Intel's 3888.9% Linux Kernel Boost, Isn't What It Seems

Рет қаралды 15,387

Brodie Robertson

Күн бұрын

Пікірлер: 89

@lePirateMan 2 күн бұрын

I have intel so I have to watch at 1/39 times speed for it to be normal

@hk_8084 2 күн бұрын

Just press the turbo button

@s0ygecko Күн бұрын

@@hk_8084 that thing is even wired now?

@Tessil666 Күн бұрын

🤣

@-yttrium-1187 2 күн бұрын

This time the % increase in performance isn't really interesting because it was a regression fix. No one actually saw any function complete a thousand times faster. What I'd really like to mention is that there is no greater joy than a developer fixing awful O(n^2) algorithms and posting their million-fold improvement even if the function almost never runs and the wound was self-inflicted.

@phill6859 2 күн бұрын

And nothing worse than a customer deciding that it's perfectly fine to increase their data flows by orders of magnitude without testing or even asking you whether it's been tested. When the system is down and they can't do deliveries, you will feel stressed while trying to fix it

@-yttrium-1187 2 күн бұрын

@@phill6859 Ah yes, Design what someone asks and you'll fail to deliver what they meant. But if you actually take your time then you're not delivering what you've been asked to do.

@pip5528 Күн бұрын

Could that be why I was recently bottlenecking on CachyOS and suddenly am not? My CPU was stuck around 2 GHz or so when it has a boost clock of over 4 GHZ. Both it and the GPU had strangely low temperatures and clock speeds.

@yeetdeets 23 сағат бұрын

Note that 4000% would only be 40 times faster. And with the 600% regression beforehand, it would only be ~7 times faster than before the regression.

@Y4K3D0 2 күн бұрын

0:25 The only people saying that are in fact Userbenchmark's alt accounts

@phill6859 2 күн бұрын

I optimised a piece of code by 1,000,000% and it sent be back in time to 1955. I hitched a ride back and vowed never to do it again

@genstian 2 күн бұрын

Noobs. I turned a list to a set in python and speed up code by 1000x.

@moarjank Күн бұрын

You joke but I did this in some commercial software that was reading a huge linked about 10x every second and it had significant impact. O(1) v O(n) can be huge

@no_name4796 Күн бұрын

I qm pretty sure lists in python, in most cases, are faster then sets. After all lists are just contiguous space in memory, while sets do require some hashing math to access the values

@TailRecursive Күн бұрын

@@no_name4796 I think it depends... sets are probably faster for lookup. The fragmentation might be a downside, but idk if the hashing takes that much time

@no_name4796 Күн бұрын

@@TailRecursive "i am pretty sure lists in python, IN MOST CASES, are faster then sets" congrats man, you just said what i said lol

@gardian06_85 Күн бұрын

can't you do the same thing by unwrapping a for loop?

@mactan_sc 2 күн бұрын

this is like the AVX benchmark again isn't it

@45545videos Күн бұрын

I game on Linux for the 3,800% performance boost

@EwanMarshall 2 күн бұрын

Saw the article, skimmed it, saw the intel test bot found it, thought cool, a kernel bug fixed for some workload or another. And then went on with my life, at no point did I consider it would or wouldn't be Intel only, hell, there is a chance it could even effect ARM, PowerPC... and all the other architectures the kernel support.

@geertb4341 Күн бұрын

Ok, what does "a 600%" decrease mean? If -100% slows it down to 0, will -600% make it run 5x as fast backwards? Silly. I assume they saw a 6x slowdown. This is 1/6 or 17% of the original speed, so an 83% slowdown. NOT 600.

@reimner Күн бұрын

yeah right? i am confused too

@gr-lf9ul Күн бұрын

I imagine it is a "lower is better" benchmark, but even then, "6x slower" would be "500% worse", not 600%.

@ChrisAzure Күн бұрын

Probably, it reaches the point where the machine, instead of drawing power, it generates power to the grid.

@ahettinger525 2 күн бұрын

"technically not wrong, but not even remotely relevant to most people" is pretty much the definition of a click-bait headline. I mean, there's nothing stopping them from saying "only really hit by one edge case" as part of the headline, but that wouldn't get as many clicks.

@insu_na 2 күн бұрын

Man, you're ruining UserBenchmark's day :(

@szaszm_ 2 күн бұрын

That Phoronix article title _is_ clickbait. It's not wrong, but it's misleading. There was a Veritasium video some time ago about clickbait, and how there's often a tradeoff between being truthful and being clickable. I think this Phoronix title is just barely truthful, but misleading, and very clickbaity.

@dashcharger24 Күн бұрын

Phoronix has a bit of a trademark on it. I do enjoy reading articles on Phoronix, but I found the bcachefs articles a bit optimistic (it would make everything else obsolete).

@szaszm_ Күн бұрын

@@dashcharger24 well, he takes the author's _very_ optimistic (delusional?) words at face value

@Vitis-n2v 2 күн бұрын

People commenting stupid stuff always reminds me of the DRM panic articles. People don't even read about what features that actually bring and only resort to insults saying we don't need BSOD in Linux when that's not even the main point of DRM panic. For those who don't know the DRM panic allows the kernel to draw kernel panic screen even over a stuck GUI(X11 or wayland) session which without DRM panic is impossible and in case of kernel panic you end up with stuck last frame of your GUI and can't see the errors making debugging difficult. The ability to show error QR code is just an extra feature on top of that.

@maclypse Күн бұрын

What always astonishes me is that people can see "4000% performance improvement" and think "Sounds legit. I have understood this correctly."

@peterjansen4826 2 күн бұрын

Kernel 6.11 also has a weird bug. On many systems (confirmed by several other users) when you compile dwm in a graphical environmen the code compiles succesfully but as soon as it is done compiling then X11 crashes and you get dropped to the tty. Fun. I just quickly asked a few people and so far I have it confirmed for at least 3 other users, mixed hardware (AMD, Intel, Nvidia). For some weird reason a segfault gets triggered and causes a crash of X11 and this exclusively happens with kernel 6.11 and 100% of the times.

@kyrylmelekhin2667 Күн бұрын

Yeah, well not sure if it's a bug or intentional, but this happens because when dwm gets installed to /usr/local/bin/dwm the inode is updated but the kernel keeps a reference to the old one and any program will crash under such circumstance. Try hello world application in C as example

@kyrylmelekhin2667 Күн бұрын

As a work around, just add rm /usr/local/bin/dwm command to install target.

@peterjansen4826 Күн бұрын

@@kyrylmelekhin2667 I will try it out, thanks. I have no idea why this got introduced in kernel 6.11.

@IlluminatiBG 2 күн бұрын

My favorite kind of dark magic programming! Days upon days of investigation, reading, and reverse engineering, only to discover the fix is single line of code.

@dashcharger24 Күн бұрын

It's actually possible. For instance because you call the same statement over-and-over.

@spambot8896 Күн бұрын

I've been primed not to expect any noticeable changes whenever I hear something like this

@marsovac 23 сағат бұрын

That Phoronix title is accurate and misleading at the same time so that it could be a clickbait. Instead of "Intel" the title should have "Benchmark bot", so to not suggest that the improvement is only on Intel, but on the kernel itself. But then it would not be as interesting to fanboys of each or the two flavors.

@thatoneannoyingtornadosire8755 2 күн бұрын

the corn bots are intel fans, apparently

@R4d1o4ct1v3_ Күн бұрын

My usual reflex is to shame the big tech company simps for their stupidity, but it's hard to be mean to the Intel ones anymore. They just keep taking L's from all sides these days.

@annieworroll4373 Күн бұрын

I saw the headline, hadn't gotten around to reading the article but I was assuming it was probably real but only under some circumstances most people won't care about, or was like you mentioned a thing that takes up so little time to begin with that a speedup is almost more trivia than actual improvement.

@KozmoPoly 2 күн бұрын

Ohh, so that's why my intel pc booted in 0.05ps (Picoseconds) earlier today

@Poldovico Күн бұрын

My boss sent me that article. He wanted to know if we should patch all the kernels in our datacenter. A little Brodie popped up in my brain to remind me to check the context. I did not patch all the kernels in our datacenter.

@HksjJkdkd 14 сағат бұрын

Why? The main point of improvements this patch has is for Servers and Datacenters, having to do scale?

@Poldovico 13 сағат бұрын

@HksjJkdkd because it affects a handful of cpus we don't use in a handful of workloads we don't perform, at a scale far larger than ours. And because maintaining custom patched kernels is a pain in the neck when you can just wait for the improvement to make it to your distro

@HksjJkdkd 6 сағат бұрын

@@Poldovico fair enough

@tranthien3932 2 күн бұрын

I think the pr0n bots have Intel CPU guys.

@siljrath Күн бұрын

Vids like this are a HUGE part of why I love your stuff, Brodie. :)

@lavavex 2 күн бұрын

I thought this was going to be a userbenchmark cope article, not an actual article lol

@MrVivichrist Күн бұрын

Data Oriented Development skills. Data locality and alignment are as important as remembering to deallocate. But they didn't teach this in Computer Science papers (BSc), I learnt C as a part of the Network Engineering curriculum. Maybe I should have actually taken that Operating Systems paper instead of AI and graphics since C++ seems to abstract away from memory allocation and alignment (In a bad and confusing way). Allocators were only a recent discovery to me, it was mentioned (but deemed too advanced?) I feel like I missed out.

@Problematist Күн бұрын

I don't read past the title, but that's why I wait for a Brodie video to tell me what opinion to have.

@SB-qm5wg Күн бұрын

Some databases don't play nice with transparent huge pages enabled.

@GormHornbori Күн бұрын

Cache aliasing issues is just one of those things that can have crazy big performance impact on modern/big systems. Not surprised it the crazy high percentage was found on a system with hundreds of cores, and multiple physical CPUs. (High latency for cache syncronization.) They are very hard to reason about, so cache aliasing problems will just seem like black magic. And a fix for one application/benchmark are likely cause a problem for other applications/benchmarks.

@Blueeeeeee 21 сағат бұрын

I know this is completely off-topic, but I feel like it would be interesting to cover opensuse again at some point. From what I remember they're going through changes recently, but how is that going ? I often heard of it as the "best rolling release" yet it's never talked about on channels like yours, this is very disorienting. With more and more popping up, seeing a comparison of atomic distributions and the technologies behind them in the coming months would be very interesting, too.

@autohmae Күн бұрын

As someone who has been following Linux kernel development for a long time,... a large improvement often means, just a small area which might translate to a huge improvement in real life - for a select number of people. "technically not wrong" - technically correct is the best correct ? Well, if people know the context. I remember Rick van Riel being a new kernel developer, memory management is a really hard topic and he has made many many improvements/rewrites, but also can cause issues for some people.

@GoodVolition 22 сағат бұрын

Well that's great Brodie, but I just sold my investors hundreds of Intel CPUs promising just under 40x performance. What do you plan to do about it?

@LordHonkInc 2 күн бұрын

Hey, that almost makes up for all the performance loss due to Intel's security flaw mitigations :P

@dashcharger24 Күн бұрын

Until they find another security hole.

@KomradeMikhail Күн бұрын

Computer nerds that skipped their math classes, always make mistakes with adding and subtracting percentages.

@aande1 9 сағат бұрын

Such headlines don't faze me. I've read the Phoronix article days ago. Translated: One random internal Kernel thing got much faster/a regression was fixed. Real-world-workload-benchmarks will show if this has an overall measurable performance-impact. HOWEVER, keep in mind, even if this ends up improving the overall performance only by 1%, having - over time - multiple improvements of this kind, these improvements add up and "suddenly" the kernel overall is 10% faster. Every bit counts.

@kxuydhj Күн бұрын

the first thing i think when i see stuff like this is "this is definitely not what it seems". if it were it would easily make mainstream headlines. and even if it did it would still be a 50/50 chance because mainstream media likes clickbait too.

@Linuxdirk Күн бұрын

I recently removed a function from my code. This function is ∞% faster now because instead of 0.2 seconds it now runs with no seconds at all.

@emacsking4310 Күн бұрын

What do you guys think of the new EMACS KING profile picture?

@kitsune-chan6897 Күн бұрын

THE NUMBERS MASON. WHAT DO THEY MEAN.

@WilReid 2 күн бұрын

What's wrong with 192 GB on a desktop? I only have 32 threads, but I'm back up to 192GB after getting my 14900k RMA'd. Also, if you're going to talk about Amdahl's Law, you should mention that you're talking about Amdahl's Law.