Scheduler Changes For Linux 5.15 - Still No Sign Of Any Intel Thread Director Optimizations

[phoronix]

Phoronix: Scheduler Changes For Linux 5.15 - Still No Sign Of Any Intel Thread Director Optimizations

Ingo Molnar began sending in his pull requests bright and early as usual for the just-opened Linux 5.15 merge window. With the scheduler changes for this next kernel version there are some improvements worth mentioning but also worth mentioning is what hasn't found its way to the kernel yet: any software optimizations around Intel Thread Director for upcoming Alder Lake processors...

https://www.phoronix.com/scan.php?page=news_item&px=Linux-5.15-Scheduler

[chuckula]

At HotChips, Intel did confirm that Thread Director support is coming to Linux (not a big surprise). However, they did not have a specific timeline on when it would land.

Given that Alder Lake will start first on the gaming desktop segment and eventually work its way into notebooks and lower-end servers with Xeon-enabled silicon, I would guess support will land early in 2022 as Alder Lake goes more mainstream. Of course TD support isn't needed to get Linux to running on the chip, but it should be a nice feature when support is completed.

[uid313]

The Alder Lake CPU is interesting because it has 8 powerful cores and 8 energy efficient cores for a total of 16 cores, so now desktop x86 has more cores than ever and an interesting heterogeneous architecture similar to the ones long used in ARM processors on smartphones.

But rumors is that Alder Lake is very power hungry, that it uses up to 250 W which is rather insane!

[waxhead]

There is one thing I have never understood and that is why processes seem to jump between cores on Linux (and Windows) machines even if there is low load. I don't see any reason for load balancing a task UNLESS all the CPU time is used up. Would it not be better if the task stays on the CPU and new tasks is allocated to the CPU cores with the least amount of load / biggest timeslice budget left.

[tomtomme]

@waxhead: equal heat distribution?

[ms178]

Maybe some CPU designs might benefit where the internal memory access is different, my Xeon 2678v3 is a great example where it is not 2 x 6 cores per memory bus but 1 x 8 and 1 x 4. That means that memory latencies are not the same for each core. The scheduler could favor the core complex which shares the same memory latencies for the workload which fits into a core complex.

[tildearrow]

The Alder Lake CPU is interesting because it has 8 powerful cores and 8 energy efficient cores for a total of 16 cores, so now desktop x86 has more cores than ever and an interesting heterogeneous architecture similar to the ones long used in ARM processors on smartphones.

But rumors is that Alder Lake is very power hungry, that it uses up to 250 W which is rather insane!

How could it be 250W?! A lot of cores I may understand, but otherwise... what?

[intelfx]

How could it be 250W?! A lot of cores I may understand, but otherwise... what?

My 5950X system (overclocked the shit out of it, of course) routinely eats up to 270W when I'm compiling stuff on all cores.

[zerothruster]

@waxhead: equal heat distribution?

As I understand things, that was the original reason. But does it still make sense on desktop or laptop machines ? For big servers working flat out 247, perhaps (I have no idea). But on many zen2 (and apparently some lower-end zen3 cezannes, if what I read is correct) there is latency and cache change when changing to a different CCX. On a fast enough machine, it probably isn't measurable except in benchmarks, but it certainly looks weird when I watch 'top'.

[jaxa]

The Alder Lake CPU is interesting because it has 8 powerful cores and 8 energy efficient cores for a total of 16 cores, so now desktop x86 has more cores than ever

AMD already did that with mainstream Ryzen 9 3950X and 5950X CPUs.

How could it be 250W?! A lot of cores I may understand, but otherwise... what?

They are pushing it hard. Probably 5.3 GHz for the big cores, and 3.9 GHz for the small cores. That's high even for the small cores, as I'm seeing 3.3 GHz max for Tremont. This is also power draw typical of the flagship Core i9-12900K. Other Alder Lake CPUs will use less.

It looks like AMD will also raise the TDP for CPUs on the AM5 socket. Instead of 105W for the 12-16 core parts, 120W for 12-core, 170W for 16-core. TDP is almost a marketing term but you can see where things are going. Chips auto-overclock themselves based on the cooling situation and can turbo higher for very brief intervals. If you can cool it, why not?