Mikko, This is pretty much what I expected, thanks for re-affirming this. I agree that from my experience, because I'm doing pure random playouts with very little lookups on patterns, MCTS trees, etc my CPU utilization is higher until I start adding these features so I will see less of a gain when using HT.
On Mon, May 12, 2014 at 6:54 PM, Mikko Aarnos <[email protected]>wrote: > There is a big difference here: Ellis's program can only do light > playouts. He doesn't have MCTS or patterns. That is parallelized extremely > simply by just giving each thread an internal board state, doing a playout > from that, resetting the board state to the original, doing a playout etc. > There are no bottlenecks there, and that shouldn't get any increase in > performance from HT as far as I know(also see the first sentence of > Schmicker's comment). On the other hand, your programs all have MCTS and > patterns(with emphasis being on patterns) and they both need constant > memory reads. Thanks to that it's not a huge surprise that HT works better. > Still, that doesn't change the fact that I was a bit off. I never actually > expected that there was so much memory reading that it would actually make > HT work. Guess I should implement patterns and see if I get similar results. > > Regards, > > Mikko Aarnos > > PS. Of course, all this rests on the assumption that you don't do your > playouts exactly like Ellis, and if you do I am really, REALLY surprised > with the performance of HT. > > PPS. And on the assumption that the 40%-50% performance increase was from > going from 4 threads to 8 threads with HT on all the time, not from going > from 4 threads with no HT to 8 threads with HT. Here as well, if the latter > is true I am again honestly surprised. > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
