Olivier Teytaud wrote: >> It looks like we have a clear trend now. Light play-outs do not scale >> as well as heavy play-outs. > > I don't know if the data are sufficient for this conclusion, ... I don't know how there could be any doubt. The gap between level 1 and level 13 is:
heavy - 1930 light - 1410 So the heavy play-out version improved by 520 ELO more than the light play-out version. If you assumed that each player has 100 ELO of statistical error in the appropriate direction to nullify this conclusion, you would still show 120 ELO more improvement for the heavy play-outs. 100 ELO or more statistical error for all 4 players is not something that is reasonably likely. LightLeela_13 is showing about 60 ELO plus or minus with 95% confidence and it's the worst of the 4. - Don > but another > element is that heavy playouts are seemingly easier to parallelize > than light playouts. This is tested clearly in multi-core > parallelization, but I guess it should also hold for > mpi-parallelization (I've not sufficiently > tested to be very sure of that). > > By the way, for 19x19, I have not yet found bounds on the increase > of efficiency of mogo as we increase the number of nodes. If someone > has a huge cluster with good nodes and a good switch, I would be very > happy to have a try on mogo on, say 10 000 cores, if someone has > something like that. > > The current code > almost sacrifices one core for MPI, but we can improve that for > clusters with small numbers of cores per node. With 64 4-core nodes, > we have 95% winning rate against one node. I guess it is better with > more cores as one core is lost for MPI (on N cores, we loose nearly > 1/N for the MPI parallelization). > > Perhaps MoGo stops improving beyond 64 nodes, but perhaps not :-) > Olivier > _______________________________________________ > 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/
