also having lazysmp trouble

[flok]

Hi,

I also have trouble with my lazysmp implementation.
I use a lockless tt, the 2nd thread (it currently only has 2 threads) starts with depth + 1.
Yet:

Code: Select all

info depth 13 score cp 24 nodes 3869732 time 1220 nps 3171911 pv e2e4 e7e5 g1f3 g8f6 b1c3 b8c6 d2d4 f8b4 d4d5 f6e4 d5c6 e4c3 b2c3
info depth 13 score cp 24 nodes 5104411 time 1569 nps 3253289 pv e2e4 e7e5 g1f3 g8f6 b1c3 b8c6 d2d4 f8b4 d4d5 f6e4 d5c6 e4c3 b2c3 b4c3
info depth 14 score cp 24 nodes 5902896 time 1796 nps 3286690 pv e2e4 e7e5 g1f3 g8f6 b1c3 b8c6 d2d4 f8b4 d4d5 f6e4 d5c6 e4c3 b2c3 b4c3
info depth 14 score cp 24 nodes 5928469 time 1802 nps 3289938 pv e2e4 e7e5 g1f3 g8f6 b1c3 b8c6 d2d4 f8b4 d4d5 f6e4 d5c6 e4c3 b2c3 b4c3 c1d2
info depth 15 score cp 24 nodes 8518619 time 2521 nps 3379063 pv e2e4 e7e5 g1f3 g8f6 b1c3 b8c6 d2d4 f8b4 d4d5 f6e4 d5c6 e4c3 b2c3 b4c3 c1d2

...the two threads are doing mostly the same and the elo-rating does not get higher of this at all.

What could be the cause of this?

[smatovic]

Elo gain by different, modern Lazy SMP methods depend on engine, like selective search heuristics, Bob used to say "they just widen the search", so if you prune too aggressive Lazy SMP catches again, I tried different depth+1 methods by myself, could not make it work either, meanwhile I prefer ABDADA with a move hash table as described by Tom Kerrigan.

--
Srdja