Deep HIARCS 14 WCSC and hyper-threading

[Peter Grayson]

Just compared Deep HIARCS 14 WCSC with hyperthreading enabled on my 16 core machine to give the extra 16 virtual cores. DH14 reported a maximum of 24 cores. Testing on the Richter.epd Position 7 for example

r6r/p2Rbk1p/2p2p2/q6b/8/2N2N1P/P1PP1PP1/2BQ2K1 w - - 1 0

it indicated about an extra 10% kN/s count but curiously its primary and secondary search depths were 1 and 2 ply down respectively at the same time interval when compared to being given the true 16-cores with hyperthreading disabled. With HT enabled, is the engine matching 1 true core with 1 hyperthreaded core and if that is the case I can understand why 24 cores would then be slower in actual search depth but what is not clear is why the kN/s should show higher with a reduced search depth?

Best regards,
PeterG

[Carl Bicknell]

Just compared Deep HIARCS 14 WCSC with hyperthreading enabled on my 16 core machine to give the extra 16 virtual cores. DH14 reported a maximum of 24 cores. Testing on the Richter.epd Position 7 for example

r6r/p2Rbk1p/2p2p2/q6b/8/2N2N1P/P1PP1PP1/2BQ2K1 w - - 1 0

it indicated about an extra 10% kN/s count but curiously its primary and secondary search depths were 1 and 2 ply down respectively at the same time interval when compared to being given the true 16-cores with hyperthreading disabled. With HT enabled, is the engine matching 1 true core with 1 hyperthreaded core and if that is the case I can understand why 24 cores would then be slower in actual search depth but what is not clear is why the kN/s should show higher with a reduced search depth?

Best regards,
PeterG

A couple of points:

I'm guessing it would be 16 physical cores + 8 virtual, it would be very strange if it was anything other than that.

Secondly, you should see a higher NPS with HT ON and a lower search depth in a given time, so this is exactly what you'd expect. As core count increases (be it virtual or physical) the search efficiency will DECREASE.
Usually if a doubling of threads can't increase NPS by +30% the actual speed is going down.

In your case you are going from 16 to 24 cores, so you'd need around 15% extra NPS just to break even.

[Peter Grayson]

Thanks for reply Carl. Thought I had read somewhere that HIARCS could benefit from hyper-threading but does not seem that way in practice so I have turned off hyper-threading in the BIOS now to get rid of any potential issues surrounding hyper threading.

Regards,
Peter

[Carl Bicknell]

I've now done some experiments...unless you have really good reasons to have HT on, I'd turn it off.

I tested a new dual xeon workstation (24 cores, 48 threads with HT) with HT ON and 48 instances of Komodo running in IDeA. Although it seemed to be working fine, task manager was not at 100%.

I tried SMP but those results seemed to fluctuate. When I disabled HT in the BIOS everything ran normally again.

[Peter Grayson]

I've now done some experiments...unless you have really good reasons to have HT on, I'd turn it off.

I tested a new dual xeon workstation (24 cores, 48 threads with HT) with HT ON and 48 instances of Komodo running in IDeA. Although it seemed to be working fine, task manager was not at 100%.

I tried SMP but those results seemed to fluctuate. When I disabled HT in the BIOS everything ran normally again.

Thanks for update Carl. I have left HT switched off in the BIOS.
Regards,
Peter

[Mark Uniacke]

There is a balance with hyper threading with chess engines, using hyperthreading may result in higher nps rates because more threads are working at the same time. The down side is there may be too many workers to do the job as efficiently, to the overall impact may be stronger or weaker depending on different circumstances.

A good analogy is if you had your kitchen changed with 4 workers, if instead you have 8 people working on it you may well find they are less efficient. It all depends on how interdependent they become in doing their work.