Compare Stockfish Hyper off vs on

[kgburcham]

I have always had difficult time getting a repeat with smp.
No I dont mean exact----I mean too much difference.
If everyone else has the same problem and with all programs, then I wonder how many lower level moves were played in tournaments because of time control.
Several times a program has posted a quick solution that I cannot get it to repeat.
Anyway I am not getting a big increase in kns with 12 threads vs 6 threads.
I did prove number of threads with Task manager.
Please dont tell me Robert proved in 1920 that hyperthreading doesnt work-----everyone here knows that. I was just trying to show close to exact results----I guess I could test 10 times and take an average.

Here I was trying to compare Hyperthreading off vs on with Stockfish.

Notice mate in 13 with 425mn and mate in 13 with 1750mn

[d] 2K1k1br/3p1n1r/2p2pN1/P2p1N2/2P4P/5p2/P2P4/8 w - -

6 core Xeon 6x4gig
6 threads Hyperthreading off

Stockfish 1.7 JA 64bit (2048 MB)
(Shredder Interface)

23.01 0:20 -11.07++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c5 8.a5 c3 9.dxc3 (222.229.678) 11034
23.01 0:20 -10.74++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c5 8.a5 c3 9.dxc3 (224.213.018) 11030
23.01 0:20 -10.10++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c5 8.a5 c3 9.dxc3 (228.383.934) 11007
23.01 0:22 -8.80++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c5 8.a5 c3 9.dxc3 c4 10.a6 (253.248.033) 11058
23.01 0:24 -6.22++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c5 8.a5 c3 9.dxc3 c4 10.a6 (268.002.295) 11076
23.01 0:25 -1.05++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c5 8.a5 c3 9.dxc3 c4 10.a6 (287.351.922) 11083
23.01 0:26 +9.29++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c5 8.a5 c3 9.dxc3 c4 10.a6 (289.779.255) 11076
23.01 0:26 +29.97++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c5 8.a5 c3 9.dxc3 c4 10.a6 (290.404.496) 11074
23.01 0:26 +71.35++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c5 8.a5 c3 9.dxc3 c4 10.a6 (290.414.453) 11074
23.01 0:38 +M13 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.a4 d5 7.Kc8 c3 8.dxc3 c5 9.a5 c4 10.a6 d4 11.a7 dxc3 12.a8Q Nd8 13.Nd6+ (425.277.319) 11014



6 core Xeon
6 thread Hyperthreading off

Stockfish 1.7 JA 64bit (2048 MB)
(Shredder Interface)

27.01 1:01 -10.38++ 10.a6 f2 11.a7 f1Q 12.a8N Qf4 13.Nc7+ Qxc7+ 14.Kxc7 dxc4 15.a4 d5 16.Kc8 d4 17.a5 c3 18.a6 (748.388.011) 12142
27.01 1:03 -10.30++ 10.a6 f2 11.a7 f1Q 12.a8N Qf4 13.Nc7+ Qxc7+ 14.Kxc7 dxc4 15.a4 d5 16.Kc8 d4 17.a5 c3 18.a6 (769.179.703) 12129
27.01 1:05 -10.14++ 10.a6 f2 11.a7 f1Q 12.a8N Qf4 13.Nc7+ Qxc7+ 14.Kxc7 dxc4 15.a4 d5 16.Kc8 d4 17.a5 c3 18.a6 (792.917.628) 12110
27.01 1:07 -9.81++ 10.a6 f2 11.a7 f1Q 12.a8N Qf4 13.Nc7+ Qxc7+ 14.Kxc7 dxc4 15.a4 d5 16.Kc8 d4 17.a5 c3 18.a6 (820.002.902) 12106
27.01 1:17 -9.17++ 10.a6 f2 11.a7 f1Q 12.a8N Qf4 13.Nc7+ Qxc7+ 14.Kxc7 dxc4 15.a4 d5 16.Kc8 d4 17.a5 c3 18.a6 (938.902.662) 12092
27.01 1:28 -7.87++ 10.a6 f2 11.a7 f1Q 12.a8N Qf4 13.Nc7+ Qxc7+ 14.Kxc7 dxc4 15.a4 d5 16.Kc8 d4 17.a5 c3 18.a6 (1.069.164.035) 12074
27.01 1:55 -5.29++ 10.a6 dxc4 11.a7 (1.391.462.731) 12082
27.01 2:08 -0.12++ 10.a6 dxc4 11.a7 (1.553.213.614) 12077
27.01 2:08 +10.22++ 10.a6 dxc4 11.a7 (1.555.897.710) 12076
27.01 2:08 +30.90++ 10.a6 dxc4 11.a7 (1.557.043.850) 12076
27.01 2:08 +72.28++ 10.a6 dxc4 11.a7 (1.557.047.387) 12074
27.01 2:25 +M13 10.a6 f2 11.a7 f1Q 12.a8N Qf4 13.Nc7+ Qxc7+ 14.Kxc7 dxc4 15.a4 d5 16.Kc8 c3 17.dxc3 c5 18.a5 d4 19.a6 d3 20.a7 d2 21.a8N d1Q 22.Nc7+ 1.751.427.122) 12016



6 core Xeon
12 threads Hyperthreading on

Stockfish 1.7.1-16 JA 64bit (2048 MB)
(Shredder Interface)

24.01 0:24 -11.03++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (317.779.116) 13007
24.01 0:24 -10.82++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (320.390.618) 12982
24.01 0:25 -10.42++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (326.440.673) 12948
24.01 0:26 -9.61++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (339.738.528) 12932
24.01 0:29 -8.00++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (382.313.953) 12959
24.01 0:32 -4.76++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (416.307.783) 12998
24.01 0:32 +1.69++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (424.812.386) 12986
24.01 0:32 +14.62++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (427.182.669) 12977
24.01 0:33 +40.48++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (428.423.432) 12966
24.01 0:33 +92.20++ 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 d4 8.a5 c3 9.a6 cxd2 10.a7 d1Q 11.a8N (428.459.102) 12961
24.01 0:54 +M13 1.a6 f2 2.a7 f1Q 3.a8N Qf4 4.Nc7+ Qxc7+ 5.Kxc7 dxc4 6.Kc8 d5 7.a4 c3 8.dxc3 c5 9.a5 d4 10.a6 d3 11.a7 d2 12.a8N d1Q 13.Nc7+ (711.906.877) 13068
best move: a5-a6 time: 1:01.184 min n/s: 13.029.614 nodes: 797.190.899

Not sure how clear this pic will be. Shows 12 threads @13000kns
About the same with 6 threads in these type of positions.

[lmader]

I'm probably being dense, so apologies in advance, but I don't understand the goal of your experiment (HT on/off). Are you trying to evaluate whether HT is creating a net loss in the overall engine performance/search-outcome vs whether it's a wash (no real difference)?

[Eelco de Groot]

Well against the expectations, on an Atom netbook hyperthreading could be shown to have some positive influence on running the internal benchmark. But going from 6 to 12 threads would be very different from hyperthreading an Atom. That would probably be harder to show a net gain, just because of randomness in SMP, because of less efficiency with so many threads, the used parallel algorithm may not work as well with many fysical threads in the first place so with hypothetical hyperthreaded threads it may be even less. In theory binding 6 threads to the six processors in BIOS should be best I believe, but I don't know much about it. I also don't know if it would be possible to run a ponder off match between a hyperthreaded Stockfish against a Stockfish using 6 threads on one machine? Probably not because you would have to switch hyperthreading on and off after every turn. Anyway personally I would not expect that much of a gain is possible, but it is good to have some real data, in either case. Nodes per second is no good that tells you nothing about a real speed up.

Eelco

[kgburcham]

Good day to you Lar. There was a post the other day talking about Stockfish and hyperthreading. Not sure if the post concluded anything. I wanted to see the numbers myself, with the new 16 thread Stockfish, which allowed me to hyperthread. The other version is limited to eight threads.
Anyway I am not going to run so many tests and then average.
Maybe you are suggesting it does not matter anyway. I agree.

gotta do something
cant go the firing range----its raining
cant work on the race car----sold it
cant pull weeds in my yard---there are no weeds
cant clean the shop----it stays tour ready
cant run an old test position through the pc----I am out of test positions
cant assemble my new SR2----have everything except MB

On my machine with that program, Hyper is not making much difference.
Of course like I said that is old news.

[lmader]

Ah, I see. In other words, you're doing it for the best reason: Because it's there! (or, Because you can)