Intel Updates Cluster Scheduling Linux Patches For Hybrid CPUs

[phoronix]

Phoronix: Intel Updates Cluster Scheduling Linux Patches For Hybrid CPUs

Intel engineers have been working on new cluster scheduling code for the Linux kernel to better help with process scheduling for their modern hybrid processors. An updated version of these patches have now been posted for attempting to help with the Linux performance of Alder Lake CPUs and newer...

https://www.phoronix.com/news/Intel-Cluster-Sched-Hybrid-V2

[kylew77]

Looks like Intel thread director is not needed on Unix-like systems as much as Windows needs it to perform well. IF implementing thread director actually hurts performance (in some cases) then why optimize for it? I was reading from some OpenBSD developers on someone who wanted to use an X1 Carbon 10th Gen with Alder Lake processing cores with OpenBSD 7.3 and it was mentioned that the OS has not optimizations for P and E cores but tends to "just work." Looks like Linux would be best served by leaving a "just works" method for the CPU cores.

[Grinness]

Looks like Intel thread director is not needed on Unix-like systems as much as Windows needs it to perform well. IF implementing thread director actually hurts performance (in some cases) then why optimize for it? I was reading from some OpenBSD developers on someone who wanted to use an X1 Carbon 10th Gen with Alder Lake processing cores with OpenBSD 7.3 and it was mentioned that the OS has not optimizations for P and E cores but tends to "just work." Looks like Linux would be best served by leaving a "just works" method for the CPU cores.

I assume the issue is with applications that span only a handful of threads and run for long time (e.g. games) and are badly affected in their throughput (e.g. FPS) by the bouncing across P and E cores done by the scheduler (e,g. for managing temperature across the die)

If the above is correct, I do not see the meaning of those benchmarks -- unless they mimic the above scenario
In 'burst' scenario -- where an application span a lot of threads that run in short time -- the higher number of cores for the P/E platform compensate (plus high IPC) - that's what I remember from phoronix article looking a blender performance Alder vs Ryzen

Cheers

[sophisticles]

I suspect Intel will eventually move to an all E-core design for most of their processors, which should resolve this problem.

[skeevy420]

I suspect Intel will eventually move to an all E-core design for most of their processors, which should resolve this problem.

That defeats the purpose of P and E cores. I suspect that in the future compilers, kernels, and operating systems will account for multiple types of cores better. Asymmetrical computing from multiple sources using shared resources is the way of the future. Their all E-core designs, and they will make these, will be the Atoms of the future and will become the butts of our jokes. The more things change, the more they stay the same.

That reminds me, I wonder whatever happened with AMD's dual-architecture arm/x86 stuff they were talking about doing around a decade ago. Is the horribly performing PS4 hypervisor/OS all that ever amounted to?

[NobodyXu]

Their all E-core designs, and they will make these, will be the Atoms of the future and will become the butts of our jokes.

For consumer products, this is true.
However, for server products, this is false.

Servers mostly don't care about single-core performance much and instead want as much cores as possible with lower power consumption, that's why AMD currently outperforms Intel in server market: Their chip is more efficient and delivers more cores, despite Intel has better single-perf plus a bunch of accelerators for niche use cases that will require a lot of software updates to take advantages of.