I did exactly this experiment with Oakfoam and I termed the potential loss of strength, given a fixed number of playouts and a different amount of parallel hardware, the "parallel effect." You can look at my work on parallelisation here: http://leafcloud.com/publications/mcts-parallelisation/ and http://leafcloud.com/publications/igs-talk-2012
I found that while root parallelisation can scale well, it was not close to the scaling of tree parallelisation for multi-core systems. -- Francois van Niekerk Email: flash.sl...@gmail.com | Twitter: @francoisvn Cell: +2784 0350 214 | Website: http://leafcloud.com On Sat, Aug 11, 2012 at 9:46 AM, David Fotland <fotl...@smart-games.com> wrote: > I'm happy with MFGO's scaling. I'm running a scaling test now, 4 threads vs > 8 threads, fixed 32K total playouts per move, 19x19, no pondering. Ideally > the win rate should be 50%, since the total playouts are the same. Has > anyone tried this kind of scaling experiment, and is willing to share > results? > > David > >> -----Original Message----- >> From: computer-go-boun...@dvandva.org [mailto:computer-go- >> boun...@dvandva.org] On Behalf Of Petr Baudis >> Sent: Friday, August 10, 2012 12:47 PM >> To: computer-go@dvandva.org >> Subject: Re: [Computer-go] Kas Cup - results and prizes >> >> On Fri, Aug 10, 2012 at 09:26:31AM -0700, David Fotland wrote: >> > Because my current approach seems to work just as well (or maybe >> > better), and I haven't had time to code up a shared try and tune it up >> > to validate that assumption. Chaslot's paper indicates perhaps that >> > not having a shared tree is stronger. My guess is that they are about >> > the same, so it's not worth the effort to change. >> >> In Pachi, having a shared tree makes all the difference when scaling up >> to more threads. See the graph (really awful one, sorry, it's old!) at >> >> http://pachi.or.cz/root-vs-shared.png >> >> If you have some information sharing near the root, I imagine it might >> be similar to Pachi's distributed engine performance (or just slightly >> better). But that is still far behind in scaling compared to the shared >> tree in our experience. >> >> P.S.: There are two important things, virtual loss (not necessarily 1 >> simulation but possibly more) and mainly lockless updates. The latter >> also means that sane code should be really easy to modify to use single >> shared tree instead of multiple trees. >> >> Petr "Pasky" Baudis >> _______________________________________________ >> Computer-go mailing list >> Computer-go@dvandva.org >> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > > _______________________________________________ > Computer-go mailing list > Computer-go@dvandva.org > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
