Number of joob on very modern CPU

Message

fbeth · Post by **fbeth** » Mon Sep 02, 2024 10:20 am

Hello everybody,

After 8 years with my old laptop I decided it was time for an upgrade, I received the new computer today that has an Intel Core Ultra 7 155H. I looked at the specs online and it says that this CPU has 2 different type of cores : 6 Performance-cores and 8 Efficient-cores, for a total of 22 threads. I'm wondering if it's better to compile on 12 threads to only use the performance-cores or 22 to use all of them. If someone has experience with that kind of CPU I would be delighted if you could share about how you handle compilation. Please note that the computer has 32G of RAM so it won't be a problem for most packages.

Post by **Banana** » Mon Sep 02, 2024 10:39 am

Not with the CPU, but keep in mind the RAM usage

But another item to consider is RAM usage. Recent gcc versions have been known to take 1.5 GB to 2 GB of RAM per job. If the system has the 8 logical CPUs from the previous example, but only 4 GB RAM, the MAKEOPTS value should be lowered to -j2. This is so that the system has RAM to run the basics as well as compile without hitting swap very often slowing things down.

https://wiki.gentoo.org/wiki/MAKEOPTS
https://wiki.gentoo.org/wiki/Knowledge_ ... _of_memory

New at Gentoo · Post by **pietinger** » Mon Sep 02, 2024 1:05 pm

fbeth,

Your CPU has 6 P-cores (12 logical cores with Intel "Hyperthreading" == SMT), 8 E-cores and 2 Low Power E-cores:
https://www.intel.com/content/www/us/en ... tions.html

Normally the Linux kernel decides on which core it starts a new thread. Since I also have an i9-13900k (with 8 P-cores and 16 E-cores), I am very interested in how the kernel handles this. From the looks of it, it SHOULD be better under 6.11:
https://www.phoronix.com/news/Intel-P-S ... metric-Cap

fbeth · Post by **fbeth** » Tue Sep 03, 2024 6:41 pm

I've settled with 20 threads, everything seems good. I'm using kernel 6.11 currently because it's the first version that supports my sound card. I might do a few benchmarks later, but no promise.

spica · Post by **spica** » Tue Sep 03, 2024 10:05 pm

fbeth,

If you run -march=native, try to re-emerge gcc. If it succeds you're good now, otherwise you'll see something like this one https://www.reddit.com/r/Gentoo/comment ... _12th_gen/

New at Gentoo · Post by **pietinger** » Tue Sep 03, 2024 10:25 pm

spica wrote:fbeth,

If you run -march=native, try to re-emerge gcc. If it succeds you're good now, otherwise you'll see something like this one https://www.reddit.com/r/Gentoo/comment ... _12th_gen/

When using GCC 13.3 I think it would be better for PO to use: -march=meteorlake. (I use =raptorlake)
(using this was a recommendation of our developer)

(see also in chapter: IA-32/x86-64 of https://gcc.gnu.org/gcc-13/changes.html )

fbeth · Post by **fbeth** » Wed Sep 04, 2024 11:27 am

So I did a small benchmark, nothing to fancy or complex. The CPU is an Intel Core Ultra 7 155H and I compiled my current kernel (6.11.0-rc6) two times, the first time with -j12 and the second time with -j20.
I was only running alacritty inside KDE Plasma 6 with the ppower cord plugged in. If you need more info or if you have any ideas in order to make better tests, feel free to share it with me.

-j12: 6min 34s
-j20: 5min 55s

That's almost a 10% time difference, it's not huge considering I used 8 more cores, but it shows that the E-cores are not that useful during heavy CPU tasks. pietinger mentioned that the scheduler should be better under kernel 6.11, it looks like the scheduler is prioritizing the P-cores.