Re: RMO - Randomized Move Order - yet another Lazy SMP derivate
Posted: Sun Jan 05, 2020 2:27 am
Peter, have you seen my latest results. Stockfish mulit-threaded (8-threads) even when constrained to single-core thrashes 1-thread stockfish.
There is clearly a "bug" in fixed depth tests with LazySMP, because it happens to sneak to depth d+1 is the current most likely theory.
Are you sure you took care of that for your depth-limited tests? You said it is a mix of ABDADA + Lazy so that d/d+1 is a potential problem ...
Also, my tests with my implementation of ABDADA give absolutely zero elo comparing 8-thread vs 1-thread at fixed depth.
I repost here the fixed depth = 14 result
On the other hand, my ABDADA implementation beats even YBW on timed games even with 12 cores. Since it doesn't widen search as shown above
the only conclusion is it must have deepened search instead (at least in most critical positions is my guess)!
Here ABDADA > YBW > SHT without any of them showing any sign of widening.
There is clearly a "bug" in fixed depth tests with LazySMP, because it happens to sneak to depth d+1 is the current most likely theory.
Are you sure you took care of that for your depth-limited tests? You said it is a mix of ABDADA + Lazy so that d/d+1 is a potential problem ...
Also, my tests with my implementation of ABDADA give absolutely zero elo comparing 8-thread vs 1-thread at fixed depth.
I repost here the fixed depth = 14 result
Code: Select all
Score of scorpio-sht-8-thread vs scorpio-sht-1-thread: 78 - 90 - 132 [0.480] 300
Score of scorpio-abdada-8-thread vs scorpio-abdada-1-thread: 111 - 107 - 106 [0.506] 324
Score of scorpio-ybw-8-thread vs scorpio-ybw-1-thread: 92 - 94 - 114 [0.496] 300
the only conclusion is it must have deepened search instead (at least in most critical positions is my guess)!
Code: Select all
Score of scorpio-ybw vs scorpio-sht: 44 - 18 - 70 [0.598] 132
Score of scorpio-ybw vs scorpio-abdada: 32 - 67 - 101 [0.412] 200