Intel i9-12900K Alder Lake Linux Performance In Different P/E Core Configurations

[Michael]

Hey Michael, did you find the problem? I have the same issue with the new article "5850U - Windows vs. Linux". Basically any benchmark article that has the graphs from openbenchmarking.org in it loads super slow with any browser (10 to 30 sec.) and doesn't show the graphs if loaded.
Firefox behaves the same, if i deactivate all plugins (with plugins it needed 10 min but showed the graphs).
Calling openbenchmarking.org gives a "ERR_EMPTY_RESPONSE" and the page doesn't load. So the problem seems to be openbenchmarking.org.

The traceroute to phoronix is short and fast:

 7 23 ms 39 ms 12 ms de-cix-frankfurt.as13335.net [80.81.194.180]
8 8 ms 8 ms 8 ms 104.21.42.68

But to ob.org:

 7 8 ms 8 ms 8 ms ffm-b5-link.ip.twelve99.net [213.248.97.40]
8 128 ms 128 ms 128 ms ffm-bb2-link.ip.twelve99.net [62.115.114.90]
9 128 ms 128 ms 128 ms prs-bb2-link.ip.twelve99.net [62.115.122.138]
10 128 ms 128 ms 128 ms rest-bb1-link.ip.twelve99.net [62.115.122.159]
11 * * * Zeitüberschreitung der Anforderung.
12 129 ms 130 ms 130 ms nash-bb1-link.ip.twelve99.net [62.115.137.55]
13 129 ms * 128 ms dls-bb1-link.ip.twelve99.net [62.115.137.45]
14 128 ms 128 ms 128 ms dls-b23-link.ip.twelve99.net [62.115.136.119]
15 * 132 ms 128 ms dls-b1-link.ip.twelve99.net [62.115.113.85]
16 126 ms 127 ms 126 ms hivelocity-svc070168-ic355947.ip.twelve99-cust.net [213.248.75.13]
17 * * * Zeitüberschreitung der Anforderung.
18 * * * Zeitüberschreitung der Anforderung.
19 * * * Zeitüberschreitung der Anforderung.
20 * * * Zeitüberschreitung der Anforderung.
21 * * * Zeitüberschreitung der Anforderung.
22 * * * Zeitüberschreitung der Anforderung.
23 * * * Zeitüberschreitung der Anforderung.
24 * * * Zeitüberschreitung der Anforderung.
25 * * * Zeitüberschreitung der Anforderung.
26 * * * Zeitüberschreitung der Anforderung.
27 * * * Zeitüberschreitung der Anforderung.
28 * * * Zeitüberschreitung der Anforderung.
29 * * * Zeitüberschreitung der Anforderung.
30 * * * Zeitüberschreitung der Anforderung.

Haven't been able to reproduce, but now that you indicate an OpenBenchmarking connectivity problem.... If you try now does it work? Just flushed some firewall blocks in case something like that happened...

[Anux]

... in Windows, UI, Explorer.exe and other system components are tightly connected and integrated, so the Windows kernel knows or gets hints about what applications are running in the foreground and it can adjust the process CPU cores affinity accordingly.

I'm totaly with you, I already fantasized about a system, that would give user input the highest priority, its responses the second highest prority, then audio in general and then video. Linux comming from the server space never put much into this concept but for a UI there is nothing more importend than giving the user control and feedback.

[Anux]

Loading fine here... what isn't loading for you.

Hey Michael, did you find the problem? I have the same issue with the new article "5850U - Windows vs. Linux". Basically any benchmark article that has the graphs from openbenchmarking.org in it loads super slow with any browser (10 to 30 sec.) and doesn't show the graphs if loaded.
Firefox behaves the same, if i deactivate all plugins (with plugins it needed 10 min but showed the graphs).
Calling openbenchmarking.org gives a "ERR_EMPTY_RESPONSE" and the page doesn't load. So the problem seems to be openbenchmarking.org.

The traceroute to phoronix is short and fast:

 7 23 ms 39 ms 12 ms de-cix-frankfurt.as13335.net [80.81.194.180]
8 8 ms 8 ms 8 ms 104.21.42.68

But to ob.org:

 7 8 ms 8 ms 8 ms ffm-b5-link.ip.twelve99.net [213.248.97.40]
8 128 ms 128 ms 128 ms ffm-bb2-link.ip.twelve99.net [62.115.114.90]
9 128 ms 128 ms 128 ms prs-bb2-link.ip.twelve99.net [62.115.122.138]
10 128 ms 128 ms 128 ms rest-bb1-link.ip.twelve99.net [62.115.122.159]
11 * * * Zeitüberschreitung der Anforderung.
12 129 ms 130 ms 130 ms nash-bb1-link.ip.twelve99.net [62.115.137.55]
13 129 ms * 128 ms dls-bb1-link.ip.twelve99.net [62.115.137.45]
14 128 ms 128 ms 128 ms dls-b23-link.ip.twelve99.net [62.115.136.119]
15 * 132 ms 128 ms dls-b1-link.ip.twelve99.net [62.115.113.85]
16 126 ms 127 ms 126 ms hivelocity-svc070168-ic355947.ip.twelve99-cust.net [213.248.75.13]
17 * * * Zeitüberschreitung der Anforderung.
18 * * * Zeitüberschreitung der Anforderung.
19 * * * Zeitüberschreitung der Anforderung.
20 * * * Zeitüberschreitung der Anforderung.
21 * * * Zeitüberschreitung der Anforderung.
22 * * * Zeitüberschreitung der Anforderung.
23 * * * Zeitüberschreitung der Anforderung.
24 * * * Zeitüberschreitung der Anforderung.
25 * * * Zeitüberschreitung der Anforderung.
26 * * * Zeitüberschreitung der Anforderung.
27 * * * Zeitüberschreitung der Anforderung.
28 * * * Zeitüberschreitung der Anforderung.
29 * * * Zeitüberschreitung der Anforderung.
30 * * * Zeitüberschreitung der Anforderung.

[birdie]

The saddest thing about the whole drama about ADL and its Linux support is that in Windows, UI, Explorer.exe and other system components are tightly connected and integrated, so the Windows kernel knows or gets hints about what applications are running in the foreground and it can adjust the process CPU cores affinity accordingly.

In Linux on the other hand we have the kernel all by itself, the Xorg/WM or Wayland Compositor by themselves and running applications. All of them are not aware of one another altogether.

This is further exacerbated by the fact that in Linux you can increase your process priority ("niceness"), say 19, but you can never lower it back to the original value, e.g. 0. This sounds almost idiotic to think about that. Why can't you renice it back to 0? In the end you can simply restart it and circumvent this "restriction".

[sdack]

I am getting the impression that the biggest issue appears to be Intel trying to provide a solution for something that from at least my OS studies back at uni is not really solvable, i.e. automagic scheduling on big little design that generally works better than the alternative. ...

There is no unsolvable issue here. It is just a mess that still needs sorting out. An OS can certainly decide, depending on its energy setting, to prefer performance cores over efficiency cores and vice versa, and further also automatically migrate workloads depending on whether these are compute- or I/O-bound. It does not have to be perfect, because there will always be edge cases. Only what should not be is the need for users to go into their BIOS settings to control it, possibly disabling all their efficiency cores, but they should either not have to bother about it or at least be able to control it from within the OS. This will already please the majority of users and one can improve it further from there.

[Michael]

Michael you say it gives the option to enable AVX 512 but you dont specify if you did or not, could you clarify?

None of the tests in this article were AVX-512. See the linked article from there if wanting AVX-512 ADL data. Was simply mentioning when all E cores are disabled, AVX-512 is possible. AVX-512 was out of scope for this article especially with many workloads not being relevant for AVX-512, this article was just about core/thread comparison.

[MadCatX]

It's not an OOM problem at all. The laptop has access to 8GB of memory and 4GB of swap, and even when the applications were stalling free never reports more than 5GB in use at any time. And it was a 5.15 kernel, not the dinosaur 5.10 kernel that got bundled with Bullseye.

Was it 5 GiB with or without the page cache? Swap is only helpful if you have a lot of pages that you really don't need for a long time. In most cases you might actually be better off without swap and having the system kill the process that runs out of memory. Managing to keep the interactive processes in RAM is one area where Windows does a lot better than Linux.

Which was the point I was trying to make. Windows 10 does not have a scheduler specially for Alder Lake but they have experience on BIG.little architectures because of their work on Windows RT and the ARM64 versions of WIndows 10, where all hardware use BIG.little.

I seriously doubt this is the case. Win10 scheduler has had numerous issues with far simpler CPU topologies, take first Ryzens and Threadrippers for example. What I think is more likely is the fact that the simple scheduling logic of Win10 is less likely to take a wrong guess and mess things up. With predictable workloads which spawn fixed number of threads that do the same kind of work all the time it might work out better than attempts at sophisticated guesswork. Linux on ADL still beats Win11 in tasks like physics simulations, DNN stuff and 3D rendering because the scheduler just assigns works to CPUs and probably doesn't move things around very much then. Things get dicey when you have asymmetric tasks that require a lot of CPU-to-CPU synchronization but then again even Win11 doesn't seem to be conclusively better in this area either.

[Anux]

I'm also not sure if there is much to gain by compiling the kernel yourself, apart from newer drivers. Atleast not, if you don't heavily modify it to a specific task. I bet one can find a benchmark here on Phoronix.

[mdedetrich]

Which was the point I was trying to make. Windows 10 does not have a scheduler specially for Alder Lake but they have experience on BIG.little architectures because of their work on Windows RT and the ARM64 versions of WIndows 10, where all hardware use BIG.little. It's practically a forgone conclusion that this experience factored into their continuous work on the scheduler to the point where Windows 10 for x64 is able to handle Alder Lake as-is.

Well firstly I don't think there is much cross over between ARM64 BIG little and alder lake, or enough for it to be useful

Secondly the mere fact that Windows 10 with Alder Lake is beating Linux is probably demonstrating that we are talking about the difference between general schedulers and/or CPU support rather than hybrid core scheduler specifically.

And its not like Linux doesn't have experience with big little, pretty much all android phones also follow a big/little design.

[Grinness]

-march=native not good enough for you?

I am fine with Arch, compile what I need (e.g. ROCm, pytorch and related packages) with default flags.
It is you who is complaining/showing off that you are able to compile thus you have superior understanding ...
Do what you want, I don't care