SPCC: Testrun of KomodoDragon 1.0 MCTS finished

[pohl4711]

Testrun of KomodoDragon 1.0 MCTS finished.

https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

[Vinvin]

Hi Stefan,
I see your methodology on your page "Since 20/07/21 AMD Ryzen 3900 12-core (24 threads) notebook with 32GB RAM. Now, 20 games are played simultaneously".
I think there's a problem because some threads could use hyperthreading and some others can use a full core to run.
Don't you think so ?

[Ras]

I think there's a problem because some threads could use hyperthreading and some others can use a full core to run.

A hyperthreaded core is worth about 1.5 real cores. Means, adding a second thread to a core will reduce the core performance to 0.75 cores per thread. However, the OS will switch the threads round and round unless you set core affinity per process. That happens so quickly that it will average itself out. The idea from the OS side is to even out the core usage to avoid thermal imbalances.

[mar]

I think there's a problem because some threads could use hyperthreading and some others can use a full core to run.
Don't you think so ?

Yes I think so. People shouldn't run more games at once than what their physical core count is, unless noise is desired.

I typically get a speedup of about 20-30% when I parallelize some task with hyperthreading, so I only get about 1.3x speedup when using 16 threads vs 8 on a 8-core machine.
I'm talking about tasks where I only care about how long it takes to get the result, I don't care if some random threads get preempted or have to wait otherwise, as long as the total runtime goes down (typical chess-related example would be Texel tuning).

This doesn't apply to playing games, however. When testing I only use n cores where n = number of physical cores, because I want reliable results (especially at hyperbullet where I don't want random time losses/low search depths).

So when someone is running say 47 games on a 24 core CPU, unless half the engines wait for memory and the other half is actually doing some useful work, many won't be able to utilize the core fully and this will only add noise to the results. IMO - YMMV.

[pohl4711]

I think there's a problem because some threads could use hyperthreading and some others can use a full core to run.

the OS will switch the threads round and round unless you set core affinity per process. That happens so quickly that it will average itself out. The idea from the OS side is to even out the core usage to avoid thermal imbalances.

Exactly!