Thank you Jean, Did those experiments use what network devices? Very popular GbE or very expensive 10 GbE such as InfiniBand? #I guess GbE :) Using just MPI, simple TCP, or UDP broadcasting?
I'd like to agree that we need new algorithms to get more scalability, anyway. One possibility could be not only reducing statistical errors (by averaging) but also growing trees on each simulation (not realistic?) or a bunch of simulations. Hideki Jean-loup Gailly: <[email protected]>: >> have you evaluated the improvement of strength on your 40-core pc cluster? > >I already published a scalability analysis of Pachi on a single machine: >http://comments.gmane.org/gmane.games.devel.go/20925 > >I wrote the multi-machine (message passing) code for Pachi. It uses slow >tree >parallelization as described in http://hal.inria.fr/inria-00512854/ >On a single machine doubling the number of playouts gains approximately 100 >elo and I did not observe reduced scalability at high number of playouts. >On multiple machines, a doubling gains approximately 50 elo initially, but >scalability quickly drops to a point where increasing the number of machines >doesn't improve strength and can actually reduce it. This is consistent >with the observations of the above paper inria-00512854. > >This analysis was done with a constant opponent (Fuego running on a >single machine with 7 cores) in 19x19. 3662 games were played in >total. pachi was running on up to 40 machines using 7 cores >each. This is not Petr's cluster, but Petr also measured roughly 50 >elo per initial doubling on his cluster. In practice pachi cannot >currently make good use of more than 10 machines. I'm confident that >over time we will learn how to push scalability further, but this >will require new algorithmic insights. > >[image: pachi-scalability.png] > >Jean-loup >---- inline file >_______________________________________________ >Computer-go mailing list >[email protected] >http://dvandva.org/cgi-bin/mailman/listinfo/computer-go -- Hideki Kato <mailto:[email protected]> _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
