo was migrated during the last days. But now to some more specific parts worth discussing:

A) Testing on TDP-limited hardware

As you've stated in your post, you did your testing on a 28W TDP-limited Broadwell-NUC, if I interpreted the Intel Ark page to the stated model number correctly (https://www.intel.de/content/www/de/...fications.html). If so, that would be a rather uncommon setup as I wouldn't qualify that as a typical desktop machine which usually come with a lot more TDP headroom. While this is still a legit data point for people with more power-limited setups, the findings on that particular NUC could be misleading because older Intel and AMD generations had a worse AVX2-implementation where power usage increased with AVX2 usage while the operating frequency even decreased which could also decrease the overall performance.

However this is a CPU microarchitecture-specific behavior and no longer the case on the latest CPU generations of both Intel and AMD. Therefore this odd behavior will fade away over time as CPUs with these older microarchitectures will get replaced. For this reason alone I would not draw any generic conclusions from your findings as the x86-64-v3 packages could yield more benefits on newer CPUs or higher power TDP models (e.g. I use a Haswell-EP Xeon for my own testing). As CachyOS is targeting desktops and enthusiast user's in particular which usually own CPUs within the 65 - 125W TDP range, it would be more representative to use such a more common desktop setup for further testing.

B) Choice of tested packages

We need far more data points to get a bigger picture. Also from the software perspective, the testing was limited to a few packages only that might or might not be representative for the whole repo. Also you cannot exactly attribute all the performance improvements to the ISA alone as CachyOS comes with more alterations, especially in one of the offered more advanced Kernels (CachyOS-BORE is the current recommendation) while staying compatible with Arch which limits some further optimization opportunities. You'd need to test v1 packages vs v3 packages on CachyOS for such comparisons (and not vs. vanilla-Arch).

C) Compiler Flags discussion

I think the O2/O3-discussion is a seperate topic entirely and not specific to CachyOS, the most interesting part is that x86-64-v3 enables the usage of some newer ISA instructions and that these provide some benefits where either the software or the compiler (with auto-vectorization) can make use of them. This is no news, however. Also the rather narrow scope of the older vector ISAs. With AVX-512 carrying some more general-purpose improvements with it, I'd expect improvements with a broader set of packages with x86-64-v4 in the near future."

[timofonic]

FTR, Ikey is Lead and Founder, so he has the final say.

I know COSMIC is on Ikey's radar. We'll see how it pans out.

Oh. I just saw you and asked.

I hope collaboration happens soon! Thanks a lot!

[sunnyflunk]

Fair enough, but I think this statement on the CachyOS website wasn't meant as a scientific accurate statement in the first place but more so as a rule of thumb. You made a case of "false advertising" out of it, which is misleading, too.

10% gains are fairly rare given the number of packages in a distribution so "rule of thumb" seems pretty dishonest as well. If you go back the closest I got was asking "but is it true?" (which I never actually answered) so you are adding a whole lot of issues that simply weren't in the post.

I have commented in more detail in the reddit thread on your article, I didn't get any response to the valid points I've raised there to this date. And to be honest, producing headlines of your own which were based on the testing of a very small sample size of packages, isn't fair to the CachyOS project at all. Quoting my reddit post:

Are you sure you're not shadow banned or something, cause I can't see your post on the thread (there is one comment that I see is deleted). Also I made it pretty clear what I was testing by laying out the methodology. The entire point of the comparison (the impact of x86-64-v3) was to not switch kernels or any other settings as only the packages with different flags were changed which I happened to use the CachyOS packages as they are built with those optimizations (made bold claims) and were easy to install. " x86-64-v3: Mixed Bag of Performance" is correct and doesn't reference CachyOS in any way (they simply make up the methodology), particularly when people expect it to always be faster. You have somehow managed to personalize this when no one else mentioned the results being reflective of the project (from reading and understanding the methodology no doubt).

[rmfx]

The moss package system looks very interesting for sure. However to me such a project should be decoupled from the distro because that leads only to even more fragmentation.

I really hope that someday, the linux world will freaking agree on ONE unified software package system and move forward together because that’s ridiculous currently.

Damn I wish Linus cared more about OSes and not just kernels.

[cj.wijtmans]

The moss package system looks very interesting for sure. However to me such a project should be decoupled from the distro because that leads only to even more fragmentation.

I really hope that someday, the linux world will freaking agree on ONE unified software package system and move forward together because that’s ridiculous currently.

Damn I wish Linus cared more about OSes and not just kernels.

How about no thanks. I like innovation not stagnation. If you want something universal use flathub or something.

[hoohoo]

I don't suppose it uses something other than systemd? Hoping...