Stockfish Prioritizing E-Cores on Hybrid CPU Architectures

[dav1312]

Describe the issue

On systems with hybrid CPU architectures (e.g., Intel 12th+ Gen with P-cores and E-cores), Stockfish appears to sometimes schedule its threads predominantly on the less powerful E-cores, even when P-cores are available.

This behavior significantly impacts performance. While workarounds exist (like manually increasing the process priority), it would be beneficial if Stockfish could inherently hint to the OS scheduler for a more optimal core assignment or provide a built-in mechanism to prioritize P-cores.

Expected behavior

Hint to the OS scheduler to prioritize P-cores.

Steps to reproduce

Run Stockfish on a system with a hybrid core architecture, then check the thread assignment in Task Manager or similar software.

Anything else?

No response

Operating system

All

Stockfish version

Latest

[vondele]

Maybe something like calling SetPriorityClass(GetCurrentProcess(), HIGH_PRIORITY_CLASS); on windows? idk.

[dav1312]

Looks like in Win10+ there is SetThreadInformation with the ThreadQualityOfService information class.
In the case of using thread priority I could check if "above normal" instead of "high" is enough to avoid making the whole system sluggish.

[Disservin]

So when does this? Even with just 1 thread ? Or threads > number of performance threads/cores ?

[Sopel97]

common issue https://www.reddit.com/r/XMG_gg/comments/vlqn6d/psa_rendering_tasks_are_moved_to_ecores_when/

[dav1312]

So when does this? Even with just 1 thread ? Or threads > number of performance threads/cores ?

Yes, even with just 1 thread

[dav1312]

To give some more info.
In an i5-12600KF (6P + 6HT + 4E) if I tell Stockfish to use all 16 threads, it will mainly use just the 4E cores.

results of speedtest <threads> <threads>*128 30

Threads	Normal	Above Normal
1	518 935	1 169 358
2	965 566	2 368 502
3	1 247 521	3 452 754
4	1 421 129	4 635 617
5	1 480 480	5 758 079
6	1 499 579	6 656 757
16	1 700 203	9 606 779

[mstembera]

I thought this would be handled using affinity not priority?

[Oqarshi]

I thought this would be handled using affinity not priority?

Yes, that’s true i think. one problem is that when your PC shuts down, windows removes the affinity for some reason, so I just use a third party program called processlasso to always force Stockfish to use the P-cores.

[vondele]

if it can be fixed purely with affinity you might be able to set the NumaPolicy option to a suitable value. For example, for the abovei5-12600KF, I could imagine that setoption name NumaPolicy value 0-11:12-15 and running with 16 threads would utilize all cores. While for up to 12 threads you'd probably be able to use setoption name NumaPolicy value 0-11. Would need testing to be sure though.

[dav1312]

setoption name NumaPolicy value 0-11 does prevent the E cores from being used, but it also prevents HT threads from being used.
There is no HT usage going from 6 to 12 threads.

[vondele]

so where do the additional threads go?

[vondele]

and maybe you can 'assist' by setting them as a domain as well?
setoption name NumaPolicy value 0-5:6-11
alternatively
setoption name NumaPolicy value 0,2,4,6,8,10:1,3,5,7,9,11

[dav1312]

Nope. Still using 6 threads. Also, speed decreased noticeably with both of those settings.

0-11	7.1Mnps
0-5:6-11	6.5Mnps
0,2,4,6,8,10:1,3,5,7,9,11	6.7Mnps
auto + prio	9.4Mnps
auto + prio + aff	9.1Mnps

[vondele]

ok, speed decrease can be understood (using two 'numa domains', so duplicates net access). Could it be that it is the core frequency that is different between using 0-11 and auto + prio ?

[dav1312]

With 0-11 it uses just 6 threads

With auto + prio it distributes the load across all threads, even E-cores

With neither, it uses the 4 E-cores. I guess Windows assumes it has to be some sort of antivirus program or something like that...