Cluster Scheduling For Intel Hybrid CPUs Looks Like It Will Be Ready For Linux 6.6

[phoronix]

Phoronix: Cluster Scheduling For Intel Hybrid CPUs Looks Like It Will Be Ready For Linux 6.6

For the past few months Intel has been working on a new cluster scheduling implementation for their hybrid CPUs. This rework was due to their earlier cluster scheduling code not working out so well for the likes of Alder Lake and Raptor Lake processors while this new patch series can at least help some workloads in the ~1% range...

https://www.phoronix.com/news/Cluster-Scheduling-Hybrid-6.6

[mangeek]

I know there are great minds doing lots of math and testing on this stuff, but I always thought a good way to deal with common desktop/laptop CPU schedulers on Alder Lake might be to do something like "every process starts on an E core, but if it uses more than 80% of a core for more than one full second, it migrates to a P core and stays there for twenty seconds."

[V1tol]

but I always thought a good way to deal with common desktop/laptop CPU schedulers on Alder Lake

is to buy AMD /fixed

sorry, couldn't resist

[Kjell]

E cores is just Intel being cheap.
AMD is the way

[Alex Doe]

I know there are great minds doing lots of math and testing on this stuff, but I always thought a good way to deal with common desktop/laptop CPU schedulers on Alder Lake might be to do something like "every process starts on an E core, but if it uses more than 80% of a core for more than one full second, it migrates to a P core and stays there for twenty seconds."

I don't think it's a good way. You will be loosing some performance all(!) the time for no reason. Every task that requiers big core will have degraded performance at the begining and many task are not like 100% cpu consuming 3 hours straight. It can be many 3 sec bursts.

btw I'm pretty happy with my Intel 13600kf powering Fedora 38. Faster and cooler than I expected.

[mangeek]

I don't think it's a good way. You will be loosing some performance all(!) the time for no reason.

What they're actually doing is similar to my idea!

There are a small group of E-cores on the SoC chiplet that 'always has to run anyways', and jobs start there. If the jobs exceed the SoC chiplet's E-Cores, the CPU chiplet is powered-up and the jobs start moving to CPU chiplet E-cores and then P-cores.

I know you're skeptical, but I think you'll see that this provides a lot of CPU horsepower, but it doesn't just apply it when it's not needed. In most use-cases (like, 99% of my day), there's really very little reason for anything to happen on a P-core. In this new design, little background processes and desktoppy stuff will likely run on the SOC E-cores, but by the time your compute job or game is half a second deep in CPU, it'll be migrated to P-cores

https://www.techpowerup.com/review/i...ep-dive/3.html.