I share all uct-nodes with more than N visits, where N is currently 100, but performance doesn't seem very sensitive to N.
Does Mogo share RAVE values as well over MPI? I agree that low scaling is a problem, and I don't understand why. It might be the MFGO bias. With low numbers of playouts MFGO bias boosts the strength since it provides a lot of go knowledge. With high playouts the brittleness and missing pieces of the MFGO knowledge are exposed so the strength plateaus. Perhaps with more nodes the strength would start increasing again, but we didn't test it. It might be due to a bug in my MPI implementation, or any number of other possible bugs. If a bug causes it to lose a small number of games no matter how many playouts are made, it would look like a lack of scaling. My focus this year has been making the single node version as strong as possible since that's what I'm selling, so I haven't spent much effort in fixing this MPI scaling problem (and I don't have a lot of time available on a cluster with more than 4 nodes). David > -----Original Message----- > From: computer-go-boun...@computer-go.org [mailto:computer-go- > boun...@computer-go.org] On Behalf Of Brian Sheppard > Sent: Friday, October 30, 2009 7:49 AM > To: computer-go@computer-go.org > Subject: [computer-go] MPI vs Thread-safe > > >I only share UCT wins and visits, and the MPI version only > >scales well to 4 nodes. > > The scalability limit seems very low. > > Just curious: what is the policy for deciding what to synchronize? I > recall > that MoGo shared only the root node. > > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
