lazy smp questions

[elcabesa]

Hi, I'm just trying to implement Lazy SMP in Vajolet and finally I was able to make it not crash

my first test has 2 thread searching he same root position, with early termination, the first one searching at depth N and the second at depth n+1. I always use the result of the first thread. The only shared data is the TT table.

the test are running and the first result looks promising.
vajolet 2 thread vs vajolet 1 thread at 1''+0.05 ( very bullet time ) has gained 29 Elo point after 500 games.

what performance should I expect from lazy smp with 2, 3,4 core andn so on? my engine seems to have a speedup lower than 2 ( i still have to measure it whit precision.)

and now a question to Martin Sedlak, if he read this post, why syncrnize History Table after each search?

thank you all

[mar]

and now a question to Martin Sedlak, if he read this post, why syncrnize History Table after each search?

Hi Marco, I just described what I do in cheng. When I wrote it, I simply wanted to sync everything
(in fact sync is done once at the beginning of ID loop, not inside it - at least in my current dev version, where I also fixed some minor smp-related bugs
and refactored some parts).
In the case of history table I'm pretty sure you don't have to do it
Only the important things (game state, i.e. board, repetitions etc. ...) must be synced.
Otherwise, congratulations for making it work
As for selfplay elo per cores, I don't quite remember, but IIRC there was a thread where Peter tested smp scaling of cheng and Texel with various cores, maybe you can look it up.
But I think you should get more than 30 (if it's selfplay), IIRC in the first test I got around 50 elo (2 vs 1, 10"+0.01 or something like that),
but that's maybe because of the TC, no idea.
As long as it's not worse then it works

[elcabesa]

I'm now tryng some more tests

[elcabesa]

new test with cutechess cli, no concurency, 12+0.1s and INTEL TURBO BOOST disabled:
2core vs 1core -> +52 elo so far. ( 108W-46L-262D : 416 games)